CN114936708A

CN114936708A - Fault diagnosis optimization method based on edge cloud collaborative task unloading and electronic equipment

Info

Publication number: CN114936708A
Application number: CN202210650906.2A
Authority: CN
Inventors: 刘晶; 石志鹏; 季海鹏; 赵佳; 徐伟杰
Original assignee: Hebei University of Technology
Current assignee: Hebei University of Technology
Priority date: 2022-06-10
Filing date: 2022-06-10
Publication date: 2022-08-23

Abstract

The invention provides a fault diagnosis optimization method based on edge cloud cooperative task unloading and an electronic device, belongs to the technical field of industrial device fault diagnosis and edge cloud cooperative task unloading, and solves the problem of high diagnosis delay in the prior art. The method comprises the following steps: aiming at different tolerant time delays of all diagnosis tasks, a fault diagnosis model of edge cloud DNN multi-branch is established; the fault diagnosis model comprises a plurality of fault diagnosis model branches, and each fault diagnosis model branch consists of a corresponding DNN network layer; dividing a fault diagnosis model into a plurality of parts according to the diagnosis model structure and the layers as granularity according to the FLOPs, wherein each part forms a diagnosis task and is modeled as a DAG; analyzing unloading time delay and unloading energy consumption of a diagnosis task of the fault diagnosis model, and determining a multi-objective optimization function of task unloading; and solving the minimum solution of the multi-objective optimization function by using an APAM-DRQN algorithm.

Description

Fault diagnosis optimization method based on edge cloud cooperative task unloading and electronic equipment

Technical Field

The invention relates to the technical field of fault diagnosis of industrial equipment and unloading of edge cloud cooperative tasks, in particular to a fault diagnosis optimization method based on the unloading of the edge cloud cooperative tasks and electronic equipment.

Background

With the development of artificial intelligence, internet of things and industrial internet technology, the industrial manufacturing industry is moving to digitization and intellectualization, and mechanical equipment is also moving towards increasingly complex and integrated directions. Once the precise key mechanical parts with complex structures break down, the normal operation of mechanical equipment is seriously influenced, and a great loss is caused. Therefore, the predictive diagnosis and prevention of the failure of large-scale machinery is an important subject for the development of the industrial manufacturing industry.

The real-time performance and energy consumption overhead of equipment fault diagnosis directly influence the efficiency and safety of actual production, and are key indexes for determining the quality of fault diagnosis service. However, most of the current fault diagnosis methods have high diagnosis delay and energy consumption, and it is difficult to provide high-quality diagnosis service and meet the requirements of intelligent manufacturing scenarios. In order to solve the above problems, scholars introduce a technique of edge cloud collaboration. The edge cloud cooperation mainly comprises three cooperation modes of resource cooperation, data cooperation and service cooperation.

The resource cooperation is that the edge terminal close to the data source provides basic computational power and storage resources, local resources can be automatically scheduled, delay sensitive tasks are processed, and the central cloud is used as a resource general scheduling center and can provide resource scheduling decisions for the edge terminal. An article [ Wang S, Tuor T, Salonidis T, et al. When edge media learning: adaptive control for resource-constrained distributed Computer learning [ C ]. IEEE INFOCOM 2018-IEEE Conference on Computer communications. IEEE 2018:63-71 ] proposes a distributed model training algorithm for parameter aggregation by using a central parameter server, which uploads local model parameters of different distributed nodes for aggregation without uploading original data to a central cloud. At the same time, a trade-off is made between local updates and global parameter aggregation, and a control algorithm is proposed that minimizes a given resource budget. The article [ Li Y, Shen B, Zhang J, et al, offload in HCNs: coherent-aware network selection and user implicit design [ J ]. IEEE Transactions on Wireless Communications,2017,16(10):6479 and 6492 ] designs a resource scheduling mechanism under congestion processing and trust perception aiming at the problem of cooperative multi-hop caused by the heterogeneity of resources in a cloud edge scene, and realizes efficient task scheduling and global resource sharing on the premise of ensuring network stability and load balance.

And performing data cooperation, namely performing original processing on the acquired data by the edge end, then sending corresponding data and a processing result to the central cloud, collecting mass data by the central cloud, performing data mining and model training, and sinking the model to the edge node. For the research of Edge cloud data cooperation, an article [ Zhang X, Qiao M, Liu L, et al, colorful closed-end calculation for personalized driving behavior module [ C ]. Proceedings of the 4th ACM/IEEE Symposium on Edge calculation.2019: 209-221 ] proposes a personalized driving behavior analysis method based on cloud Edge cooperation. Experiments prove that the method can meet the real-time requirement.

The service cooperation, namely the center cloud, is used as a management center to provide a management scheme of life cycles such as application deployment, service starting, service ending and the like for the edge nodes; and the center cloud with rich resources is responsible for model training, and sinks the model to the edge node for identification and classification. For the research of the edge cloud Service cooperation, an article [ Ma X, Zhang S, Li W, et al, cost-effective workload scheduling in closed associated mobile computing [ C ].2017IEEE/ACM 25th International Symposium on Quality of Service (IWQoS) IEEE,2017:1-10 ] designs a cloud edge Cooperation (CAME) framework, and simultaneously provides a load balancing and cloud outsourcing strategy, wherein the CAME framework models and analyzes the system delay and is constructed as a Service Quality maximization problem, so that the delay and cost are balanced, and the Service Quality is improved. The results show that the framework can reduce latency as desired. The article [ Kang Y, Hauswald J, Gao C, et al. Neurosgeon: colour incidence between the layer and mobile edge [ J ]. ACM SIGARCH Computer Architecture News,2017,45(1): 615:. Neurosgeon frame was designed, according to the type of each layer of DNN (Deep Neural Network), using layer as granularity, the frame performs model division between mobile equipment and central cloud, finally returns the optimal model division point to reduce delay, provides real-time and high-efficiency service.

The development of the edge cloud cooperation technology brings opportunity for falling to the ground of scenes such as a smart factory. In an intelligent factory scene, a large number of sensors are used for collecting equipment production state data in real time, equipment fault diagnosis has high requirements on diagnosis instantaneity, data safety and resource management efficiency, and the traditional equipment fault diagnosis method is difficult to meet the requirements of industrial production. Facing to an equipment fault diagnosis scene, cloud computing is an important support technology for realizing the equipment fault diagnosis method based on deep learning, and feature extraction and data mining are carried out on mass data by utilizing the abundant computing power and storage resources of a center cloud. However, due to the rapid development of the internet of things technology, data generated by equipment is exponentially and explosively increased, huge network transmission resource overhead is caused by uploading of mass data, and the central cloud is far away from a data source, so that the real-time performance of fault diagnosis cannot be guaranteed. In the edge computing mode, data are stored and preprocessed at edge nodes close to a data source, and are not required to be directly uploaded to a remote cloud, so that the diagnosis response time is greatly shortened, and the privacy protection of the data is facilitated. An article [ Qian G, Lu S, Pan D, et al. edge computing: A simulating frame for real-time fault diagnosis and dynamic control of rotating machinery using multi-sensor data [ J ]. IEEE Sensors Journal,2019,19(11): 4211-4. 4220 ] proposes a method for diagnosing faults of rotating machinery based on edge computing, firstly, an edge end collects original signal data of a motor, and then characteristic extraction and fusion are carried out on vibration signal data, finally classification diagnosis is realized, and experimental results prove that the edge end can realize high-precision fault diagnosis and make effective fault handling measures; an article [ Tharmakularsingam S, Lu S, Phung B T, et al. Sustainable deep learning at grid for real-time high impedance fault detection [ J ]. IEEE Transactions on Sustainable Computing,2018,2879960:1-12 ] proposes a fault diagnosis method for high-voltage electrical equipment based on DNN neural network and edge calculation, which is used for reducing time delay of high-impedance fault diagnosis of high-voltage electrical wires. The documents have good effects, but the central cloud does not participate in a fault diagnosis scene, so that the method cannot be applied to an intelligent factory scene with massive heterogeneous state data. An article [ Wang Y, Gao L, Zheng P, et al.A smart surface inspection system using a device R-CNN in closed-edge computing environment [ J ]. Advanced Engineering information, 2020,43: 101037-. Zhang wenlong et al improves the adaptability of the fault diagnosis method to personalized data and reduces the time delay of fault diagnosis based on a side cloud collaborative architecture and in combination with a transfer learning technology. However, the above documents do not perform fine-grained decomposition on the diagnosis task, so that the diagnosis task does not obtain an optimal edge cloud deployment scheme, and the problem of high diagnosis delay still exists.

Disclosure of Invention

The invention aims to provide a fault diagnosis optimization method based on edge cloud cooperative task unloading and an electronic device, and solves the problem of high diagnosis delay in the prior art.

In a first aspect, the present invention provides a fault diagnosis optimization method based on edge cloud collaborative task offloading, including:

aiming at different tolerant time delays of all diagnosis tasks, a fault diagnosis model of edge cloud DNN multi-branch is established; the fault diagnosis model comprises a plurality of fault diagnosis model branches, and each fault diagnosis model branch consists of a corresponding DNN network layer;

dividing a fault diagnosis model into a plurality of parts according to a diagnosis model structure, taking a layer as granularity, and floating point Operations Per Second (FLOPs), wherein each part forms a diagnosis task and is modeled into a Directed Acyclic Graph (DAG) to represent the execution sequence among the parts;

analyzing unloading time delay and unloading energy consumption of a diagnosis task of the fault diagnosis model, and determining a multi-objective optimization function of task unloading;

and solving the minimum solution of the multi-objective optimization function by using a DRQN algorithm of the average network parameters based on the attention mechanism.

In a second aspect, the present invention further provides an electronic device, which includes a memory and a processor, where the memory stores a computer program operable on the processor, and the processor implements the steps of the above method when executing the computer program.

In a third aspect, the present invention also provides a computer readable storage medium having stored thereon machine executable instructions which, when invoked and executed by a processor, cause the processor to perform the method described above.

The invention provides a fault diagnosis optimization method based on edge cloud collaborative task unloading, which aims at the problem that the DRQN (Deep Current Q-Network, based on an improved Deep cycle Q Network) algorithm uses an epsilon-greedy strategy to cause lower environmental exploration efficiency, and provides a DRQN algorithm based on average Network parameters to carry out multi-target optimization, thereby providing an optimal fault diagnosis edge cloud task unloading scheme, reducing the time delay and energy consumption of fault diagnosis, relieving the problem of higher diagnosis time delay in the prior art and improving the diagnosis service quality.

Drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below, and it is obvious that the drawings in the following description are some embodiments of the present invention, and other drawings can be obtained by those skilled in the art without creative efforts.

Fig. 1 is a flowchart of a fault diagnosis optimization method based on edge cloud collaborative task offloading according to an embodiment of the present invention;

FIG. 2 is a block diagram of a branch model for diagnosing equipment failure according to an embodiment of the present invention;

FIG. 3 is a first branch model DAG diagram according to an embodiment of the present invention;

FIG. 4 is a second branch model DAG diagram according to an embodiment of the present invention;

FIG. 5 is a third branch model DAG diagram according to an embodiment of the present invention;

FIG. 6 is a flow chart of the APAM-DRQN algorithm according to an embodiment of the present invention;

FIG. 7 is a comparison graph of long term average jackpot values for the DQN, DRQN and APAM-DRQN algorithms of embodiments of the present invention;

FIG. 8 is a comparison graph of the time delay and weighted sum of energy consumption for 5 algorithms with different task numbers;

fig. 9 is a comparison graph of time delay and energy consumption weighted sum of 5 algorithms under different computing capabilities of the cloud server;

FIG. 10 is a comparison graph of time delay and weighted sum of energy consumption for 5 algorithms with different task data amounts;

FIG. 11 is a graph of the average latency of task execution under different algorithms;

FIG. 12 is a graph of average cost of total energy consumption under different algorithms.

Detailed Description

To make the objects, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings, and it is apparent that the described embodiments are some, but not all embodiments of the present invention. All other embodiments, which can be obtained by a person skilled in the art without making any creative effort based on the embodiments in the present invention, belong to the protection scope of the present invention.

The terms "comprising" and "having," and any variations thereof, as referred to in embodiments of the present invention, are intended to cover non-exclusive inclusions. For example, a process, method, system, article, or apparatus that comprises a list of steps or elements is not limited to only those steps or elements but may alternatively include other steps or elements not expressly listed or inherent to such process, method, article, or apparatus.

As shown in fig. 1, an embodiment of the present invention provides a Fault diagnosis optimization method (a Fault diagnosis method on edge-closed collaborative task offloading, FDBECTO method for short) based on edge cloud collaborative task offloading, including:

s1: and establishing a side cloud DNN multi-branch fault diagnosis model aiming at different delay tolerances of the diagnosis tasks.

The fault diagnosis model comprises a plurality of fault diagnosis model branches and two fully connected layers (FC), and each fault diagnosis model branch consists of a corresponding DNN network layer. The complexity of each branch model gradually rises, and the generalization and diagnosis delay of the model sequentially rise. During fault diagnosis, according to the tolerant time delay of the diagnosis task, a proper branch is selected, and the branch structure of the equipment fault diagnosis model is shown in fig. 2. A first fault diagnosis model branch of the plurality of fault diagnosis model branches comprises 4 convolutional layers and 1 pooling layer; the second fault diagnosis model branch comprises 5 convolutional layers and 1 pooling layer; the third fault diagnosis model branch comprises 3 convolutional layers, 3 Long Short-Term Memory network Layers (LSTM) and 1 pooling layer.

The convolution layer extracts local features from input sample data, the pooling layer performs down-sampling processing on the features, and the two network layers form a group of fault feature extraction modules. Through the series connection of a plurality of groups of feature extraction modules, the fault features are extracted from shallow to deep. And finally, carrying out fault mapping by the full connection layer to realize classification diagnosis. The first layer CNN network of the branch 1 model adopts convolution kernels with larger sizes, and the rest CNN network adopts convolution kernels with smaller sizes to obtain local features so as to improve the diagnosis precision; a CNN network layer, a pooling layer and other hidden layers are added to the branch 2 model, so that the diagnosis precision of the diagnosis model is improved; the branch 3 comprises three convolution layers, three pooling layers and other hidden layers and 3 LSTM networks. Feature extraction is carried out by the convolution layer, abstract identification is carried out on fault signal data, and then identification processing is carried out on time sequence fault diagnosis data by the LSTM layer, so that the generalization and the diagnosis precision of the diagnosis model are improved. The fault diagnosis precision and the execution time delay of the three branches are sequentially increased.

The first layer of CNN network of the branch 1 fault diagnosis model is a large convolution kernel of 64 x 1, and the following convolution layer is a small convolution kernel of 3 x 1. The large convolution layer can obtain a large enough receptive field, the other small convolution layers can ensure that the parameters of the network are less, and the branch 2 fault diagnosis model is added with a hidden layer such as a CNN layer and a pooling layer. The branch 1 fault diagnosis model is a hidden layer such as a four-layer CNN and a four-layer pooling layer, and the branch 2 fault diagnosis model is a hidden layer such as a five-layer CNN and a five-layer pooling layer. The structural parameters of the branch 1 fault diagnosis model are shown in table 1, and the structural parameters of the branch 2 fault diagnosis model are shown in table 2.

TABLE 1 Branch 1 model network architecture parameters

TABLE 2 Branch 2 model network architecture parameters

LSTM networks are suitable for processing time-sequential correlated data, however training speeds are slow. Hidden layers such as a CNN layer and a pooling layer firstly extract abstract features of fault signal data, the CNN layer extracts vibration signal features, and the pooling layer performs dimension reduction processing, so that model training time is shortened, model diagnosis precision is improved, and diagnosis model structure parameters of the branch 3 are shown in a table 3.

Key parameter items:

conv (convolution) layer and pooling (pooling) layer: convolution kernel size/step size

LSTM layer: number of hidden layer units

TABLE 3 Branch 3 model network architecture parameters

S2: according to the diagnosis model structure, layers are taken as granularity, a fault diagnosis model is divided into a plurality of parts according to Floating Point Operations Per Second (FLOPs), each part forms a diagnosis task, and the diagnosis model is modeled into a directed acyclic graph to represent the execution sequence of the parts.

The process of establishing the DAG directed acyclic graph of the fault diagnosis model is described as follows:

for the equipment fault diagnosis model of the DNN structure, due to the fact that types, input data and calculation characteristics of all layers of the model are different, the running time of different layers of the diagnosis model has great heterogeneity. The execution time of the DNN diagnostic model may be estimated by calculating the number of floating point operations (FLOPs). Wherein, the convolution layer and the full connection layer are layers with higher FLOPs, and the other neural network layers have lower FLOPs.

According to the diagnostic model structure, layers are taken as granularity, the neural network is divided into multiple parts according to FLOPs, each part forms a diagnostic task, and the neural network is modeled into a DAG directed acyclic graph to represent the sequential execution sequence of the parts, and the DAG is represented by G ═ V, E. Wherein, V ═ V (V) ₁ ,v ₂ ,…,v _N ) Each vertex is a part of the model and can only be processed on one computing resource of the cloud end or the edge end as a diagnosis task. E represents a set of directed edges, E ═ v _i ,v _j ) E.g. E, represents the diagnostic model diagnostic task v _j At v _i The task can be executed after the execution is finished.

The DNN-based fault diagnosis branch model builds DAG maps as shown in FIGS. 3-5, respectively. Each task node v _i The definition of epsilon V is shown as formula (1):

v _i ＝(ω _i ,dataSize _i ,In _i ,Out _i ) (1)

wherein, ω is _i Presentation of diagnostic tasks task _i Including task _i The calculation amount executed by the task and the calculation amount output by the predecessor task node; DataSize _i Representing task _i The data size of (2), including task _i The data volume of the task and the data volume output by the predecessor task node; in _i Representing task _i The predecessor task node set of (1) only if task _i After all the predecessor tasks are executed, task _i Can be executed; out _i Representing task _i The subsequent task node set of (1) only if task _i After the execution is finished, the subsequent task node can execute the program.

In order to describe the context dependency relationship between the diagnosis tasks, a task DAG graph is established, and a priority list between the tasks is established. Defining the Task class and the TaskList class:

the member variables of the Task class are used to describe the relevant attributes of the Task.

Class Task：{

taskId; // task Attribute ID

index; // task indexing in priority list

flops: // number of floating-point operations of task, reflecting the amount of task computation

dataSize: // amount of data for task

A preList; // predecessor task list of tasks

postList; // task successor list of tasks

Status; the execution state of the task, 1 represents that the task is executed; 0 indicates that the task is ready to be executed.

}

The member variable of the TaskList class is an array containing Map key value pairs, the array is subjected to task priority descending sorting, the priority of the task is lower than that of a precursor task and higher than that of a subsequent task, each Map element key is task Id, and value is task.

Class TasksList：{

Array < Map < taskId, task > (Sorted); // sorting in descending order of priority,

}

after the task priority list is built, when the tasks in the DAG are executed, the tasks are executed according to the priority sequence, the tasks with higher priority are executed first, and then the subsequent tasks are executed, so that the tasks can be executed correctly.

S3: and analyzing unloading time delay and unloading energy consumption of the diagnosis task of the fault diagnosis model, and determining a multi-objective optimization function of task unloading. The method comprises the following specific steps:

1-1) dividing the unloading time of the diagnosis task into calculation time and transmission time, and analyzing the unloading time of the diagnosis task to obtain total running time T.

And taking each part of the diagnosis model as a diagnosis task to carry out cloud-side task unloading, wherein the total time comprises the calculation time of the task and the transmission time of the task in a cloud-side resource environment.

(1) Calculating time

After the structure of the diagnostic model is determined, diagnostic task E can be derived by computing FLOPs _i Load W of _i And the output data size Out of the partial model can be counted _i . When task E _i Is allocated to resource R _i When E is greater _i Calculated time of

Represented by formula (2):

in the formula (2), the reaction mixture is,

the task may be placed in the cloud or edge for execution as determined by the task's offload decision. R _i ∈R _cloud Representing task being placed in cloud for execution, f _i ^cloud Representing the execution speed of the task Ei in the central cloud; r _i ∈R _edge Representing tasks being placed to edgesEnd execution, f _i ^edge Represent task E _i Execution speed at the edge end; r is _i ∈R _end Indicating that the task is executed on the device side, f _i ^end Representing task E _i Execution speed on the device side.

(2) Time of flight

(E _i ，E _j ) Representing task E _i And E _j Before and after the execution of the relation, i.e. task E _i As task E _j The pre-layer diagnostic task of (1). Time of flight

Represented by formula (3).

Wherein, Out _i As task E _i Is simultaneously task E _j The input data of (2). B _upload And B _download Respectively representing the uploading network bandwidth from the edge end to the cloud end and the returning network bandwidth from the cloud end to the edge end. When two adjacent tasks are placed on the same computing resource, namely, the two tasks are placed on the cloud end, the edge end or the equipment end, at the moment, the task E _i Is directly input to task E _j No transmission over the network is required; when task E _i Put to the device side, task E _j Is offloaded to the edge end, task E _i The output result of the task needs to be uploaded to the edge terminal through the network as a task E _j When the network bandwidth is B _upload (ii) a When task E _i Is offloaded to the edge end, task E _j Put to the device side, task E _i The output result of (a) needs to be transmitted back to the edge end through the network as a task E _j At the time the network bandwidth is B _download (ii) a When task E _i Put to the edge end, task E _j Is offloaded to the cloud, task E _i The output result of (1) needs to be uploaded to the cloud end through the network as a task E _j At the time the network bandwidth is B _upload (ii) a When task E _i Is offloaded to the cloud, task E _j Put to the edge end, task E _i The output result of (1) needs to be transmitted back to the cloud end through the network as a task E _j When the network bandwidth is B _download 。

(3) Total run time

The total running time T includes the computing time of the task and the transmission time of the task in the cloud-side resource environment, as shown in formula (4).

Wherein N represents the total number of tasks divided by the DNN diagnostic model, and the total running time is equal to the cumulative sum of the calculation time and the network transmission time of each task.

1-2) dividing the unloading energy consumption of the diagnosis task into calculation energy consumption and transmission energy consumption, and analyzing the unloading energy consumption of the diagnosis task to obtain total energy consumption E.

The total energy consumption E of the terminal device includes the task execution energy consumption of the device side, the energy consumption of uploading the task data to the cloud server, and the receiving energy consumption of returning the task execution result to the device side, and the total energy consumption is as shown in formula (5).

Wherein, P _comp 、P _upload 、P _download Respectively representing task execution power, data uploading power and data receiving power of a device end; t is _comp 、T _upload 、T _download Respectively representing the task execution time, the data uploading time and the data receiving time.

Wherein, in the formulas (6) to (8), task E _i Is placed at the device side, if E _j The data is unloaded to an edge end, and the equipment end is a data sending end; if task E _i Is placed at the edge end, E _j The data receiving end is placed at the equipment end which is a data receiving end; if E _j The data is unloaded to the cloud end, and the equipment end is a data sending end; if task E _i Placed in the cloud, E _j The data receiving end is placed at the equipment end, and the equipment end is a data receiving end.

1-3) establishing a weight factor alpha of task time delay and a weight factor beta of task energy consumption, and establishing a multi-objective optimization function G ═ alpha x T + beta x E of task unloading, so that the problem is converted into minimization of an objective function value.

The invention aims to provide a good unloading decision and resource allocation scheme for a diagnosis task split by a DNN fault diagnosis model under a cloud edge architecture, and finally, the weighted sum of the time and the energy consumption overhead of the diagnosis task is minimized. The objective function G is a weighted sum of the time and energy consumption of all diagnostic tasks.

G＝α*T+β*E (9)

Where α + β is 1, α represents a weighting factor for task latency, and β represents a weighting factor for energy consumption of a task under cloud-edge architecture. T represents the accumulated sum of time delay of each part of the diagnosis model under the edge cloud architecture, and E represents the accumulated sum of energy consumption of each part of the diagnosis model under the edge cloud architecture.

And under the resource constraint condition of the cloud edge system, taking the DNN diagnosis model as a diagnosis task, and unloading the task and distributing resources. From the foregoing modeling of the problem, the problem is transformed into a minimization of an objective function value, which is expressed as:

min(α*T+β*E) (10)

task if _i Is offloaded to an edge node edge _j Upper, object letterThe constraint on the number is:

in the formula (11), rf _ij Indicating the resources allocated by the task taski, which indicates the offloading to the edge node edge _j The sum of the resources allocated by the upper task is less than or equal to the edge node edge _j The sum of the resources of (c).

Task if _i And unloading the target function to a cloud cluster, wherein the constraint conditions of the target function are as follows:

in the formula (12), rf _ij Representing task _i And the formula represents that the sum of the resources distributed by the tasks unloaded to the cloud is less than or equal to the sum of the resources of the cloud.

S4: and (3) solving the minimum solution of the multi-objective optimization function by using an Average Network parameter DRQN algorithm (APAM-DRQN) Based on an Attention Mechanism.

The method utilizes APAM-DRQN to solve the optimal solution of the problem to minimize the objective function value G, and comprises the following specific steps:

2-1) setting three important elements, namely three parameters of State (State), Action (Action) and Reward (Reward):

(1) status of state

State S at time t _t Is defined as:

S _t ＝[task _i ，RA，RF] (13)

wherein, task _i Representing the execution priority, data volume, load and other related information of the ith DAG task; vector RA ═ RA ₀ ，ra ₁ ，…，ra _N ]，ra _i E (0, 1), wherein 0 represents that the task is executed at the edge end, and 1 represents that the task is executed at the cloud end. N is the total number of tasks; vector RF ═ RF ₀ ，rf ₁ ，…，rf _N ]，rf _i Representing the computational resources allocated by the ith diagnostic task.

(2) Movement of

The action determines the offload decision and the allocated computing resources for the current task, thereby adjusting the system state. The action is composed of three parts, namely an unloading decision from the equipment end to the optimal edge node, an unloading decision from the edge node to the cloud end and a computing resource distributed by the current task.

Action _i ＝[rx _i ，ry _i ，rf _i ] (14)

Wherein, rx _i Representing task _i Offloading scheme vectors from the device side to the edge node,

rx _i ＝(x _i，1 ，x _i，2 ，x _i，3 ，...，x _i，n ) (15)

x _i，j e {0, 1}, 0 represents the task _i Executed on the local equipment side, 1 stands for task _i At edge node j. ry _i Task representation _i Offload solution vector from edge node to cloud, ry _i E {0, 1}, 0 represents task _i Executing at the current edge node, 1 stands for task _i Is further offloaded to cloud execution.

rf _i Representing task _i The resource allocation scheme of (1).

(3) Reward

The reward value reward can evaluate the quality degree of the scheme obtained by executing the action in the current state, and the DRQN algorithm aims to obtain the maximum reward value and is associated with the target function to optimize and obtain the minimum value G.

reward＝(G _t -G _t-1 )/G _edge (16)

Wherein G is _t Indicating the target value at time t, G _t+1 Indicating action taken _t The rear state is represented by _t Becomes s _t+1 Target value of latter, G _edge And performing objective function values at the edge terminal for all tasks. Since a minimum target value needs to be found, the smaller the target value G,the higher the prize won. When G is _t -G _t+1 > 0, a positive reward may be obtained indicating that an action is taken _t The rear state is represented by _t Becomes s _t+1 Then a better objective function value can be obtained; otherwise, a negative reward is obtained.

2-2) inputting the state, the action and the reward into a fault diagnosis model, and iteratively calculating a task time delay weight factor alpha and a task energy consumption weight factor beta of a minimum solution obtained by a multi-objective optimization function G by using a DRQN algorithm of an average network parameter of an attention mechanism.

The embodiment of the invention improves the DRQN algorithm and provides an attention mechanism-based DRQN algorithm (APAM-DRQN) of average network parameters. Compared with the original DRQN algorithm, the algorithm provided by the invention is improved mainly in two parts. Firstly, aiming at the problem that the DRQN algorithm uses an epsilon-greedy strategy to cause lower environment exploration efficiency, the DRQN algorithm based on average network parameters is provided; secondly, the neural network structure and the output part of the network Q value in the DRQN algorithm are improved.

Compared with the original DRQN algorithm, the algorithm provided by the invention is improved mainly in two parts. Firstly, aiming at the problem that the DRQN algorithm uses an epsilon-greedy strategy to cause lower environment exploration efficiency, the DRQN algorithm based on average network parameters is provided; secondly, the neural network structure and the output part of the network Q value in the DRQN algorithm are improved.

(1) DRQN algorithm based on average network parameters

Compared with the DRQN algorithm, the APAM-DRQN algorithm provided by the invention needs to average a plurality of values of network parameters learned by the intelligent agent in the front when each round of training starts, so as to obtain an average disturbance network parameter, as shown in formula (17) and formula (18). And selecting a strategy according to the action made by the parameter, thereby improving the exploration efficiency of the algorithm on the environment.

Where k represents the number of value networks,

a layer i neural network weight value matrix representing a kth value network,

a layer i neural network bias value matrix representing a kth value network, and W represents an average disturbance network Q _p B represents the average disturbance network parameter Q _p A matrix of bias values. Next, the best action is filtered by the average perturbation network using a greedy strategy.

The network parameters in different periods have larger difference, and a new network parameter is obtained through average processing. Due to small variation of parameters in the network, linkage variation is generated in a plurality of time steps in the future, and the variation can influence the action selection strategy of the next stage of the APAM-DRQN algorithm, so that the exploration capability of the APAM-DRQN algorithm is improved. And because the average disturbance network obtained by the APAM-DRQN algorithm has high-efficiency exploration capability, a greedy strategy is adopted to replace an epsilon-greedy strategy.

(2) Network structure and output part for improving DRQN algorithm

FIG. 6 shows the algorithm flow of AMAP-DRQN, first interacting the environment with the agent to obtain samples (z) _t ，a _t ，r _t ，z _t+1 ) And storing the samples into an experience playback unit, then randomly sampling and splitting the samples, and respectively training a current value network and a target value network by using the split two parts. The 2 networks have the same structure and are composed of 1 LSTM layer and 2 full-connection layers, and the target value network carries out model optimization on parameters of the copied current value network every C iterations. The 2 networks are first paired by the LSTM layer to the current state s _t And the next state s _t+1 And calculating and deducing, and finally updating the weight of the current value network reversely according to the gradient value.

The neural network structure of the DRQN algorithm is improved and designed. Firstly, an attention mechanism is introduced, so that the system adaptively focuses attention on a target optimization index capable of reducing the weighted sum to the maximum extent, a more accurate decision is made, and the output of the system is connected with the action of the last state of the intelligent agent in series to serve as the input of the next layer. And then adding an LSTM layer to enable the model to have memory capacity on time sequence data, store previous state information, and be more beneficial to the time sequence reasoning of the state, so that the model has a better learning effect, wherein the unit number of the LSTM layer is the product of the edge node number n and the edge node server resource number F. Finally, two full connection layers are introduced, the number of units of the first full connection layer is 256, and the number of units of the second full connection layer is the product of the number n of edge nodes and the number F of edge node resources. And continuously adjusting the model learning weight through multiple iterations to obtain an optimal solution.

The output of the DRQN model is a Q value corresponding to each action, and then the Q value of network prediction is calculated by a greedy algorithm. Aiming at the problems, the output part of the DRQN model is improved, and the model directly outputs Q value scalars corresponding to a state space and an action space, so that the calculated amount is reduced, and the back propagation of MSE loss is more convenient. All Q values output by the DRQN model and the motion space a selected in the current state _t Dot multiplication yields an estimate Q (s, a) which is given a long term return by the Bellman equation of equation (19).

Q(s，a)＝Q(s，a)+k(R(s，a)+γmaxQ′(s′，a′)-Q(s，a)) (19)

Where Q (s, a) represents the predicted Q value of the network, and R (s, a) + γ maxQ ' (s ', a ') represents the target Q value. According to the above description, the flow of the APAM-DRQN algorithm proposed by the present invention is shown as algorithm 1.

Firstly, establishing a DAG graph for each part of the split DNN diagnostic model branch, and establishing a front-back dependency relationship among tasks to obtain a task priority list tasks of the DAG graph G. And inputting the task priority list tasks into an APAM-DRQN task unloading and resource allocation algorithm to obtain a corresponding unloading and resource allocation scheme, so that the task execution delay and the energy consumption are reduced, and the algorithm flow is shown as an algorithm 2.

The DRQN algorithm is only suitable for processing discrete actions, so that when the invention carries out side cloud task unloading and resource allocation, the action space value of each iteration allocation is a discrete value. For the unloading of the edge cloud task, the unloading vectors x and y take the discrete values of 0 or 1, and respectively represent the local execution or unloading of the task; for edge cloud resource allocation, the allocation value of the computing resource is theoretically a continuous real value, so that a resource allocation vector rf _i The value of (a) needs to be discretized.

If edge node edge _i Has a total resource value of

Discretizing the integer 1 as precision to obtain a resource allocation vector rf _i A value range of

Based on the steps, in the FDBECTO method provided by the invention, firstly, in a fault diagnosis multi-objective optimization module for edge cloud task unloading, according to the characteristics of each layer of a diagnosis model, the layer is divided by taking the layer as granularity, and a DAG directed acyclic graph is established to ensure the front-back dependency relationship among each part of the model. Meanwhile, fault diagnosis optimization indexes are determined, a multi-objective optimization function is determined, and a theoretical basis is provided for the cloud task unloading scheme at the fault diagnosis side; and then, in a fault diagnosis side cloud task unloading module based on the improved DRQN, multi-target optimization is carried out by the proposed APAM-DRQN algorithm, an optimal fault diagnosis side cloud task unloading scheme is provided, the time delay and the energy consumption of fault diagnosis are further reduced, and the diagnosis service quality is improved.

The invention discloses a test verification of a fault diagnosis optimization method based on edge cloud collaborative task unloading, which comprises the following steps:

the rolling bearing is a critical and easily-damaged precise mechanical part, and the rolling bearing is taken as a research object to perform experimental demonstration on the fault diagnosis optimization method based on the edge cloud cooperative task unloading.

1. Test environment

In the comparison experiment of the APAM-DRQN algorithm, a simulation experiment program is written by Python language, and the size of output data of a model layer and the task load of each part of the model are obtained according to a DNN diagnosis model structure. The computing power of a CPU (central processing unit) at the edge end is set to be 0.5GHz/s, the computing power of a cloud server is set to be 5GHz/s, the computing power, the data uploading power and the data receiving power of the edge end are respectively 0.8W, 0.1W and 0.025W, the uplink and downlink network bandwidths from the cloud end to the edge end are respectively 10Mbps and 8Mbps, the data volume of each DNN fault diagnosis model diagnosis task is Unif (300, 500) kbits, and the distance between each edge node and the central cloud is Unif (200, 2000) km.

The hyper-parameters of the APAM-DRQN algorithm are shown in Table 4.

TABLE 4 hyper-parameters of the APAM-DRQN algorithm

Test one: APAM-DRQN model training

The reward value can evaluate the quality degree of the current strategy, and the effectiveness of the algorithm is evaluated by adopting the average accumulated reward value.

Wherein,

representing the average jackpot value achieved for the ith iteration, r (i) representing the jackpot value achieved for the ith iteration,

the formula (2) is shown in formula (20).

In order to verify the effectiveness of the APAM-DRQN algorithm provided by the invention, a DQN (Deep Q-Network) algorithm and a DRQN algorithm are selected as baseline algorithms, a side cloud unloading and resource allocation scheme is provided for a DNN diagnosis model task by the three algorithms, and the obtained average accumulated reward value is shown in FIG. 7. When the DQN algorithm is trained for about 280 rounds, the obtained reward value is remarkably reduced, and the stability of the algorithm is poor; after 300 rounds of training of the DRQN algorithm, the model tends to converge; compared with the two algorithms, the APAM-DRQN algorithm provided by the invention has the advantages that on one hand, the convergence speed is higher, in addition, after the model is converged, the fluctuation of the reward value is smaller, and the stability of the algorithm is better. On the other hand, the algorithm achieves a higher average jackpot value than the DQN and DRQN algorithms. The effectiveness of the APAM-DRQN algorithm provided by the invention is verified.

And (2) test II: fault diagnosis multi-objective optimization contrast experiment

The APAM-DRQN algorithm can provide a side cloud unloading strategy and a resource allocation scheme for each part of the equipment fault diagnosis model, so that the diagnosis real-time performance is improved, and the energy consumption is reduced. The section compares 5 task unloading and resource allocation algorithms in terms of task quantity, cloud server CPU computing capacity and task data quantity. The 5 algorithms comprise an edge terminal non-uninstalling Algorithm (ALE), a cloud terminal based complete uninstalling Algorithm (ALC), a DQN algorithm, a DRQN algorithm and an APAM-DRQN algorithm provided by the invention. The DQN algorithm, the DRQN algorithm and the APAM-DRQN algorithm provided by the invention provide a side cloud unloading and resource allocation scheme for the DNN diagnosis model task.

The computing capacity of the central cloud server is set to be 5GHz/s, and the number of DNN diagnostic model diagnostic tasks is continuously increased. Fig. 8 shows the average weighted sum of the execution delay and the energy consumption of the diagnostic task under 5 algorithms as the number of tasks increases. From the graph, as the number of tasks is increased under the edge cloud architecture, the weighted sum values corresponding to the 5 algorithms are in an ascending trend. The reason is that under the edge cloud architecture, the time delay and energy consumption of the edge end and the cloud end for processing respective tasks are increased due to the increase of the number of the tasks, and the weighted sum is continuously increased. In addition, under the same number of tasks, the weighted sum corresponding to the ALE algorithm is the maximum, while the weighted sum corresponding to the time delay and the energy consumption of the APAM-DRQN algorithm provided by the invention is the minimum. This is because when the ALE algorithm is used, the edge is responsible for performing all tasks, and the computation power consumption of the edge is large, resulting in a large weighted sum. When the ALC algorithm is adopted, all tasks need to be uploaded to the cloud from the edge end, so that the time delay is increased, and the weighted sum is still high. The APAM-DRQN algorithm provided by the invention schedules tasks among side clouds, and the weighted sum obtained under different numbers of tasks is superior to the former two algorithms and superior to the DQN and DRQN side cloud cooperative unloading algorithm.

The DNN diagnosis model diagnosis task is divided into 5 parts, so that the computing capacity of the central cloud server is continuously improved. Fig. 9 shows the average weighted sum of the execution delay and the energy consumption of the diagnostic task under 5 algorithms as the computing power of the cloud server increases. According to the graph, with the continuous increase of the computing capacity of the cloud server under the edge cloud architecture, the corresponding weighted sum under the ALE algorithm is basically kept unchanged, and the weighted sum values corresponding to the other 4 algorithms are in a descending trend. When the ALE algorithm is adopted, all tasks are executed at the edge end, and the time delay and the energy consumption for executing the tasks are not changed due to the enhancement of the computing capacity of the cloud server; for the tasks unloaded to the cloud end, the computing time delay of the tasks is greatly reduced due to the enhancement of the computing capability of the cloud server, so that the weighted sum is in a descending trend. In addition, with the enhancement of the computing capability of the cloud server, the weighted sum obtained by the APAM-DRQN edge cloud collaborative unloading algorithm provided by the invention is superior to the other 4 algorithms.

The DNN diagnosis model diagnosis tasks are divided into 5 parts, the computing capacity of the central cloud server is set to be 5GHz/s, and the input data volume of the DNN diagnosis model diagnosis tasks is increased continuously. Fig. 10 shows the average weighted sum of the execution delay and the energy consumption of the diagnostic task under 5 algorithms as the data volume of the task is increased. From the graph, the weighted sum values corresponding to the 5 algorithms are in an ascending trend along with the continuous increase of the task data volume under the edge cloud architecture. This is because the larger the data amount of the task is, the larger the delay and power consumption of the task processing are. By contrast, the APAM-DRQN edge cloud collaborative unloading algorithm provided by the invention has the advantage that the weighted sum obtained under different task data amounts is superior to the other 4 algorithms.

In summary, by comparing the 5 algorithms in terms of the number of tasks, the computing capacity of the CPU of the cloud server, and the amount of task data, it can be obtained that the weighted sum of the delay and the energy consumption obtained by the APAM-DRQN edge cloud collaborative offload algorithm provided by the present invention is superior to the other 4 algorithms, and the effectiveness of the algorithm in reducing the diagnosis delay and the energy consumption is proved.

And (3) test III: comparison experiment for diagnosis time delay and energy consumption overhead

And dividing the diagnosis task of the diagnosis model into 5 parts, and setting the computing capacity of the central cloud server to be 5 GHz/s. Under 5 task unloading and resource allocation algorithms, the section compares the average time delay of the execution of the fault diagnosis task with the energy consumption overhead. The 5 algorithms include an Edge-based non-offload algorithm (ALE Locate on the Edge, abbreviated as ALE), a Cloud-based complete offload algorithm (ALE Locate on the Cloud, abbreviated as ALC), a DQN (Deep Q-Learning) algorithm, a DRQN algorithm, and an APAM-DRQN algorithm provided by the present invention. The DQN algorithm, the DRQN algorithm and the APAM-DRQN algorithm provided by the invention provide a side cloud unloading and resource allocation scheme for the DNN diagnosis model task. The cloud end has massive computing power and storage resources, so that the processing energy consumption cost of tasks in the cloud end is ignored in the experiment.

Fig. 11 shows the average delay of the execution of the diagnostic task under different algorithms, and the total delay includes the execution delay of the task and the network transmission delay. In the aspect of total time delay overhead, the sequence from large to small is as follows: ALC algorithm, ALE algorithm, DQN algorithm, DRQN algorithm and APAM-DRQN algorithm provided by the invention. The ALE algorithm total time delay only comprises task execution time delay, and the tasks do not need to be unloaded to a central cloud, so that network transmission is not needed to be carried out on the tasks; in the ALC algorithm, the DQN algorithm, the DRQN algorithm and the APAM-DRQN algorithm proposed by the present invention, the ratio of the transmission delay of the task in the network to the total delay is: 82.52%, 63.45%, 64.16% and 62.13%.

Fig. 12 shows the average cost of total energy consumption under different algorithms, and the total energy consumption cost includes the execution energy consumption of tasks and the network transmission energy consumption. In the aspect of total energy consumption overhead, the method comprises the following steps from large to small: ALE algorithm, ALC algorithm, DQN algorithm, DRQN algorithm and APAM-DRQN algorithm provided by the invention. The ALE algorithm total energy consumption only comprises task execution energy consumption, and the tasks do not need to be unloaded to a central cloud, so that network transmission is not needed to be carried out on the tasks; the ALC total energy consumption overhead only comprises task transmission energy consumption, as all tasks are unloaded to the central cloud, the cloud has massive computing resources, and the execution energy consumption of the tasks at the cloud is ignored; in the DQN algorithm, the DRQN algorithm and the APAM-DRQN algorithm proposed by the present invention, the ratio of transmission energy consumption to total energy consumption overhead of a task in a network is respectively: 58.80%, 58.75% and 59.43%. Experimental results show that the time delay and the energy consumption of the APAM-DRQN edge cloud collaborative unloading algorithm provided by the invention are superior to those of other 4 algorithms, and the effectiveness of the algorithm in reducing the diagnosis time delay and the energy consumption is proved.

Conclusion

The invention provides a fault diagnosis optimization method based on edge cloud collaborative task unloading. Firstly, in a fault diagnosis multi-objective optimization module for unloading edge cloud tasks, according to the characteristics of each layer of a diagnosis model, the layers are divided by taking the layers as granularity, and a DAG directed acyclic graph is established to ensure the front-back dependency relationship among all parts of the model. Meanwhile, fault diagnosis optimization indexes are determined, a multi-objective optimization function is determined, and a theoretical basis is provided for the fault diagnosis side cloud task unloading scheme; and then, in a fault diagnosis side cloud task unloading module based on the improved DRQN, multi-target optimization is carried out by the proposed APAM-DRQN algorithm, an optimal fault diagnosis side cloud task unloading scheme is provided, the time delay and the energy consumption of fault diagnosis are further reduced, and the diagnosis service quality is improved.

The experimental results show that: comparing the 5 task unloading and resource allocation algorithms in terms of the number of tasks, the computing capacity of a CPU of the cloud server and the data volume of the tasks, the time delay and the energy consumption weighted sum obtained by improving the DRQN algorithm provided by the invention are lower than those of other 4 algorithms, and the effectiveness of the algorithm in reducing the diagnosis time delay and the energy consumption overhead is verified. In future work, indexes such as cost and load balance are further considered, and reliability and stability of diagnosis service are improved.

The invention further provides an electronic device, which includes a memory and a processor, wherein the memory stores a computer program operable on the processor, and the processor implements the steps of the method when executing the computer program.

The functions, if implemented in the form of software functional units and sold or used as a stand-alone product, may be stored in a computer readable storage medium. Based on such understanding, the technical solution of the present invention may be embodied in the form of a software product, which is stored in a storage medium and includes instructions for causing a computer device (which may be a personal computer, a server, or a network device) to execute all or part of the steps of the method according to the embodiments of the present invention. And the aforementioned storage medium includes: various media capable of storing program codes, such as a usb disk, a removable hard disk, a Read-Only Memory (ROM), a Random Access Memory (RAM), a magnetic disk, or an optical disk.

Finally, it should be noted that: the above-mentioned embodiments are only specific embodiments of the present invention, which are used for illustrating the technical solutions of the present invention and not for limiting the same, and the protection scope of the present invention is not limited thereto, although the present invention is described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: any person skilled in the art can modify or easily conceive the technical solutions described in the foregoing embodiments or equivalent substitutes for some technical features within the technical scope of the present disclosure; and the modifications, changes or substitutions do not make the essence of the corresponding technical solutions depart from the scope of the technical solutions of the embodiments of the present invention. Are intended to be covered by the scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A fault diagnosis optimization method based on edge cloud collaborative task unloading is characterized by comprising the following steps:

dividing a fault diagnosis model into a plurality of parts according to a diagnosis model structure, taking a layer as granularity, and floating point Operations Per Second (FLOPs), wherein each part forms a diagnosis task and is modeled as a Directed Acyclic Graph (DAG) to represent the execution sequence among the parts;

2. The method of claim 1, wherein a first fault diagnosis model branch of the plurality of fault diagnosis model branches comprises 4 convolutional layers and 1 pooling layer.

3. The method of claim 2, wherein a second fault diagnosis model branch of the plurality of fault diagnosis model branches comprises 5 convolutional layers and 1 pooling layer.

4. The method of claim 3, wherein a third fault diagnosis model branch of the plurality of fault diagnosis model branches comprises 3 convolutional layers, 3 long short term memory network layers, and 1 pooling layer.

5. The method of claim 1, wherein the step of analyzing the unloading delay and unloading energy consumption of the diagnostic task of the fault diagnosis model and establishing a multi-objective optimization function for task unloading comprises:

dividing the unloading time of the diagnosis task into calculation time and transmission time, and analyzing the unloading time of the diagnosis task to obtain total operation time T;

dividing the unloading energy consumption of the diagnosis task into calculation energy consumption and transmission energy consumption, and analyzing the unloading energy consumption of the diagnosis task to obtain total energy consumption E;

and establishing a weight factor alpha of task time delay and a weight factor beta of task energy consumption, and establishing a multi-objective optimization function G of task unloading, wherein the function G is alpha T + beta E.

6. The method of claim 5, wherein the step of solving the minimum solution for the multi-objective optimization function using a DRQN algorithm based on average network parameters of an attention mechanism comprises:

setting three parameters of state, action and reward;

and inputting the state, the action and the reward into a fault diagnosis model, and performing iterative computation by using a DRQN algorithm of average network parameters of an attention mechanism to obtain a weight factor alpha of task time delay and a weight factor beta of task energy consumption of a minimum solution of the multi-objective optimization function G.

7. An electronic device comprising a memory and a processor, wherein the memory stores a computer program operable on the processor, and wherein the processor implements the steps of the method of any of claims 1 to 6 when executing the computer program.

8. A computer readable storage medium having stored thereon machine executable instructions which, when invoked and executed by a processor, cause the processor to execute the method of any of claims 1 to 6.