CN114742166A - Communication network field maintenance model migration method based on time delay optimization - Google Patents
Communication network field maintenance model migration method based on time delay optimization Download PDFInfo
- Publication number
- CN114742166A CN114742166A CN202210399528.5A CN202210399528A CN114742166A CN 114742166 A CN114742166 A CN 114742166A CN 202210399528 A CN202210399528 A CN 202210399528A CN 114742166 A CN114742166 A CN 114742166A
- Authority
- CN
- China
- Prior art keywords
- similarity
- task
- model
- migration
- particle
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Software Systems (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Physics & Mathematics (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Medical Informatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Data Exchanges In Wide-Area Networks (AREA)
Abstract
The invention discloses a communication network field maintenance model migration method based on time delay optimization, which comprises the following steps: firstly, analyzing and calculating type similarity, name similarity and interpretation information similarity respectively based on service requirements of the similarity, reducing influence of irrelevant information, improving service matching accuracy, and finally calculating and obtaining comprehensive similarity according to the weight of each similarity; secondly, migration modes under different conditions are selected according to the similarity of service requirements, and model migration is carried out by using modes of direct migration, freezing of a backbone network and multi-stage migration respectively, so that the problems of complex intelligent operation and maintenance environment and low migration efficiency of a communication network are solved; finally, considering various indexes of task delay, edge node resources and task resource requirements, and using a multi-priority queue model and an improved discrete particle swarm algorithm to finish the migration task and the maintenance task unloading in the scene of field maintenance of the communication network, so that the aims of minimizing the task delay and improving the task completion efficiency are fulfilled.
Description
Technical Field
The invention relates to the technical field of communication network field maintenance, in particular to a communication network field maintenance model migration method based on time delay optimization.
Background
With the continuous expansion and the increasing complexity of the communication network, the problems of large workload of field maintenance, more management problems, lack of automatic supporting tools and the like of the communication network are increasingly highlighted. In response to these problems, the field maintenance architecture of the communication network also starts to adopt the mobile edge computing architecture to improve the quality and efficiency of field maintenance of the communication network. In the scene, the model used by the field maintenance of the communication network is trained by different edge nodes, and the model is relatively independent. However, intelligent terminal devices (such as wearable devices, inspection unmanned planes, and inspection robots) accessing edge nodes in field maintenance work of a communication network all have mobility, which causes changes of the edge nodes accessible to the intelligent terminal devices. However, these image recognition and fault diagnosis models are not deployed to all edge nodes, and the models are difficult to be directly migrated to other edge nodes for use, and if the trained models are directly migrated, the model migration precision may be too poor or even the models cannot be used. In addition, training of the model is involved when the edge node performs model migration, a large amount of computing resources are consumed, and execution of the model migration affects execution of normal maintenance tasks.
In order to solve the above problems, there are methods related to model migration in the prior art, such as:
the technical scheme 1: CN202111206993.4 proposes a method and an apparatus for model migration, where the method includes: acquiring a target model, a verification data set and a parameter fine tuning data set; knowledge distillation processing is carried out on the target model to obtain a migration model, and optimization processing is carried out on the migration model according to the verification data set and an error function corresponding to the knowledge distillation processing in the knowledge distillation processing process of the target model; and performing self-supervision training on the migration model by using the parameter fine-tuning data set so as to perform fine tuning on the parameters of the migration model. By adopting the technical means, the problem that a robot model migration method is lacked in the prior art is solved.
The technical scheme 2 is as follows: CN202110510119.3 proposes a model migration method and apparatus, where the model is a score card model, and the method includes: training to obtain a first machine learning model corresponding to the source scene based on a first training sample of the source scene and a sample label corresponding to the first training sample data; acquiring a second unlabeled training sample of the target scene; wherein the second training sample has the same feature space as the first training sample; and calibrating the trained model parameters of the first machine learning model based on adjusting that the characteristic distribution of the second training sample approaches the characteristic distribution of the first training sample, so as to obtain a second machine learning model after model parameter calibration, and thus completing model migration.
Technical scheme 3: CN202010219449.2 provides a model migration training method, device, equipment and storage medium, relating to the field of artificial intelligence. The specific implementation scheme is as follows: taking the network parameters of at least two migration layers in the source model as the initial parameters of the associated migration layers in the target model; constructing an objective function according to the distance between the training parameters associated with the at least two migration layers and the initial parameters; and training an object model comprising initial parameters based on the object function. Finally, the inheritance of the information of the source model and the self-adaptation of the target model are realized, the over-fitting phenomenon in the model migration training process is avoided, and the generalization capability of the target model is improved.
In the existing model migration method, the migration model is obtained by knowledge distillation of the target model in the technical scheme 1, and the model is finely adjusted by using a data set, but the knowledge distillation method is harsh in application conditions and sensitive to parameters. In the technical scheme 2, the trained model parameters of the first machine learning model are used for calibration to obtain the second machine learning model after the model parameters are calibrated, so that model migration is completed. According to the technical scheme 3, by introducing the distance between the training parameters and the initial parameters of the migration layers, the model migration and training conditions of each migration layer are considered in the model training process, but only the relation between the parameters is concerned, and no research is carried out on the service requirements of the source model and the target model. In addition, the technical schemes 1, 2 and 3 do not consider the influence of model migration on the execution of a common maintenance task in a communication network field maintenance scene, and the schemes are not suitable for the communication network field maintenance scene.
Disclosure of Invention
Aiming at the technical problems, the invention provides a model migration method based on service requirement similarity and time delay optimization and suitable for field maintenance of a communication network, which comprehensively considers a migration task and a maintenance task, can efficiently complete model migration, effectively reduces task completion time delay and ensures task completion rate.
In order to achieve the above purpose, the invention provides the following technical scheme:
a communication network field maintenance model migration method based on time delay optimization comprises the following steps:
s1, analyzing and calculating type similarity, name similarity and interpretation information similarity respectively based on the service requirements of the similarity, and finally calculating to obtain comprehensive similarity according to the weight of each similarity;
s2, measuring the correlation of the model by using the service requirement similarity, and selecting the migration modes under different conditions according to the service requirement similarity; when the service requirement of the new model is completely the same as that of the model to be migrated, directly migrating the model; freezing the backbone network when the service requirement of the new model is similar to that of the model to be migrated; when the similarity between the service requirement of the new model and the service requirement of the model to be migrated is poor, migrating the model by adopting multi-stage migration learning;
s3, taking into account various indexes of task delay, edge node resources and task resource requirements, and using a multi-priority queue model and an improved discrete particle swarm algorithm to finish the migration task and the maintenance task unloading in the scene of field maintenance of the communication network, so as to minimize the task delay.
Further, the method for calculating the type similarity in step S1 includes: classifying and marking different data characteristic spaces, and assuming that a source model RC and a target model TC are model types, wherein the type similarity is as follows:
if the data feature spaces of the source model and the target model are the same, the type similarity is 1, otherwise, the type similarity is 0.
Further, the method for calculating the name similarity in step S1 includes:
first, a name vector is calculated according to equation (2):
where name _ vec is the name vector, veciIs the ith word vector in the name;
secondly, the similarity of two name vectors is calculated according to the formula (3):
wherein, SimnameFor name similarity, name _ vec1 is the first name vector and name _ vec2 is the second name vector.
Further, the method for calculating the similarity of the interpretation information in step S1 includes: firstly, performing word segmentation analysis on explanatory information by using a BilSTM-CRF algorithm according to the field maintenance characteristics of a communication network, classifying texts by using a trained model, judging the influence of entity types on the similarity of the model, extracting texts with larger influence for similarity calculation, and then performing similarity calculation on words with the same type.
Further, step S1 extracts the device element and the text of the failure to perform similarity calculation, the method includes: analyzing the interpretation information of the two models to obtain an equipment element set and a fault set of the interpretation information of the two models and a service requirement S1The information of m equipment elements is analyzed to be E1＝{e11，e12，...，e1mF, and p failure messages, F1＝{f11，f12，...，f1p}; service S2The information of n equipment elements is analyzed to be E2＝{e21，e22，...，e2nF, and q failure messages, F2＝{f21，f22，...，f2q}；
Calculating E according to equation (4)1，E2The similarity matrix of (2):
according to the formula (4), traversing the matrix to take the maximum value of each column to obtain a maximum value set P ═ P1，p2，…，plWhere l ═ min (m, n);
calculating device element similarity S according to formula (5)e：
Calculating F according to equation (6)1，F2Wherein each item is the similarity of two fault information:
according to the formula (6), traversing the matrix to obtain the maximum value of each column to obtain a maximum value set Q { Q }1，q2，…，qkWhere k ═ min (p, q);
calculating the fault similarity S according to the formula (7)f：
And (3) obtaining the similarity of the interpretation information according to a formula (8) by combining the similarity of the equipment elements and the similarity of the faults:
Simexplain＝ω1Se+ω2Sf (8)
in the formula (8), SexplainTo interpret information similarity, SeAs device component similarity, SfIs the fault similarity; omega1，ω2Is a weight coefficient, and ω1+ω2＝1。
Further, the technical method for integrating the similarity in step S1 includes: on the basis that the type similarity is 1, the comprehensive similarity is obtained according to the formula (9):
S＝Simcategory(ω1·Simname+ω2·Simexplain) (9)
in the formula (9), ω1，ω2The name similarity and the weight of the interpretation information similarity respectively satisfy omega1+ω2＝1。
Further, the multi-priority queue model in step S3 includes three queues, which are an F pile, an M first-in first-out queue, and an S first-in first-out queue, respectively; firstly, dividing the task into three grades according to the importance degree of the task and the time delay sensitivity of the task, wherein P1 is the task with the highest priority, P2 is the task with the medium priority, and P3 is the task with lower time delay requirement; then adding the tasks at the P1 level into an F heap, adding the tasks at the P2 level into an M queue, and adding the tasks at the P3 level into an S queue; f is a large top heap, the priority is determined according to the task deadline, the task with the highest top heap priority is executed each time the task is selected for execution, and when the top heap task does not meet the execution requirement, namely the task start time is greater than the current time, the first task in the M queue is selected for execution; when a new task is added into the F heap, calculating the priority of the task and inserting the task into the heap; if the task in the M queue approaches the deadline, the task needs to be immediately allocated to an MEC (Mobile Edge computing) service node for execution, otherwise, the task is discarded, and the priority of the task is promoted at this time and the task is inserted into an F heap; and if the task in the S queue approaches to the deadline time, the priority of the task is promoted and inserted into the tail of the M queue to wait for execution.
Further, the modified discrete particle swarm algorithm in step S3 is:
firstly, setting n sub-populations, wherein each sub-population has m particles, and in each sub-population, two types of particles, namely an exploration type and a navigation type, exist; in each iteration process, the algorithm updates the position and the speed of each particle; at the ith iteration, the velocity of the kth particle in the population j in the d dimension is updated as:
in the formula (27), ω is the inertia coefficient of the particle, c1Is a dynamic learning factor of the particle, c2For the sub-population dynamic learning factor, c3Is a population dynamic learning factor, r1，r2And r3Is a random coefficient;
He Wei
respectively representing the current optimal positions of particles, sub-populations and populations; wherein, type1 represents that the particle is a navigation particle, and the optimal solution of the individual and the population of the particle is considered; the type2 represents that the particle is an exploration type particle, and the optimal solution of the individual particle and the sub-population is considered;
the variation of the position is represented by probability, the value is-1, 0,1, and the probability function is:
mapping the velocity value to an interval [0,1] through a probability function, and obtaining a particle position updating formula by using the direction of the particle velocity as follows:
equation (29) is a position update equation in the d-dimension for the kth particle, and r is a random number in [0,1 ].
Further, after diversity and discretization processing are performed on the particle swarm algorithm, the inertial weight and the learning factor are redefined, and firstly, the inertial weight ω adopts a nonlinear decreasing function:
wherein, ω ismax＝0.9,ωminWhen omega ranges from 0.9 to 0.4, the algorithm reaches the optimal value, I is the current iteration number, and I is the maximum iteration number;
in an iterative process, a dynamic learning factor, an individual learning factor c, is used1Gradually decreasing, population learning factor c2And c3Gradually increasing, at the i-th iteration c1、c2And c3Is shown in equations (31) (32) (33):
wherein I is the current iteration number, and I is the integral iteration number.
Further, the method of step S3 is: screening the MEC service nodes which can be distributed to each migration task according to the similarity of the service requirementsObtaining an initial position matrix X and an initial speed matrix V of particles according to a task set, unloading the task according to the value of each particle in the population, queuing each MEC service node by using a multi-priority queue model, working out the fitness of the particle in the iteration, updating the historical maximum fitness and the historical maximum fitness position of the particle, updating the sub-population, the global maximum fitness and the maximum fitness position, updating the inertia weight according to a formula (30), updating the learning factor according to formulas (31), (32) and (33), if the particle is a navigation type, updating the particle speed according to type1 in the formula (27), if the particle is an exploration type, updating the particle speed according to type2 in the formula (27), obtaining a probability matrix P according to the particle speed, updating the particle position X according to the formula (29), and calculating the minimum average time delay TminCalculating an optimal task allocation vector
And repeating the iteration until the iteration times are reached.
Compared with the prior art, the invention has the beneficial effects that:
the communication network field maintenance model migration method based on the time delay optimization, provided by the invention, designs a model migration strategy according to the service requirements of a source model and a target model, and minimizes the completion time delay of a migration task and a maintenance task.
Firstly, pre-screening communication network field maintenance service before model migration, and calculating type similarity, name similarity and interpretation information similarity. When the similarity of the interpretation information is calculated, a BilSTM-CRF algorithm is used for identifying the key entity type of the communication network, and the similarity calculation is carried out on the specific entity type, so that the influence of irrelevant information is reduced, and the accuracy of service matching is improved. And finally, calculating according to each similarity weight to obtain the comprehensive similarity.
Secondly, aiming at the problem of model migration, because part of edge nodes have the problem of less collected data, the invention adopts a migration learning method. And selecting migration modes under different conditions according to the similarity of service requirements, and performing model migration by using direct migration, freezing of a backbone network and multi-stage migration respectively, so that the problems of complex intelligent operation and maintenance environment and low migration efficiency of a communication network are solved.
Finally, as the communication network maintenance tasks are complex and part of the tasks are urgent, the invention considers the unloading of the migration tasks and the maintenance tasks at the same time, fully considers various indexes such as task time delay, edge node resources, task resource requirements and the like, uses an improved discrete particle swarm algorithm to finish the unloading of the migration tasks and the maintenance tasks in the scene of field maintenance of the communication network, and achieves the purposes of minimizing the task time delay and improving the task completion efficiency.
In addition, the method and the system respectively carry out simulation experiments on the service demand similarity model, the migration strategy based on the service demand similarity model and the task unloading strategy of minimizing time delay. Simulation results show that the model migration strategy based on the similarity of service requirements can reduce the blindness of model migration, relieve the problem of limited field maintenance data volume of a communication network, and realize efficient model migration according to the service requirements. Meanwhile, the task unloading algorithm with the minimized time delay can ensure that the migration task and the maintenance task are executed efficiently at the same time, effectively reduces the time delay of the completion of the maintenance task while completing the model migration, and improves the task completion rate.
Drawings
In order to more clearly illustrate the embodiments of the present application or technical solutions in the prior art, the drawings required in the embodiments will be briefly described below, it is obvious that the drawings in the following description are only some embodiments described in the present invention, and other drawings can be obtained by those skilled in the art according to these drawings.
Fig. 1 is a model of a communication network field maintenance system according to an embodiment of the present invention.
Fig. 2 is a comparison graph of the multi-stage migration effect provided by the embodiment of the present invention.
Fig. 3 is a diagram illustrating task latency for different algorithms according to an embodiment of the present invention.
Fig. 4 is a diagram illustrating an effect of the number of migration tasks on the total time for completion of the maintenance task according to an embodiment of the present invention.
Fig. 5 illustrates an influence of the number of migration tasks on the task completion rate of the maintenance task according to an embodiment of the present invention.
Fig. 6 is a diagram illustrating an influence of the number of migration tasks on the average completion time of the migration tasks according to an embodiment of the present invention.
Detailed Description
In order to make the technical solutions of the present invention better understood by those skilled in the art, the present invention will be further described in detail with reference to the accompanying drawings and examples.
The invention aims at the scene of communication network field maintenance, designs a model migration strategy according to the service requirements of a source model and a target model, and establishes a model migration method suitable for communication network field maintenance by solving the problem of minimizing the completion time delay of a migration task and a maintenance task. The model migration method comprises the following steps:
similarity-based service demand analysis
In order to avoid resource burden and cost overhead caused by frequent migration of a model in communication network field maintenance, the invention designs a service requirement similarity analysis method in a communication network field maintenance scene, and aims to screen MEC service nodes with similar service requirements for a trained model for migration, reduce the migration blindness and improve the migration efficiency. And analyzing the service requirements based on the similarity to respectively calculate the type similarity, the name similarity and the interpretation information similarity, and finally calculating according to the weight of each similarity to obtain the comprehensive similarity.
(1) Similarity of type
In the field maintenance process of the communication network, the data characteristic spaces of different models have diversity and comprise image data, voice data, text data and the like. When the feature spaces of the source domain and the target domain are the same, for example, both the source domain and the target domain are images, the homogeneous transfer learning is performed, and otherwise, the heterogeneous transfer learning is performed. Since the heterogeneous transfer learning process is complex and relies too much on a priori knowledge, only isomorphic transfer learning is concerned. Based on the above, in the field maintenance scene of the communication network, different data feature spaces are classified and labeled, for example, image data is 1, voice data is 2, and text data is 3. Because only isomorphic transfer learning is considered, the data feature space types of the source model and the target model need to be completely matched for transfer. Assuming that the source model RC and the destination model TC are model types, the similarity of the types is as follows:
if the data feature spaces of the source model and the target model are the same, the type similarity is 1, otherwise, the type similarity is 0.
(2) Similarity of names
The name information of the communication network field maintenance model is stored in a text mode, word vectors are calculated by using a word2vec method, sentence vectors are calculated according to the word vectors, and then cosine similarity of the sentence vectors is calculated to obtain name similarity. First, a name vector is calculated by equation (2), name _ vec being the name vector, veciIs the ith word vector in the name.
Second, the similarity of the two name vectors is calculated. SimnameFor name similarity, name _ vec is the first name vector, name _ vec2 is the second name vector, and the similarity between the two vectors is calculated according to formula (3), i.e., the name similarity.
(3) Similarity of interpretation information
The explanation information of the model is described by a piece of text, which is brief introduction and explanation of the model. In order to improve the pertinence of the similarity of the interpretation information, the invention firstly uses the BilSTM-CRF algorithm to carry out word segmentation analysis on the interpretation information according to the field maintenance characteristics of the communication network, and then carries out similarity calculation on the words of the same type, thereby improving the accuracy of the calculation.
TABLE 1 communication network field maintenance entity types and description thereof
As shown in table 1, the entity types maintained on site in the communication network are divided into 9 types. Based on entity type classification, the invention trains communication network field maintenance data by using a BilSTM-CRF model. And when the similarity of the interpretation information is calculated, classifying the texts by the trained model. Since many entity types have little influence on judgment of the similarity of the model in the interpretation information text, such as roles, time, weather and the like, only the text of the equipment element and the fault is extracted for similarity calculation.
According to the method, the interpretation information of the two models is analyzed to obtain the equipment element set and the fault set of the two models of interpretation information. Service requirement S1The information of m equipment elements is analyzed to be E1＝{e11，w12，...，e1mF, and p failure messages, F1＝{f11，f12，...，f1p}. Service S2The information of n equipment elements is analyzed to be E2＝{e21，e22，...，e2nF, and q failure messages, F2＝{f21，f22，...，f2q}。
First, calculate E1，E2As shown in equation (4), where each term is the similarity of the two pieces of equipment component information.
According to the formula (4), traversing the matrix to take the maximum value of each column to obtain a maximum value set P ═ P1，p2，…，plWhere l ═ min (m, n). Computing device element similarity Se，SeThe larger the size the similarityThe greater the degree.
Similarly, F can be calculated according to equation (6)1，F2The similarity matrix of (2). Each of which is the similarity of two fault messages.
According to the formula (6), traversing the matrix to obtain the maximum value of each column to obtain a maximum value set Q { Q }1，q2，…，qkWhere k is min (p, q). Calculating fault similarity Sf，SfThe larger the similarity is.
And finally, the similarity of the interpretation information can be obtained by combining the similarity of the equipment elements and the similarity of the faults. In the formula (8), SexplainTo interpret information similarity, SeAs device component similarity, SfIs the fault similarity. Omega1，ω2Is a weight coefficient, and ω1+ω2＝1。
Simexplain＝ω1Se+ω2Sf (8)
(4) Integrated similarity
The type similarity ensures that the data characteristic space of the source model and the data characteristic space of the target model are isomorphic resources, the model can be migrated, and the name similarity and the interpretation information similarity ensure that the service requirements of the source model and the target model have similarity. Therefore, when the service requirement comprehensive similarity is evaluated, the influence of the name similarity and the interpretation information similarity needs to be comprehensively considered on the basis that the type similarity is 1. Finally, the comprehensive similarity is obtained as follows:
S＝Simcategory(ω1·Simname+ω2·Simexplain) (9)
as in formula (9), ω1，ω2The name similarity and the weight of the interpretation information similarity respectively satisfy omega1+ω2＝1。
(II) model migration strategy based on service demand similarity
Aiming at the service demand analysis algorithm based on the similarity, the invention provides a model migration strategy based on the similarity of the service demands. In a communication network field maintenance scenario, MEC service nodes at different locations of a communication network often have the same or similar service requirements. And the training of large models such as YOLOv3 is time-consuming and requires a large amount of data support. Therefore, after the MEC service nodes complete the training deployment of the model, if other MEC service nodes have the same or similar service requirements, the service model can be migrated to other MEC service nodes, so that the training time of other nodes is reduced. Because the size of the training data volume and the difference between the correlation of the new data and the previous data are directly related to the selection of the migration mode, the method uses the service requirement similarity to measure the correlation of the model, and selects the migration mode under different conditions according to the service requirement similarity.
Firstly, when the service requirement of the new model is completely the same as that of the model to be migrated, the model is migrated directly without additional training. In this case, the model can be migrated directly from the source MEC service node to the destination MEC service node without additional training.
Secondly, when the service requirements of the new model are similar to those of the model to be migrated, in this case, because the service requirements are highly similar, the basic model of the source model can be used as a feature extractor of the new model, so that the new model can benefit from the learned features of the source model, a good detection effect can be achieved only by training part of the network, and the time consumed by model training can be greatly reduced. The specific scheme used by the invention is that all convolutional layers in a backbone network of a YOLOv3 model are frozen, the model to be migrated and the weight parameters thereof are loaded, the model is trained, and only the parameter values of the unfrozen layers are updated.
And finally, when the similarity between the service requirement of the new model and the service requirement of the model to be migrated is poor, the model is migrated by adopting multi-stage migration learning. Under the condition of poor similarity of service requirements, the method provided by the invention firstly extracts the universality characteristics of the image by using a method of freezing partial convolution layer, prevents over-fitting, reduces the model training time, and then carries out fine adjustment on the basis, thereby improving the model precision. The specific scheme is that the number of selected layers is frozen in the first step, and parameters after the layers are frozen are trained. And secondly, fine tuning is carried out, and all parameters in the network are trained. Different learning rates are respectively adopted in the two stages, the first stage adopts a larger learning rate, and the fine adjustment adopts a smaller learning rate.
Task unloading strategy for minimizing time delay
(1) System model
Fig. 1 is a model of a field maintenance system for a communication network. Under the scene of communication network field maintenance, after the maintenance task is obtained, the operation and maintenance personnel wearing the intelligent wearable equipment, the inspection unmanned aerial vehicle and the inspection robot carry out maintenance work. The intelligent terminal equipment can upload the collected field audio and video data to the MEC service node and the cloud center to perform deep learning model training, and the MEC service node with similar service requirements can be selected to perform model migration after the model training is completed, so that model multiplexing is completed. In the working process, the intelligent terminal equipment can process the tasks in real time through the deployed deep learning model and can also unload the tasks to the MEC service node and the cloud center for processing. And finally, after the maintenance task is completed, updating a task completion result to the MEC service node and the cloud center to realize resource synchronization.
a) Problem formalization
Definition 1: t isn，m，qTime delay for completing maintenance tasks. And the maintenance task n generated by the intelligent terminal device m unloads the calculation task to the MEC service node k through the channel q for processing. The time delay of the maintenance task under the scene of field maintenance of the communication network comprises queuing time delay, unloading time delay, execution time delay andand returning the time delay by the calculation result. Since the downlink propagation speed is fast and the data volume of the calculation result is small, the time for returning the calculation result is very short, and therefore, the part of the time delay is ignored. The delay of the maintenance task can then be calculated by the following formula:
referring to shannon's theorem in the calculation of the uploading speed of the intelligent terminal device m, the uploading speed of the terminal device m for task unloading on the transmission channel q is as follows:
wherein, BqFor the bandwidth of channel q, Pm，qOffloading data Power for an Intelligent terminal device m over a channel q, Gm，qIn order to obtain the gain of the channel,
is additive white gaussian noise power.
According to Vm，qComputing
Wherein d isnIndicating the data size of task n.
Assuming that an MEC service node has multiple channels for terminal devices to perform task offloading, each channel serves one intelligent terminal device, and computing resources of the MEC service node are shared by maintenance tasks and migration tasks offloaded to the node. Then calculate
An，q，k1, indicating that maintenance task n is offloading the compute unit to MEC node k over channel q, an，q，kI.e. indicating that channel n is in an idle state, 0. A. thek1, a migration task is being performed on MEC service node kkAnd 0, that is, it means that the migration task is not executed on the current MEC node k. By
And calculating the number of intelligent terminal equipment users served by the MEC service node k. FnIs the computational load of task n.
Finally, the total time delay of the maintenance task is:
definition 2: t isn，k′Delay for completing migration tasks. MEC node k' offloads the migration task to MEC service node k. The time delay of the migration task under the scene of field maintenance of the communication network is related to queuing time delay, transmission time delay and execution time delay.
We calculate below
Wherein d isnData size, V, representing migration task nk′，kServing node k' data transmission rate for MEC:
the transmission rate from MEC service node k' to MEC service node k is Vk′，k。
Is calculated next
The calculation method is the same as that of the maintenance task, but the calculation resource of the migration task at the same time is twice that of the maintenance task.
The total latency of the migration task is:
definition 3: em，nIs the energy consumption of the intelligent terminal device m when executing the task n. In a communication network field maintenance scenario, the MEC service node typically has sufficient energy to perform tasks, and therefore, the consumption of MEC service node resources is not considered. However, since the energy of the intelligent terminal device is limited, the energy consumption of the intelligent terminal device is limited, and the energy calculation formula of the terminal device is as follows:
wherein, Pm，qThe power of the terminal device(s),
is the task transmission time.
According to the communication network field maintenance system model, unloading decisions are made for maintenance tasks generated by intelligent terminal equipment and migration tasks generated by MEC service nodes, and the purpose is to minimize the time delay for completing all tasks. Thus, the optimization objective is:
wherein the constraint conditions are as follows:
(1) the task needs to be executed before a tolerant time, otherwise the task is discarded.
C1：
(2) The migration task must select MEC service nodes with similar service requirements for migration. Calculating the service requirement s of the current model of the migration taskmodelService requirement set S with MEC service node k1＝{s1，s2，...，siAnd if the maximum value meets a threshold value, selecting the current MEC service node k for migration.
C2：
(3) MEC service node k can only execute one migration task at the same time, AnWhen 1, the ith task is a migration task, AnThe ith task is a maintenance task when 0.
C3:
(4) One intelligent terminal device can only be connected with one channel at the same time. m isqA value of 1 indicates that the intelligent terminal m is unloaded through the channel q.
C4:
(5) One channel can only be connected with one intelligent terminal device at the same time. A. them，qA value of 1 indicates that the intelligent terminal m is unloaded through the channel q.
C5:
(6) The total energy consumption of the intelligent terminal does not exceed a specified threshold value within a period of time.
C6:Em≤Ethreshold (26)
(2) Multi-priority queue model
Different queuing algorithms adopted for the tasks on the MEC service nodes have different influences on task completion delay, so that a multi-priority queue model suitable for field maintenance of a communication network needs to be designed. The invention divides tasks into a migration task and a maintenance task, wherein the migration task is generated by an MEC service node, and the maintenance task is generated by intelligent terminal equipment, wherein the maintenance task comprises an inspection image processing task, a voice recognition task, a video processing task and the like.
Because the field maintenance task types of the communication network are complex, the invention is divided into three grades according to the importance degree of the tasks and the task delay sensitivity. The P1 is the task with the highest priority, and is the migration task in the present invention, because whether the migration task is completed or not will directly affect the completion of the subsequent maintenance task. P2 is a medium priority task with strict delay requirements and needs to be completed as soon as possible, for example, an image task that needs to be processed in real time during field inspection for field maintenance of a communication network. P3 is a task with low delay requirement, and has a certain tolerance time for more resources required in the processing process, such as a video processing task in the polling process. Based on the above, the invention designs a multi-priority queue model.
Firstly, the invention establishes three queues, namely an F pile, an M first-in first-out queue and an S first-in first-out queue. And adding the tasks at the P1 level into the F heap, adding the tasks at the P2 level into the M queue, and adding the tasks at the P3 level into the S queue. And F, determining the priority according to the task deadline, executing the task with the highest priority from the top of the heap each time the task is selected to be executed, and selecting the first task in the queue M to execute when the task at the top of the heap does not meet the execution requirement, namely the task start time is greater than the current time. When a new task is added into the F heap, the priority of the task is calculated and inserted into the heap. If the task in the M queue approaches the deadline, the task needs to be immediately allocated to the MEC service node for execution, otherwise, the task is discarded, and the priority of the task is promoted at this time, and the task is inserted into the F heap. And if the task in the S queue approaches the deadline, the priority of the task is promoted and inserted into the tail of the M queue to wait for execution. In addition, to ensure the processing capability of the edge node, only one migration task can be executed on one MEC service node at the same time, so when the migration task is executed, the start time of other migration tasks under the same edge node needs to be updated.
(3) Discrete particle swarm algorithm
The invention solves a d-dimensional vector by taking task time delay minimization as an optimization target
Make a function
Is as small as possible, wherein d is the dimension of the feasible solution space, i.e., the number of tasks. And finally, solving the obtained solution into MEC service nodes to which different tasks are distributed.
In the classical particle swarm optimization, the velocities of the particles are calculated according to the same formula. In the iterative process, the diversity of the population is gradually lost, the particle motion range is limited, and the local optimal solution is difficult to jump out. Meanwhile, task unloading in a communication network field maintenance scene is discrete, and a discrete particle swarm algorithm is needed for calculation. Accordingly, the present invention improves upon the above problems.
First, n sub-populations are set, where each sub-population has a total of m particles. Within each sub-population, there are two types of particles, exploratory and navigatory. During each iteration, the algorithm will update the position, velocity of each particle. At the ith iteration, the velocity of the kth particle in the population j in the d dimension is updated as:
in the formula (27), ω is the inertia coefficient of the particle, c1Is a dynamic learning factor of the particle, c2For the sub-population dynamic learning factor, c3Is a population dynamic learning factor, r1，r2And r3Are random coefficients.
He Wei
Respectively the current optimal positions of the particles, the sub-populations and the populations. Wherein, type ═ 1 indicates that the particle is a navigation type particle, and the particle personal and population optimal solutions are considered. type2 indicates that the particle is an exploratory particle, and the optimal solution of the individual particle and the sub-population is considered.
In addition, as the task unloading is discrete, the intelligent terminal device and the MEC service node can communicate with a plurality of MEC service nodes in the communication range. Therefore, a new probability function is designed according to the actual situation, the variation of the position is represented by the probability, and the value is-1, 0 and 1. The probability function is:
the velocity values are mapped into the interval [0,1] by a probability function, and by using the direction of the particle velocity, the particle position update formula can be obtained as follows:
the above equation is the position update equation in the d dimension for the k particle. r is a random number on [0,1 ].
After the particle swarm algorithm is subjected to diversity and discretization processing, the inertial weight and the learning factor are redefined. First, the inertial weight determines the effect of the particle's previous velocity on this iteration. A nonlinear decreasing function is adopted, omega is larger at the initial stage of iteration and is beneficial to global optimization, and omega is smaller at the later stage and is beneficial to local optimization:
wherein, ω ismax＝0.9,ωminWhen ω is from 0.9 to 0.4, the algorithm is optimal, I is the current iteration number and I is the maximum iteration number.
The learning factor can reflect the information exchange condition among the populations in the particle swarm optimization. When the learning factor is always the same value, the particle will not change under the influence of individuals and populations, and therefore, the algorithm considers using the dynamic learning factor. In an iterative process, an individual learns the factor c1Gradually decreasing, population learning factor c2And c3And gradually increasing, enabling the particles to have more diversity in the former iteration, and prompting the particles to optimally obtain the optimal solution according to the population in the later iteration.
At the i-th iteration c1、c2And c3The value of (a) is shown in the formula, wherein I is the current iteration number, and I is the overall iteration number.
Based on the discrete particle swarm algorithm and the multi-priority queue model, the invention designs a task unloading strategy for minimizing time delay, and pseudo codes are as follows:
in order to evaluate the model migration method applicable to field maintenance of the communication network, which is provided by the invention, the following simulation experiment is carried out.
Model migration strategy based on service demand similarity
The method first verifies the model migration strategy based on the similarity of the service requirements. When the service requirements of the two models are similar, migration is performed by freezing all convolutional layers in the backbone network. According to the experimental result of the similarity-based service demand analysis, a pair of similar demands is selected for experiment, the service demands are insulator fault diagnosis, and the similarity reaches over 0.8. Therefore, the invention adopts two different insulator sub data sets to carry out experiments, and firstly trains according to the insulator sub data set with larger data volume to obtain an original model. And then the migration strategy provided by the invention is used for carrying out model migration by using the insulator data set with smaller data volume. The basic model used in the present invention is the YOLOv3 model. Table 2 compares the mAP values of models in different migration modes, and if the original model is directly used for identification, the final mAP can reach 80.2% because the service requirements of the two models are similar. And the mAP directly trained by using the new data set is only 74.7 percent, because the data volume is small, and the structure of the Yolov3 model is complex, so that the best training effect cannot be achieved. If the strategy provided by the invention is adopted, and the migration mode of freezing the backbone network is adopted, the characteristics which can be learned from the source model are benefited, and the model training is carried out according to the new data set, so that the final mAP can reach 86.3%.
TABLE 2 comparison of model accuracy under different migration strategies
However, when the service requirements of the two models are not similar, the above method cannot be used, for example, the model to be migrated is a communication network field maintenance target identification model, and the target model is an insulator fault detection model. If the migration method is adopted in this case, the recognition effect of the final target model will be poor due to the large difference between the two model data sets. Thus, the present invention uses a multi-stage model migration method. In order to verify the feasibility of the strategy, migration is performed by adopting a YOLOv3 model trained on a COCO training set in a simulation mode, and the insulator fault recognition is completed. As shown in fig. 2, the present invention firstly compares the migration effect at different freezing layer numbers, which are selected according to the YOLOv3 structure. With the increase of the number of frozen layers, the training calculation amount is reduced, and the model training time is gradually reduced. When the number of frozen layers is 43, the accuracy and the recall rate of the model reach the highest, and reach more than 88%. When the number of frozen layers is 0, the model accuracy is low because the model has an overfitting problem. On the other hand, when the number of frozen layers is too high, for example, 52 layers, the model accuracy gradually decreases, and the model features learned in the deep layer of YOLOv3 have no versatility and cannot be used directly. Therefore, when the similarity of the service requirements of the model is low, the method selects and freezes 43 layers of convolutional layers to carry out multi-stage migration, and the migration strategy is more efficient at the moment and is respectively embodied in two aspects of model training speed and model precision.
Task unloading strategy for minimizing time delay
In the task unloading strategy simulation of the minimized time delay, 8 wireless access base stations are provided, each base station can serve 4 terminal devices at the same time, and the calculation formula of the wireless channel gain in the simulation is Gm，q127+30 logd. Other parameters are shown in table 3.
TABLE 3 simulation parameters
The invention compares the proposed Multi-priority Task Scheduling (MPTS) with a First-in First-out queue (FCFS) and an Earliest Deadline First-out queue (EDF). As shown in fig. 3, when the amount of tasks is small, the algorithm proposed by the present invention has a smaller gap compared to other algorithms, because the queue is shorter and most tasks can be processed quickly. As the number of tasks increases, the FCFS queue blocks, and new short tasks will block behind long tasks, causing an increase in overall queuing delay. For the EDF queue, the deadline time exists in the task queue later, but the task reaches an earlier task, so that the average waiting time delay is difficult to reduce although the completion rate of the task is guaranteed by using an EDF algorithm, and the EDF queue needs to queue according to the deadline time in the queuing process, so that the time complexity is high. The MPTS algorithm provided by the invention divides the tasks into three levels, firstly processes the migration task with the highest priority, and takes the arrival time and the latest completion time of the tasks under different levels into consideration, so as to timely adjust the priority of the tasks, preferentially process the urgent small tasks, then process the large tasks with higher delay tolerance, and finally reduce the task waiting delay.
Finally, the invention compares the task Offloading policy of minimizing the delay with other algorithms, including Random Computing Offloading (RCO) algorithm and discrete Particle Swarm Optimization (BPSO) algorithm. First, as shown in fig. 4 and 5, the total time for completing the maintenance task and the task completion rate of the three algorithms under different migration task numbers are compared. As shown in fig. 4, as the number of migration tasks increases, the MEC service node calculation resources required by the migration tasks increase, and the total time for completing the maintenance tasks gradually increases. When the migration task is less, the RCO algorithm randomly selects the unloaded MEC service node, the time required for completing the maintenance task is longer than that of other algorithms, and the particle swarm algorithm is used for task unloading, so that the total time for completing the maintenance task by the BPSO and the method provided by the invention is shorter. When the migration task exceeds 4, the algorithm proposed herein is more advantageous than other algorithms as the migration task increases. The reason is that in order to ensure the efficient completion of the maintenance task, the multi-priority queue provided by the invention comprehensively considers the priority and urgency of the task for queuing, only allows the MEC service node to execute a migration task at the same time, and optimizes the total completion time of the maintenance task by using an improved discrete particle swarm algorithm.
For the task completion rate, as shown in fig. 5, as the number of migration tasks increases, the MEC service node resources are limited, and the completion rate of the maintenance task gradually decreases. The method provided by the invention considers the urgency degree of the task and the deadline time of the task, so the task completion rate is higher than that of other algorithms. In conclusion, the algorithm of the invention unloads the task by using the task unloading strategy of minimizing time delay, and simultaneously guarantees the task completion time and the task completion rate of the maintenance task.
Secondly, the invention simulates the average completion time of the migration tasks under different migration task quantities. As shown in fig. 6, when the number of migration tasks is small, the average completion time difference between the three algorithms is small, because the migration tasks are time-consuming and have high priority compared to the maintenance tasks, the occurrence of the migration tasks has a large influence on the completion of the maintenance tasks, and the migration tasks can be smoothly executed. When the number of the migration tasks is less than 12, the algorithm provided by the invention is superior to other algorithms, and as the number of the migration tasks increases, when the number of the migration tasks continues to increase, the performance of the BPSO algorithm is gradually superior to that of the algorithm provided by the invention, on the contrary, as shown in FIG. 5, the execution of the migration tasks seriously affects the completion rate of the maintenance tasks of the BPSO algorithm. In consideration of the fact that in a communication network field maintenance scene, the training time of the model is long, and a large number of migration tasks cannot occur in a short time, the algorithm provided by the invention still has advantages in the completion of the migration tasks compared with other algorithms in the communication network field maintenance scene.
Simulation results show that the model migration strategy based on the similarity of service requirements can reduce the blindness of model migration, relieve the problem of limited field maintenance data volume of a communication network, and realize efficient model migration according to the service requirements. Meanwhile, the task unloading algorithm for minimizing the time delay can ensure that the migration task and the maintenance task are efficiently executed at the same time, the time delay for completing the maintenance task is effectively reduced while the model migration is completed, and the task completion rate is improved.
The above examples are only intended to illustrate the technical solution of the present invention, but not to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: it is to be understood that modifications may be made to the technical solutions described in the foregoing embodiments, or equivalents may be substituted for some of the technical features thereof, but such modifications or substitutions do not depart from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims (10)
1. A communication network field maintenance model migration method based on time delay optimization is characterized by comprising the following steps:
s1, analyzing and calculating type similarity, name similarity and interpretation information similarity respectively based on the service requirements of the similarity, and finally calculating to obtain comprehensive similarity according to the weight of each similarity;
s2, measuring the correlation of the model by using the service requirement similarity, and selecting the migration modes under different conditions according to the service requirement similarity; when the service requirement of the new model is completely the same as that of the model to be migrated, directly migrating the model; freezing the backbone network when the service requirement of the new model is similar to that of the model to be migrated; when the similarity between the service requirement of the new model and the service requirement of the model to be migrated is poor, migrating the model by adopting multi-stage migration learning;
s3, taking into account various indexes of task delay, edge node resources and task resource requirements, and using a multi-priority queue model and an improved discrete particle swarm algorithm to finish the migration task and the maintenance task unloading in the scene of field maintenance of the communication network, so as to minimize the task delay.
2. The communication network field maintenance model migration method based on time delay optimization according to claim 1, wherein the type similarity calculation method in step S1 is as follows: classifying and marking different data characteristic spaces, assuming that a source model RC and a target model TC are model types, and the type similarity is as follows:
if the data feature spaces of the source model and the target model are the same, the type similarity is 1, otherwise, the type similarity is 0.
3. The communication network field maintenance model migration method based on time delay optimization according to claim 1, wherein the name similarity calculation method in step S1 is as follows:
first, a name vector is calculated according to equation (2):
where name _ vec is the name vector, veciIs the ith word vector in the name;
secondly, the similarity of two name vectors is calculated according to the formula (3):
wherein, SimnameFor name similarity, name _ vec1 is the first name vector and name _ vec2 is the second name vector.
4. The communication network field maintenance model migration method based on time delay optimization according to claim 1, wherein the calculation method of the interpretation information similarity in step S1 is as follows: firstly, performing word segmentation analysis on explanatory information by using a BilSTM-CRF algorithm according to the field maintenance characteristics of a communication network, classifying texts by using a trained model, judging the influence of entity types on the similarity of the model, extracting texts with larger influence for similarity calculation, and then performing similarity calculation on the words with the same type.
5. The communication network field maintenance model migration method based on time delay optimization according to claim 4, wherein step S1 is to extract the device element and the text of the fault for similarity calculation, and the method is: analyzing the interpretation information of the two models to obtain an equipment element set and a fault set of the interpretation information of the two models and a service requirement S1The information of m equipment elements is analyzed to be E1＝{e11，e12，...，e1mF, and p failure messages, F1＝{f11，f12，...，f1p}; service S2The information of n equipment elements is analyzed to be E2＝{e21，e22，...，e2nF, and q failure messages, F2＝{f21，f22，...，f2q}；
Calculating E according to equation (4)1，E2The similarity matrix of (2):
according to the formula (4), traversing the matrix to take the maximum value of each column to obtain a maximum value set P ═ P1，p2，…，PlH, where l ═ min (m, n);
calculating the device element similarity Se according to equation (5):
calculating F according to equation (6)1，F2Wherein each item is the similarity of two fault information:
according to the formula (6), traversing the matrix to take the maximum value of each column to obtain a maximum value set Q ═ Q1，q2，…，qkWhere k ═ min (p, q);
calculating the fault similarity S according to the formula (7)f：
And (3) obtaining the similarity of the interpretation information according to a formula (8) by combining the similarity of the equipment elements and the similarity of the faults:
Simexplain＝ω1Se+ω2Sf (8)
in the formula (8), SexplainTo interpret information similarity, SeAs device component similarity, SfIs the fault similarity; omega1，ω2Is a weight coefficient, and ω1+ω2＝1。
6. The communication network field maintenance model migration method based on time delay optimization according to claim 1, wherein the technical method for synthesizing similarity in step S1 is as follows: on the basis that the type similarity is 1, the comprehensive similarity is obtained according to the formula (9):
S＝Simcategory(ω1·Simname+ω2·Simexplain) (9)
in the formula (9), ω1，ω2The name similarity and the weight of the interpretation information similarity respectively satisfy omega1+ω2＝1。
7. The communication network field maintenance model migration method based on delay optimization according to claim 1, wherein the multi-priority queue model in step S3 includes three queues, which are F-pile, M first-in first-out queue, S first-in first-out queue; firstly, dividing the task into three grades according to the importance degree of the task and the time delay sensitivity of the task, wherein P1 is the task with the highest priority, P2 is the task with the medium priority, and P3 is the task with lower time delay requirement; then adding the tasks of the P1 level into the F heap, adding the tasks of the P2 level into the M queue, and adding the tasks of the P3 level into the S queue; f is a large top heap, the priority is determined according to the task deadline, the task with the highest top heap priority is executed each time the task is selected for execution, and when the top heap task does not meet the execution requirement, namely the task start time is greater than the current time, the first task in the M queue is selected for execution; when a new task is added into the F heap, calculating the priority of the task and inserting the task into the heap; if the task in the M queue approaches the deadline, the task needs to be immediately allocated to an MEC service node for execution, otherwise, the task is discarded, and the priority of the task is promoted at this time and the task is inserted into an F heap; and if the task in the S queue approaches the deadline, the priority of the task is promoted and inserted into the tail of the M queue to wait for execution.
8. The communication network field maintenance model migration method based on time delay optimization according to claim 1, wherein the modified discrete particle swarm algorithm in step S3 is:
firstly, setting n sub-populations, wherein each sub-population has m particles, and in each sub-population, two types of particles, namely an exploration type and a navigation type, exist; in each iteration process, the algorithm updates the position and the speed of each particle; at the ith iteration, the velocity of the kth particle in the population j in the d dimension is updated as:
in the formula (27)) In, ω is the inertia coefficient of the particle, c1Is a dynamic learning factor of the particle, c2For the sub-population dynamic learning factor, c3Is a population dynamic learning factor, r1，r2And r3Is a random coefficient;
He Wei
respectively representing the current optimal positions of particles, sub-populations and populations; wherein, type1 represents that the particle is a navigation particle, and the optimal solution of the particle individual and the population is considered; type2 represents that the particle is an exploration type particle, and the optimal solution of the individual particle and the sub-population is considered;
the variation of the position is represented by probability, the value is-1, 0,1, and the probability function is:
mapping the velocity value to an interval [0,1] through a probability function, and obtaining a particle position updating formula by using the direction of the particle velocity as follows:
equation (29) is a position update equation in the d-dimension for the kth particle, and r is a random number in [0,1 ].
9. The communication network field maintenance model migration method based on time delay optimization according to claim 8, wherein after diversity and discretization processing are performed on the particle swarm optimization, the inertial weight and the learning factor are redefined, and first, the inertial weight ω adopts a nonlinear decreasing function:
wherein, ω ismax＝0.9，ωminWhen ω is from 0.9 to 0.4, the algorithm reaches the optimum, I is the current iteration number, and I is the maximum iteration number;
in an iterative process, a dynamic learning factor, an individual learning factor c, is used1Gradually decreasing, population learning factor c2And c3Gradually increasing, at the i-th iteration c1、c2And c3Is shown in equations (31) (32) (33):
wherein I is the current iteration number, and I is the integral iteration number.
10. The communication network field maintenance model migration method based on time delay optimization according to claim 9, wherein the method of step S3 is: screening the MEC service nodes which can be distributed to each migration task according to the service requirement similarity, obtaining an initial position matrix X and an initial speed matrix V of the particles according to a task set, unloading the tasks according to the value of each particle in the population, queuing each MEC service node by using a multi-priority queue model, calculating the fitness of the particle in the iteration, updating the historical maximum fitness and the historical maximum fitness of the particleDegree position, updating the sub-population and global maximum fitness and maximum fitness position, updating the inertial weight according to formula (30), updating the learning factor according to formulas (31), (32) and (33), updating the particle velocity according to type1 in formula (27) if the particle is of navigation type, updating the particle velocity according to type2 in formula (27) if the particle is of exploration type, obtaining a probability matrix P according to the particle velocity, updating the particle position X according to formula (29), and calculating the minimum average delay TminCalculating an optimal task allocation vector
And repeating the iteration until the iteration times are reached.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210399528.5A CN114742166A (en) | 2022-04-15 | 2022-04-15 | Communication network field maintenance model migration method based on time delay optimization |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210399528.5A CN114742166A (en) | 2022-04-15 | 2022-04-15 | Communication network field maintenance model migration method based on time delay optimization |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN114742166A true CN114742166A (en) | 2022-07-12 |
Family
ID=82280680
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210399528.5A Pending CN114742166A (en) | 2022-04-15 | 2022-04-15 | Communication network field maintenance model migration method based on time delay optimization |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114742166A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116483585A (en) * | 2023-06-16 | 2023-07-25 | 江西师范大学 | Multi-service pre-deployment method based on edge computing |
-
2022
- 2022-04-15 CN CN202210399528.5A patent/CN114742166A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116483585A (en) * | 2023-06-16 | 2023-07-25 | 江西师范大学 | Multi-service pre-deployment method based on edge computing |
| CN116483585B (en) * | 2023-06-16 | 2023-09-15 | 江西师范大学 | Multi-service pre-deployment method based on edge computing |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111240701B (en) | Task unloading optimization method for end-side-cloud collaborative computing | |
| CN110968426B (en) | Edge cloud collaborative k-means clustering model optimization method based on online learning | |
| CN113469325B (en) | Hierarchical federation learning method for edge aggregation interval self-adaptive control, computer equipment and storage medium | |
| CN110955463B (en) | Internet of things multi-user computing unloading method supporting edge computing | |
| CN110471621B (en) | Edge collaborative storage method for real-time data processing application | |
| CN112511336B (en) | Online service placement method in edge computing system | |
| CN113642700B (en) | Cross-platform multi-mode public opinion analysis method based on federal learning and edge calculation | |
| CN115659803A (en) | Intelligent unloading method for computing tasks under unmanned aerial vehicle twin network mapping error condition | |
| WO2022242468A1 (en) | Task offloading method and apparatus, scheduling optimization method and apparatus, electronic device, and storage medium | |
| CN111836284B (en) | Energy consumption optimization calculation and unloading method and system based on mobile edge calculation | |
| CN112887905B (en) | Task unloading method based on periodic resource scheduling in Internet of vehicles | |
| CN109639833A (en) | A kind of method for scheduling task based on wireless MAN thin cloud load balancing | |
| CN111988787A (en) | Method and system for selecting network access and service placement positions of tasks | |
| CN116546559B (en) | Distributed multi-target space-ground combined track planning and unloading scheduling method and system | |
| CN113573320A (en) | SFC deployment method based on improved actor-critic algorithm in edge network | |
| CN115967990A (en) | Classification and prediction-based border collaborative service unloading method | |
| CN109375999A (en) | A kind of MEC Random Task moving method based on Bayesian network | |
| CN114742166A (en) | Communication network field maintenance model migration method based on time delay optimization | |
| CN115499875A (en) | Satellite internet task unloading method and system and readable storage medium | |
| CN111930435B (en) | Task unloading decision method based on PD-BPSO technology | |
| CN112559078B (en) | Method and system for hierarchically unloading tasks of mobile edge computing server | |
| CN109981372A (en) | Streaming big data processing method and system based on edge calculations | |
| CN117579701A (en) | Mobile edge network computing and unloading method and system | |
| CN114928893B (en) | Architecture based on intelligent reflecting surface and task unloading method | |
| CN114615705B (en) | Single-user resource allocation strategy method based on 5G network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |