Service migration method and apparatus for anomaly perception

Abstract

The application provides a service migration method and device for abnormal perception, and relates to the technical field of communication services. The method comprises the following steps: obtaining relevant data of a target service at different time steps, wherein the relevant data comprises performance data of multiple network nodes in a physical network where the target service is located, performance data of physical links between different network nodes, running data and log data of a target service combination to which the target service belongs; inputting the relevant data of the target service at different time steps into a preset Informer model to obtain a detection result of the target service; in the case that the detection result is an abnormal service, generating a service migration strategy corresponding to the target service based on the environmental state of each network node, and performing a migration operation on the target service based on the service migration strategy, thereby solving the problem that the timeliness of service migration is poor in the prior art because migration is performed only after the target service fails abnormally, and effectively improving the timeliness of service migration.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of communication services, and particularly relates to a service migration method and device for abnormality perception. BACKGROUND

[0002] In modern cloud computing and service-oriented architecture, service composition has become a key technology, enabling independent services from different sources to be integrated into complex business processes to meet more extensive user needs. Service composition involves the integration and coordination of multiple services, which may be distributed in different physical locations and run on independent platforms and network environments.

[0003] In service composition, each service needs to find suitable bearing resources in the physical network, which not only involves the allocation of computing resources, but also includes network bandwidth, delay, network load, and dependency, and other multi-dimensional considerations. In the case of failure of a certain service, the service migration method for abnormality perception can be moved between different physical resources to adapt to changes in network conditions, optimize resource use efficiency, and improve service performance and reliability.

[0004] However, the existing service migration method for abnormality perception mostly migrates after a certain service fails, which has serious hysteresis, resulting in poor timeliness of service migration. SUMMARY

[0005] The present application provides a service migration method and device for abnormality perception, which can effectively improve the timeliness of service migration.

[0006] The present application provides a service migration method for abnormality perception, comprising: acquiring related data of a target service at different time steps, the related data including performance data of each network node in a physical network where the target service is located, performance data of physical links between different network nodes, running data and log data of a target service composition to which the target service belongs; inputting the related data of the target service at different time steps into a preset Informer model to obtain a detection result of the target service; in the case that the detection result is an abnormal service, generating a service migration strategy corresponding to the target service based on the environmental state of each network node, and performing a migration operation on the target service based on the service migration strategy.

[0007] According to the service migration method for abnormality perception provided in the embodiment of the present application, the encoder includes a multi-head self-attention layer improved based on a time decay mechanism, the related data of the target service at different time steps is input into the preset Informer model, and a detection result of the target service is obtained, including: inputting the related data of the target service at different time steps into the encoder, combining the multi-head self-attention layer improved based on the time decay mechanism, and extracting high-dimensional feature vectors at different time steps; and inputting the high-dimensional feature vectors into the decoder to obtain the detection result of the target service.

[0008] According to the service migration method for abnormality perception provided in the embodiment of the present application, the encoder further includes an embedding layer, a sparse self-attention layer, a feedforward neural network layer, a residual and normalization layer; the related data of the target service at different time steps is input into the encoder, combining the multi-head self-attention layer improved based on the time decay mechanism, and extracting high-dimensional feature vectors at different time steps, including: inputting the related data of the target service at different time steps into the embedding layer, and extracting corresponding initial feature vectors through the embedding layer; inputting the initial feature vectors into the sparse self-attention layer, calculating attention scores of different time steps through the sparse self-attention layer, and obtaining first self-attention feature vectors based on the attention scores of different time steps and the initial feature vectors; inputting the first self-attention feature vectors into the multi-head self-attention layer, mapping the first self-attention feature vectors into different subspaces through the multi-head self-attention layer to obtain self-attention weights containing a time decay mechanism, and obtaining second self-attention feature vectors based on the self-attention weights containing the time decay mechanism and the first self-attention feature vectors; inputting the second self-attention feature vectors into the feedforward neural network layer, sequentially performing linear transformation and activation function processing on the second self-attention feature vectors through the feedforward neural network layer to obtain converted feature vectors; and inputting the converted feature vectors into the residual and normalization layer, sequentially performing residual connection and normalization processing on the converted feature vectors through the residual and normalization layer to obtain the high-dimensional feature vectors.

[0009] According to the service migration method for abnormality perception provided in the embodiment of the present application, the service migration strategy corresponding to the target service is generated based on the environment state of each network node, including: using a multi-agent reinforcement learning algorithm to generate a migration action corresponding to the target service based on the environment state of each network node; and generating the service migration strategy corresponding to the target service based on the migration action corresponding to the target service.

[0010] The method further includes: determining a preset negative value as the reward function value corresponding to the target service in a case where the target service migration fails or a total migration time delay of the target service combination is greater than an end-to-end delay threshold of the target service combination; determining the reward function value corresponding to the target service based on a storage resource requirement of a service entity of the target service, a bandwidth of a physical link between the different network nodes, an average power of the physical link between the different network nodes when transmitting data, an actual migration time of the service entity, a central processing unit (CPU) utilization of a target network node where the target service combination is located, and a storage resource utilization of the target network node in a case where the target service migration succeeds and the total migration time delay of the target service combination is less than or equal to the end-to-end delay threshold of the target service combination; determining a predicted cumulative reward value corresponding to the target service based on an environment state and an action of a plurality of agents and a Critic network, the plurality of agents including an agent corresponding to the service combination; and determining an evaluation value of the service migration strategy based on the reward function value corresponding to the target service and the predicted cumulative reward value.

[0011] The method further includes: determining a preset negative value as the reward function value corresponding to the target service in a case where the target service migration fails or a total migration time delay of the target service combination is greater than an end-to-end delay threshold of the target service combination; determining the reward function value corresponding to the target service based on a storage resource requirement of a service entity of the target service, a bandwidth of a physical link between the different network nodes, an average power of the physical link between the different network nodes when transmitting data, an actual migration time of the service entity, a central processing unit (CPU) utilization of a target network node where the target service combination is located, and a storage resource utilization of the target network node in a case where the target service migration succeeds and the total migration time delay of the target service combination is less than or equal to the end-to-end delay threshold of the target service combination; determining a predicted cumulative reward value corresponding to the target service based on an environment state and an action of a plurality of agents and a Critic network, the plurality of agents including an agent corresponding to the service combination; and determining an evaluation value of the service migration strategy based on the reward function value corresponding to the target service and the predicted cumulative reward value.

[0012] According to the service migration method for abnormality perception provided in the embodiment of the present application, the migration overhead corresponding to the service entity is determined based on the storage resource requirement of the service entity of the target service, the bandwidth of the physical link between different network nodes, the average power when the physical link between different network nodes transmits data, and the actual migration time of the service entity, comprising: determining the transmission delay of the service entity based on the storage resource requirement of the service entity of the target service and the bandwidth of the physical link between different network nodes; determining the migration energy consumption of the service entity based on the average power when the physical link between different network nodes transmits data, and the actual migration time of the service entity; determining the migration overhead corresponding to the service entity based on the transmission delay of the service entity and the migration energy consumption of the service entity.

[0013] According to the service migration method for abnormality perception provided in the embodiment of the present application, the predicted cumulative reward value corresponding to the target service is determined based on the environment state and action of a plurality of agents and a Critic network, comprising: inputting the environment state and action of the plurality of agents into a first multi-layer perceptron (MLP) in the Critic network to obtain a relationship feature matrix corresponding to the plurality of agents; inputting the relationship feature matrix into an encoder of a graph autoencoder of the Critic network to obtain an intermediate layer feature matrix corresponding to the plurality of agents; inputting the intermediate layer feature matrix into a second multi-layer perceptron (MLP) in the Critic network to obtain the predicted cumulative reward value.

[0014] According to the service migration method for abnormality perception provided in the embodiment of the present application, the method further comprises: inputting the intermediate layer feature matrix into a decoder of the graph autoencoder to obtain a reconstructed adjacency matrix.

[0015] Based on the reconstructed adjacency matrix and an initial adjacency matrix in the encoder of the graph autoencoder before the service migration is performed, a target adjacency matrix is determined; the initial adjacency matrix in the encoder of the graph autoencoder is updated based on the target adjacency matrix.

[0016] The application further provides a service migration device for abnormality perception, comprising: an acquisition unit configured to acquire relevant data of a target service at different time steps, the relevant data comprising performance data of each of a plurality of network nodes in a physical network in which the target service is located, performance data of physical links between different network nodes, running data and log data of a target service combination to which the target service belongs; a first processing unit configured to input the relevant data of the target service at the different time steps into a preset Informer model to obtain a detection result of the target service; and a second processing unit configured to, in a case where the detection result is an abnormal service, generate a service migration strategy corresponding to the target service based on an environmental state of each of the plurality of network nodes, and perform a migration operation on the target service based on the service migration strategy.

[0017] The application further provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the service migration method for abnormality perception according to any one of the above when executing the program.

[0018] The application further provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program is executable by a processor to implement the service migration method for abnormality perception according to any one of the above.

[0019] The application further provides a computer program product comprising a computer program, wherein the computer program is executable by a processor to implement the service migration method for abnormality perception according to any one of the above.

[0020] The service migration method and device for abnormality perception provided by the application, by acquiring relevant data of a target service at different time steps, the relevant data comprising performance data of each of a plurality of network nodes in a physical network in which the target service is located, performance data of physical links between different network nodes, running data and log data of a target service combination to which the target service belongs; inputting the relevant data of the target service at the different time steps into a preset Informer model to obtain a detection result of the target service; in a case where the detection result is an abnormal service, generating a service migration strategy corresponding to the target service based on an environmental state of each of the plurality of network nodes, and performing a migration operation on the target service based on the service migration strategy. In this way, the target service is predicted to be abnormal in advance by combining the relevant data of the target service at the different time steps and the Informer model, and in a case where the target service is predicted to be abnormal, a migration operation is performed on the target service based on the service migration strategy, that is, the migration operation is performed on the target service before the target service fails, thereby solving the problem of poor timeliness of service migration in the prior art because the migration is performed after the target service abnormally fails, and effectively improving the timeliness of service migration. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart illustrating a service migration method for anomaly detection provided in an embodiment of this application.

[0023] Figure 2 This is a flowchart illustrating a process for determining the evaluation value of a service migration strategy, as provided in an embodiment of this application.

[0024] Figure 3 This is a schematic diagram of the structure of a Critic network provided in an embodiment of this application.

[0025] Figure 4 This is a schematic diagram of a service migration device for anomaly detection provided in an embodiment of this application.

[0026] Figure 5 This is a schematic diagram of the physical structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0028] In the embodiments of this application, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone, where A and B can be singular or plural. In the textual description of this application, the character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0029] In the case of service failure, the abnormality-aware service migration method can be moved between different physical resources to adapt to the changes in network conditions, optimize resource use efficiency, and improve the performance and reliability of services. Existing abnormality-aware service migration methods mostly migrate after a service failure, which has serious lag, resulting in poor timeliness of service migration.

[0030] To improve the timeliness of service migration, the embodiments of the present application provide an abnormality-aware service migration method. In the following, the abnormality-aware service migration method provided by the present application will be described in detail through the following specific embodiments. It can be understood that the following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in some embodiments.

[0031] Figure 1 A flowchart of an abnormality-aware service migration method provided by an embodiment of the present application is shown in the figure. The method can be executed by software and / or hardware devices. For example, as shown in the figure, the abnormality-aware service migration method can include: Figure 1

[0032] S101, obtaining related data of the target service at different time steps, the related data including performance data of each network node in the physical network where the target service is located, performance data of the physical link between different network nodes, running data and log data of the target service combination to which the target service belongs.

[0033] It can be understood that the related data can include other related data of the target service in addition to the above data. Here, the embodiments of the present application are only described by taking the related data including the performance data of each network node in the physical network where the target service is located, the performance data of the physical link between different network nodes, the running data and log data of the target service combination to which the target service belongs as an example, but it does not mean that the embodiments of the present application are limited to this.

[0034] After obtaining the related data of the target service at different time steps, the related data of the target service at different time steps can be input into the preset Informer model as the basis for target service detection, i.e., performing the following S102.

[0035] S102, inputting the related data of the target service at different time steps into the preset Informer model to obtain the detection result of the target service.

[0036] ​For example, in the embodiments of the present application, the Informer model includes an encoder and a decoder, and the encoder includes a multi-head self-attention layer improved based on a time decay mechanism. In this way, the multi-head self-attention layer in the Informer model is improved by adding the time decay mechanism, which can better enhance the response capability of the Informer model to recent events.

[0037] For example, in the embodiments of the present application, the time decay mechanism can be implemented by a time decay matrix D, D nm represents the (n, m)th element in the time decay matrix D. n and m represent the position index of the relevant data of the target service at different time steps. When n is greater than or equal to m, D nm decreases with the increase of the distance between n and m, so that the influence of distant data on the current time can be weakened; when n is less than m, D nm is 0, which ensures that the Informer model will not know the future information in advance when processing the position n.

[0038] For the (n, m)th element D nm in the time decay matrix D, the Decay_attention function of the multi-head self-attention layer improved based on the time decay mechanism can be seen from the following formula 1 and formula 2.

[0039]

[0040]

[0041] wherein λ represents a time decay factor, and the value is in the range of (0, 1); d represents the multi-head self-attention layer improved based on the time decay mechanism, the vector dimension Q, K, and V of the extracted high-dimensional feature vector respectively represent the query vector, the key vector, and the value vector, represents the selected first u query vectors, and Decay_Attention(Q, K, V) represents the multi-head self-attention weight improved based on the time decay mechanism.

[0042] In the case of including an encoder and a decoder in the Informer model, when inputting the related data of the target service at different time steps into the preset Informer model to obtain the detection result of the target service, the related data of the target service at different time steps can be first input into the encoder, the high-dimensional feature vectors at different time steps are extracted by combining the multi-head self-attention layer improved based on the time decay mechanism, so as to capture the dependency between different time steps through the high-dimensional feature vectors at different time steps; and the high-dimensional feature vectors are input into the decoder to obtain the detection result of the target service. In this way, the Informer model based on the time decay mechanism performs service anomaly detection, gives greater weight to the recent events in the sequence through the time decay mechanism, thereby ignoring or reducing the influence of long-term events, makes the Informer model more effective in capturing the local patterns of time series data, and more sensitively responds to the recent state changes, thereby effectively improving the anomaly detection capability of the Informer model and the accuracy of the prediction result.

[0043] For example, in the embodiments of the present application, the encoder further includes an embedding layer, a sparse self-attention layer, a feed-forward neural network layer, a residual and normalization layer, and when inputting the related data of the target service at different time steps into the encoder and extracting the high-dimensional feature vectors at different time steps by combining the multi-head self-attention layer improved based on the time decay mechanism, the related data of the target service at different time steps can be first input into the embedding layer, and the corresponding initial feature vectors are extracted through the embedding layer; the initial feature vectors are input into the sparse self-attention layer, the attention scores of different time steps are calculated through the sparse self-attention layer, and the first self-attention feature vectors are obtained based on the attention scores of different time steps and the initial feature vectors; the first self-attention feature vectors are input into the multi-head self-attention layer, the first self-attention feature vectors are mapped into different subspaces through the multi-head self-attention layer, the self-attention weights containing the time decay mechanism are obtained, and the second self-attention feature vectors are obtained based on the self-attention weights containing the time decay mechanism and the first self-attention feature vectors; the second self-attention feature vectors are input into the feed-forward neural network layer, the second self-attention feature vectors are sequentially linearly transformed and processed by the activation function through the feed-forward neural network layer, and the converted feature vectors are obtained; the converted feature vectors are input into the residual and normalization layer, the converted feature vectors are sequentially residual connected and normalized through the residual and normalization layer, and the high-dimensional feature vectors are obtained, which can better capture the dependency between different time steps.

[0044] After the high-dimensional feature vector is extracted by the encoder, the high-dimensional feature vector can be input into the decoder to obtain a detection result of the target service. In a case where the detection result is a normal service, a migration operation does not need to be performed on the target service; conversely, in a case where the detection result is an abnormal service, S103 is performed.

[0045] S103. In a case where the detection result is an abnormal service, a service migration strategy corresponding to the target service is generated based on the respective environment states of the plurality of network nodes, and a migration operation is performed on the target service based on the service migration strategy.

[0046] For example, in the embodiment of the present application, when the service migration strategy corresponding to the target service is generated based on the respective environment states of the plurality of network nodes, a multi-agent reinforcement learning algorithm can be used to generate a migration action corresponding to the target service based on the respective environment states of the plurality of network nodes, and to generate the service migration strategy corresponding to the target service based on the migration action corresponding to the target service. In this way, the network and service changes can be dynamically adapted, and an optimized service migration strategy can be generated for the target service, so that the migration operation of the target service is performed based on the service migration strategy.

[0047] As can be seen, in the embodiment of the present application, the related data of the target service at different time steps is obtained, the related data including the performance data of the plurality of network nodes in the physical network where the target service is located, the performance data of the physical links between different network nodes, the running data and log data of the target service combination to which the target service belongs; the related data of the target service at different time steps is input into a preset Informer model to obtain a detection result of the target service; in a case where the detection result is an abnormal service, a service migration strategy corresponding to the target service is generated based on the respective environment states of the plurality of network nodes, and a migration operation is performed on the target service based on the service migration strategy. In this way, the Informer model is combined with the related data of the target service at different time steps in advance to predict the target service, and in a case where the target service is predicted to be an abnormal service, a migration operation is performed on the target service based on the service migration strategy, i.e., the migration operation is performed on the target service before the target service fails, thereby solving the problem of poor timeliness of service migration in the prior art caused by migration after the target service fails, and effectively improving the timeliness of service migration.

[0048] Based on the above Figure 1 As shown in the embodiment, after the migration operation of the target service is performed based on the service migration strategy, the service migration strategy can be evaluated to determine the evaluation value of the service migration strategy, which can be referred to the embodiment described below. Figure 2

[0049] Figure 2 ​A flowchart for determining the evaluation value of the service migration strategy provided in the embodiments of the present application, which can also be executed by software and / or hardware devices, for example, please refer to Figure 2 The service migration method for abnormal perception can include:

[0050] S201, in the case of target service migration failure or total migration time delay of the target service combination being greater than the end-to-end delay threshold of the target service combination, a preset negative value is determined as the reward function value corresponding to the target service.

[0051] For example, the preset negative number can be represented by M, then in the case of migration failure or , r c (t) = M. Where M is a small negative number, when the migration fails or the migration decision cannot meet the end-to-end performance constraints of the resource or service combination, the agent is given a larger punishment, T c represents the total migration time delay of the target service combination c, represents the end-to-end delay threshold of the target service combination c, which can be set according to actual needs.

[0052] S202, in the case of target service migration success and total migration time delay of the target service combination being less than or equal to the end-to-end delay threshold of the target service combination, the reward function value corresponding to the target service is determined based on the storage resource requirement of the service entity of the target service, the bandwidth of the physical link between different network nodes, the average power when the physical link between different network nodes transmits data, the actual migration time of the service entity, the CPU utilization of the target network node where the target service combination is located, and the storage resource utilization of the target network node.

[0053] For example, in the embodiments of the present application, when the reward function value corresponding to the target service is determined based on the storage resource requirement of the service entity of the target service, the bandwidth of the physical link between different network nodes, the average power when the physical link between different network nodes transmits data, the actual migration time of the service entity, the CPU utilization of the target network node where the target service combination is located, and the storage resource utilization of the target network node, the migration overhead corresponding to the service entity can be determined based on the storage resource requirement of the service entity of the target service, the bandwidth of the physical link between different network nodes, the average power when the physical link between different network nodes transmits data, and the actual migration time of the service entity; and the load value of the service entity is determined based on the CPU utilization of the target network node where the target service combination is located and the storage resource utilization of the target network node; then the reward function value is determined based on the preset reward factor, the migration overhead corresponding to the service entity, and the load value of the service entity.

[0054] For example, in the embodiments of the present application, when determining the migration overhead corresponding to the service entity of the target service based on the storage resource requirement of the service entity of the target service, the bandwidth of the physical link between different network nodes, the average power when the physical link between different network nodes transmits data, and the actual migration time of the service entity, the transmission delay of the service entity can be first determined based on the storage resource requirement of the service entity of the target service and the bandwidth of the physical link between different network nodes; and the migration energy consumption of the service entity can be determined based on the average power when the physical link between different network nodes transmits data and the actual migration time of the service entity; and then the migration overhead corresponding to the service entity can be determined based on the transmission delay of the service entity and the migration energy consumption of the service entity.

[0055] For example, in the embodiments of the present application, the network service system and the cost model can be modeled first. The network of the physical link is defined as G P P P , V P is a set of physical network nodes, E P is a set of physical links between nodes. The service dependency graph is defined as G S S S , V S is a set of service entities, E S is a set of virtual links between service nodes. Assuming that the service entity of the target service is the jth service entity in the ith service combination, the transmission delay of the service entity can be denoted as T i，j , and the migration energy consumption of the service entity can be denoted as E i，j , when determining the transmission delay of the service entity based on the storage resource requirement of the service entity of the target service and the bandwidth of the physical link between different network nodes, the following formula 3 can be referred to.

[0056]

[0057] , wherein, represents the storage resource requirement of the jth service entity in the ith service combination, b P represents the bandwidth of the physical link between different network nodes.

[0058] When determining the migration energy consumption of the service entity based on the average power when the physical link between different network nodes transmits data and the actual migration time of the service entity, the following formula 4 can be referred to.

[0059]

[0060] , wherein, P trans represents the average power when the physical link between different network nodes transmits data, ​​​​represents the actual migration time of the jth service entity in the ith service combination.

[0061] After the transmission delay of the service entity and the migration energy consumption of the service entity are determined respectively, the migration cost corresponding to the service entity can be determined based on the transmission delay of the service entity and the migration energy consumption of the service entity, which can be seen from the following formula 5.

[0062] cost i，j = ω1·T i，j + ω2·E i，j + ω3·T i，j · E i，j Formula 5

[0063] Wherein, cost i，j represents the migration cost of the jth service entity in the ith service combination, which is the weighted sum of the transmission delay, the migration energy consumption and their product, the product is a nonlinear term used to capture the interaction relationship between the migration delay and the energy consumption, ω1, ω2 and ω3 all represent weight coefficients, used to adjust the proportion of the transmission delay and the migration energy consumption in the migration cost, the values of ω1, ω2 and ω3 are all in the range of (0, 1).

[0064] For example, in the embodiment of the present application, when the load value of the service entity is determined based on the CPU utilization of the target network node where the target service combination is located and the storage resource utilization of the target network node, which can be seen from the following formula 6.

[0065]

[0066] Wherein, load represents the load value of the service entity, N P represents the number of physical nodes, represents the CPU utilization of the target network node where the ith service combination is located, respectively represent the storage resource utilization of the target network node where the ith service combination is located, β i and Y i The values of β and Y are in the range of (0, 1), which respectively represent the weight of the CPU utilization and the storage resource utilization of the ith node in the network load score.

[0067] In combination with the above description, after the migration cost corresponding to the service entity and the load value of the service entity are obtained respectively, the reward function value can be determined based on the preset reward factor, the migration cost corresponding to the service entity and the load value of the service entity, which can be seen from the following formula 7.

[0068] ​

[0069] wherein μ represents a preset reward factor, which can be set according to actual needs.

[0070] Based on the above description, for example, in the embodiments of the present application, a Markov decision model for service combination migration is constructed, and each service combination is regarded as an agent responsible for the migration of services in the service combination. For a service combination c, the smaller the time and energy consumption of the migration action taken by the system when the services in c are abnormal, the more balanced the network load, and the greater the reward obtained. Therefore, in the embodiments of the present application, the reward function value r c (t) can be seen from the following formula 8.

[0071]

[0072] It can be understood that in actual application process, the above S201 and S202 will be executed alternatively, in the case that the target service migration fails or the total migration time delay of the target service combination is greater than the end-to-end delay threshold of the target service combination, only the above S201 is executed, without executing the above S202, to obtain the reward function value corresponding to the target service; in the case that the target service migration succeeds and the total migration time delay of the target service combination is less than or equal to the end-to-end delay threshold of the target service combination, only the above S202 is executed, without executing the above S201, to obtain the reward function value corresponding to the target service.

[0073] In determining the evaluation value of the service migration strategy, in addition to the reward function value r c (t) corresponding to the target service, the predicted cumulative reward value corresponding to the target service can also be determined based on the Critic network, that is, the following S203 is executed.

[0074] S203, determining the predicted cumulative reward value corresponding to the target service based on the environment state and action of the plurality of agents and the Critic network; the plurality of agents include the agent corresponding to the service combination.

[0075] For example, in the embodiments of the present application, when determining the predicted cumulative reward value corresponding to the target service based on the environment state and action of the plurality of agents and the Critic network, the Critic network can be combined for implementation, and the network structure of the Critic network can be seen from Figure 3 , Figure 3 is a structural schematic diagram of a Critic network provided in the embodiments of the present application.

[0076] The environmental states and actions of multiple agents are input into the first multilayer perceptron (MLP) of the Critic network to obtain relational feature matrices for each agent. These relational feature matrices are then input into the encoder of the graph autoencoder in the Critic network to obtain intermediate layer feature matrices for each agent. Finally, these intermediate layer feature matrices are input into the second multilayer perceptron (MLP) of the Critic network to obtain the predicted cumulative reward value. This approach, by introducing a graph autoencoder to learn the dependencies between agents, uses a global perspective when evaluating service migration strategies, dynamically adapting to changes in the physical network and service composition, and thus making better decisions regarding migration problems.

[0077] For example, the environmental state of an agent can be represented as o = (o1, o2, ..., o...). N () indicates that o1 represents the environmental state of the first agent, o2 represents the environmental state of the second agent, and o N This represents the environmental state of the Nth agent. The agent's actions can be represented as a = (a1, a2, ..., a...). N Let a1 represent the action of the first agent, a2 represent the action of the second agent, and a2 represent the action of the third agent. N This represents the action of the Nth agent.

[0078] In combination with the above Figure 3 The Critic network shown takes the environmental states o = (o1, o2, ..., o3) of multiple agents as input. N ) and action a = (a1, a2, ..., a N For each agent i, the predicted cumulative reward value Q corresponding to the target service must be determined based on its own environmental state and actions, and in combination with the contributions of other agents. i (o, a), see Formula 9 below.

[0079]

[0080] in, and Indicates a fully connected layer network, x i This represents the i-th feature vector among the N feature vectors output by the graph autoencoder.

[0081] The relational feature matrix obtained from the first MLP for multiple agents can be found in [reference]. Figure 3 X, X∈R N×d The feature matrix represents the relationships between agents, consisting of the environmental states and actions of all agents. In the encoder stage of the graph autoencoder, a graph convolutional neural network is used to learn the feature vector x for each agent.i The obtained intermediate layer feature matrix can be seen from Figure 3 The initial adjacency matrix in the encoder of the graph autoencoder can be seen from Figure 3 A, A e R N×N is an adjacency matrix representing the relationship of the agents, A mn represents the correlation between agent m and agent n, and is initialized as the distance between the two service combinations after normalization.

[0082] For example, in the embodiments of the present application, the initial adjacency matrix in the encoder of the graph autoencoder can be further updated, which can specifically include: inputting the intermediate layer feature matrix into the decoder of the graph autoencoder to obtain a reconstructed adjacency matrix; and determining a target adjacency matrix based on the reconstructed adjacency matrix and the initial adjacency matrix in the encoder of the graph autoencoder before the service migration is performed; and updating the initial adjacency matrix in the encoder of the graph autoencoder based on the target adjacency matrix, so as to realize the update of the initial adjacency matrix in the encoder of the graph autoencoder.

[0083] For example, the intermediate layer feature matrix Z is decoded by the decoder of the graph autoencoder to reconstruct the adjacency matrix of the agent. The reconstructed adjacency matrix is combined with the initial adjacency matrix A to obtain a target adjacency matrix A new , which can be seen from the following formula 10, and the initial adjacency matrix A is replaced by the target adjacency matrix A new , so that the Critic network can further optimize the decision while retaining the original agent relationship by using the learned interaction information.

[0084]

[0085] wherein a represents a parameter for controlling the update amplitude.

[0086] After the target service corresponding reward function value and the predicted cumulative reward value are determined, the evaluation value of the service migration strategy can be determined based on the target service corresponding reward function value and the predicted cumulative reward value, that is, the following S204 is executed.

[0087] S204, determining the evaluation value of the service migration strategy based on the target service corresponding reward function value and the predicted cumulative reward value.

[0088] It can be seen that in the embodiments of the present application, after performing the migration operation on the target service based on the service migration strategy, the reward function value corresponding to the target service can be further determined, and the predicted cumulative reward value corresponding to the target service can be determined based on the environment state and action of the plurality of agents and the Critic network; and then the evaluation value of the service migration strategy is determined based on the reward function value and the predicted cumulative reward value corresponding to the target service, so as to realize the evaluation of the service migration strategy.

[0089] For example, in the examples of the present application, the integration of the collected physical network and service combination data is used for analysis, and the analysis steps are as follows:

[0090] S1, a physical network and a service system thereon are constructed, the number of network nodes of the physical network is N P = 32, the number of service combinations N = 16, and each service combination contains no more than 10 services. The physical network and service combination data within one month are collected from the system, including the mapping relationship between the physical nodes and the services. Among them, the physical node attributes include CPU utilization, storage resource consumption, fault alarm and other data, the physical link attributes include bandwidth utilization, delay, packet loss rate and traffic data, and the service combination attributes include fault log, service delay, service completion rate and other data.

[0091] S2, the data collected in S1 is normalized and input into the preset Informer model, wherein the embedding layer dimension of the preset Informer model is 512, the number of multi-head self-attention layers is 8, and the number of encoder layers is 6. In the training process, λ = 0.8 is taken, and the time decay matrix The evaluation standard of the Informer model is that as long as at least one abnormal point can be detected within the abnormal time period of 10 predefined time, the time period is considered to be correctly identified. It is predicted that in the next time period, the 2nd service of service combination 3 and the 5th service of service combination 9 will be abnormal.

[0092] S3, migrating the predicted services in S2, each service combination using an agent to make action decisions. Through multiple rounds of training, observing the current environment state and action, the service migration strategy of the second service is to migrate from node 8 to node 7, and the service migration strategy of the fifth service is to migrate from node 10 to node 15. After learning the environment state and action of all agents by the graph autoencoder in the Critic network, the eigenvalue of agent 3 is [0.5835, 0.21129, …, 0.62048]. In the encoder stage of the graph autoencoder, the intermediate layer feature matrix learned by the graph convolutional neural network is reconstructed by decoding, and the reconstructed adjacency matrix is combined with the initial adjacency matrix to obtain a new target adjacency matrix, and the value of a is 0.9. After evaluation by the Critic network, the reward values of the service migration strategies taken by the second service and the fifth service are 3.764 and 4.235, respectively. Not only can a better migration effect be achieved, but also the service migration operation is performed before the service fails, solving the problem of poor timeliness of service migration in the prior art after the service abnormally fails, thereby effectively improving the timeliness of service migration.

[0093] The abnormality-aware service migration device provided by the present application is described below. The abnormality-aware service migration device described below can be referred to in correspondence with the abnormality-aware service migration method described above.

[0094] Figure 4 The abnormality-aware service migration device provided by the present application is described below. The abnormality-aware service migration device described below can be referred to in correspondence with the abnormality-aware service migration method described above. Figure 4 As shown in the figure, the abnormality-aware service migration device 40 can include:

[0095] The acquisition unit 401 is configured to acquire relevant data of a target service at different time steps, wherein the relevant data includes performance data of each network node in a physical network where the target service is located, performance data of a physical link between different network nodes, running data and log data of a target service combination to which the target service belongs.

[0096] The first processing unit 402 is configured to input the relevant data of the target service at different time steps into a preset Informer model to obtain a detection result of the target service.

[0097] The second processing unit 403 is configured to, in a case where the detection result is an abnormal service, generate a service migration strategy corresponding to the target service based on the environment state of each network node, and perform a migration operation on the target service based on the service migration strategy.

[0098] For example, in the embodiments of the present application, the Informer model comprises an encoder and a decoder, the encoder comprises a multi-head self-attention layer improved based on a time decay mechanism, the first processing unit 402 is configured to input the relevant data of the target service at different time steps into the preset Informer model to obtain a detection result of the target service, including: inputting the relevant data of the target service at different time steps into the encoder, combining the multi-head self-attention layer improved based on the time decay mechanism, and extracting high-dimensional feature vectors at different time steps; and inputting the high-dimensional feature vectors into the decoder to obtain the detection result of the target service.

[0099] For example, in the embodiments of the present application, the encoder further comprises an embedding layer, a sparse self-attention layer, a feed-forward neural network layer, a residual and normalization layer; the first processing unit 402 is configured to input the relevant data of the target service at different time steps into the encoder, combine the multi-head self-attention layer improved based on the time decay mechanism, and extract high-dimensional feature vectors at different time steps, including: inputting the relevant data of the target service at different time steps into the embedding layer to extract corresponding initial feature vectors through the embedding layer; inputting the initial feature vectors into the sparse self-attention layer, calculating attention scores of different time steps through the sparse self-attention layer, and obtaining first self-attention feature vectors based on the attention scores of different time steps and the initial feature vectors; inputting the first self-attention feature vectors into the multi-head self-attention layer, mapping the first self-attention feature vectors into different subspaces through the multi-head self-attention layer to obtain self-attention weights containing a time decay mechanism, and obtaining second self-attention feature vectors based on the self-attention weights containing the time decay mechanism and the first self-attention feature vectors; inputting the second self-attention feature vectors into the feed-forward neural network layer, sequentially performing linear transformation and activation function processing on the second self-attention feature vectors through the feed-forward neural network layer to obtain converted feature vectors; and inputting the converted feature vectors into the residual and normalization layer, sequentially performing residual connection and normalization processing on the converted feature vectors through the residual and normalization layer to obtain the high-dimensional feature vectors.

[0100] For example, in the embodiments of the present application, the second processing unit 403 is configured to generate a service migration strategy corresponding to the target service based on the environment state of each network node, including: using a multi-agent reinforcement learning algorithm to generate a migration action corresponding to the target service based on the environment state of each network node; and generating the service migration strategy corresponding to the target service based on the migration action corresponding to the target service.

[0101] For example, in the embodiment of the present application, the service migration device 40 for abnormal perception comprises a third processing unit configured to determine a preset negative value as the reward function value corresponding to the target service in the case that the target service migration fails or the total migration time delay of the target service combination is greater than the end-to-end delay threshold of the target service combination; a fourth processing unit configured to determine the reward function value corresponding to the target service based on the storage resource requirement of the service entity of the target service, the bandwidth of the physical link between the different network nodes, the average power of the physical link between the different network nodes when transmitting data, the actual migration time of the service entity, the CPU utilization of the target network node where the target service combination is located, and the storage resource utilization of the target network node in the case that the target service migration succeeds and the total migration time delay of the target service combination is less than or equal to the end-to-end delay threshold of the target service combination; a fifth processing unit configured to determine the predicted cumulative reward value corresponding to the target service based on the environment state and action of a plurality of agents and a Critic network, wherein the plurality of agents comprise the agent corresponding to the service combination; and a sixth processing unit configured to determine the evaluation value of the service migration strategy based on the reward function value corresponding to the target service and the predicted cumulative reward value.

[0102] For example, in the embodiment of the present application, the fourth processing unit is configured to determine the reward function value corresponding to the target service based on the storage resource requirement of the service entity of the target service, the bandwidth of the physical link between the different network nodes, the average power of the physical link between the different network nodes when transmitting data, the actual migration time of the service entity, the CPU utilization of the target network node where the target service combination is located, and the storage resource utilization of the target network node, and comprises: determining the migration overhead corresponding to the service entity based on the storage resource requirement of the service entity of the target service, the bandwidth of the physical link between the different network nodes, the average power of the physical link between the different network nodes when transmitting data, and the actual migration time of the service entity; determining the load value of the service entity based on the CPU utilization of the target network node where the target service combination is located and the storage resource utilization of the target network node; and determining the reward function value based on a preset reward factor, the migration overhead corresponding to the service entity, and the load value of the service entity.

[0103] For example, in the embodiment of the present application, the fourth processing unit is configured to determine the migration overhead corresponding to the service entity based on the storage resource requirement of the service entity of the target service, the bandwidth of the physical link between the different network nodes, the average power of the physical link between the different network nodes when transmitting data, and the actual migration time of the service entity, including: determining the transmission delay of the service entity based on the storage resource requirement of the service entity of the target service and the bandwidth of the physical link between the different network nodes; determining the migration energy consumption of the service entity based on the average power of the physical link between the different network nodes when transmitting data, and the actual migration time of the service entity; and determining the migration overhead corresponding to the service entity based on the transmission delay of the service entity and the migration energy consumption of the service entity.

[0104] For example, in the embodiment of the present application, the fifth processing unit is configured to determine the predicted cumulative reward value corresponding to the target service based on the environment state and action of the plurality of agents and the Critic network, including: inputting the environment state and action of the plurality of agents into a first multi-layer perceptron (MLP) in the Critic network to obtain a relationship feature matrix corresponding to the plurality of agents; inputting the relationship feature matrix into an encoder of a graph autoencoder in the Critic network to obtain an intermediate layer feature matrix corresponding to the plurality of agents; and inputting the intermediate layer feature matrix into a second multi-layer perceptron (MLP) in the Critic network to obtain the predicted cumulative reward value.

[0105] For example, in the embodiment of the present application, the service migration device 40 for abnormal perception further includes: a seventh processing unit configured to input the intermediate layer feature matrix into a decoder of the graph autoencoder to obtain a reconstructed adjacency matrix; an eighth processing unit configured to determine a target adjacency matrix based on the reconstructed adjacency matrix and an initial adjacency matrix in the encoder of the graph autoencoder before the service migration is performed; and a ninth processing unit configured to update the initial adjacency matrix in the encoder of the graph autoencoder based on the target adjacency matrix.

[0106] The service migration device 40 for abnormal perception provided by the embodiment of the present application can execute the technical solutions of the service migration method for abnormal perception in any of the above embodiments, and the implementation principles and beneficial effects thereof are similar to those of the service migration method for abnormal perception. For details, refer to the implementation principles and beneficial effects of the service migration method for abnormal perception, which will not be described here.

[0107] Figure 5 An entity structure schematic diagram of an electronic device provided by the embodiment of the present application is shown in FIG. 1. Figure 5As shown, the electronic device can include a processor 510, a communications interface 520, a memory 530, and a communications bus 540, wherein the processor 510, the communications interface 520, and the memory 530 complete mutual communication through the communications bus 540. The processor 510 can invoke a logic instruction in the memory 530 to execute an exception-aware service migration method, which includes: obtaining relevant data of a target service at different time steps, the relevant data including performance data of each network node in a physical network where the target service is located, performance data of a physical link between different network nodes, running data and log data of a target service combination to which the target service belongs; inputting the relevant data of the target service at the different time steps into a preset Informer model to obtain a detection result of the target service; in the case that the detection result is an abnormal service, generating a service migration strategy corresponding to the target service based on an environment state of each network node, and performing a migration operation on the target service based on the service migration strategy.

[0108] In addition, the logic instruction in the memory 530 described above can be implemented in the form of a software functional unit and sold or used as an independent product, and can be stored in a computer-readable storage medium. Based on such understanding, the technical solutions of the present application essentially or the part that contributes to the prior art or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0109] In another aspect, the present application also provides a computer program product comprising a computer program, the computer program being stored in a non-transitory computer-readable storage medium, and the computer program being executable by a processor to enable a computer to perform the method of service migration based on anomaly awareness, the method comprising: obtaining relevant data of a target service at different time steps, the relevant data comprising performance data of each of a plurality of network nodes in a physical network in which the target service is located, performance data of physical links between different network nodes, running data and log data of a target service combination to which the target service belongs; inputting the relevant data of the target service at the different time steps into a preset Informer model to obtain a detection result of the target service; and in a case where the detection result is an abnormal service, generating a service migration strategy corresponding to the target service based on an environmental state of each of the plurality of network nodes, and performing a migration operation on the target service based on the service migration strategy.

[0110] In another aspect, the present application also provides a computer program product comprising a computer program, the computer program being stored in a non-transitory computer-readable storage medium, and the computer program being executable by a processor to enable a computer to perform the method of service migration based on anomaly awareness, the method comprising: obtaining relevant data of a target service at different time steps, the relevant data comprising performance data of each of a plurality of network nodes in a physical network in which the target service is located, performance data of physical links between different network nodes, running data and log data of a target service combination to which the target service belongs; inputting the relevant data of the target service at the different time steps into a preset Informer model to obtain a detection result of the target service; and in a case where the detection result is an abnormal service, generating a service migration strategy corresponding to the target service based on an environmental state of each of the plurality of network nodes, and performing a migration operation on the target service based on the service migration strategy.

[0111] The device embodiments described above are merely illustrative, wherein the units described as separate components can or can not be physically separate, and the components displayed as units can or can not be physical units, i.e., they can be located in one place, or distributed on multiple network units. Part or all of the modules can be selected to achieve the purpose of the present embodiment scheme according to actual needs. Those skilled in the art can understand and implement it without creative labor.

[0112] Those skilled in the art can clearly understand the implementation of the various embodiments by means of software and necessary general hardware platforms through the description of the above embodiments, and of course, the embodiments can also be implemented by hardware. Based on such understanding, the above technical solutions, essentially or in other words, the part of the prior art that contributes to the technical solutions can be embodied in the form of a software product. The computer software product can be stored in a computer readable storage medium, such as a ROM / RAM, a magnetic disk, an optical disk, and the like, and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0113] Finally, it should be noted that: the above examples are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing examples, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A service migration method oriented towards anomaly detection, characterized in that, include: Obtain relevant data of the target service at different time steps. The relevant data includes the performance data of multiple network nodes in the physical network where the target service is located, the performance data of the physical links between different network nodes, and the operation data and log data of the target service combination to which the target service belongs. The relevant data of the target service at different time steps are input into the preset Informer model to obtain the detection results of the target service; If the detection result indicates an abnormal service, a service migration strategy corresponding to the target service is generated based on the environmental status of each of the multiple network nodes, and a migration operation is performed on the target service based on the service migration strategy. The Informer model includes an encoder and a decoder. The encoder includes a multi-head self-attention layer improved based on a time decay mechanism. The relevant data of the target service at different time steps are input into the preset Informer model to obtain the detection result of the target service, including: The relevant data of the target service at different time steps are input into the encoder, and the high-dimensional feature vectors at different time steps are extracted by combining the multi-head self-attention layer based on the time decay mechanism. The high-dimensional feature vector is input into the decoder to obtain the detection result of the target service.

2. The service migration method for anomaly detection according to claim 1, characterized in that, The encoder further includes an embedding layer, a sparse self-attention layer, a feedforward neural network layer, and a residual and normalization layer; The step of inputting relevant data of the target service at different time steps into the encoder, and combining it with the multi-head self-attention layer improved based on the time decay mechanism, to extract high-dimensional feature vectors at different time steps includes: The relevant data of the target service at different time steps are input into the embedding layer, and the corresponding initial feature vector is extracted through the embedding layer. The initial feature vector is input into the sparse self-attention layer, and the attention score at each time step is calculated through the sparse self-attention layer. Based on the attention score at each time step and the initial feature vector, the first self-attention feature vector is obtained. The first self-attention feature vector is input into the multi-head self-attention layer, and the first self-attention feature vector is mapped to different subspaces through the multi-head self-attention layer to obtain self-attention weights with time decay mechanism. Based on the self-attention weights with time decay mechanism and the first self-attention feature vector, the second self-attention feature vector is obtained. The second self-attention feature vector is input into the feedforward neural network layer, and the feedforward neural network layer sequentially performs linear transformation and activation function processing on the second self-attention feature vector to obtain the transformed feature vector. The transformed feature vector is input into the residual and normalization layer. The transformed feature vector is then subjected to residual concatenation and normalization processes sequentially through the residual and normalization layer to obtain the high-dimensional feature vector.

3. The service migration method for anomaly detection according to claim 1 or 2, characterized in that, The process of generating the service migration strategy corresponding to the target service based on the environmental states of the multiple network nodes includes: A multi-agent reinforcement learning algorithm is used to generate migration actions corresponding to the target service based on the environmental states of the multiple network nodes. Based on the migration action corresponding to the target service, a service migration strategy corresponding to the target service is generated.

4. The service migration method for anomaly detection according to claim 1 or 2, characterized in that, The method further includes: If the migration of the target service fails, or if the total migration latency of the target service combination exceeds the end-to-end latency threshold of the target service combination, a preset negative value will be determined as the reward function value corresponding to the target service. If the target service migration is successful and the total migration latency of the target service combination is less than or equal to the end-to-end latency threshold of the target service combination, the reward function value corresponding to the target service is determined based on the storage resource requirements of the service entity of the target service, the bandwidth of the physical link between the different network nodes, the average power of the physical link between the different network nodes when transmitting data, the actual migration time of the service entity, the CPU utilization of the target network node where the target service combination is located, and the storage resource utilization of the target network node. Based on the environmental states and actions of multiple agents and a Critic network, the predicted cumulative reward value corresponding to the target service is determined; the multiple agents include the agents corresponding to the service combination. The evaluation value of the service migration strategy is determined based on the reward function value corresponding to the target service and the predicted cumulative reward value.

5. The service migration method for anomaly detection according to claim 4, characterized in that, The reward function value corresponding to the target service is determined based on the storage resource requirements of the service entity of the target service, the bandwidth of the physical links between different network nodes, the average power of the physical links between different network nodes when transmitting data, the actual migration time of the service entity, the CPU utilization of the target network node where the target service combination is located, and the storage resource utilization of the target network node. Based on the storage resource requirements of the service entity of the target service, the bandwidth of the physical link between the different network nodes, the average power of the physical link between the different network nodes when transmitting data, and the actual migration time of the service entity, the migration overhead corresponding to the service entity is determined. The load value of the service entity is determined based on the CPU utilization and storage resource utilization of the target network node where the target service combination is located. The reward function value is determined based on a preset reward factor, the migration cost corresponding to the service entity, and the load value of the service entity.

6. The service migration method for anomaly detection according to claim 5, characterized in that, The migration overhead for the service entity is determined based on the storage resource requirements of the service entity, the bandwidth of the physical links between different network nodes, the average power of the physical links between different network nodes when transmitting data, and the actual migration time of the service entity. This includes: Based on the storage resource requirements of the service entity of the target service and the bandwidth of the physical links between the different network nodes, the transmission latency of the service entity is determined. The migration energy consumption of the service entity is determined based on the average power of data transmission through the physical links between the different network nodes and the actual migration time of the service entity. Based on the transmission latency and migration energy consumption of the service entity, the migration overhead corresponding to the service entity is determined.

7. The service migration method for anomaly detection according to claim 4, characterized in that, The determination of the predicted cumulative reward value corresponding to the target service based on the environmental state and actions of multiple agents and a Critic network includes: The environmental state and actions of the multiple agents are input into the first multilayer perceptron (MLP) in the Critic network to obtain the relation feature matrix corresponding to the multiple agents. The relation feature matrix is ​​input into the encoder of the graph autoencoder of the Critic network to obtain the intermediate layer feature matrix corresponding to the multiple agents; The intermediate layer feature matrix is ​​input into the second multilayer perceptron (MLP) in the Critic network to obtain the predicted cumulative reward value.

8. The service migration method for anomaly detection according to claim 7, characterized in that, The method further includes: The intermediate layer feature matrix is ​​input into the decoder of the graph autoencoder to obtain the reconstructed adjacency matrix; Based on the reconstructed adjacency matrix and the initial adjacency matrix in the encoder of the graph autoencoder before performing the service migration, the target adjacency matrix is ​​determined; The initial adjacency matrix in the encoder of the graph autoencoder is updated based on the target adjacency matrix.

9. A service migration device for anomaly detection, characterized in that, include: The acquisition unit is used to acquire relevant data of the target service at different time steps. The relevant data includes the performance data of multiple network nodes in the physical network where the target service is located, the performance data of the physical links between different network nodes, and the operation data and log data of the target service combination to which the target service belongs. The first processing unit is used to input the relevant data of the target service at different time steps into a preset Informer model to obtain the detection result of the target service; The second processing unit is configured to, when the detection result is an abnormal service, generate a service migration strategy corresponding to the target service based on the environmental state of each of the multiple network nodes, and perform a migration operation on the target service based on the service migration strategy. The Informer model includes an encoder and a decoder, and the encoder includes a multi-head self-attention layer based on a time decay mechanism. The first processing unit is specifically used for: The relevant data of the target service at different time steps are input into the encoder, and the high-dimensional feature vectors at different time steps are extracted by combining the multi-head self-attention layer based on the time decay mechanism. The high-dimensional feature vector is input into the decoder to obtain the detection result of the target service.

Images