Cloud service combination processing method, device and equipment

Abstract

The invention provides a cloud service combination processing method, device and equipment. The method comprises the following steps: acquiring position information and service quality index information of a plurality of target objects; wherein each target object can provide various types of cloud services; performing feature extraction on the position information and the service quality index information of each target object through a feature extraction network to obtain a target feature vector representing mutual influence between the target objects; obtaining target tasks of a plurality of users; performing multi-head attention calculation on the task feature vector of each target task and the target feature vector to obtain state information under each target task; and through a preset strategy network, obtaining a target object matched with each target task under the state information of each target task. According to the scheme, mutual influence among different target objects can be fully considered, and the reasonable target object is matched for the manufacturing task of the user in a large number of candidate target objects.

Description

Technical Field

[0001] The present invention relates to the field of computer information technology processing technology, and in particular to a cloud service combination processing method, device and equipment. Background Art

[0002] The intelligent manufacturing model uses technologies such as artificial intelligence, cloud computing, and the Internet of Things to gather the production needs of enterprises, encapsulate manufacturing resources into manufacturing service modules and publish them on the intelligent manufacturing platform. It uses intelligent algorithms to select suitable service providers from a large number of manufacturing services to complete the manufacturing tasks of enterprises. This not only improves the utilization rate of manufacturing resources, but also helps enterprises optimize resource allocation, improve product quality, and reduce costs. At the same time, the optimization of intelligent manufacturing cloud services also faces a series of challenges, including differences in cost, time, and service quality among many service providers; the manufacturing tasks of service demanders require the combination of different types of manufacturing services, which are distributed in different physical locations and have cooperative or competitive relationships, and the correlation between manufacturing service providers needs to be fully considered; the security and data privacy of enterprise manufacturing tasks are also very important factors, which need to be fully considered in the service combination optimization process. Under these circumstances, it is necessary to select a set of service combination solutions with high-quality services. Therefore, there is an urgent need for a combination optimization method with good combination optimization performance, short-time response capability, applicability, and scalability.

[0003] At present, traditional combination optimization methods usually select service providers according to certain standards. Although they can quickly form a combination optimization plan, these methods do not take into account the relationships between service providers. These relationships mainly include: (1) Service providers are located in different physical locations, and there are logistics relationships between enterprises. (2) When service providers jointly handle a task, there is a cooperative relationship. (3) There is a competitive relationship between service providers that provide the same service type. With the continuous development of cloud services, there is mutual influence between different service providers. When performing cloud service combination optimization, this correlation influence should not be ignored. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a cloud service combination processing method, device and equipment, which can fully consider the mutual influence between different target objects and match reasonable target objects for users' different manufacturing tasks among a large number of candidate target objects.

[0005] In order to solve the above technical problems, the technical solution of the present invention is as follows:

[0006] A cloud service combination processing method, comprising:

[0007] Obtain location information and service quality indicator information of multiple target objects; each target object can provide multiple types of cloud services;

[0008] The location information and service quality index information of each target object are extracted through a feature extraction network to obtain a target feature vector representing the mutual influence between target objects;

[0009] Obtain target tasks of multiple users; each of the target tasks requires a type of cloud service;

[0010] Perform multi-head attention calculation on the task feature vector and the target feature vector of each target task to obtain state information under each target task;

[0011] Through the preset policy network, a target object matched for each target task under the state information of each target task is obtained, and the target object can provide a cloud service of the target type required by the target task.

[0012] Optionally, feature extraction is performed on the location information and service quality indicator information of each target object through a feature extraction network to obtain a target feature vector representing the mutual influence between the target objects, including:

[0013] The location information and service quality index information of each target object are represented as a feature vector;

[0014] Inputting the feature vector into a linear neural network layer for linear calculation to obtain an original feature vector;

[0015] Perform multiple bias attention calculations on the original feature vector in sequence to obtain a feature vector after each bias attention calculation;

[0016] The feature vectors calculated after each biased attention calculation are connected to obtain a target feature vector representing the mutual influence between target objects.

[0017] Optionally, the original feature vector is subjected to multiple biased attention calculations in sequence to obtain a feature vector after each biased attention calculation, including:

[0018] Perform self-attention calculation on the original feature vector to obtain a high-level feature vector;

[0019] Taking a difference between the high-level feature vector and the original feature vector to obtain an attention offset;

[0020] Adding the attention offset to the original feature vector to obtain a feature vector after a biased attention calculation;

[0021] The feature vector after the biased attention calculation is repeatedly subjected to multiple biased attention calculations to obtain the feature vector after each biased attention calculation.

[0022] Optionally, multi-head attention calculation is performed on the task feature vector and the target feature vector of each target task to obtain state information under each target task, including:

[0023] Obtain a query vector according to the target feature vector, the target type cloud service required by the target task in the task feature vector, the carrying capacity information of the target object, and the target type cloud service information required by the target task;

[0024] Processing the target feature vector with a layer of linear neural network to obtain a key vector and a value vector;

[0025] The query vector, key vector and value vector are subjected to multi-head attention calculation to obtain state information under each target task.

[0026] Optionally, a target object matched for each target task under the state information of each target task is obtained through a preset strategy network, including:

[0027] Through the preset strategy network, the probability value of selecting each target object under the current state information is obtained;

[0028] The target object with the largest probability value is taken as the target object matched by the current target task;

[0029] According to the target object matched by the current target task, the environment is changed, and the state information of the next target task is obtained according to the changed environment;

[0030] According to the state information, matching the target object for the next target task through a preset strategy network;

[0031] Repeatedly calculate the state information under each target task and calculate the probability value through the preset strategy network to obtain the target object matched by each target task.

[0032] Optionally, a probability value of selecting each target object under the current state information is obtained through a preset strategy network, including:

[0033] Through the preset strategy network, according to the formula

[0034] Perform probability calculation to obtain the probability value of selecting each target object under the current state information; where Pi is the probability of selecting the i-th target object, W is the neural network weight, and State t is the status information at step t, Key iis the high-level feature representation of the i-th target object, d k Key i , Mask is the formulated action masking scheme, and softmax is the normalization function.

[0035] Optionally, training the feature extraction network and the strategy network includes: establishing an objective function.

[0036] Optionally, the objective function includes:

[0037] After assigning a target object to each target task, the time, cost, load balancing and energy consumption required to complete the target task are minimized; and after assigning a target object to each target task, the service reliability, service quality, security and resource utilization achieved by completing the target task are maximized.

[0038] The present invention also provides a cloud service combination processing device, comprising:

[0039] An acquisition module, used to acquire location information and service quality indicator information of multiple target objects; each target object can provide multiple types of cloud services;

[0040] The processing module is used to extract the location information and service quality indicator information of each target object through a feature extraction network to obtain a target feature vector representing the mutual influence between the target objects; obtain target tasks of multiple users; each of the target tasks requires a type of cloud service; perform multi-head attention calculation on the task feature vector and the target feature vector of each target task to obtain the state information under each target task; obtain a target object matched for each target task under the state information of each target task through a preset strategy network, and the target object can provide the cloud service of the target type required by the target task.

[0041] The present invention also provides a computing device, comprising: a processor and a memory storing a computer program, wherein when the computer program is run by the processor, the method as described above is executed.

[0042] The above solution of the present invention includes at least the following beneficial effects:

[0043] The above scheme of the present invention obtains the location information and service quality index information of multiple target objects; each target object can provide multiple types of cloud services; the location information and service quality index information of each target object are subjected to feature extraction through a feature extraction network to obtain a target feature vector representing the mutual influence between target objects; the target tasks of multiple users are obtained; each of the target tasks requires a type of cloud service; the task feature vector and the target feature vector of each target task are subjected to multi-head attention calculation to obtain the state information under each target task; and the target object matched for each target task under the state information of each target task is obtained through a preset strategy network, and the target object can provide the target type of cloud service required by the target task. The mutual influence between different target objects can be fully considered, and reasonable target objects can be matched for different manufacturing tasks of users among a large number of candidate target objects. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 is a flowchart of a cloud service composition processing method according to an embodiment of the present invention;

[0045] Figure 2 It is a smart manufacturing cloud service platform architecture diagram of the cloud service combination processing method according to an embodiment of the present invention;

[0046] Figure 3 is an architecture diagram of a cloud service composition processing method according to an embodiment of the present invention;

[0047] Figure 4 It is an overall framework diagram of the cloud service composition processing method according to an embodiment of the present invention;

[0048] Figure 5 It is a structural diagram of a cloud service combination processing device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0049] The exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided in order to enable a more thorough understanding of the present invention and to enable the scope of the present invention to be fully communicated to those skilled in the art.

[0050] like Figure 1 As shown, an embodiment of the present invention provides a cloud service composition processing method, including:

[0051] Step 11, obtaining location information and service quality indicator information of multiple target objects; each target object can provide multiple types of cloud services;

[0052] Here, the target object refers to a cloud service provider, and the cloud service provider can provide multiple types of cloud services, and the cloud service refers to an abstract manufacturing resource. In the present invention, there are multiple cloud service providers, each of which is located in a different location, and each cloud service provider can provide different types of cloud services.

[0053] In this embodiment, on the intelligent manufacturing cloud service platform, physical manufacturing resources form manufacturing cloud services through resource virtualization technology. A target object (i.e., cloud service provider) can provide multiple manufacturing cloud services. Assuming there are S target objects, they can be represented as SP1, SP2, SP3...SP s For the target object SP p (provides a set of manufacturing cloud services) has K p cloud service, it can be expressed as Target object SP p It can be expressed as: {loc, fc, tm i , ct i ,rel i ,qua i , sec i , enc i}. Where loc represents the location information of the target object, fc represents the cloud service set owned by the target object, and tm i , ct i ,rel i ,qua i , sec i , enc i They represent the time, cost, reliability, quality, safety, and energy consumption of the i-th manufacturing service of the target object. i , ct i ,rel i ,qua i , sec i , enc i Together they constitute the service quality indicator information of the target object.

[0054] Step 12, extracting the location information and service quality index information of each target object through a feature extraction network to obtain a target feature vector representing the mutual influence between the target objects;

[0055] Step 13, obtaining target tasks of multiple users; each of the target tasks requires a type of cloud service;

[0056] Here, the target task refers to the secondary subtask obtained by decomposing the user's requirements. Specifically, in this embodiment, the requirements submitted by the user are decomposed into different numbers of primary subtasks {T1, T2, T3...T n}, and then decompose it into multiple secondary subtasks according to the functional characteristics of the task That is, the target task. And confirm the structure of the manufacturing subtask, where n is the number of decomposed tasks. Taking the first-level subtask as an example, the first-level subtasks can be parallel. Each first-level subtask starts from the origin, and the task transfer between enterprises depends on logistics transportation, and returns to the origin after completing the task. Here, the origin is the user location, and the enterprise is the target object, that is, the cloud service provider. Each second-level subtask requires a type of manufacturing cloud service.

[0057] Step 14, performing multi-head attention calculation on the task feature vector and the target feature vector of each target task to obtain state information under each target task;

[0058] Here, the attention mechanism is used to integrate the user's target task and the cloud service information that the target object can provide to obtain a state. The state is used to describe and represent the current task and optional resources. It should be noted that since there are multiple target tasks and each target task affects each other, the choice of the target object matched by one target task will affect the target object that can be selected by the next target task. Therefore, as the target task changes, the state information is also constantly changing. When matching the target object for each target task, a new state information is faced.

[0059] Step 15, obtaining a target object matched for each target task under the state information of each target task through a preset policy network, wherein the target object can provide a cloud service of a target type required by the target task.

[0060] Here, different target objects can provide different cloud services, and each target task requires a specific type of cloud service. Through the preset policy network, each target task is matched with a target object that can meet its cloud service requirements and optimize the cost, service quality, etc.

[0061] In this embodiment, Figure 2 As shown in the figure, multiple manufacturing tasks of service demanders (i.e. users) are decomposed into multiple secondary subtasks, i.e. target tasks. The cloud services that can be provided by multiple target objects (i.e. cloud service providers) are abstractly aggregated in the cloud service pool. According to the service type required by the target task, candidate services are selected from the cloud service pool to form a candidate cloud service pool. In the service optimization, the optimal target cloud service of the target object is matched for each secondary subtask to complete the service combination optimization.

[0062] In an optional embodiment of the present invention, step 12 may include:

[0063] Step 121, representing the location information and service quality indicator information of each target object as a feature vector;

[0064] Here, the eigenvector O i It can be expressed as:

[0065]

[0066] Among them, S is the number of target objects, loc0 represents the initial position, loc i Indicates the geographical location of the target object, t i represents the manufacturing time distribution of the target object, c i represents the manufacturing cost distribution of the target object, r i represents the reliability distribution of the target object, q i represents the service quality distribution of the target object, s i represents the safety performance distribution of the target object, e i Represents the energy consumption distribution of the target object.

[0067] Step 122, inputting the feature vector into the neural network layer for linear calculation to obtain the original feature vector;

[0068] Here, these feature vectors are horizontally connected and projected to the same latitude through a layer of neural network to obtain set H0, which is the original feature vector. Set H0 can be expressed as:

[0069] H0=W(O i )+b,i∈[0,S]

[0070] Where W is the weight matrix and b is the bias term.

[0071] Step 123, performing multiple biased attention calculations on the original feature vector in sequence to obtain a feature vector after each biased attention calculation;

[0072] Here, if Figure 3 As shown, preferably, H0 is fed as the initial input into four bias attention layers.

[0073] Step 124, connect the feature vectors calculated after each biased attention calculation to obtain a target feature vector representing the mutual influence between target objects.

[0074] Here, the four bias attention layers extract the features of the interaction between the target objects from the original feature vector, and use the target feature vector Hout The target feature vector H out It can be expressed as:

[0075]

[0076] H out =W·concat(H1, H2, H3, H4)

[0077] Among them, H i is the feature vector output by the i-th bias attention layer, W is the weight matrix, and concat is the function.

[0078] In an optional embodiment of the present invention, step 123 may include:

[0079] Step 1231, performing self-attention calculation on the original feature vector to obtain a high-level feature vector;

[0080] Step 1232, taking a difference between the high-level feature vector and the original feature vector to obtain an attention offset;

[0081] Step 1233, adding the attention offset to the original feature vector to obtain a feature vector after a biased attention calculation;

[0082] Step 1234, repeatedly perform biased attention calculations on the feature vector after the biased attention calculation, to obtain the feature vector after each biased attention calculation.

[0083] Here, if Figure 3 As shown, the calculation process of each biased attention layer includes: multiplying the input original feature vector with the preset Q matrix, K matrix and V matrix respectively to obtain Q vector, K vector and V vector. The Q matrix, K matrix and V matrix are obtained by neural network training. The training process of the neural network is described in detail below and will not be repeated here. After multiplying the Q vector and the K vector, they are normalized by the softmax function to obtain the attention weight. According to the attention weight and the V vector, the weighted output of each element is obtained. The weighted outputs of all elements are summarized to obtain a high-level feature vector after self-attention calculation. The high-level feature vector after self-attention calculation is subtracted from the original feature vector to obtain the attention offset, and the attention offset and the original feature vector are added to obtain the feature vector after biased attention calculation. The biased attention calculation method in this embodiment can make the obtained attention weight more detailed, thereby reducing the impact of noise.

[0084] In an optional embodiment of the present invention, step 14 may include:

[0085] Step 141, obtaining a query vector according to the target feature vector, the target type cloud service required by the target task in the task feature vector, the carrying capacity information of the target object, and the target type cloud service information required by the target task;

[0086] Step 142, subjecting the target feature vector to a layer of linear neural network processing to obtain a key vector and a value vector;

[0087] Step 143, performing multi-head attention calculation on the query vector, key vector and value vector to obtain state information under each target task.

[0088] In this embodiment, the multi-head attention layer is used to integrate the task features into the target features to obtain the state State. Specifically, the current task features and all target object features and other information are used as the query vector, H out After passing through a layer of linear neural network as Key (key vector) and Value (value vector), the output is obtained through multi-head attention calculation to obtain State (state), so that the agent can obtain more comprehensive information. State can be expressed as follows:

[0089]

[0090] Key, Value = W·H out

[0091] State=MHA{Query, Key, Value}

[0092] in, H out is the high-level representation set of all target feature vectors, A t-1 is the location of the choice made by the strategy in the previous step, Cap is the carrying capacity of the service type required by the current task on all target objects, and Ser is the service type required by the current task.

[0093] In an optional embodiment of the present invention, step 15 may include:

[0094] Step 151, obtaining the probability value of selecting each target object under the current state information through a preset strategy network;

[0095] Specifically, step 151 may include: using a preset strategy network according to a formula Probability calculation is performed to obtain the probability value of selecting each target object under the current state information; where P i is the probability of selecting the i-th target object, W is the neural network weight, State t is the status information at step t, Keyi is the high-level feature representation of the i-th target object, d k Key i , Mask is the formulated action masking scheme, and softmax is the normalization function.

[0096] In the preset strategy network, the number of target objects is used as the action space, and the strategy network calculates the probability P of the agent selecting each target object based on the obtained state information. i , select the action with the largest probability Pi as action a t , as follows:

[0097] a t =max(Pi)

[0098] The agent reaches the new target location according to action a, and the state information changes. Action a is the behavior that changes the environment in one iteration.

[0099] In this embodiment, since the cloud service that a certain target object can provide may not meet the needs of the current target task, an action masking scheme is formulated to express that the model contains some constraints when selecting the target object. If these constraints are left to the agent to learn, the learning efficiency will be reduced. In this embodiment, a masking scheme is designed according to the model constraints to ensure that the infeasible target objects will be masked, ensuring the correctness of the final generated solution.

[0100] Step 152, taking the target object with the largest probability value as the target object matched by the current target task;

[0101] Here, the target object selected by action a is the target object matched by the current target task.

[0102] Step 153, according to the target object matched by the current target task, the environment changes, and the state information of the next target task is obtained according to the changed environment;

[0103] Here, after the execution of action a, the environment is changed, and a new state State is calculated based on the new environment. The policy network gives action a based on the newly obtained State. And so on, until all manufacturing tasks complete the service optimization.

[0104] Step 154, matching a target object for a next target task through a preset strategy network according to the state information;

[0105] Step 155, repeatedly calculating the state information under each target task and calculating the probability value through a preset strategy network to obtain the target object matched by each target task.

[0106] In this embodiment, Figure 3 As shown, by collecting the action a of each step, we can get an action trajectory {a1a2...a n}. Among them a1a2...a n Corresponding to the target objects matched by n secondary subtasks, a1a2…a n Together they form the target object sequence matched by the first-level subtask Tt.

[0107] When training the feature extraction network and the policy network, an objective function is set to ensure that the extracted features are the most accurate and the choices made by the policy network are the most accurate. The reward given to the agent in one iteration is given according to the objective function to ensure that the results given by the feature extraction network and the policy network are most in line with actual needs and can maximize benefits, that is, the objective function value is the largest. Specifically, the reward can be expressed as:

[0108]

[0109] where f i There are 8 objective functions in total. They represent that after assigning a target object to each target task, the time, cost, load balance and energy consumption required to complete the target task are minimized; and after assigning a target object to each target task, the service reliability, service quality, security and resource utilization achieved by completing the target task are maximized. That is:

[0110] Min F={f1, f2, f7, f8}; Max F={f3, f4, f5, f6}

[0111] f1=min(T m +T l +T w )

[0112] f2=min(C m +C l +C c )

[0113]

[0114]

[0115] f5=max(safetyper-LS)

[0116]

[0117] f7=minσ(Load p ), p∈S

[0118] f8=min(SE+LE)

[0119]

[0120] Among them, f1 is the cloud service time indicator, f2 is the cloud service cost indicator, f3 is the cloud service reliability indicator, f4 is the cloud service quality indicator, f5 is the cloud service security indicator, f6 is the cloud service resource utilization indicator, f7 is the cloud service load balancing indicator, f8 is the cloud service energy consumption indicator, T m is the manufacturing time, T l is the logistics time, T w is the waiting time, C m is the manufacturing cost, C l is the logistics cost, C c is the manufacturing platform service fee, T is the number of first-level subtasks, J is the number of second-level subtasks, S is the number of target objects, K p is the number of service types owned by the target object, Is the flag of whether the service k of the target object p executes the subtask, is the reliability index of the target object, is the service quality indicator of the target object, safetyper is the safety index of the target object, LS is the safety indicator of the logistics transportation process, is the time consumed by the target object on this subtask, STIME is the sum of the actual available manufacturing time of all target objects, Load p is the current workload of the pth target object, SE is the energy consumed in manufacturing the target object, LE is the energy consumed in transporting the target object, and w1-w8 are the weight coefficients of each objective function.

[0121] The loss function is calculated based on the reward obtained, and the expected value of the cumulative reward is maximized through the policy gradient. Then the parameters of the policy network are updated. The update rules are as follows:

[0122]

[0123] in, represents the probability of solution S, Gπ(S) represents the cumulative reward obtained by the current policy model, and baseline represents the baseline value. The baseline can effectively reduce the variance to accelerate learning and improve the stability and convergence of the algorithm. In this embodiment, the strategy with the best historical performance is selected as the baseline model. At the end of each training round, the t-test is used to test whether the training strategy is better than the baseline model. If the performance of the policy model in this round is better than the baseline model, the baseline model is replaced, otherwise it remains unchanged. Repeat the above process until the policy network converges.

[0124] The implementation process of the above method is further described below with a specific embodiment:

[0125] Example 1

[0126] Step S100, establishing a smart manufacturing platform cloud service combination optimization model, mainly includes the following steps:

[0127] S101 decomposes the overall manufacturing task into multiple manufacturing subtasks and determines the task structure.

[0128] Decompose the requirements submitted by users into different numbers of first-level subtasks {T1, T2, T3…T n}, and then decompose it into multiple secondary subtasks according to the functional characteristics of the task And confirm the structure of manufacturing subtasks, where n is the number of decomposed tasks. Taking the first-level subtasks as an example, the first-level subtasks can be carried out in parallel. Each first-level subtask starts from the origin, and the task transfer between enterprises depends on logistics transportation, and returns to the origin after completing the task. Each second-level subtask requires a type of manufacturing cloud service.

[0129] The optimization problems of intelligent manufacturing cloud service combination are all relatively complex problems. In the actual implementation process, it is necessary to gradually decompose the user's demand items to obtain subtask sequences. The structure of subtask sequences includes sequential structure, parallel structure, selection structure and loop structure.

[0130] S102 On the intelligent manufacturing cloud service platform, physical manufacturing resources form manufacturing cloud services through resource virtualization technology. A service provider can provide multiple manufacturing cloud services. Assuming there are S service providers, they can be represented as SP1, SP2, SP3…SP s For Service Providers SP p (provides a set of manufacturing cloud services) has K p manufacturing service, it can be expressed as For Service Providers (SP) p It can be expressed as: {loc, fc, tm i , ct i ,rel i ,qua i ,Sec i , enc i}. Where loc represents the physical location of the service provider, fc represents the service set owned by the service provider, and tm i , ct i ,rel i ,qua i , sec i , enc iRespectively represent the time and cost consumed by the service provider's i-th manufacturing service, as well as the service reliability, service quality, safety, and energy consumption attribute values. For example, 20 service providers and 16 different service types are defined for the application scenario of automobile engine parts production. Each service provider randomly provides a certain number of manufacturing services among the 16 types of manufacturing services, but there must be at least one service provider for each of the 16 service types. The detailed information of the service provider is shown in the following table.

[0131] Table 1 Detailed information of service providers

[0132] property Number of Services Service load Service Hours Cost of service scope 1-16 8-15 15-100 300-3000 property Service Reliability Quality of Service Service Security Energy consumption scope 0.85-0.99 0.85-0.99 0.85-0.99 10-45

[0133] S103 determines, according to the decomposed subtask sequence, candidate cloud services required to implement corresponding subtasks from the cloud service pool of the cloud server.

[0134] Since the scope and dimension of each optimization objective function are different, the QoS (quality of service) attributes of all candidate services are normalized so that the QoS attribute value of each candidate service is distributed between 0 and 1. Among them, the objective function is divided into two types: positive indicators (such as service quality and reliability) and negative indicators (time and cost). The larger the value of the former, the higher the satisfaction of the service demander, and vice versa for the latter. Therefore, it is necessary to normalize the two types of indicators according to the following formula.

[0135]

[0136]

[0137] in, Represents the normalization of positive indicators; represents the normalization of negative indicators; q represents the QoS attribute value of the candidate service; min q represents the minimum value of the QoS attribute value of all candidate services; max q represents the corresponding maximum value;

[0138] S104 constructs a multi-objective optimization model with the optimization objectives of minimizing time, cost, load balancing, energy consumption and maximizing reliability, service quality, security, and resource utilization. The optimization objectives are shown in the following formula:

[0139] Min F={f1, f2, f7, f8}; Max F={f3, f4, f5, f6}

[0140] f1=min(T m +T l +T w )

[0141] f2=min(C m +Cl +C c )

[0142]

[0143]

[0144] f5=max(safetyper-LS)

[0145]

[0146] f7=minσ(Load p ), p∈S

[0147] f8=min(SE+LE)

[0148]

[0149] Among them, f1 is the cloud service time indicator; f2 is the cloud service cost indicator; f3 is the cloud service reliability indicator; f4 is the cloud service quality indicator; f5 is the cloud service security indicator; f6 is the cloud service resource utilization indicator; f7 is the cloud service load balancing indicator; f8 is the cloud service energy consumption indicator; T m is the manufacturing time; T l is the logistics time; T w is the waiting time; C m is the manufacturing cost; C l is the logistics cost; C c is the manufacturing platform service fee; T is the number of first-level subtasks; J is the number of second-level subtasks; S is the number of service providers; K p is the number of service types that the service provider has; is a flag indicating whether service k of service provider p executes the subtask; It is an indicator of the reliability of the service provider; is the service quality indicator of the service provider; safetyper is the safety index of the service provider; LS is the safety indicator of the logistics transportation process; The time consumed by the service provider on this subtask; STIME is the sum of the actual available manufacturing time of all service providers; Load p is the current workload of the pth service provider; SE is the energy consumed by the service provider for manufacturing; LE is the energy consumed by the service provider for transportation; w1-w8 are the weight coefficients of each objective function.

[0150] Step S200, feature extraction, mainly includes the following steps:

[0151] S201 first uses the location information and QoS indicators of the service provider as the input of the entire neural network model. i

[0152]

[0153] Among them, loc0 represents the initial position, loc i Indicates the geographical location of the enterprise, t i represents the manufacturing time distribution of the service provider, c i represents the manufacturing cost distribution of the service provider, r i represents the reliability distribution of service providers, q i represents the service quality distribution of service providers, s i represents the security performance distribution of service providers, e i Represents the energy consumption distribution of the service provider.

[0154] S202 connects these feature vectors horizontally and projects them to the same latitude through a layer of neural network to obtain set H0.

[0155] H0=W(O i )+b,i∈[0,S]

[0156] The attention offset is obtained by subtracting the original feature from the high-level feature processed by the self-attention neural network, and then added to the original feature to obtain the input feature. This can make the obtained attention weight more detailed, thereby reducing the impact of noise.

[0157] H0 is sent into the neural network as the initial input, and the mutual influence H between enterprise information is extracted from these feature vectors. out .

[0158]

[0159] H out =W·concat(H1, H2, H3, H4)

[0160] Step S300, constructing a policy network, and optimizing service combinations based on enterprise characteristics, specifically includes:

[0161] After S301 converts the original enterprise features into high-level representations, it uses a multi-head attention mechanism to integrate the task features into the enterprise features to obtain the state State.

[0162] Specifically, the current task characteristics, all enterprise characteristics and other information are combined as the query, H outAfter passing through a layer of linear neural network as Key and Value, the State is output through the multi-head attention mechanism, so that the agent can obtain more comprehensive information. State can be expressed as follows:

[0163]

[0164] Key,Value=W·H out

[0165] State=MHA{Query, Key, Value}

[0166] in, H out is a high-level representation set of all enterprise features, A t-1 It is the location of the choice made by the strategy in the previous step, Cap is the carrying capacity of the service type required by the current task in all enterprises, and Ser is the service type required by the current task.

[0167] S302 takes the number of service providers as the action space, and the policy network gives action a as the behavior of changing the environment in one iteration based on the obtained State. Calculate the probability P of selecting each service provider i The specific formula is as follows:

[0168]

[0169] Where W is the neural network weight, State t Represents the state input at step t, Key i represents the high-level feature representation of the i-th service provider, d k Indicates Key i The dimension of Mask is the action masking scheme. The model contains some constraints when selecting service providers. If these constraints are left to the agent to learn, the learning efficiency will decrease. Therefore, we designed a masking scheme based on the model constraints to ensure that unfeasible companies will be masked, ensuring the correctness of the final generated solution. The agent selects the action with the highest probability as action a. t , as follows:

[0170] a t =max(P i )

[0171] Step S400, neural network training is performed according to the obtained service combination optimization result, such as Figure 4 As shown, it mainly includes the following steps:

[0172] S401 The policy network calculates action a according to the current state State, and the environment executes action a to change to a new environment. According to steps S200 and S300, a new state State is calculated, and the policy network gives the next action a according to the newly obtained State. This process is repeated until all manufacturing tasks complete the service optimization work.

[0173] S402 collects the action a of each step, obtains an action trajectory Solution, and calculates the reward Reward based on the trajectory and the optimization objective function, as shown in the following formula:

[0174]

[0175] S403 calculates the Loss function based on the obtained Reward, maximizes the expected value of the cumulative reward through the policy gradient, and then updates the parameters of the policy network. The update rule is as follows:

[0176]

[0177] in, represents the probability of solution S, G π (S) represents the cumulative reward obtained by the current strategy model, and baseline represents the baseline value. The baseline can effectively reduce the variance to accelerate learning and improve the stability and convergence of the algorithm. The appropriate baseline function has an important impact on the performance of the algorithm. The present invention selects the strategy with the best historical performance as the baseline model. At the end of each training round, the t-test is used to test whether the training strategy is better than the baseline model. If the performance of the strategy model in this round is better than the baseline model, the baseline model is replaced, otherwise it remains unchanged.

[0178] S404 repeats the above process until the policy network converges or reaches the maximum number of iterations, and finally saves the trained network model as the final method model.

[0179] S405: Input the decomposed subtask sequence into the final method model, and output a suitable cloud service combination solution.

[0180] The above embodiment of the present invention extracts the feature vector of mutual influence between enterprises, and then uses the multi-head attention mechanism to integrate the task features into the enterprise feature vector to obtain the state State. The State combined with the mutual influence features between enterprises can provide the policy network with richer information, and the policy network can better select the appropriate manufacturing cloud service from a large number of candidate services, and the final cloud service combination optimization scheme has a better overall effect.

[0181] like Figure 5 As shown, the embodiment of the present invention further provides a cloud service composition processing device 50, including:

[0182] An acquisition module 51 is used to acquire location information and service quality indicator information of multiple target objects; each target object can provide multiple types of cloud services;

[0183] The processing module 52 is used to extract the location information and service quality indicator information of each target object through a feature extraction network to obtain a target feature vector representing the mutual influence between the target objects; obtain target tasks of multiple users; each of the target tasks requires a type of cloud service; perform multi-head attention calculations on the task feature vector and the target feature vector of each target task to obtain the state information under each target task; obtain a target object that matches each target task under the state information of each target task through a preset strategy network, and the target object can provide the cloud service of the target type required by the target task.

[0184] Optionally, feature extraction is performed on the location information and service quality indicator information of each target object through a feature extraction network to obtain a target feature vector representing the mutual influence between the target objects, including:

[0185] The location information and service quality index information of each target object are represented as a feature vector;

[0186] Inputting the feature vector into the neural network layer for linear calculation to obtain the original feature vector;

[0187] Perform multiple bias attention calculations on the original feature vector in sequence to obtain a feature vector after each bias attention calculation;

[0188] The feature vectors calculated after each biased attention calculation are connected to obtain a target feature vector representing the mutual influence between target objects.

[0189] Optionally, the original feature vector is subjected to multiple biased attention calculations in sequence to obtain a feature vector after each biased attention calculation, including:

[0190] Perform self-attention calculation on the original feature vector to obtain a high-level feature vector;

[0191] Taking a difference between the high-level feature vector and the original feature vector to obtain an attention offset;

[0192] Adding the attention offset to the original feature vector to obtain a feature vector after a biased attention calculation;

[0193] The feature vector after the biased attention calculation is repeatedly subjected to multiple biased attention calculations to obtain the feature vector after each biased attention calculation.

[0194] Optionally, multi-head attention calculation is performed on the task feature vector and the target feature vector of each target task to obtain state information under each target task, including:

[0195] Obtain a query vector according to the target feature vector, the target type cloud service required by the target task in the task feature vector, the carrying capacity information of the target object, and the target type cloud service information required by the target task;

[0196] Processing the target feature vector with a layer of linear neural network to obtain a key vector and a value vector;

[0197] The query vector, key vector and value vector are subjected to multi-head attention calculation to obtain state information under each target task.

[0198] Optionally, a target object matched for each target task under the state information of each target task is obtained through a preset strategy network, including:

[0199] Through the preset strategy network, the probability value of selecting each target object under the current state information is obtained;

[0200] The target object with the largest probability value is taken as the target object matched by the current target task;

[0201] According to the target object matched by the current target task, the environment changes, and the state information of the next target task is obtained according to the changed environment;

[0202] According to the state information, matching the target object for the next target task through a preset strategy network;

[0203] Repeatedly calculate the state information under each target task and calculate the probability value through the preset strategy network to obtain the target object matched by each target task.

[0204] Optionally, a probability value of selecting each target object under the current state information is obtained through a preset strategy network, including:

[0205] Through the preset strategy network according to the formula

[0206] Probability calculation is performed to obtain the probability value of selecting each target object under the current state information; where P i is the probability of selecting the i-th target object, W is the neural network weight, State t is the status information at step t, Key i is the high-level feature representation of the i-th target object, d k Key i, Mask is the formulated action masking scheme, and softmax is the normalization function.

[0207] Optionally, training the feature extraction network and the strategy network includes: establishing an objective function.

[0208] Optionally, the objective function includes:

[0209] After assigning a target object to each target task, the time, cost, load balancing and energy consumption required to complete the target task are minimized; and after assigning a target object to each target task, the service reliability, service quality, security and resource utilization achieved by completing the target task are maximized.

[0210] It should be noted that the device is a device corresponding to the above method, and all implementation methods in the above method embodiments are applicable to the embodiments of the device and can achieve the same technical effects.

[0211] The embodiment of the present invention further provides a computing device, comprising: a processor, a memory storing a computer program, wherein when the computer program is executed by the processor, the method described above is executed. All implementations in the above method embodiment are applicable to this embodiment and can achieve the same technical effect.

[0212] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0213] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0214] In the embodiments provided by the present invention, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0215] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0216] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0217] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as USB flash drives, mobile hard disks, ROM, RAM, magnetic disks, or optical disks.

[0218] In addition, it should be noted that in the apparatus and method of the present invention, it is obvious that each component or each step can be decomposed and / or recombined. These decompositions and / or recombinations should be regarded as equivalent schemes of the present invention. Moreover, the steps of performing the above-mentioned series of processing can naturally be performed in chronological order according to the order of description, but it is not necessary to perform them in chronological order, and some steps can be performed in parallel or independently of each other. For those of ordinary skill in the art, it is understood that all or any steps or components of the method and apparatus of the present invention can be implemented in any computing device (including processors, storage media, etc.) or a network of computing devices in hardware, firmware, software or a combination thereof, which can be achieved by those of ordinary skill in the art using their basic programming skills after reading the description of the present invention.

[0219] Therefore, the purpose of the present invention can also be achieved by running a program or a group of programs on any computing device. The computing device can be a well-known general device. Therefore, the purpose of the present invention can also be achieved by simply providing a program product containing a program code that implements the method or device. That is to say, such a program product also constitutes the present invention, and the storage medium storing such a program product also constitutes the present invention. Obviously, the storage medium can be any well-known storage medium or any storage medium developed in the future. It should also be pointed out that in the device and method of the present invention, it is obvious that each component or each step can be decomposed and / or recombined. These decompositions and / or recombinations should be regarded as equivalent schemes of the present invention. In addition, the steps of performing the above-mentioned series of processing can naturally be performed in chronological order according to the order of description, but it is not necessary to perform them in chronological order. Some steps can be performed in parallel or independently of each other.

[0220] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A cloud service composition processing method, characterized in that: include: Obtain location information and service quality indicator information of multiple target objects; Each target object can provide multiple types of cloud services; The location information and service quality index information of each target object are extracted through a feature extraction network to obtain a target feature vector representing the mutual influence between target objects; Obtain target tasks of multiple users; each of the target tasks requires a type of cloud service; Perform multi-head attention calculation on the task feature vector and the target feature vector of each target task to obtain state information under each target task; Through the preset policy network, a target object matched for each target task under the state information of each target task is obtained, and the target object can provide a cloud service of the target type required by the target task.

2. The cloud service composition processing method according to claim 1, characterized in that: The location information and service quality index information of each target object are extracted through the feature extraction network to obtain the target feature vector representing the mutual influence between the target objects, including: The location information and service quality index information of each target object are represented as a feature vector; Input the feature vector into a linear neural network layer for calculation to obtain an original feature vector; Perform multiple bias attention calculations on the original feature vector in sequence to obtain a feature vector; The feature vectors calculated after each biased attention calculation are connected to obtain a target feature vector representing the mutual influence between target objects.

3. The cloud service composition processing method according to claim 2, characterized in that: The original feature vector is subjected to multiple bias attention calculations in sequence to obtain a feature vector, including: Perform self-attention calculation on the original feature vector to obtain a high-level feature vector; Taking a difference between the high-level feature vector and the original feature vector to obtain an attention offset; Adding the attention offset to the original feature vector to obtain a feature vector after a biased attention calculation; Perform multiple biased attention calculations to obtain the feature vector after each biased attention calculation.

4. The cloud service composition processing method according to claim 1, characterized in that: Perform multi-head attention calculation on the task feature vector and target feature vector of each target task to obtain state information under each target task, including: Obtain a query vector according to the target feature vector, the target type cloud service required by the target task in the task feature vector, the carrying capacity information of the target object, and the target type cloud service information required by the target task; Processing the target feature vector with a layer of linear neural network to obtain a key vector and a value vector; The query vector, key vector and value vector are subjected to multi-head attention calculation to obtain state information under each target task.

5. The cloud service composition processing method according to claim 1, characterized in that: Through the preset strategy network, the target object matched for each target task under the state information of each target task is obtained, including: Through the preset strategy network, the probability value of selecting each target object under the current state information is obtained; The target object with the largest probability value is taken as the target object matched by the current target task; According to the target object matched by the current target task, the environment is changed, and the state information of the next target task is obtained according to the changed environment; According to the state information, matching the target object for the next target task through a preset strategy network; Repeatedly calculate the state information under each target task and calculate the probability value through the preset strategy network to obtain the target object matched by each target task.

6. The cloud service composition processing method according to claim 5, characterized in that: Through the preset strategy network, the probability value of selecting each target object under the current state information is obtained, including: Through the preset strategy network, according to the formula Calculate the probability and get the probability value of selecting each target object in the current state; where P i is the probability of selecting the i-th target object, W is the neural network weight, State t is the status information at step t, Key i is the high-level feature representation of the i-th target object, d k Key i , Mask is the formulated action masking scheme, and softmax is the normalization function.

7. The cloud service composition processing method according to claim 1, characterized in that: Training the feature extraction network and the strategy network includes: establishing an objective function.

8. The cloud service composition processing method according to claim 7, characterized in that: The objective function includes: After assigning a target object to each target task, the time, cost, load balancing and energy consumption required to complete the target task are minimized; and after assigning a target object to each target task, the service reliability, service quality, security and resource utilization achieved by completing the target task are maximized.

9. A cloud service combination processing device, characterized in that: include: An acquisition module, used to acquire location information and service quality indicator information of multiple target objects; Each target object can provide multiple types of cloud services; The processing module is used to extract the location information and service quality indicator information of each target object through a feature extraction network to obtain a target feature vector representing the mutual influence between the target objects; obtain target tasks of multiple users; each of the target tasks requires a type of cloud service; perform multi-head attention calculation on the task feature vector and the target feature vector of each target task to obtain the state information under each target task; obtain a target object matched for each target task under the state information of each target task through a preset strategy network, and the target object can provide the cloud service of the target type required by the target task.

10. A computing device, characterized in that include: A processor and a memory storing a computer program, wherein when the computer program is executed by the processor, the method according to any one of claims 1 to 8 is performed.

Images