A method for clustering optimization design of system actions

Abstract

The present invention discloses a system action clustering optimization design method, the method comprising the following steps: decomposing and obtaining detailed actions required to realize system functions through function-behavior-action analysis; constructing action label vectors, clustering actions based on the similarity of action label vectors, and obtaining action subsets to be integrated; adopting an integrated action combination optimization method based on branch and bound to design the actions in an integrated manner, thereby correspondingly mapping and obtaining integrated physical domain design parameters / components. In the present invention, the relationship between functional requirements, behaviors, and actions in the system design process can be fully analyzed, and the integration of physical domain design parameters can be achieved through clustering of similar actions, effectively reducing design redundancy; at the same time, the integrated optimization process based on branch and bound can consider the constraints in the design process and the cost brought by the integrated design, thereby selecting a design scheme with the best comprehensive performance.

Description

Technical Field

[0001] The invention belongs to the field of system engineering and relates to a system action clustering optimization design method. Background Art

[0002] Axiomatic design is a new product design concept theory, the core content of which is the mapping design process between user domain, functional domain, physical domain and process domain. It provides a set of standards and specifications for the system design process through the role of design axioms, which can effectively reduce the influence of human factors of designers in the design process and guide the production of innovative designs. It is currently a research hotspot in the fields of system engineering and industrial design.

[0003] At present, the research on axiomatic design methods includes the expansion of design domains, the combination with other design methods, and the application research of axiomatic design theory. The above research has further expanded the scope of application of axiomatic design, but there are still certain problems. For example, in the axiomatic design process, the functional requirements and the physical domain design parameters have a one-to-one correspondence. When the system functional requirements are numerous and there are overlaps, it will lead to a large amount of design redundancy, which is not conducive to the full utilization of design resources and space. Therefore, an innovative design method is needed to fully analyze the interaction between the subsystems and components in the system and cluster and optimize similar actions to achieve integrated design. Summary of the invention

[0004] In order to solve the above problems existing in the background technology, the present invention provides a system action clustering optimization design method.

[0005] The objective of the present invention is achieved through the following technical solutions:

[0006] A system action clustering optimization design method includes the following steps:

[0007] Step 1: Through function-behavior-action analysis, decompose the detailed actions required to realize the system functions, providing a basis for action clustering optimization. The specific steps are as follows:

[0008] Step 1.1, Requirements-Function Analysis:

[0009] Starting from the top-level user needs, decompose from top to bottom to get the implementation path of the function - behavior;

[0010] Step 1.2, Behavior-Motion Analysis:

[0011] For the obtained behavior, an "input-output" black box model is established. Based on the functional system analysis technology, the implementation path of the behavior is decomposed to obtain the system action;

[0012] Step 2: Construct an action label vector, cluster actions based on the similarity of the action label vectors, and obtain a subset of actions to be integrated. The specific steps are as follows:

[0013] Step 2.1: Construction of action label vector:

[0014] Based on the results of the function-behavior-action analysis, the types of action labels are summarized and analyzed;

[0015] Step 2.2, calculation of action similarity:

[0016] ① Normalize the number of labels;

[0017] ② Use word frequency to indicate the frequency of the label in the action. The word frequency tf corresponding to the label w in a certain action w The specific calculation formula is:

[0018]

[0019] ③ Calculate the inverse text frequency idf of a specific tag. The specific calculation formula is:

[0020]

[0021] Where: D is the total number of actions; |{j:t i ∈d j}| represents the number of actions containing label t;

[0022] ④ For an action j, the TF-IDF value w of label i ij for:

[0023] w ij =tf ij ·idf i

[0024] After weighting the label vector of the action, a weighted label vector is obtained;

[0025] ⑤ Based on the cosine similarity calculation formula, calculate the similarity between two actions;

[0026] Step 2.3: Action clustering based on K-means algorithm:

[0027] ① Select the initial value k of the number of action clusters;

[0028] ② Select k cluster centers and perform action clustering based on the traditional k-means algorithm;

[0029] ③ Calculate the maximum Euclidean distance between the actions in each cluster and the center point in the action clustering results;

[0030] ④ Determine whether the change in the maximum Euclidean distance between two adjacent values ​​is less than the threshold. If so, go to ⑤; otherwise, set k = k + 1 and go to ②;

[0031] ⑤After the action clustering is completed, the action clustering results are output;

[0032] Step 3: For the action subsets obtained by clustering, an integrated action combination optimization method based on branch and bound is used to design the actions in an integrated manner, thereby correspondingly mapping the integrated physical domain design parameters / components. The specific steps are as follows:

[0033] Step 3.1: Construction of action integrated optimization model:

[0034] ① According to the action TF-IDF label vector, after clustering based on label similarity, the action clustering with certain similarity is realized, and several action subsets are obtained. The actions in each action subset have the possibility of integrated design;

[0035] ② Carry out integrated optimization design for the actions in each action subset, and transform this type of integrated design problem into a mixed integer programming problem: solve the integrated combination of actions that satisfies certain constraints and makes the given objective function take the extreme value, that is:

[0036] Find x=[x1,x2,…,x n ],(x i =0 / 1)

[0037] min J=αM+βC+ηW

[0038]

[0039] φ i (X i )≤0

[0040] Where: The design parameter x = [x1, x2, ..., x n ] is the characteristic variable of whether the action is merged, and n means that there are n components in the parameters to be designed:

[0041]

[0042] Where: i is the action number, M is the structural mass, C is the cost, W is the power, α, β, η are the weights, is the sum of multiple component indicators, φ i (X i ) is a single component constraint;

[0043] Step 3.2: Solve the mixed integer programming based on branch and bound:

[0044] Pruning occurs when:

[0045] Pruning situation 1: The constraints corresponding to the actions conflict and cannot be merged, so the branch is discarded;

[0046] Pruning situation 2: The merging cost of the action is too high, which causes the performance index value to increase, so the branch is discarded;

[0047] Pruning situation 3: The performance index does not drop significantly, so the branch is temporarily retained or discarded after a certain number of iterations.

[0048] Compared with the prior art, the present invention has the following advantages:

[0049] In the present invention, the relationship between functional requirements, behaviors, and actions in the system design process can be fully analyzed, and the integration of physical domain design parameters can be achieved through clustering of similar actions, effectively reducing design redundancy; at the same time, the integrated optimization process based on branch and bound can consider the constraints in the design process and the cost brought by the integrated design, so as to select the design scheme with the best comprehensive performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 The overall scheme diagram of the system action clustering optimization design method;

[0051] Figure 2 This is a schematic diagram of action branching and delimitation;

[0052] Figure 3 Schematic diagram of branch and bound for subset 1 actions. DETAILED DESCRIPTION

[0053] The technical solution of the present invention is further described below in conjunction with the accompanying drawings, but is not limited thereto. Any modification or equivalent replacement of the technical solution of the present invention without departing from the spirit and scope of the technical solution of the present invention should be included in the protection scope of the present invention.

[0054] The present invention provides a system action clustering optimization design method, such as Figure 1 As shown, the method comprises the following steps:

[0055] Step 1: Through function-behavior-action analysis, decompose the detailed actions required to realize the system functions, providing a basis for action clustering optimization. The specific steps are as follows:

[0056] Step 1.1, Requirements-Function Analysis:

[0057] Starting from the top-level user needs, we decompose them from top to bottom to get the way to implement the functions - behaviors.

[0058] Step 1.2, Behavior-Motion Analysis:

[0059] For the obtained behavior, an "input-output" black box model is established. Based on the functional system analysis technology, the implementation path of the behavior is decomposed to obtain the system action.

[0060] Step 2: Construct an action label vector, perform action clustering based on the similarity of the action label vector, and obtain a subset of actions to be integrated. The specific steps are as follows:

[0061] Step 2.1: Construction of action label vector:

[0062] According to the results of function-behavior-action analysis, the types of action labels are summarized and analyzed. Action labels include input type, output type, constraints, etc. For a certain action, its corresponding label vector is:

[0063] {Input information, input energy, output information, output energy...constraints (data volume)}

[0064] The elements in the label vector should cover all the characteristics of the action. The element type can be freely defined and determined by the designer according to the specific application scenario. When the element in the label vector is 1, it means that the action has this label. For example, when the label vector is {1, 0, 1, 0, 1}, it means that its input type is "information", the output type is "information", the constraint type is "data volume", and the input and output do not contain "energy".

[0065] Step 2.2, calculation of action similarity:

[0066] Based on the action label vector, each action in the design process can be expressed in a unified form. For the label vector of each action, if the label exists in most actions, there is no need for clustering; if the label only exists in a few actions, it has a good ability to distinguish categories and can reflect the characteristics of the action. In order to more objectively judge the similarity between actions, it is necessary to measure the importance of the label and weight it in the similarity calculation process.

[0067] First, we use word frequency (tf) to represent the frequency of the tag (keyword) in the action. Therefore, we first normalize the number of tags (number of times a tag appears / total number of tags) to prevent it from being biased towards actions with more tags. The word frequency tf corresponding to the tag w in a certain action w The specific calculation formula is:

[0068]

[0069] Next, we calculate the inverse text frequency (idf) of a specific tag by dividing the total number of actions by the number of actions containing the tag and taking the logarithm of the quotient.

[0070] If there are fewer documents containing tag t and the idf is larger, it means that the tag has good category distinction ability. The specific calculation formula is:

[0071]

[0072] Where: D is the total number of actions; |{j:t i ∈d j}| represents the number of actions containing label t.

[0073] A high label frequency within a specific action and a low label frequency of the label in the entire action set can produce a high-weighted TF-IDF value. Therefore, the TF-IDF value tends to filter out common labels and retain important labels. For an action j, the TF-IDF value w of label i is ij for:

[0074] w ij =tf ij ·idf i

[0075] After weighting the action label vector, the weighted label vector can be obtained. The similarity between two actions is further calculated based on the cosine similarity calculation formula.

[0076] Step 2.3: Action clustering based on K-means algorithm:

[0077] ① Select the initial value k of the number of action clusters;

[0078] ② Select k cluster centers and perform action clustering based on the traditional k-means algorithm;

[0079] ③ Calculate the maximum Euclidean distance between the actions in each cluster and the center point in the action clustering results;

[0080] ④ Determine whether the change in the maximum Euclidean distance between two adjacent values ​​is less than the threshold. If so, go to ⑤; otherwise, set k = k + 1 and go to ②;

[0081] ⑤After the action clustering is completed, the action clustering results are output.

[0082] Step 3: For the action subsets obtained through clustering, an integrated action combination optimization method based on branch-and-bound is used to design the actions in an integrated manner, thereby obtaining integrated physical domain design parameters / components correspondingly mapped. The specific steps are as follows:

[0083] Step 3.1: Construction of action integration optimization model:

[0084] According to the action TF-IDF label vector, after clustering based on label similarity, we realized the clustering of actions with certain similarity and obtained several action subsets. The actions in each action subset have the possibility of integrated design. Next, we conduct integrated optimization design for the actions in each action subset. This type of integrated design problem is transformed into a mixed integer programming problem: solve the integrated combination of actions that satisfies certain constraints and makes the given objective function take the extreme value. That is:

[0085]

[0086] Where: The design parameter x = [x1, x2, ..., x n ] is the characteristic variable of whether the action is merged:

[0087]

[0088] Where i is the action number. The purpose of integrated design is light structure, low cost, low power, etc. Therefore, the objective function is the weighted structure mass M, cost C, and power W, where the weights α, β, and η are predetermined by the designer. In addition, the constraints of the optimization problem include combined inequality constraints (the sum of multiple component indicators ) and the single-component constraint φ i (X i )(guarantee normal operation).

[0089] Step 3.2: Solve the mixed integer programming based on branch and bound:

[0090] Combination Figure 2 Taking the integrated optimization design process of three actions as an example, an integrated optimization design method is given. Among them, action 1 (1) means that action 1 is integrated with other actions, and vice versa, action 1 (0) means that action 1 is not integrated with other actions.

[0091] First, determine whether action 1 and action 2 should be integrated. When action 1 and action 2 are integrated, map the physical domain components and calculate the performance index. The performance index value is better than the case where action 1 and action 2 are not integrated. Therefore, cut off the branch where action 1 and action 2 are not integrated. Further, take action 2 (1) as the starting point to determine whether action 2 and action 3 should be integrated. When action 2 and action 3 are integrated, the design cost is too high, so the branch is discarded, and the integration of action 1 and action 2 is the current optimal design solution.

[0092] In summary, pruning is performed when the following situations occur:

[0093] Pruning situation 1: The constraints corresponding to the actions conflict and cannot be merged, so the branch is discarded;

[0094] Pruning situation 2: The merging cost of the action is too high, which causes the performance index value to increase, so the branch is discarded;

[0095] Pruning situation 3: The performance index does not drop significantly, so the branch is temporarily retained or discarded after a certain number of iterations.

[0096] Example:

[0097] This embodiment takes the design of a spacecraft power system including the functions of ascent, orbit entry, and on-orbit operation as an example to illustrate the specific steps of the method:

[0098] Step 1: Through function-behavior-action analysis, decompose the detailed actions required to achieve the system function of "providing power for ascent, orbit entry, and on-orbit operation", providing a basis for action clustering optimization. The specific steps are as follows:

[0099] Step 1.1, Requirements-Function Analysis:

[0100] Starting from the top-level user demand of "being able to provide power for ascent, orbit entry, and on-orbit operation", three specific functional requirements are obtained: {FR1: the system can provide power for attitude and orbit control during ascent; FR2: the system can provide power for attitude and orbit control during orbit entry; FR3: the system can provide power for attitude and orbit control during on-orbit operation}. FR represents the functional requirement.

[0101] Decomposed from top to bottom, we can get the implementation path of each function - behavior: {BE1: provide attitude and orbit control power during the ascent phase; BE2: provide attitude and orbit control power during the entry phase; BE3: provide attitude and orbit control power during the on-orbit phase}. Among them, BE represents behavior.

[0102] Step 1.2, Behavior-Motion Analysis:

[0103] For the obtained behavior, an "input-output" black box model is established. Based on the functional system analysis technology, the implementation path of the behavior is decomposed, that is, the system actions are as follows:

[0104] The corresponding actions of BE1 "providing ascending stage attitude and trajectory control power" are as follows: {AC1-1: output ascending stage trajectory control thrust; AC1-2: output ascending stage attitude control torque}, where AC represents the action.

[0105] The corresponding actions of BE2 "providing trajectory control power for the orbital entry phase" are as follows: {AC2-1: output trajectory control thrust for the orbital entry phase; AC2-2: output attitude control torque for the orbital entry phase}.

[0106] The corresponding actions of BE3 "Providing on-orbit attitude and trajectory control power" are as follows: {AC3-1: output trajectory change control thrust; AC3-2: output attitude maneuvering torque}.

[0107] Step 2: Construct an action label vector, perform action clustering based on the similarity of the action label vector, and obtain a subset of actions to be integrated. The specific steps are as follows:

[0108] Step 2.1: Construction of action label vector:

[0109] According to the results of function-behavior-action analysis, the action label types of the power system are summarized and analyzed. Each action of the power system involves the conversion of information and energy at different stages. Therefore, the time period in which the action occurs and the type of information energy conversion are used as action labels. The labels obtained are as follows: {time-ascending segment; input information type-thrust command; time-on-orbit segment; output-energy; time-on-orbit segment; output energy type-torque; input-information; output energy type-thrust; input information type-torque command}.

[0110] The label vector for each action is obtained as follows:

[0111]

[0112] Each row represents an action, and when the element is 1, it means that the action has this label.

[0113] Step 2.2, calculate action similarity:

[0114] For the label vectors of each action of the dynamic system, if the label exists in most actions, there is no need for clustering; if the label only exists in a few actions, it has a good ability to distinguish categories and can reflect the characteristics of the action. In order to more objectively judge the similarity between actions, it is necessary to measure the importance of the label and weight it in the similarity calculation process.

[0115] 1) Calculate the word frequency (tf) of the tag (keyword). The word frequency tf corresponding to the tag w in each action w for:

[0116]

[0117] 2) Calculate the inverse text frequency idf for each specific tag:

[0118]

[0119] Where: D is the total number of actions; |{j:t i ∈d j}| represents the number of actions containing label t.

[0120] 3) For each action, calculate the TF-IDF value of the label (i.e., the weighted label vector). For an action j, the TF-IDF value w of label i is ij for:

[0121] w ij =tf ij ·idf i

[0122] For this system, the TF-IDF value is calculated as:

[0123]

[0124] Step 2.3: Action clustering based on K-means algorithm:

[0125] ① Select the initial value k=1 for the number of action clusters, and set the threshold of cluster calculation to 1%;

[0126] ② Select k cluster centers and perform action clustering based on the traditional k-means algorithm;

[0127] ③ Calculate the maximum Euclidean distance between the actions in each cluster and the center point in the action clustering results;

[0128] ④ Determine whether the change of the maximum Euclidean distance between two adjacent values ​​is less than the set threshold value 1%. If so, go to ⑤; otherwise, set k = k + 1 and go to ②;

[0129] ⑤After the action clustering is completed, the action clustering results are output.

[0130] During the calculation process, the distances obtained in each iteration were 0.3618, 0.3545, and 0.3542, respectively. When iterating to the third time, the distance change from the second time was within 1%, and the calculation jumped out. At this time, there were 2 cluster types. Therefore, two action sets were obtained by clustering: {Set 1: Output ascent segment trajectory control thrust (AC1); Output orbit change control thrust (AC5); Output orbit entry segment trajectory control thrust (AC3)}{Set 2: Output ascent segment control torque (AC2); Output orbit entry segment control torque (AC4); Output attitude maneuvering torque (AC6)}.

[0131] Step 3: For the subset of power system actions obtained by clustering, an integrated action combination optimization method based on branch and bound is used to design the actions in an integrated manner, thereby obtaining integrated physical domain design parameters / components correspondingly mapped. Subset 1 is used as an example for explanation.

[0132] Step 3.1: Construction of action integrated optimization model:

[0133] For the power system, the purpose of integrated design is mainly to make the structure light, so the weighting of mass is α=1, and other weights are 0. Then for the power system, the optimization model is:

[0134]

[0135] The parameter to be designed x = [x1, x2, ..., x6] is the characteristic variable of whether the action is merged:

[0136]

[0137] Where: i is the action number. i (X i )≤0 indicates whether the force / torque provided by the action meets the spacecraft requirements.

[0138] Step 3.2: Solve the mixed integer programming based on branch and bound:

[0139] Combination Figure 3 , taking the integrated optimization design process of subset 1 as an example, an integrated optimization design method is given. First, determine whether actions 1 and 3 are integrated. When actions 1 and 3 are integrated, only one propulsion system is used throughout the ascent and orbital stages, which can save structural mass. However, due to the large difference in thrust constraints between the two flight stages, the existing technology cannot achieve such a large range of variable thrust, which meets the pruning situation 1. Therefore, actions 1 and 3 are not merged, and the merged branches are pruned. Similarly, actions 1 and 5 should not be merged. Continue to take action 3 as the starting point to determine whether action 5 can be merged with it. After actions 3 and 5 are merged, the required physical domain components are reduced from 2 propulsion systems to 1 set, the mass is reduced from 28kg to 14kg, and the performance index J is better, so the design of actions 3 and 5 is merged.

[0140] In summary, the design results are as follows: the output of the ascent trajectory control thrust (AC1) is realized by one component, and the output of the orbit entry trajectory control thrust (AC3) and the output of the orbit change control thrust (AC5) are realized by one component.

Claims

1. A system action clustering optimization design method, characterized by The method comprises the following steps: Step 1: Through function-behavior-action analysis, decompose the detailed actions required to realize the system functions, providing a basis for action clustering optimization; Step 2: Construct an action label vector, cluster actions based on the similarity of the action label vectors, and obtain a subset of actions to be integrated. The specific steps are as follows: Step 2.1: Construction of action label vector: Based on the results of the function-behavior-action analysis, the types of action labels are summarized and analyzed; Step 2.2, calculation of action similarity: ① Normalize the number of labels; ② Use word frequency to indicate the frequency of the label in the action. The word frequency tf corresponding to the label w in a certain action w The specific calculation formula is: ③Calculate the inverse text frequency idf of a specific tag; ④ For an action j, the TF-IDF value w of label i ij for: w ij =tf ij ·idf i After weighting the label vector of the action, a weighted label vector is obtained; ⑤ Based on the cosine similarity calculation formula, calculate the similarity between two actions; Step 2.3: Action clustering based on K-means algorithm: ① Select the initial value k of the number of action clusters; ② Select k cluster centers and perform action clustering based on the traditional k-means algorithm; ③ Calculate the maximum Euclidean distance between the actions in each cluster and the center point in the action clustering results; ④ Determine whether the change in the maximum Euclidean distance between two adjacent values ​​is less than the threshold. If so, go to ⑤; otherwise, set k = k + 1 and go to ②; ⑤After the action clustering is completed, the action clustering results are output; Step 3: For the action subsets obtained by clustering, an integrated action combination optimization method based on branch and bound is used to design the actions in an integrated manner, so as to obtain the integrated physical domain design parameters / components through corresponding mapping.

2. The system action clustering optimization design method according to claim 1 is characterized in that The specific steps of step 1 are as follows: Step 1.1, Requirements-Function Analysis: Starting from the top-level user needs, decompose from top to bottom to get the implementation path of the function - behavior; Step 1.2, Behavior-Motion Analysis: For the obtained behavior, an "input-output" black box model is established. Based on the functional system analysis technology, the implementation path of the behavior is decomposed to obtain the system action.

3. The system action clustering optimization design method according to claim 1 is characterized in that The calculation formula of the inverse text frequency idf is: Where: D is the total number of actions; |{j:t i ∈d j }| represents the number of actions containing label t.

4. The system action clustering optimization design method according to claim 1 is characterized in that The specific steps of step 3 are as follows: Step 3.1: Construction of action integrated optimization model: ① According to the action TF-IDF label vector, after clustering based on label similarity, the action clustering with certain similarity is realized, and several action subsets are obtained. The actions in each action subset have the possibility of integrated design; ② Carry out integrated optimization design for the actions in each action subset, and transform this type of integrated design problem into a mixed integer programming problem: solve the integrated combination of actions that satisfies certain constraints and makes the given objective function take the extreme value, that is: Find x=[x1,x2,…,x n ],(x i =0 / 1) min J=αM+βC+ηW f i (X i )≤0 Where: The design parameter x = [x1, x2, ..., x n ] is the characteristic variable of whether the action is merged, and n means that there are n components in the parameters to be designed: Where: i is the action number, M is the structural mass, C is the cost, W is the power, α, β, η are the weights, is the sum of multiple component indicators, φ i (X i ) is a single component constraint; Step 3.2: Solve the mixed integer programming based on branch and bound: Pruning occurs when: Pruning situation 1: The constraints corresponding to the actions conflict and cannot be merged, so the branch is discarded; Pruning situation 2: The merging cost of the action is too high, which causes the performance index value to increase, so the branch is discarded; Pruning situation 3: The performance index does not drop significantly, so the branch is temporarily retained or discarded after a certain number of iterations.

Images