A complex product intelligent design method based on case-based reasoning and reinforcement learning

Abstract

The present invention discloses a complex product intelligent design method based on case reasoning and reinforcement learning, which belongs to the complex product design technical solution, and the steps include: case characterization, structurally characterizing the case through framework theory, accurately describing the characteristics and module composition of the case; case matching, retrieving similar cases in the case library according to the needs, and obtaining relevant experience and solutions; case reconstruction, using a method based on reinforcement learning and weighted sequential cross reconstruction to expand the case library and generate more possible design solutions; case optimization, using a multi-attribute decision-making method to evaluate and optimize the expanded solution set. The present invention adopts the above-mentioned complex product intelligent design method based on case reasoning and reinforcement learning to meet the needs of the engineering design field for efficient and high-quality design solutions, quickly and diversely provide design solutions to designers, and achieve the effect of shortening the design cycle and reducing manufacturing costs.

Description

Technical Field

[0001] The present invention relates to the technical field of complex product design, and in particular to a complex product intelligent design method based on case reasoning and reinforcement learning. Background Art

[0002] With the advancement of science and technology and the diversification of market demand, the high demand for the customized market has driven the transformation of production methods from mass production to small-batch customized production, which means that companies are facing increasingly severe challenges in shortening production cycles and reducing manufacturing costs. In order to achieve efficient and agile design, designers usually draw on and apply historical experience cases, and reconstruct and generate new solutions on this basis, thereby improving design efficiency and reducing costs to meet changing needs. However, this brings two difficulties: how to quickly obtain historical cases that are close to design requirements, and how to quickly modify these cases to meet flexible design requirements.

[0003] In order to quickly retrieve similar historical cases, the case-based reasoning method (CBR) was created. This method provides solutions to new design problems by reusing the experience of similar cases and has been widely used in engineering design and innovative design. However, traditional CBR methods often only focus on individual cases, while ignoring the relationship between modules in the case and lacking support for modular design. In addition, the adjustment and modification of cases is also a challenge. Many CBR systems only stay at the case retrieval stage, lack efficient and high-quality case reconstruction strategies, and cannot make full use of the information of existing cases for effective expansion.

[0004] Modular design enables the rapid design of complex products. It can effectively cope with the diverse and competitive global market. This method can generate new solutions by configuring different modules for previous cases. However, it is a challenge to configure modules effectively and efficiently while considering product characteristics and user needs. Usually, innovative design or optimized design is iterated on the basis of existing cases. In the modular design of complex products, configuration efficiency and configuration quality are often difficult to solve at the same time, so how to balance the two becomes a key issue in configuration design. In this context, a method for complex product design is needed to meet the needs of efficient and high-quality design solutions. Summary of the invention

[0005] The purpose of this invention is to provide a complex product intelligent design method based on case reasoning and reinforcement learning, which solves the demand for efficient and high-quality design solutions in the field of engineering design, and can provide designers with design solutions quickly and in a variety of ways, thereby shortening the design cycle and reducing manufacturing costs.

[0006] To achieve the above object, the present invention provides a complex product intelligent design method based on case reasoning and reinforcement learning, the steps comprising:

[0007] S1. Case characterization: structurally characterize the case based on framework theory to accurately describe the characteristics and module composition of the case;

[0008] S2, case matching, searching similar cases in the case library according to needs to obtain relevant experience and solutions;

[0009] S3, case reconstruction, expands the case library based on reinforcement learning and weighted sequential cross-reconstruction method, and generates multiple design solutions;

[0010] S4, case optimization, using multi-attribute decision-making method to evaluate and optimize the expansion plan set.

[0011] Preferably, the structural characterization of the case in step S1 includes characterizing case features and case modules;

[0012] Case characteristics: Based on the various attributes, characteristics and performance indicators of the product or system, comprehensively describe the characteristics and functions of the product or system in different aspects;

[0013] Case module, based on the various components that make up the product or system, describes the role and characteristics of the module in the product.

[0014] Preferably, the step S2 specifically includes:

[0015] S21. Generate an index by human-computer interaction. According to the problem description of the product, select specific attributes from multiple features of the product as index information. The system grabs keywords and matches instances that meet the retrieval information from the case library. Different similarity calculation methods are used according to different feature objects to calculate the local similarity between the instance and the retrieval requirement.

[0016] S22, performing weighted calculation of the global similarity between the instance and the search requirement according to the weights of each attribute, wherein the weight is a weight combination value adjusted by the user;

[0017] S23. Select instances with high similarity to generate an initial case set.

[0018] Preferably, the similarity calculation method in step S21 includes five types of feature similarity calculation types, including text type, numerical type, interval type, fuzzy type and inclusion type.

[0019] Text

[0020] like

[0021] The calculation formula is:

[0022]

[0023] in, is the resource instance feature concept vector, is the characteristic concept vector of the target solution, and its value is obtained through the vector space model VSM;

[0024] Numeric

[0025] For example, x=a,y=b

[0026] The calculation formula is:

[0027]

[0028] Where a is the attribute value of the resource instance feature x, b is the attribute value of the target solution feature y, α and β are the lower and upper bounds of the partitions respectively;

[0029] Interval type

[0030] Such as x∈[a, a′], y=[b, b′]

[0031] The calculation formula is:

[0032]

[0033] Where a, a′ are interval attribute values ​​of resource instance feature x, b, b′ are interval attribute values ​​of target solution feature y;

[0034] Fuzziness

[0035] For example, a≥max(f) or a≤min(f)

[0036] The calculation formula is:

[0037]

[0038] Where a is the target value of the fuzzy attribute feature, x is the true value of the fuzzy attribute feature, f is the feature attribute, min(f) and max(f) are the lowest and highest domain values ​​of the fuzzy attribute value respectively;

[0039] Inclusive

[0040] For example, x={a0,a1,...,a m}, y = {b0, b1, ..., b n}

[0041] The calculation formula is:

[0042]

[0043]

[0044] Among them, a i is the attribute value of the resource instance feature x, b j is the attribute value of the target solution feature y, Count(x→y) represents the sum of the number of times the elements in x appear in y, count ij represents the indicator function, where when the attribute value a in x i and the attribute value b in y j The value is 1 if they are the same, and 0 if they are different.

[0045] Preferably, the step S3 specifically includes:

[0046] S31, merging modules with the same function in similar cases obtained through the case matching into one module, forming a module set representing different functions, and constituting a module library for solution reconstruction;

[0047] S32, calculating the priority of each module according to a priority calculation method based on reinforcement learning;

[0048] S33. Combine the weighted sequential cross-reconstruction method to perform case reconstruction and generate multiple design solutions.

[0049] Preferably, the step S32 specifically includes:

[0050] S321, using the Markov decision process to model the problem and determine the state, reward, and action of the module; the module state is the currently selected module s i ; The module action is the behavior of the designer selecting the next module to design a ij ; The reward is a reward function based on module relevance and module value r a (s i ,s j ), used to measure the ij From the state i Transfer to state s j The immediate profit is calculated as follows:

[0051]

[0052] Among them, L(s i ,s j ) is the module relevance return, indicating that module s i With modules j The correlation between them, V(s j ) is the module value return, indicating the selection of module s j The value brought by a (s i ,s j ) indicates that from state si Transfer to state s j The reward values ​​obtained, ω1 and ω2 represent L(s i ,s j ) and V(s j ) and satisfies ω1+ω2=1;

[0053] S322. Using the Monte Carlo reinforcement learning method, random sampling is performed, and the average state value function Q is calculated based on the sampling results, thereby generating an action update strategy, solving the Markov decision process model, and ending the loop iteration according to the convergence criterion.

[0054] Preferably, the module relevance report in step S321 includes the connection relevance H(s i ,s j ), assembly association M(s i ,s j ), shape and position correlation G(s i ,s j ) and physical relevance P(s i ,s j ) Four types of indicators, calculated as follows:

[0055]

[0056] Among them, ω H ,ω M ,ω G and ω P are the weight values ​​of various indicators, satisfying their sum to be 1, H ij 、M ij , G ij , P ij are the reward values ​​for each type of association.

[0057] The module value return is based on the product characteristics and design requirements, combined with expert knowledge, user feedback, market demand and other information for analysis and decision-making.

[0058] Preferably, the step S4 specifically includes:

[0059] S41, defining evaluation indicators and weights according to various schemes reconstructed according to the case;

[0060] S42: According to the evaluation index and weight, a method of approximating to an ideal point is used to determine a solution that best meets the user's needs.

[0061] Preferably, the step S42 specifically includes:

[0062] S421, classify the measurement indicators and make the evaluation indicator S positive;

[0063] S422, standardize the evaluation indicators after positive transformation;

[0064] S423, determine the optimal solution vector and the worst solution vector of the evaluation index. If the user has expected data, the user defines the optimal solution vector. The calculation formula is:

[0065]

[0066] in, represents the optimal solution vector, represents the worst solution vector, r ij Indicates the reward received;

[0067] S424. Calculate the distance and relative proximity to the optimal solution vector and the worst solution vector. If the solution is an ideal solution, the value is 1. Otherwise, the closer it is to the negative ideal solution, the closer it is to 0.

[0068] Therefore, the present invention adopts a complex product intelligent design method based on case reasoning and reinforcement learning with the above structure, which has the following beneficial effects:

[0069] (1) The present invention can provide designers with design solutions quickly, flexibly and in a variety of ways, expand the case library, and provide strong support for engineering design;

[0070] (2) In the case characterization and retrieval stage proposed by the present invention, the cases stored in the constructed case library include their respective features and component modules, so that similar cases can be quickly matched according to user needs;

[0071] (3) In the case reconstruction stage, the present invention uses a reinforcement learning method to calculate the priority of each module, providing a clear basis for users to reconstruct cases. By determining the priority of the modules, users can quickly and flexibly generate new alternatives to meet their needs.

[0072] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 It is an overall flow chart of an embodiment of a complex product intelligent design method based on case reasoning and reinforcement learning of the present invention;

[0074] Figure 2 A case representation form diagram based on framework theory of an embodiment of a complex product intelligent design method based on case reasoning and reinforcement learning of the present invention;

[0075] Figure 3 A flowchart of case matching based on similarity retrieval in an embodiment of a complex product intelligent design method based on case reasoning and reinforcement learning of the present invention;

[0076] Figure 4 A schematic diagram of the priority of a reinforcement learning calculation module of an embodiment of a complex product intelligent design method based on case reasoning and reinforcement learning of the present invention;

[0077] Figure 5 A definition diagram of states and actions of an embodiment of a complex product intelligent design method based on case reasoning and reinforcement learning of the present invention;

[0078] Figure 6 A flowchart of a priority algorithm of reinforcement learning in an embodiment of a complex product intelligent design method based on case reasoning and reinforcement learning of the present invention;

[0079] Figure 7 A schematic diagram of a single-point cross-reconstruction scheme of an embodiment of a complex product intelligent design method based on case reasoning and reinforcement learning of the present invention;

[0080] Figure 8 Modeling a radar case library for an embodiment of a complex product intelligent design method based on case reasoning and reinforcement learning of the present invention;

[0081] Fig. 9 A diagram showing the iteration results of a reinforcement learning strategy of an embodiment of a complex product intelligent design method based on case reasoning and reinforcement learning according to the present invention;

[0082] Fig.10 This is the radar solution reconstruction result of an embodiment of a complex product intelligent design method based on case reasoning and reinforcement learning of the present invention. DETAILED DESCRIPTION

[0083] Example

[0084] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0085] Reference Figure 1-10 , a method for intelligent design of complex products based on case-based reasoning and reinforcement learning, e.g. Figure 1 As shown, the steps include:

[0086] S1. Case representation: Based on the framework theory, the case is structured and represented, e.g. Figure 2 As shown, it includes case features and case modules.

[0087] Case characteristics comprehensively describe the characteristics and functions of a product or system in different aspects based on its various attributes, features and performance indicators, including indicators such as quality, cost, efficiency, etc.

[0088] Case module, based on the various components that make up the product or system, describes the role and characteristics of the module in the product.

[0089] S2, case matching, as shown in 3, retrieves similar cases in the case library according to the needs and obtains relevant experience and solutions, including:

[0090] S21. The index is generated by human-computer interaction. The user provides a description of the product problem based on the overall needs and selects specific attributes from multiple features of the product as index information. The system captures keywords and matches instances that meet the retrieval information from the case library. Different similarity calculation methods are used according to different feature objects to calculate the local similarity between the instance and the retrieval needs. The similarity calculation method includes five types of feature similarity calculation types, including text type, numerical type, interval type, fuzzy type and inclusion type.

[0091] Text

[0092] like

[0093] The calculation formula is:

[0094]

[0095] in, is the resource instance feature concept vector, is the characteristic concept vector of the target solution, and its value is obtained through the vector space model VSM;

[0096] Numeric

[0097] For example, x=a,y=b

[0098] The calculation formula is:

[0099]

[0100] Where a is the attribute value of the resource instance feature x, b is the attribute value of the target solution feature y, α and β are the lower and upper bounds of the partitions respectively;

[0101] Interval type

[0102] Such as x∈[a, a′], y=[b, b′]

[0103] The calculation formula is:

[0104]

[0105] Where a, a′ are interval attribute values ​​of resource instance feature x, b, b′ are interval attribute values ​​of target solution feature y;

[0106] Fuzziness

[0107] For example, a≥max(f) or a≤min(f)

[0108] The calculation formula is:

[0109]

[0110] Where a is the target value of the fuzzy attribute feature, x is the true value of the fuzzy attribute feature, f is the feature attribute, min(f) and max(f) are the lowest and highest domain values ​​of the fuzzy attribute value respectively;

[0111] Inclusive

[0112] For example, x={a0,a1,...,a m}, y = {b0, b1, ..., b n}

[0113] The calculation formula is:

[0114]

[0115]

[0116] Among them, a i is the attribute value of the resource instance feature x, b j is the attribute value of the target solution feature y, Count(x→y) represents the sum of the number of times the elements in x appear in y, count ij represents the indicator function, where when the attribute value a in x i and the attribute value b in y j The value is 1 if they are the same, and 0 if they are different.

[0117] S22. Calculate the global similarity between the instance and the retrieval requirement based on the weights of each attribute. The weight is the weight combination value adjusted by the user. The calculation formula is:

[0118]

[0119] Among them, ω Tex ,ω Num ,ω Int ,ω Fuz ,ω Con are the weights of their respective object features, ranging from [0, 1], and satisfying

[0120] S23. Arrange the similarity calculation results, select instances with high similarity, and generate an initial case set.

[0121] S3, case reconstruction, expands the case library based on reinforcement learning and weighted sequential cross-reconstruction methods, and generates multiple design solutions, such as Figure 4 As shown, specifically including:

[0122] S31, merging modules with the same function in similar cases obtained through case matching into one module, forming a module set representing different functions, and constituting a module library for solution reconstruction;

[0123] S32. Calculate the priority of each module according to the priority calculation method based on reinforcement learning, specifically including:

[0124] S321, use Markov decision process to model the problem and determine the module status, reward, and action; the module status is the currently selected module s i , each time a new module i is selected, a new state s is obtained i ,like Figure 5 As shown; module action is the behavior of the designer selecting the next module to design a ij ; The reward is a reward function based on module relevance and module value r a (s i ,s j ), used to measure the ij From the state i Transfer to state s j The immediate benefit obtained can be abbreviated as r ij , calculated as:

[0125]

[0126] Among them, L(s i ,s j ) is the module relevance return, indicating that module s i With modules j The correlation between them, V(s j ) is the module value return, indicating the selection of module s j The value brought by ω1 and ω2 represents L(s i ,s j ) and V(s j ) and satisfies ω1+ω2=1.

[0127] The module value return includes the connection relevance H(s i ,s j ), assembly association M(s i ,s j), shape and position correlation G(s i ,s j ) and physical relevance P(s i ,s j ) Four types of indicators, calculated as follows:

[0128]

[0129] Among them, ω H ,ω M ,ω G and ω P are the weight values ​​of various indicators, satisfying their sum to be 1, H ij 、M ij , G ij , P ij are the reward values ​​for each type of association.

[0130] Specifically, the connection correlation H(s i ,s j ) are two modules s i and j The geometrical relationship between them, including spatial and geometric relationships. The main consideration is the ease of disassembly;

[0131] Assembly associativity M(s i ,s j ) is the difficulty of assembly between modules, the tightness of assembly, and the process rigidity of the assembly unit;

[0132] Shape-position correlation G(s i ,s j ) is whether there is a contact relationship between modules, including different forms of contact such as point contact, line contact, and surface contact;

[0133] Physical correlation P(s i ,s j ) is whether there is a physical relationship such as energy flow, information flow or material flow between modules, and mainly considers the impact of changes in one model on another model.

[0134] The module value return is based on the product characteristics and design requirements, combined with expert knowledge, user feedback, market demand and other information for analysis and decision-making.

[0135] S322, using the Monte Carlo reinforcement learning method, random sampling is performed, and the average state value function Q is calculated based on the sampling results, thereby generating an action update strategy, solving the Markov decision process model, and ending the loop iteration according to the convergence criterion, such as Figure 6 As shown, specifically including:

[0136] S3221, Monte Carlo sampling: The core of Monte Carlo reinforcement learning is to estimate the value of state-action based on random sampling. In this step, a series of states and actions are randomly selected, and multiple design processes are simulated to collect sample data. Each sampling generates a state transition chain, including the following elements: starting state s0, action a taken i , the reward r ij and the final state s N . As shown in the following formula:

[0137] <s0,a0,r 01 ;s1,a1,r 12 ;…;s t-1 , a t-1 , r (t-1)t ;s N >

[0138] By setting the number of sampled nodes T, the states, actions, and reward values ​​in the MDP are sampled according to the strategy. After M samplings, M state transfer chains are generated. The formula is:

[0139]

[0140] S3222. Calculate the average state value function Q: Calculate the state s using the first visit method i The average state value function Q(s i ), the first access means that the same state appears multiple times in a sampling sequence, and only the first result is saved when calculating the state value function; calculate the state s on the mth state transfer chain i -Action a i Value Function Q m (s i , a i ), the calculation formula is:

[0141]

[0142] Where γ is the decay factor, and its value range is [0, 1].

[0143] Calculate the state s i -Action a i The average state-action value function Q(s i , a i ), the calculation formula is:

[0144]

[0145] S3223, action strategy update: different states s obtained based on the state transition chain i -Action a iThe average state-action value function Q(s i , a i ), the ε-greedy strategy is used to adjust the strategy for the next sampling. The core idea of ​​this strategy is to make a trade-off between development and exploration; that is, in most cases, the action with the maximum estimated action value is selected, but the random action is still selected with a certain probability; the calculation formula of this strategy is as follows:

[0146]

[0147] In the formula, ε is (0, 1), A(s i ) indicates state s i The total number of actions.

[0148] S3224, judgment strategy Q(s i , a i ) converges, that is, setting the convergence threshold δ, when |Q t+1 (x, a)-Q t When (x, a)|≤δ, it is converged. If not, step S3221 is performed based on the new strategy; if converged, step S3225 is performed.

[0149] S3225, output the final strategy result μ(a i |s i ).

[0150] S33. According to the priority calculation results of the modules, the position and number of cross-reconstruction are determined, and a weighted sequential cross-reconstruction technique is used to generate a new solution.

[0151] This method reduces the blindness of reconstruction schemes and improves the quality of complex product scheme configuration by first crossing modules with high priorities and then gradually adjusting modules with lower priorities.

[0152] When reconstructing and generating new solutions, refer to the crossover method of genetic algorithms and use the single-point crossover method to reconstruct. According to the priority of the modules, the corresponding crossover positions are set in order from high to low. If a satisfactory solution is not obtained after the initial crossover, subsequent crossovers are performed layer by layer until the modules at the bottom layer are cross-reconstructed in turn, thereby achieving comprehensive solution reconstruction.

[0153] like Figure 7 As shown in the figure, it is a schematic diagram of a single-point crossover. Assuming that there are two initial solution sets, and the priority determined by reinforcement learning is "module 2>module 1>module 3", module 2 is in the first crossover position, and case 3 and case 4 are generated in the first reconstruction. If the solution obtained is not satisfactory, a second crossover is performed, and module 1 is selected for crossover to generate a new case.

[0154] S4, case optimization, using multi-attribute decision-making method to evaluate and optimize the expansion solution set, including:

[0155] S41. Define evaluation indicators and weights based on the various solutions reconstructed from the case;

[0156] S42. Based on the evaluation indicators and weights, the approach to ideal point ranking method is used to determine the solution that best meets the user's needs, including:

[0157] S421. Classify the measurement indicators and make the evaluation indicator S positive. The calculation formula is:

[0158]

[0159] Among them, S′ ij Indicates S ij The result of positive transformation, S ij represents the jth index of solution i (S j ) corresponding to the value;

[0160] A very small indicator means that the target expectation is as small as possible, such as the cost of a product; a very large indicator means that the target expectation is as large as possible, such as the reliability indicator of a product;

[0161] S422, after positive transformation, the evaluation index is standardized, and the calculation formula is:

[0162]

[0163] As shown in Table 1, suppose there are m solutions, each with A i Indicates that N (here 3) measurement indicators are proposed, which is expressed as S j , S 23 It represents the value of the indicator S3 of scheme A2; after being standardized to the interval [0, 1] by the above formula, we get r 23 , Then it is the set of maximum values ​​of each column in the standardized r in the table;

[0164] Table 1

[0165]

[0166] S423, determining the optimal solution vector and the worst solution vector. If the user has expected data, the user can customize the optimal solution vector;

[0167]

[0168] in, represents the optimal solution vector, represents the worst solution vector.

[0169] S424, calculating the distance and relative proximity to the optimal solution vector and the worst solution vector,

[0170]

[0171] in, represents the distance from the worst solution vector, Represents the distance from the optimal solution vector.

[0172] Calculate relative closeness to positive and negative ideal solutions

[0173]

[0174] If the solution is an ideal solution, the value is 1. Otherwise, the closer it is to a negative ideal solution, the closer it is to 0.

[0175] Example 1

[0176] Combined with the radar rapid design case, each step is discussed in detail:

[0177] Traditional radar design methods are time-consuming, labor-intensive, and slow to design. In the face of changing situations and complex usage scenarios, it is difficult to meet the needs of rapid design in a short period of time. Therefore, modern radar system design requires more efficient, flexible, and fast-response design methods to overcome these shortcomings.

[0178] Step S1, characterizing the radar case based on the framework theory, specifically including:

[0179] like Figure 8 As shown in the figure, the radar case library is modeled. Using the framework theory, the characterization content of the radar is divided into two parts: features and modules. Features include basic information of the radar case (such as size, weight, etc.), performance indicators (such as detection distance, detection accuracy, etc.) and other constraint information. Modules are components of the radar, including antenna elements, reflectors, TR components, cold plates, outer frames, etc.

[0180] Step S2, case matching based on similarity retrieval, specifically includes:

[0181] Enter the basic information you want to retrieve, such as overall indicators such as detection distance, accuracy, weight requirements, and temperature limits. The system calculates the similarity results based on the index information and returns the top 3 highest ranked options as alternatives.

[0182] According to the requirements, three similar cases are retrieved to form a solution set V = (V1, V2, V3). Each case includes 8 main modules, namely antenna array, reflector, TR assembly, cold plate, primary power divider, secondary power divider, outer frame, and antenna mounting frame. Taking V1 as an example, it can be recorded as:

[0183] V1=(M11 , M 12 , M 13 , M 14 , M 15 , M 16 , M 17 , M 18 )

[0184] Among them, M 11 、M 12 、M 13 、M 14 、M 15 、M 16 、M 17 、M 18 Corresponding to the 8 modules in this case.

[0185] Step S3, case reconstruction based on reinforcement learning, specifically includes:

[0186] In this embodiment, state S is the currently selected radar module; action A is the user's behavior of selecting another module for design in the next step.

[0187] First, calculate the reward value, which includes two parts: module relevance and module value. Substitute the module relevance value and module value into the formula:

[0188]

[0189] Taking ω1=0.7, ω2=0.3, the final reward value results are shown in Table 2.

[0190] Table 2 Final reward value results

[0191] M11 M12 M13 M14 M15 M16 M17 M18 M11 / 7.34 9.36 2.36 1.76 1.16 0.28 0.28 M12 2.28 / 3.22 0 0 0 5.88 0.28 M13 -0.2 3.22 / 10.68 8.26 4.16 0.56 0.28 M14 -2.4 0 2.92 / 1.2 0.6 7.4 0.28 M15 -2.1 0 1.4 -1.5 / 5.58 4.06 0.28 M16 -1.8 0 -1.8 -1.2 2.6 / 4.06 0.28 M17 -1.5 5.6 -1.5 4.42 3.78 3.78 / 5.88 M18 0 0 0 0 0 0 5.6 /

[0192] Put the defined state, action and reward value into the reinforcement learning model, take the attenuation factor as γ = 0.5, the random probability ε = 0.5, the convergence criterion as δ = 0.1, and the iterative results are as follows Fig. 9 , and reached convergence conditions at the 50th policy iteration.

[0193] According to the results, the priority order of each module is:

[0194] M 11 >M 12 =M 14 >M 13 =M 16 >M 15 >M 17 >M 18

[0195] According to the priority calculation results, the first reconstruction scheme selects the antenna array (M11 ) to cross and get the reconstruction result as Fig.10 , expanded from 3 options to 9 options.

[0196] Step S4, case selection, specifically includes:

[0197] The 9 solutions obtained through matching and reconstruction are used as the candidate case set for case optimization. Four evaluation indicators are proposed, namely detection distance, power, energy consumption and weight. Among them, detection distance and power belong to extremely large indicators, and energy consumption and weight belong to extremely small indicators. The design expectations given are: detection distance reaches 75km, power reaches 280W, power consumption is less than 570W, and weight does not exceed 7800g. TOPSIS is used for calculation, and the scores of each solution are shown in Table 3.

[0198] Table 3 TOPSIS score results of alternative solution set

[0199]

[0200] As can be seen from the table, the best solution is V4, so solution 4 is selected as the final preferred solution. However, if the user is not satisfied with V4, a second reconstruction is performed. In the second round of crossover, based on the 9 solutions obtained, M2 or M4 is selected for crossover, and the solution library can be expanded to 27. Crossover is performed in sequence until a solution that meets the requirements is found.

[0201] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A complex product intelligent design method based on case-based reasoning and reinforcement learning, characterized in that the steps include: S1, case representation, structural representation of the case based on framework theory; S2, case matching, searching similar cases in the case library according to needs to obtain relevant experience and solutions; S3, case reconstruction, expands the case library based on reinforcement learning and weighted sequential cross-reconstruction method, and generates multiple design solutions; Specifically include: S31, merging modules with the same function in similar cases obtained through the case matching into one module, forming a module set representing different functions, and constituting a module library for solution reconstruction; S32. Calculate the priority of each module according to the priority calculation method based on reinforcement learning, specifically including: S321, using Markov decision process to model the problem and determine the state, reward, and action of the module; the module state is the currently selected module s i ; Module action is the behavior of the designer selecting the next module to design a ij ; The reward is a reward function based on module relevance and module value r a (s i ,s j ), used to measure the ij From the state i Transfer to state s j The immediate profit is calculated as follows: Among them, L(s i ,s j ) is the module relevance return, indicating that module s i With modules j The correlation between them, V(s j ) is the module value return, indicating the selection of module s j The value brought by a (s i ,s j ) indicates that from state s i Transfer to state s j The reward values ​​obtained, ω1 and ω2 represent L(s i ,s j ) and V(s j ) and satisfies ω1+ω2=1; The relevance return includes the connection relevance H(s i ,s j ), assembly association M(s i ,s j ), shape and position correlation G(s i ,s j ) and physical relevance P(s i ,s j ) Four types of indicators, calculated as follows: Among them, ω H ,ω M ,ω G and ω P are the weights of various indicators, H ij 、M ij , G ij , P ij are the reward values ​​for various types of associations; The module value return is based on the product characteristics and design requirements, combined with expert knowledge, user feedback, market demand and other information for analysis and decision-making; S322, adopting the Monte Carlo reinforcement learning method, performing random sampling, calculating the average state value function Q according to the sampling results, thereby generating an action update strategy, solving the Markov decision process model, and ending the loop iteration according to the convergence criterion; S33, combining the weighted sequential cross-reconstruction method to reconstruct the case and generate multiple design solutions; When reconstructing and generating new solutions, refer to the crossover method of genetic algorithms and use the single-point crossover method to reconstruct. According to the priority of the modules, set the corresponding crossover positions in order from high to low. If a satisfactory solution is not obtained after the initial crossover, subsequent crossovers are performed layer by layer until the modules at the bottom layer are cross-reconstructed in turn, thereby realizing the reconstruction of the comprehensive solution. S4, case optimization, using multi-attribute decision-making method to evaluate and optimize the expansion plan set.

2. The complex product intelligent design method based on case-based reasoning and reinforcement learning according to claim 1, characterized in that: The structural characterization of the case in step S1 includes characterizing case features and case modules; Case characteristics: Based on the various attributes, characteristics and performance indicators of the product or system, comprehensively describe the characteristics and functions of the product or system in different aspects; Case module, based on the various components that make up the product or system, describes the role and characteristics of the module in the product.

3. The complex product intelligent design method based on case reasoning and reinforcement learning according to claim 1 is characterized in that: The step S2 specifically includes: S21, calculating the local similarity between the instance and the retrieval requirement by using different similarity calculation methods according to different feature objects; S22, performing weighted calculation of the global similarity between the instance and the search requirement according to the weights of each attribute, wherein the weight is a weight combination value adjusted by the user; S23. Select instances with high similarity to generate an initial case set.

4. The complex product intelligent design method based on case-based reasoning and reinforcement learning according to claim 3 is characterized by: The similarity calculation method in step S21 includes five types of feature similarity calculation types, including text type, numerical type, interval type, fuzzy type and inclusion type.

5. The complex product intelligent design method based on case-based reasoning and reinforcement learning according to claim 1 is characterized in that: The step S4 specifically includes: S41, defining evaluation indicators and weights according to various schemes reconstructed according to the case; S42: According to the evaluation index and weight, a method of approximating to an ideal point is used to determine a solution that best meets the user's needs.

6. The complex product intelligent design method based on case-based reasoning and reinforcement learning according to claim 5 is characterized in that: The step S42 specifically includes: S421, classify the measurement indicators and make the evaluation indicator S positive; S422, standardize the evaluation indicators after positive transformation; S423, determine the optimal solution vector and the worst solution vector of the evaluation index. If the user has expected data, the user defines the optimal solution vector. The calculation formula is: in, represents the optimal solution vector, represents the worst solution vector, r ij Indicates the reward received; S424. Calculate the distance and relative proximity to the optimal solution vector and the worst solution vector. If the solution is an ideal solution, the value is 1. Otherwise, the closer it is to the negative ideal solution, the closer it is to 0.

Images