A Modular Design Method for Manufacturing Service System Based on Intelligent Case Reasoning

Abstract

The present invention discloses a modular design method for a manufacturing service system based on intelligent case-based reasoning, belonging to the technical field of multi-source heterogeneous service-oriented manufacturing service system design, including S1, feature extraction, constructing a target case based on the description of the user's requirements for the system design scheme, extracting the key features of the target case and analyzing the modular composition of the target case; S2, case retrieval, retrieving historical cases with high similarity to the target case in the case library; S3, case reuse, based on reinforcement learning, reorganizing the modules of the retrieved historical cases with high similarity to quickly generate a system design scheme with high fit with the target case; S4, feasibility evaluation. The present invention adopts the above-mentioned modular design method for a manufacturing service system based on intelligent case-based reasoning, and realizes the modular design of the manufacturing service system by integrating case-based reasoning and machine learning technologies, helping service-oriented manufacturing enterprises reduce the system development cycle and cost.

Description

Technical Field

[0001] The present invention relates to the technical field of the design of multi-source heterogeneous service-oriented manufacturing service systems, and particularly relates to a modular design method for a manufacturing service system based on intelligent case reasoning. Background Technique

[0002] With the continuous growth of personalized needs and the increasing fierce market competition, manufacturing enterprises are facing unprecedented challenges, and the traditional manufacturing mode has been difficult to meet the rapidly changing market demands. Therefore, service-oriented manufacturing emerges as the times require, which emphasizes providing value-added services on the basis of product manufacturing. When service-oriented manufacturing enterprises design manufacturing service systems, they usually need to consider factors such as product function, performance, cost, reliability, etc. The complexity of these factors makes the system design process very complex and time-consuming.

[0003] In traditional design methods, case-based reasoning (CBR) is a commonly used technology, which solves new problems by reusing historical cases. However, case-based reasoning has limitations in dealing with large-scale case libraries and quickly adapting to new situations. In response to the above challenges, the present invention proposes a modular design method for a manufacturing service system based on intelligent case reasoning. This method combines the advantages of case-based reasoning in dealing with historical cases and the capabilities of machine learning in dealing with big data and quickly adapting to new situations. Through steps such as feature extraction, case retrieval, case reuse, and feasibility evaluation, it quickly generates a system design solution that highly matches the requirements of the target case, thereby improving the system design efficiency and quality of service-oriented manufacturing enterprises and reducing the development cycle and cost. Summary of the Invention

[0004] The purpose of the present invention is to provide a modular design method for a manufacturing service system based on intelligent case reasoning to solve the problems in the above background technique, and it can greatly shorten the system development cycle of service-oriented manufacturing enterprises.

[0005] To achieve the above purpose, the present invention provides a modular design method for a manufacturing service system based on intelligent case reasoning, including the following steps:

[0006] S1. Feature extraction: Construct a target case based on the user's description of the requirements for the system design solution, extract the key features of the target case, and analyze the modular composition of the target case;

[0007] S2. Case retrieval: Retrieve historical cases with a high similarity to the target case in the case library based on a pre-trained neural network model;

[0008] S3. Case reuse: Based on reinforcement learning, reorganize the retrieved similar historical cases into modules, and quickly generate a system design solution with the highest fit to the target case;

[0009] S4. Feasibility evaluation: The user evaluates the feasibility of the generated system design solution. If it is feasible, store the solution in the case library; otherwise, repeat steps S1 - S3 until the feasibility meets the standard or the user actively terminates.

[0010] Preferably, step S1 specifically includes:

[0011] S11. Data preprocessing: Use the description of the user's requirements for the system design solution as input data. Clean the input data to remove noise and stop words, normalize the input data and vocabulary for subsequent model learning, and split the text data into words or phrases by word segmentation;

[0012] S12. Inverse document frequency calculation (TF - IDF): Calculate the inverse document frequency of the split words or phrases in the text. This index comprehensively represents the document distinctiveness of a word or phrase in a document and is used to evaluate the importance of the word or phrase in the document set or corpus. First, calculate TF, that is where n i,j is the number of times the word or phrase appears in document d j and the denominator is the total number of times all words appear in document d j ; The inverse document frequency (IDF) of a specific word can be obtained by dividing the total number of documents by the number of documents containing the word and then taking the logarithm of the resulting quotient. Its formula expression is where |D| is the total number of documents in the corpus, and |{j:t i ∈d j}| represents the number of documents containing the word t i . Finally, the inverse document frequency can be calculated by TF * IDF, and then the weight of the candidate word is determined;

[0013] S13. Keyword extraction: After determining the weights of the candidate words, sort the weights and select the top n words as the final keyword extraction results, that is, the features of the text data;

[0014] S14. Modular composition analysis: Based on the feature information extracted from the system design solution requirements, further analyze the modular composition of the system design solution.

[0015] Preferably, step S2 specifically includes:

[0016] S21. According to the description of the product requirements, find multiple features as index information and use them as case attributes, represented by v i ;

[0017] S22. Use the averaging method to normalize v i with the formula expression as follows:

[0018] S23. Arrange the target case and the attributes of the existing cases in the case library in a fixed order to form a column vector X i ; Perform the above operations for each demand feature, and a total of n column vectors with the same dimension are obtained, denoted as {X1, X2, X3, …, X n}, which are used as feature vectors;

[0019] S24. Send the feature vectors into a BP neural network for matching; The BP neural network consists of an input layer, a hidden layer, and an output layer. Among them, the input layer is the feature vector of the case attributes, the hidden layer stores the weight data of each case attribute that has been trained in the neural network, and the output layer is the case with high similarity in the case library; The neural network reads and calls the weight data to retrieve multiple cases with relatively high similarity.

[0020] Preferably, step S3 specifically includes:

[0021] S31. Extract relevant module information from the cases with high similarity selected in S24, and integrate the modules with the same function into different modules respectively to form M modules with different functions, constituting a template library for scheme reconstruction. Each module contains modules with the same function that can replace each other;

[0022] S32. Adopt a model-free reinforcement learning method, namely Q-learning, to select modules from different modules for recombination and optimization.

[0023] Preferably, step S32 specifically includes:

[0024] S321. Initialize the Q-table between different functional modules, and assign an initial value of 0 to each element;

[0025] S322. Define the state, action, and reward of each Q-table; The state and action of each Q-table are all the modules of adjacent modules. For a Q-table, the immediate reward obtained by taking an action is a mixed reward based on the module matching quality and module value. For each Q-table, the total reward obtained by taking an action is the cumulative sum of all subsequent (including the current) immediate rewards;

[0026] For the c-th Q-table, the state s ci represents the i-th selectable module in the c-th module, and the action s (c+1)j represents the j-th selectable module in the (c + 1)-th module, where c represents the module number, and c ∈ {1, …, M - 1|c ∈ N *}, rc(c+1) (s ci ,s (c+1)j ) means in state s ci Take Actions (c+1)j Get instant rewards,

[0027] The calculation method is r c(c+1) (s ci ,s (c+1)j )=η·QU(s ci ,s (c+1)j )+β·VU(s (c+1)j )

[0028] Among them, QU(s ci ,s (c+1)j ) is module s ci and (c+1)j Module matching quality, VU(s (c+1)j ) is module s (c+1)j The value of η and β represent QU(s ci ,s (c+1)j ) and VU(s (c+1)j ) and satisfies η+β=1;

[0029] For each Q table, each action taken gets a total reward R c(c+1) It is the cumulative sum of all subsequent (including current) instant rewards, calculated as follows:

[0030]

[0031] Where γ is the discount factor and satisfies γ∈[0,1]; when c=1, the total reward R c(c+1) For R 12 , represents the total reward obtained by taking an action in the first Q table;

[0032] S323, two complete constructions of two new modular cases are a complete learning process, that is, each construction requires the information brought by the next construction to learn. According to the idea of temporal difference, these two construction processes are sampled and all Q tables are updated;

[0033] S324, in the process of case construction, the first module (the state of the first Q table) is selected by random selection with equal probability, and the remaining modules are selected according to their respective Q tables using the ∈-greedy strategy, that is, there is a probability of ∈ to select the current optimal action (that is, the action with the largest Q value), and there is a probability of 1-∈ to randomly select an action to explore new possibilities;

[0034] S325, loop iteration and optimization, until the Q table converges or meets the preset termination condition.

[0035] Preferably, the module matching quality in step S322 includes coupling degree CU(s ci , s (c+1)j ), communication efficiency CE(s ci , s (c+1)j ), risk degree RD(s ci , s (c+1)j ) and maintainability MN(s ci , s (c+1)j ) of four types of indicators, and the calculation method is:

[0036] QU(s ci , s (c+1)j ) = ω CU CU(s ci , s (c+1)j ) + ω CE CE(s ci , s (c+1)j ) + ω RD RD(s ci , s (c+1)j ) + ω MN MN(s ci , s (c+1)j )

[0037] Among them, ω CU , ω CE , ω RD and ω MN are the weight values of these four types of indicators respectively.

[0038] Preferably, step S4 specifically includes:

[0039] S41. Submit the generated system design plan to the user, and the user evaluates its feasibility according to actual requirements and conditions;

[0040] S42. If the user evaluates that the plan is feasible, then the plan will be stored in the case library for quick retrieval and use when needed in the future;

[0041] S43. If the user evaluates that the plan is not feasible, the system will need to re-design based on steps S1 - S3 until the generated system design plan is evaluated as feasible by the user, or the user decides to terminate the design process.

[0042] Therefore, the present invention adopts the above-mentioned modular design method of a manufacturing service system based on intelligent case reasoning, and has the following beneficial effects:

[0043] (1) By integrating case reasoning and machine learning technologies, the modular design of the manufacturing product service system can be realized, which can help service-oriented manufacturing enterprises reduce the system development cycle and cost;

[0044] (2) In the case retrieval stage, a pre-trained neural network is used to match historical cases with a high similarity to the target case from the case library, which can improve the efficiency and accuracy of case similarity matching;

[0045] (3) In the case reuse stage, a reinforcement learning method is used to calculate the comprehensive indicators of the combination between modules, providing clear data support for the reconstruction of cases, and then efficiently outputting new cases that meet the user's needs.

[0046] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Description of the Drawings

[0047] Figure 1 is the overall flowchart of the embodiment of the present invention;

[0048] Figure 2 is the schematic diagram of the feature extraction module of the embodiment of the present invention;

[0049] Figure 3 is the schematic diagram of the case retrieval model of the embodiment of the present invention;

[0050] Figure 4 is the case retrieval BP neural network diagram of the embodiment of the present invention;

[0051] Figure 5 is the schematic diagram of the reinforcement learning calculation module of the embodiment of the present invention;

[0052] Figure 6 is the flowchart of the reinforcement learning calculation module of the embodiment of the present invention. Detailed Embodiments

[0053] Embodiment

[0054] To make the objectives, 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 with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. Usually, the components of the embodiments of the present invention described and illustrated herein can be arranged and designed in various different configurations.

[0055] As Figure 1 shown, a modular design method for a manufacturing service system based on intelligent case reasoning includes the following steps:

[0056] S1. Feature extraction: Based on the description of the user's requirements for the system design solution, extract the key features of the target case and analyze the modular composition of the target case. As Figure 2 shown, it specifically includes:

[0057] S11. Data preprocessing: Clean the input data to remove noise and stop words, normalize the data and vocabulary for subsequent model learning, and split the text data into words or phrases by tokenization;

[0058] S12. Inverse document frequency calculation (TF-IDF): Calculate the inverse document frequency of the split words or phrases in the text. First, calculate TF, that is where n i,j is the number of times the word appears in document d j and the denominator is the total number of times all words appear in document d j ; The inverse document frequency (IDF) of a specific word can be obtained by dividing the total number of documents by the number of documents containing the word and then taking the logarithm of the resulting quotient. Its formula expression is where |D| is the total number of documents in the corpus, and |{j:t i ∈d j}| represents the number of documents containing the word t i . Finally, the inverse document frequency can be calculated by TF * IDF, and then the weight of the candidate word is determined;

[0059] S13. Keyword extraction: After determining the weights of the candidate words, sort the weights and select the top n words as the final keyword extraction results, that is, the features of the text data;

[0060] S14. Modular composition analysis: Based on the feature information extracted from the system design scheme requirements, further analyze the modular composition of the system design scheme.

[0061] S2. Case retrieval. Based on a pre-trained neural network model, efficiently retrieve historical cases with a high similarity to the target case in the case library according to the key features extracted from the target case, as Figure 3 shown, specifically including:

[0062] S21. According to the description of the product requirements, find multiple features as index information, use them as case attributes, and represent them with v i ;

[0063] S22. Use the mean method for normalization processing. Its formula expression is:

[0064] S23. Arrange the target case and the case attributes of the existing cases in the case library in a fixed order as a column vector X i ; Perform this operation for each requirement feature, and a total of n column vectors with the same dimension are obtained, denoted as {X1, X2, X3,..., X n}, which are used as feature vectors;

[0065] S24. Send the feature vector into the BP neural network as shown in Figure 4 for matching. The BP neural network consists of three parts: an input layer, a hidden layer, and an output layer. The input layer is the feature vector of the case attributes. The weight data of each case attribute that has been trained by the neural network is saved in the hidden layer. The output layer is the similar cases in the case library. When the neural network needs to identify a new case, it needs to read and call the weight matrix to retrieve multiple cases with higher similarity.

[0066] This neural network has a total of two hidden layers and uses the softmax function for classification. In the forward propagation process, different weights are assigned to each element in each input vector, and then it is output through the fully connected layer. In the backward propagation process, the weights of the parameter matrix are updated using the gradient descent algorithm, and an appropriate learning rate is selected, and it is trained with a set of feature vector samples. When the loss function converges to a smaller value, the weight matrix of the neural network is saved.

[0067] The BP neural network is a feedforward neural network that performs calculations through two processes: forward propagation and backward propagation. In the forward propagation process, the input data passes through the hidden layer layer by layer from the input layer and finally reaches the output layer. The state of each neuron in each layer only affects the state of the neurons in the next layer. If the output layer cannot produce the desired output, backward propagation will be performed, and the error signal will be returned along the original connection path. By adjusting the weights of the neurons, the error signal is minimized.

[0068] S3. Case reuse. Adopt the reinforcement learning method to reorganize similar cases and quickly obtain the system design solution with the highest degree of fit to the requirements, specifically including:

[0069] S31. Integrate the modules with the same function in the matched similar cases into different modules to form M different function modules, constituting a template library for scheme reconstruction;

[0070] S32. Adopt the model-free reinforcement learning method, i.e., Q-learning, to select modules from different modules for recombination and optimization, as shown in Figure 5 and Figure 6 . The gray dotted line with an arrow in Figure 5 represents the feasible strategy in the corresponding Q-table, and the black solid line with an arrow represents the optimal strategy in the corresponding Q-table. The numbers beside the lines represent the Q-values corresponding to the strategies, specifically including:

[0071] S321. Initialize the Q-table between different function modules, where each element is given an initial value of 0 and each Q-table corresponds to an agent;

[0072] S322. Define the states, actions, and rewards for each Q-table. The states and actions of each Q-table are all the modules of adjacent modules. For a Q-table, the immediate reward obtained by taking an action is a mixed reward based on the module matching quality and module value; for each Q-table, the total reward obtained by taking an action is the cumulative sum of all subsequent (including the current) immediate rewards.

[0073] For the c-th Q-table, the state s ci represents the i-th optional module in the c-th module, and the action s (c+1)j represents the j-th optional module in the (c + 1)-th module, where c represents the module number, and c ∈ {1, …, M - 1|c ∈ N *}}. r c(c+1) (s ci , s (c+1)j ) represents the immediate reward obtained by taking the action s ci in the state s (c+1)j , and the calculation method is: r c(c+1) (s ci , s (c+1)j ) = η · QU(s ci , s (c+1)j ) + β · VU(s (c+1)j ).

[0074] Among them, QU(s ci , s (c+1)j ) is the module matching quality of the modules s ci and s (c+1)j , VU(s (c+1)j ) is the value of the module s (c+1)j , η and β respectively represent the weights of the two, and satisfy η + β = 1.

[0075] The module matching quality includes four types of indicators: coupling degree CU(s ci , s (c+1)j ), communication efficiency CE(s ci , s (c+1)j ), risk degree RD(s ci , s (c+1)j ) and maintainability MN(s ci , s (c+1)j ), and the calculation method is:

[0076] QU(s ci , s (c+1)j )

[0077] = ω CU (s ci , s (c+1)j ) + ω CE (s ci , s (c+1)j) + ω RD (s ci , s (c+1)j ) + ω MN (s ci , s (c+1)j )

[0078] Among them, ω CU , ω CE , ω RD and ω MN are the weight values of these four indicators respectively, and their sum is 1.

[0079] Specifically, the coupling degree CU(s ci , s (c+1)j ) evaluates the coupling degree between modules s ci and s (c+1)j , that is, the dependency relationship between them. A low coupling degree means less mutual influence between modules; the communication efficiency CE(s ci , s (c+1)j ) evaluates the communication efficiency between modules s ci and s (c+1)j , including indicators such as communication delay and data transfer efficiency, to ensure good overall system performance; the risk degree RD(s ci , s (c+1)j ) evaluates whether there are potential risks in the combination of modules s ci and s (c+1)j in the new case, to ensure that the module design is within the controllable risk range; the maintainability MN(s ci , s (c+1)j ) ensures that the combination of modules s ci and s (c+1)j in the new case can be easily maintained and updated, including clear dependency relationships between modules and clear code structures. The module value VU(s (c+1)j ) makes a comprehensive decision based on system characteristics and user requirements, combining information such as knowledge bases, case bases, user feedback, and design standards.

[0080] For each Q-table, each time an action is taken, the total reward R c(c+1) is the cumulative sum of all subsequent (including the current) immediate rewards, and the calculation method is:

[0081] Among them, γ is the discount factor, and γ ∈ [0, 1]. When c = 1, the total reward R c(c+1) is R 12 , indicating the total reward obtained by taking an action once in the first Q-table.

[0082] S323. Two complete constructions of two new modular cases constitute one complete learning process, that is, each construction requires the information brought by the next construction for learning. According to the idea of temporal difference, sample the two construction processes. The complete data of one sampling includes two data chains, as follows:

[0083] The data chain obtained from the first construction:

[0084] <s 1i ,s 2j ,r 12 (s 1i ,s 2j );…;s (M-1)i ,s Mj ,r (M-1)M (s (M-1)i ,s Mj );R 12 ,…,R (M-1)M >

[0085] Similarly, the data chain obtained from the second construction is denoted as:

[0086] <s 1i＇ ,s 2j＇ ,r 12 (s 1i＇ ,s 2j＇ );…;s (M-1)i＇ ,s Mj＇ ,r (M-1) M(s (M-1)i＇ ,s Mj＇ );R' 12 ,…,R' (M-1)M >

[0087] After obtaining the complete data through sampling, immediately update all Q-tables. The way to update the c-th Q-table is shown as follows:

[0088]

[0089] where s ci is the module selected in module c during the first construction, s (c+1)j is the module selected in module c + 1 during the first construction, α is the learning rate, and α ∈ (0, 1], s ci＇ is the module selected in module c during the second construction, s (c+1)j＇ is the module that may be selected in module c + 1 during the second construction.

[0090] S324. During the construction of the case, the first module (the state of the first Q-table) is selected in an equiprobable random manner, and the remaining modules are selected using the ε-greedy strategy according to their respective Q-tables, that is, there is a probability of ε to select the current optimal action (i.e., the action with the largest Q-value), and there is a probability of 1 - ε to randomly select an action to explore new possibilities. The formula for this strategy is:

[0091]

[0092] where s (c+1)j is the action, s ci is the state, A(s ci ) is the set of all available actions in state s ci , |A(s ci )| is the number of available actions s ci in state s (c+1)j , π(s (c+1)j |s ci ) is the probability of selecting action s ci in state s (c+1)j , and ε ∈ [0, 1].

[0093] S325. Perform cyclic iteration and optimization until the Q-table converges or meets the preset termination conditions. Here, the termination condition is selected to use a fixed exploration time. Once the time is up, stop the Q-learning and output the final new case.

[0094] S4. Feasibility evaluation: The user evaluates the feasibility of the generated system design solution. If it is feasible, store the solution in the case library; otherwise, re-perform feature extraction, case retrieval, and reuse until the feasibility meets the standard or the user actively terminates.

[0095] S41. Submit the generated system design solution to the user, and the user evaluates its feasibility according to the actual requirements and conditions;

[0096] S42. If the user evaluates that the solution is feasible, then the solution will be stored in the case library for quick retrieval and use in future similar requirements.

[0097] S43. If the user evaluates that the solution is not feasible, the system will need to re-design based on steps S1 - S3 until the generated system design solution is evaluated as feasible by the user or the user decides to terminate the design process.

[0098] Therefore, the present invention adopts the above-mentioned modular design method for a manufacturing service system based on intelligent case reasoning, and realizes the modular design of the manufacturing service system by integrating case reasoning and machine learning technologies, helping service-oriented manufacturing enterprises reduce the system development cycle and cost.

[0099] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions of the present invention or make equivalent replacements, and these modifications or equivalent replacements cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A modular design method for a manufacturing service system based on intelligent case reasoning, characterized in that It includes the following steps: S1. Feature extraction: Construct a target case based on the user's description of the system design scheme requirements, extract the key features of the target case, and analyze the modular composition of the target case; S2. Case retrieval: Retrieve historical cases with high similarity to the target case in the case base based on a pre-trained neural network model; Specifically including: S21. Use the features extracted in S1 as index information and use it as a case attribute, denoted by ; S22. Use the averaging method to perform normalization processing, and the formula expression is: ; S23. Arrange the target case and the attributes of the existing cases in the case library in a fixed order to form a column vector ; Perform the above S21 - S23 for each demand characteristic, and a total of column vectors with the same dimension are obtained, denoted as , and use it as the feature vector; S24. Send the feature vector into a BP neural network for matching; The BP neural network consists of an input layer, a hidden layer, and an output layer. Among them, the input layer is the feature vector of the case attributes, the hidden layer stores the weight data of each case attribute that has been trained by the neural network, and the output layer is the cases with high similarity in the case base; The neural network reads and calls the weight data to retrieve several cases with high similarity; S3. Case reuse: Recombine the modules of the retrieved historical cases with high similarity based on reinforcement learning to quickly generate a system design scheme with high fit to the target case; Specifically including: S31. Extract relevant module information from the cases with high similarity selected in S24, and integrate the modules with the same function into different modules respectively. Each module contains modules with the same function that can be replaced with each other; S32. Use Q-learning to recombine and optimize modules with different functions; S4. Feasibility evaluation: The user evaluates the feasibility of the generated system design scheme. If it is feasible, store the scheme in the case base; otherwise, repeat steps S1 - S3 until the feasibility meets the standard or the user actively terminates.

2. The modular design method of a manufacturing service system based on intelligent case reasoning according to claim 1, characterized in that Step S1 specifically includes: S11. Data preprocessing: Use the user's description of the system design scheme requirements as input data, clean the input data to remove noise and stop words, normalize the input data and vocabulary for subsequent model learning, and split the text data into words or phrases by word segmentation; S12. Inverse document frequency calculation: Calculate the inverse document frequency of the split words or phrases in the text, and then determine the weights of the candidate words; S13. Keyword extraction: After determining the weights of the candidate words, extract the features of the text data by sorting the weights; S14. Modular composition analysis: Based on the feature information extracted from the system design scheme requirements, further analyze the modular composition of the system design scheme.

3. A modular design method for a manufacturing service system based on intelligent case reasoning according to claim 1, characterized in that Step S32 specifically includes: S321. Initialize the Q-table between different functional modules, and assign an initial value of 0 to each element; S322. Define the states, actions, and rewards of each Q-table; The states and actions of each Q-table are all the modules of adjacent modules. For a Q-table, the immediate reward obtained by taking an action is a mixed reward based on the module matching quality and module value. For each Q-table, the total reward obtained by taking an action is the cumulative sum of all subsequent immediate rewards; For the th Q-table, the state represents the th optional module in the th module, and the action represents the th optional module in the th module. Among them, represents the module number, and represents the immediate reward obtained by taking the action in the state , and the calculation method is ; Among them, is the module matching quality of modules and , is the value of module , and respectively represent and weights, and satisfy ; For each Q-table, the total reward obtained for each action taken is the cumulative sum of all subsequent immediate rewards, calculated as ； Among them, is the discount factor and satisfies ; when the total reward is , indicating the total reward obtained by taking one action in the first Q-table; S323. Construct two new modular cases completely twice as a complete learning process. Each construction requires the information brought by the next construction for learning; According to the idea of temporal difference, sample the two construction processes and update all Q-tables; S324. During the construction of the case, the first module is selected by equiprobable random selection, and the remaining modules are selected according to their respective Q-tables using the greedy strategy; S325. Loop iteration and optimization until the Q-table converges or meets the preset termination conditions.

4. A modular design method for a manufacturing service system based on intelligent case reasoning according to claim 3, characterized in that, In step S322, the module matching quality includes coupling degree , communication efficiency , risk level and maintainability These four types of indicators, and the calculation method is as follows: ； Among them, and are the weight values of these four types of indicators respectively.

5. A modular design method for a manufacturing service system based on intelligent case reasoning according to claim 1, characterized in that Step S4 specifically includes: S41. Submit the generated system design solution to the user, and let the user conduct a feasibility evaluation on it according to actual requirements and conditions; S42. If the user evaluates that the solution is feasible, then the solution will be stored in the case library for quick retrieval and use when needed in the future; S43. If the user evaluates that the solution is not feasible, the system will need to be redesigned based on steps S1 - S3 until the generated system design solution is evaluated as feasible by the user or the user decides to terminate the design process.

Images