Dynamic multi-objective optimization method and system for electromechanical products based on digital twins
By using the Gradient Boosting Regression Tree (GBRT) model and an adaptive dynamic multi-objective optimization strategy, the optimization problem of electromechanical products under changing operating conditions was solved, achieving improved accuracy and efficiency of optimization results and enhanced adaptability.
Patent Information
- Application Number
- CN202411223066.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-03
- Publication Date
- 2024-11-12
- Estimated Expiration
- 2044-09-03
AI Technical Summary
Existing optimization methods for electromechanical products cannot effectively cope with dynamic changes in operating conditions, resulting in discrepancies between optimization results and actual operation. Furthermore, digital twin technology exhibits poor correlation before and after model updates, leading to divergent convergence of optimization results, long iteration times, and poor adaptability.
A gradient boosting regression tree (GBRT) model training method is adopted to establish a digital twin optimization model. The mapping relationship is updated through an adaptive dynamic multi-objective optimization strategy, and the optimization model is dynamically adjusted to adapt to changes in working conditions by combining machine learning technology.
It improves the accuracy and efficiency of electromechanical product optimization design, ensures that the optimization results match the actual operating status, reduces redundant analysis, and improves the reusability of the model and the uniformity of data interaction.
Smart Images

Figure CN118734719B_ABST
Abstract
Description
Technical Field
[0001] The present application belongs to the technical field of electromechanical products based on digital twins, and specifically relates to a dynamic multi-objective optimization method and system for electromechanical products based on digital twins. Background Art
[0002] With the continuous advancement of technologies such as the Internet of Things, CPS, big data, and artificial intelligence, higher requirements have been placed on the optimization and upgrading of modern electromechanical products, requiring them to develop towards personalization, adaptability, and intelligence. In this context, in order to ensure the stable and reliable operation of electromechanical products during service, they must be continuously optimized and upgraded in a timely and effective manner. The goal of optimization is to accurately construct and solve the optimization model based on the operating mechanism and monitoring data of electromechanical products to improve the optimization indicators such as the operating performance and service life of electromechanical products.
[0003] For complex electromechanical products, the use of multi-objective optimization methods has become the main way to balance the contradictions between multiple optimization indicators. Current optimization methods usually assume that the optimization function or constraint conditions remain unchanged to complete the establishment and iterative solution of the optimization model. However, due to changes in actual operating conditions, the optimization function or constraint conditions of the model show a trend of dynamic change. Traditional optimization methods cannot cope with the dynamic changes of the optimization model, resulting in the optimization results not being consistent with the actual operation.
[0004] In recent years, with the development of intelligent sensing and communication technologies, and the advancement of large-scale computing and intelligent algorithms, the concept of digital twin (DT) has gradually attracted widespread attention in the field of intelligent operation and maintenance of products. Digital twin technology is characterized by the establishment of accurate digital twin models, which realizes the virtual-real mapping of the physical world and the digital world. And by applying communication technology to improve the effective collection and utilization of monitoring data, the digital twin model can be updated as the physical world changes. However, due to the randomness and dynamic variability of product operating conditions, the DT optimization model has poor correlation before and after the model is updated during the solution process, resulting in problems such as convergence divergence, long iteration time and poor adaptability of the final solution optimization result.
[0005] Therefore, in the optimization process of complex electromechanical products, it is urgent to find an effective solution to deal with the uncertainty of optimization functions or constraints under variable working conditions. Summary of the invention
[0006] Based on the above background, this application provides a dynamic multi-objective optimization method for electromechanical products based on digital twins to solve the problem of dynamic optimization of electromechanical product performance indicators under changing operating conditions. Its technical solution is:
[0007] A dynamic multi-objective optimization method for electromechanical products based on digital twins includes the following steps:
[0008] S1. Collect the physical information data of the electromechanical product under study to form multi-source data;
[0009] S2. Map the above multi-source data into digital space, perform data preprocessing and analysis, and obtain modeling parameters of working conditions and constraints;
[0010] S3. Based on the measured data set of the physical space and the simulated data set of the digital space provided, the gradient boosting regression tree (GBRT) model training method is used to train the nonlinear mapping relationship between the input parameters and the target performance indicators, and a digital twin optimization model is established;
[0011] S4. When the working conditions or constraints change, adjust the GBRT parameter structure based on the new training data to update the mapping relationship of the digital twin optimization model and provide a more accurate fitness function for parameter optimization;
[0012] S5. When the operating conditions of electromechanical products change, due to the uncertainty before and after the digital twin optimization model is updated, a digital twin adaptive dynamic multi-objective optimization strategy is constructed.
[0013] Preferably, the GBRT model training method of step S3 includes the following steps:
[0014] S3-1. Establish an optimization function for optimization solution by analyzing the target performance of the product , Represents the mapping relationship between product performance parameters and single optimization indicators, and builds an optimization model F , F It is a multidimensional mapping relationship The set of ;
[0015] S3-2. Define the training data set obtained in step S2 as ; The scheme is i Product performance parameters, The scheme is i Performance parameters The target performance indicators under the combination, where the performance parameters The combination belongs to d dimensional input vector, Belongs to a one-dimensional vector;
[0016] S3-3. For the optimization model F A single optimization function To solve , the present invention adopts the method of minimizing the expected value of the joint distribution loss function, as shown in the following formula:
[0017] ;
[0018] In the formula, , Indicates performance parameters From Scheme 1 to Scheme i The vector combination of , Represents the target performance indicator From Scheme 1 to Scheme i The vector combination of express X arrive Y The mapping relationship between Express X and Y The expected loss function of the joint distribution of .
[0019] Preferably, the updating of the mapping relationship in step S4 includes the following steps:
[0020] S4-1. When the working conditions of the product change, the new training data set As input to update the optimization model F The optimization function , Described in expanded form, as shown below:
[0021] ;
[0022] In the formula, is the expansion coefficient, It is a basic learning model, and the classification and regression tree CART is selected as the learning model;
[0023] S4-2. The number of iterations is specified as m times, divided into L Disjoint leaf nodes , each regression tree can be expressed in additive form as shown below:
[0024] ;
[0025] In the formula, It is a region The average value of the current residuals in , if ,but ,otherwise, ;
[0026] S4-3. Solving the expansion coefficient of each region of the regression tree , The area is calculated as follows:
[0027] ;
[0028] yes The initial expansion coefficient is 0.1;
[0029] S4-4. Get m After iterations As shown below:
[0030] ;
[0031] Will update Substituting into step S3-3, we can solve and obtain the updated optimization function , which is used to describe the mapping relationship between product performance parameters and optimization indicators after the working conditions change.
[0032] Preferably, the digital twin adaptive dynamic multi-objective optimization strategy of step S5 includes a dynamic update mechanism of the initial population and an improved non-dominant sorting mechanism.
[0033] Preferably, the dynamic update mechanism of the initial population in step S5 includes the following steps:
[0034] S5-11. Due to the change of product operating conditions in step S4, different correlation levels will exist after the mapping relationship is updated. The Pearson correlation coefficient is defined to quantify the correlation level of the optimization model containing the multidimensional target space, as shown in the following formula:
[0035] ;
[0036] In the formula, Indicates the number before the update n The optimization function of Indicates the updated n The optimization function of represents the representative solution set obtained by solving the original model, Indicates the correlation coefficient value calculated at the time of model update, ranging from [0, 1];
[0037] S5-12. When the correlation coefficient When There is a high correlation before and after the update. The calculation of the population centroid is defined as follows:
[0038] ;
[0039] In the formula, Before update t- The population at one moment, express t- The population size at moment 1 during the population evolution process; express t- 1 moment population individuals, ,in express t- Manifold information of individuals in the population at time 1, ,in pop Indicates the population size;
[0040] For t Population centroid at time The update is as follows:
[0041] ;
[0042] In the formula, represents the model parameters, k is the order of the autoregressive model, is a Gaussian disturbance variable with a mean of 0, where ;
[0043] For t Moment population profile The update is as follows:
[0044] ;
[0045] In the formula, ,in and The contour is related;
[0046] S5-13. When the correlation coefficient When The correlation before and after the update is low. t Population profile at time The update is as follows:
[0047] ;
[0048] In the formula, Indicates the value calculated in step S5-12 populations, Represents randomly generated populations, and Common composition t The initial population at time .
[0049] Preferably, the improved non-dominant sorting mechanism of step S5 comprises the following steps:
[0050] S5-21. For the initial population , using crossover and mutation algorithms to obtain the offspring population ;
[0051] S5-22. In the progeny population In the acquisition of Calculate the crowding distance, The calculation is as follows:
[0052] ;
[0053] In the formula, Represents the individual crowding distance, which is obtained by the crowding distance calculation method. Indicates the score calculated using the Topsis method;
[0054] S5-23. Calculation As shown below:
[0055] ;
[0056] In the formula, and It represents the Euclidean distance between the optimal solution and the worst solution, as shown in the following formula:
[0057] ;
[0058] In the formula, represents the weighted normalized decision matrix, Indicates the weight value of each indicator. Represents the original matrix Normalization of matrix The rows represent the Combination of, matrix The columns represent the combination of product performance evaluation indicators of interest. The dynamic multi-objective optimization system of electromechanical products based on digital twins includes a data acquisition module, a data processing module, a decision module, and an output module;
[0059] Data acquisition module: obtains physical information data of the electromechanical product under study through sensors, monitoring equipment or other data sources, including sensor data, operation records, and environmental parameter data;
[0060] Data processing module: Exchanges the collected multi-source data into the digital space, performs data preprocessing and analysis, and obtains key modeling parameters for working conditions and constraints;
[0061] Based on the measured data set of the physical space and the simulated data set of the digital space provided, the gradient boosting regression tree (GBRT) model training method is adopted; the nonlinear mapping relationship between the training input parameters and the target performance indicators is established to establish a digital twin optimization model;
[0062] When the working conditions or constraints change, the GBRT parameter structure is adjusted based on the new training data to update the mapping relationship of the digital twin optimization model and provide a more accurate fitness function for parameter optimization;
[0063] Decision-making module: When the operating conditions of electromechanical products change, due to the uncertainty before and after the digital twin optimization model is updated, a digital twin adaptive dynamic multi-objective optimization strategy is constructed;
[0064] Output module: Visually output the calculation results.
[0065] Compared with the prior art, the present invention has the following beneficial effects:
[0066] 1. The dynamic multi-objective optimization method for electromechanical products based on digital twins of the present invention obtains the boundary conditions and constraint data of the model from real physical entities, thereby ensuring the accuracy and reliability of the model during the dynamic optimization process. Compared with traditional methods, this technology can more accurately reflect the actual operating status of the physical system, thereby improving the effectiveness of the optimization design.
[0067] 2. This invention deeply studies the dynamic multi-objective optimization algorithm for electromechanical products under changing working conditions, and combines artificial intelligence technologies such as machine learning to improve the effectiveness of characteristic parameters related to environmental conditions. In the application of different types of electromechanical products such as construction machinery, high-speed trains, and machine tools, this method can achieve accurate optimization model establishment and optimization solution, ensuring the objectivity and accuracy of the optimization results, and providing strong support for the dynamic optimization design of complex electromechanical products.
[0068] 3. The dynamic multi-objective optimization method based on digital twins in this invention has high model reusability and avoids repeated domain knowledge analysis. Designers effectively reduce the differences in concepts and terminology by building a unified optimization framework and specification model, ensuring the uniformity and efficiency of data interaction in the dynamic optimization process, and improving the overall efficiency of design and optimization. BRIEF DESCRIPTION OF THE DRAWINGS
[0069] Figure 1 This is a framework flow chart of the dynamic multi-objective optimization method for electromechanical products based on digital twins of the present invention.
[0070] Figure 2 This is a specific implementation flow chart of the digital twin dynamic multi-objective optimization strategy of the present invention.
[0071] Figure 3 Schematic diagram of the calculation of the correlation degree of the digital twin model of the present invention, Figure 3 (a) is a schematic diagram before and after the update. Figure 3 (b) is a schematic diagram of the irrelevant state before and after the update.
[0072] Figure 4 This is a schematic diagram of the implementation principle of the improved non-dominated sorting mechanism algorithm of the present invention. DETAILED DESCRIPTION
[0073] In order to facilitate understanding of the present application, the present application will be described more fully below with reference to the relevant drawings. Embodiments of the present application are provided in the drawings. However, the present application can be implemented in many different forms and is not limited to the embodiments described herein. On the contrary, the purpose of providing these embodiments is to make the disclosure of the present application more thorough and comprehensive.
[0074] A dynamic multi-objective optimization method for electromechanical products based on digital twins. The overall framework process of the method is as follows: Figure 1 As shown, it mainly includes the following five steps:
[0075] Step S1. Physical information of the system needs to be collected, including sensor data, operation records, environmental parameters and other data. These data can be obtained through sensors, monitoring equipment or other data sources;
[0076] Step S2. Interact the above multi-source data into the digital space to analyze the working conditions and constraints of the physical system. Establish simulation models based on physical mechanisms, such as geometric models, dynamic models, and mechanical models, and obtain simulation training data sets;
[0077] Step S3. Based on the provided measured data set of the physical space and the simulated data set of the digital space, a model training method based on GBRT is used to train the nonlinear mapping relationship between the input parameters and the target performance indicators, and establish a digital twin optimization model;
[0078] Step S4. When the working conditions or constraints change, the GBRT parameter structure is adjusted based on the new training data to update the mapping relationship of the digital twin optimization model and provide a more accurate fitness function for parameter optimization;
[0079] Step S5. After the operating conditions change, due to the uncertainty of the digital twin optimization model before and after the update, this paper proposes a digital twin adaptive dynamic multi-objective optimization strategy. This strategy considers the correlation between the model before and after the update, and realizes the dynamic multi-objective solution under the adaptive update of the search population.
[0080] The digital twin model is the main part that connects the physical system with the digital space. The physical system refers to electromechanical products or physical equipment, which are composed of various physical components, structures and functions. During operation, the physical system will show different behaviors and characteristics, such as the stability, comfort and safety of product operation. This requires the collection of various types of sensor data of the physical system, such as acceleration signals, displacement signals, pressure signals, etc.
[0081] For the sensor data obtained, different types of data from different communication interfaces or protocols are uniformly packaged and transmitted to the digital space to support the establishment of digital twin optimization models and multi-objective solutions. First, the physical information of the physical system is collected and analyzed, and its physical mechanism is analyzed to establish a simulation model, including a dynamic model, a finite element model, etc.
[0082] The dynamic multi-objective optimization method is to establish a digital twin optimization model that can describe the mapping relationship of the optimization function based on the mapping relationship between performance parameters and optimization indicators and based on the constraints of the physical mechanism. Since changes in the working conditions of the physical system will change the mapping relationship or constraints, the goal of the dynamic optimization method of the present invention is to express and solve the dynamic changes of the digital twin optimization model and solve the dynamic optimization problem of changes in the working conditions of the physical system under discrete time. The specific implementation steps of the further dynamic multi-objective optimization method are as follows.
[0083] For step S1, it is first necessary to collect multi-dimensional physical information of the system, which includes various data sources such as sensor data, operation records, and environmental parameters. The acquisition of these data usually relies on high-precision sensors and monitoring equipment, covering a wide range from microscopic physical quantities to macroscopic environmental parameters. For example, sensors can record real-time data such as temperature, pressure, and vibration generated during the operation of the system, while monitoring equipment can track information such as operation logs and equipment status to ensure the comprehensiveness and continuity of the data. In addition, it is also necessary to monitor environmental parameters such as temperature and humidity in real time in order to accurately reflect the operating environment of the system.
[0084] During the data collection process, different types of data may be transmitted through different communication interfaces and protocols, such as Modbus, CAN bus, Ethernet, etc. Therefore, these data must be uniformly packaged and formatted for processing and analysis in the digital space. At the same time, the transmission efficiency and real-time performance of the data must also be considered to ensure that the digital twin model can be updated in real time based on the latest physical information. The standardization of data encapsulation can be achieved through protocol converters, middleware, or custom interfaces. These tools can effectively integrate data streams from different devices and systems to form a unified input pipeline, providing training data support for the subsequent construction of digital twin models.
[0085] For step S2, it is first necessary to analyze the physical information data from the digital space and preprocess the data, including data cleaning, noise reduction, removal of outliers and other steps to ensure the accuracy and effectiveness of the data. Then, extract key modeling parameters such as geometric parameters, dynamic parameters, and mechanical parameters. For example, geometric parameters involve the shape, size, and spatial distribution of physical entities, which are the basis for establishing geometric models; dynamic parameters are related to the motion characteristics of objects, such as speed, acceleration, mass, inertia, etc., which are used to establish dynamic models; mechanical parameters include the elastic modulus of materials, stress-strain relationships, etc., which are mainly used for mechanical modeling in finite element analysis. The established simulation model is used to provide a training data set for the construction and update of subsequent digital twin models.
[0086] The GBRT model training method in step S3 mainly includes the following steps:
[0087] S3-1. Establish an optimization function for optimization solution by analyzing the target performance of the product , Represents the mapping relationship between product performance parameters and single optimization indicators. On this basis, the optimization model is constructed F , F It is a multidimensional mapping relationship The set of .
[0088] S3-2. Define the training data set obtained in step S2 as . The scheme is i Product performance parameters, The scheme is i Performance parameters The target performance indicators under the combination, where the performance parameters The combination belongs to d dimensional input vector, Belongs to a one-dimensional vector;
[0089] For the optimization model F A single optimization function To solve , this application adopts the method of minimizing the expected value of the joint distribution loss function, as shown in the following formula:
[0090] ;
[0091] In the formula, , Indicates performance parameters From Scheme 1 to Scheme i The vector combination of , Represents the target performance indicator From Scheme 1 to Scheme i The vector combination of express X arrive Y The mapping relationship between Express X and Y The expected loss function of the joint distribution of .
[0092] The updating of the mapping relationship in step S4 mainly includes the following steps:
[0093] S4-1. When the working conditions of the product change, the new training data set As input to update the optimization model F The optimization function , at this time for S3-3 Described in expanded form, as shown below:
[0094] ;
[0095] In the formula, is the expansion coefficient, It is a basic learning model, and Classification and Regression Tree (CART) is selected as the learning model.
[0096] S4-2. The number of iterations is specified as m times, divided into L Disjoint leaf nodes , each regression tree can be expressed in additive form as shown below:
[0097] ;
[0098] In the formula, It is a region The average value of the current residuals in , if ,but ,otherwise, .
[0099] S4-3. Solving the expansion coefficient of each region of the regression tree ,by Take region as an example The calculation formula is as follows:
[0100] ;
[0101] The initial expansion coefficient The value is 0.1;
[0102] S4-4. Get m After iterations As shown below:
[0103] ;
[0104] Will update Substituting into step S3-3, we can solve and obtain the updated optimization function , which is used to describe the mapping relationship between product performance parameters and optimization indicators after the working conditions change.
[0105] In the process of solving the optimization model in step S5, commonly used multi-objective optimization algorithms such as NSGA-Ⅱ are used. It retains individuals with better characteristics in the population, improving the diversity and computational efficiency of the population. However, when the working conditions change, the mapping relationship of the optimization model in step S4 will be dynamically updated, resulting in a certain blindness in the search population of the existing optimization algorithm. This will no longer be suitable for the dynamic optimization process. To this end, the present invention proposes a digital twin adaptive dynamic multi-objective optimization strategy, and the specific implementation flowchart is as follows: Figure 2 As shown in Figure 2, this strategy mainly includes two contents, namely, the dynamic update mechanism of the initial population and the improved non-dominated sorting mechanism.
[0106] The dynamic update mechanism of the initial population described in step S5 mainly includes the following steps:
[0107] S5-11. Due to the change of product operating conditions in step S4, different correlation levels will exist after the mapping relationship is updated. The Pearson correlation coefficient is defined to quantify the correlation level of the optimization model containing the multidimensional target space, as shown in the following formula:
[0108] ;
[0109] In the formula, Indicates the number before the update n The optimization function of Indicates the updated n The optimization function of represents the representative solution set obtained by solving the original model, Indicates the correlation coefficient value calculated at the time of model update, ranging from [0, 1].
[0110] Optimization function of a two-dimensional model For example, the schematic diagram of the correlation degree calculation of the same solution set is as follows Figure 3 shown. Figure 3 (a) is a schematic diagram before and after the update. Figure 3 (b) is a schematic diagram of the irrelevant state before and after the update. Figure 3 It can be seen that there are different degrees of correlation between the models.
[0111] Since different correlation levels will cause large differences in the movement and evolution direction of the search population during the optimization process of the algorithm, continuing to use the original search population may lead to local optimality. Reinitializing the population will reduce convergence. In order to solve this problem, according to the specific situation, Threshold , which is used to determine the degree of correlation between optimization models. According to different degrees of correlation, the dynamic update mechanism of the initial population is further divided into two cases: S5-12 and S5-12.
[0112] S5-12. When the correlation coefficient When , it means that the changes before and after the model mapping relationship is updated have a high correlation. In this case, the centroid and contour of the population can be dynamically updated to improve the convergence speed of the optimization. t When the moment model changes, the center of mass of the Pareto solution set will move to a certain extent, changing the direction of population evolution convergence.
[0113] In order to accurately track the movement direction of the population, the calculation of the population center is defined as follows:
[0114] ;
[0115] In the formula, Before update t- The population at one moment, express t- The population size at moment 1 during the population evolution process; express t- 1 moment population individuals, ,in express t- Manifold information of individuals in the population at time 1, ,in pop Indicates the population size.
[0116] For t Population centroid at time The update is as follows
[0117] ;
[0118] In the formula, represents the model parameters, k is the order of the autoregressive model, is a Gaussian disturbance variable with a mean of 0. Among them, .
[0119] For t Population profile at time The update is as follows:
[0120] ;
[0121] In the formula, ,in and The contour is related.
[0122] S5-13. When the correlation coefficient , indicating that the correlation between the optimization model before and after the update is low. At this time, the movement and evolution direction of the population after the update are very different from the previous solution population. Continuing to use the previous population may not be able to jump out of the local optimal solution. Therefore, in order to improve the diversity of the population, t Moment population profile The update is as follows:
[0123] ;
[0124] In the formula, Indicates the value calculated in step S5-12 populations, Represents randomly generated populations, and Common composition t The initial population at time .
[0125] For the improved non-dominant sorting mechanism described in step S5, according to the degree of relevance in step S4 Calculate the value and use a new crowding calculation method as the comparison criterion between individuals in the population to improve the adaptability of the search population in the iterative optimization process. Figure 4 As shown, it mainly includes the following steps:
[0126] S5-21. The initial population updated from S5-11 or S5-12 , using the commonly used crossover and mutation algorithms to obtain the offspring population .
[0127] S5-22. In the S5-21 progeny population In the acquisition of the present invention, an improved metric of the congestion distance is adopted. Calculate the crowding distance and convert the above calculated correlation degree Taking into account, The calculation is as follows:
[0128] ;
[0129] In the formula, represents the individual crowding distance, which can be obtained by the commonly used crowding distance calculation method. Represents the score calculated using the Topsis method.
[0130] S5-23. Calculation As shown below:
[0131] ;
[0132] In the formula, and It represents the Euclidean distance between the optimal solution and the worst solution, as shown in the following formula:
[0133] ;
[0134] In the formula, represents the weighted normalized decision matrix; Indicates the weight value of each indicator, and its value is 1; Represents the original matrix Normalization processing, matrix The rows represent the Combination of, matrix The columns represent the combination of product performance evaluation indicators of interest.
[0135] According to the degree of relevance The difference between the initial population Dynamic updates and calculation of metrics Perform adaptive iterative optimization. When it is larger, it means the correlation is greater, and the focus is on Individuals with higher scores prompt the population to quickly jump out of the local optimum during the optimization process, thus improving the convergence speed of iterative optimization of product performance parameters under changing product operating conditions. When it is smaller, it means that the degree of correlation is smaller. At this time, more emphasis is placed on individuals with larger crowding distances, which improves the diversity in the iterative optimization process of product performance parameters.
[0136] A dynamic multi-objective optimization system for electromechanical products based on digital twins, including a data acquisition module, a data processing module, a decision module and an output module;
[0137] Data acquisition module: obtains physical information data of the electromechanical product under study through sensors, monitoring equipment or other data sources, including sensor data, operation records, and environmental parameter data;
[0138] Data processing module: Exchanges the collected multi-source data into the digital space, performs data preprocessing and analysis, and obtains key modeling parameters for working conditions and constraints;
[0139] Based on the measured data set of the physical space and the simulated data set of the digital space provided, the gradient boosting regression tree (GBRT) model training method is adopted; the nonlinear mapping relationship between the training input parameters and the target performance indicators is established to establish a digital twin optimization model;
[0140] When the working conditions or constraints change, the GBRT parameter structure is adjusted based on the new training data to update the mapping relationship of the digital twin optimization model and provide a more accurate fitness function for parameter optimization;
[0141] Decision-making module: When the operating conditions of electromechanical products change, due to the uncertainty before and after the digital twin optimization model is updated, a digital twin adaptive dynamic multi-objective optimization strategy is constructed;
[0142] Output module: Visually output the calculation results.
[0143] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A dynamic multi-objective optimization method for electromechanical products based on digital twins, It is characterized in that The following steps are involved: S1. Collect the physical information data of the electromechanical product under study to form multi-source data; S2. Map the above multi-source data into digital space, perform data preprocessing and analysis, and obtain modeling parameters of working conditions and constraints; S3. Based on the measured data set of the physical space and the simulated data set of the digital space provided, the gradient boosting regression tree (GBRT) model training method is used to train the nonlinear mapping relationship between the input parameters and the target performance indicators, and a digital twin optimization model is established; S4. When the working conditions or constraints change, adjust the GBRT parameter structure based on the new training data to update the mapping relationship of the digital twin optimization model and provide a more accurate fitness function for parameter optimization; S5. When the operating conditions of electromechanical products change, due to the uncertainty before and after the digital twin optimization model is updated, a digital twin adaptive dynamic multi-objective optimization strategy is constructed; Digital twin adaptive dynamic multi-objective optimization strategy, including dynamic update mechanism of initial population and improved non-dominated sorting mechanism; The dynamic update mechanism of the initial population in step S5 mainly includes the following steps: S5-11. Due to the change of product operating conditions in step S4, there will be different correlation levels after the mapping relationship is updated. The Pearson correlation coefficient is defined to quantify the correlation level of the optimization model containing the multidimensional target space, as shown in the following formula: In the formula, f n represents the nth optimization function before updating, represents the nth optimization function after update, θ represents the representative solution set obtained by solving the original model, and ρ represents the correlation coefficient value calculated at the time of model update, ranging from [0,1]; S5-12. When the correlation coefficient ρ<η≤1, it means that f has a high correlation before and after the update. The calculation of the population centroid is defined as follows: Where P t-1 represents the population at time t-1 before the update, |P t-1 S t-1 | represents the population size during the population evolution process at time t-1; represents the individuals in the population at time t-1, in Represents the manifold information of the individuals in the population at time t-1, j = 1, 2, ..., pop, where pop represents the population size; Then for the population centroid at time t The update is as follows: In the formula, λ={λ 0 ,λ 1 ,...,λ k } represents the model parameters, k is the order of the autoregressive model, is a Gaussian disturbance variable with a mean of 0, where Then for the population profile at time t The update is as follows: In the formula, in With P t The contour is related; S5-13. When the correlation coefficient ρ≤η, it means that the correlation degree before and after f is low, then the population profile at time t The update is as follows: In the formula, represents the ρ×pop population calculated in step S5-12, represents a randomly generated (1-ρ)×pop population, and Together they form the initial population P at time t t ; The improved non-dominated sorting mechanism of step S5 includes the following steps: S5-21. For the initial population P t , using crossover and mutation algorithms to obtain the offspring population Q t ; S5-22. In the offspring population Q t In the acquisition of a Calculate the crowding distance, V a The calculation is as follows: V a =ρ×S a +(1-ρ)×D a Where D a Represents the individual crowding distance, which is obtained by the crowding distance calculation method. a Indicates the score calculated using the Topsis method; S5-23. Calculate S a As shown below: In the formula, and It represents the Euclidean distance between the optimal solution and the worst solution, as shown in the following formula: In the formula, represents the weighted normalized decision matrix, w b Indicates the weight value of each indicator. Represents the original matrix A ab Normalization of the matrix A ab The rows represent the different populations P t The combination of matrix A ab The columns represent the combination of product performance evaluation indicators of interest.
2. According to claim 1, the dynamic multi-objective optimization method for electromechanical products based on digital twins, It is characterized in that The GBRT model training method in step S3 includes the following steps: S3-1. By analyzing the target performance of the product, an optimization function f is established for optimization solution; f represents the mapping relationship between product performance parameters and a single optimization index, and an optimization model F is constructed. F is a set of multidimensional mapping relationships f, expressed as F = {f 1 ,f 2 ,f 3 ,...,f n }; S3-2. Define the training data set obtained in step S2 as x i represents the product performance parameter of solution i, y i The performance parameter x of solution i i The target performance index under the combination, where the performance parameter x i The combination belongs to the d-dimensional input vector, y i Belongs to a one-dimensional vector; S3-3. For solving a single optimization function f in the optimization model F, the method of minimizing the expected value of the joint distribution loss function is adopted, as shown in the following formula: In the formula, X represents performance parameter x i The vector combination from scheme 1 to scheme i, Y represents the target performance index y i The vector combination from scheme 1 to scheme i, f(X) represents the mapping relationship between X and Y, E Y,X ψ(Y,f(X)) represents the expected loss function of the joint distribution of X and Y.
3. According to claim 2, the dynamic multi-objective optimization method for electromechanical products based on digital twins, It is characterized in that The mapping relationship update in step S4 includes the following steps: S4-1. When the working conditions of the product change, the new training data set The optimization function f, f(X) used as input to update the optimization model F is described in the form of an expanded form as shown below: In the formula, β m is the expansion coefficient, h represents the base learning model, and the present invention selects the classification and regression tree CART as the learning model; S4-2. Set the number of iterations to m and divide into L non-intersecting leaf nodes Each regression tree can be expressed in additive form as shown below: In the formula, It is region R lm The average value of the current residuals in , if Then ξ=1, otherwise, ξ=0; S4-3. Solve the expansion coefficient β of each region of the regression tree m , taking region l as an example lm The calculation formula is as follows, where the initial expansion coefficient β is taken as 0.
1. S4-4. After every m iterations, f(X) is obtained as follows: Substituting the updated f(X) into step S3-3, we can obtain the updated optimization function Used to describe the mapping relationship between product performance parameters and optimization indicators after the working conditions change.
4. A system for the dynamic multi-objective optimization method of electromechanical products based on digital twins according to any one of claims 1 to 3, It is characterized in that It includes data acquisition module, data processing module, decision module and output module; Data acquisition module: obtains physical information data of the electromechanical product under study through sensors, monitoring equipment or other data sources, including sensor data, operation records, and environmental parameter data; Data processing module: Exchanges the collected multi-source data into the digital space, performs data preprocessing and analysis, and obtains key modeling parameters for working conditions and constraints; Based on the measured data set of the physical space and the simulated data set of the digital space provided, the gradient boosting regression tree (GBRT) model training method is adopted; Train the nonlinear mapping relationship between input parameters and target performance indicators to establish a digital twin optimization model; When the working conditions or constraints change, the GBRT parameter structure is adjusted based on the new training data to update the mapping relationship of the digital twin optimization model and provide a more accurate fitness function for parameter optimization; Decision-making module: When the operating conditions of electromechanical products change, due to the uncertainty before and after the digital twin optimization model is updated, a digital twin adaptive dynamic multi-objective optimization strategy is constructed; Output module: Visually output the calculation results.
