Assembly process monitoring and precision prediction method and system based on digital twinning

Abstract

The invention discloses an assembly workshop monitoring and assembly precision prediction method based on digital twinning. The method comprises the steps that a virtual assembly workshop is built; based on a nested fuzzy finite state machine model, the assembly process is subdivided into a plurality of states, and switching conditions among the states are a set; determining a monitoring project by comprehensively utilizing an analytic hierarchy process and an entropy weight method; based on measured data, a particle swarm-SG filtering data processing method is provided to improve the credibility of the data; establishing an assembly precision prediction model, and predicting the assembly precision based on Monte Carlo simulation by optimizing a parameter optimization process of a support vector machine model; and an assembly process optimization knowledge base is constructed, and when the prediction precision does not reach the standard, the system automatically throws a corresponding solution. Through the digital twinning technology and the machine learning model, assembly workshop monitoring and assembly precision prediction are realized, smooth completion of the assembly process is guaranteed, and the assembly efficiency and the intelligent level are improved.

Description

Technical Field

[0001] The present invention relates to the field of assembly workshop monitoring, and particularly to a monitoring of a marine diesel engine cylinder head assembly workshop based on digital twin. Background Technique

[0002] Assembly is to combine scattered parts and components into a product according to specified precision standards and technical requirements, using a certain assembly process flow and various necessary assembly methods. It is one of the most critical links in the product life cycle. If problems occur during the assembly process, it will prolong the assembly time, reduce production efficiency, increase the production cost of the product additionally, and directly affect the production cycle and product quality of the whole product. For complex assemblies such as marine diesel engine cylinder heads, the assembly precision requirements for each link are very strict. However, because its component parts are complex, problems occurring during the assembly process are difficult to detect, and the production quality and assembly precision are difficult to predict, which restricts the smooth completion of high-precision assembly work. Therefore, the monitoring of the assembly process and the prediction of assembly precision have become the main research contents in the assembly field of complex assemblies such as marine diesel engine cylinder heads.

[0003] In the traditional marine diesel engine cylinder head assembly workshop, the production efficiency is low, the production process is opaque, the production data is scattered, and problems cannot be sensed in time, making it difficult to manage integrally. This not only wastes time but also increases the possibility of errors and repetitive work. To sense the production status of the cylinder head assembly workshop in time, realize the centralized and unified management of workshop data, and make the production data of the workshop traceable, it is necessary to carry out research on the assembly monitoring of the cylinder head workshop, establish an overall control over the workshop behavior, and ensure the assembly precision. In the prediction of assembly precision, machine learning can analyze a large amount of historical assembly data, identify the key factors affecting the precision, and establish an accurate prediction model. This technology can not only improve the accuracy of prediction but also adapt to the changes in the workshop environment and assembly process by updating the model in real time, thereby providing more reliable support for the assembly process, improving the overall product quality and production efficiency. When abnormalities occur in the assembly process or the assembly precision does not meet the standard, through the constructed assembly process optimization knowledge base, the solution strategy is automatically pushed to assist the workshop personnel to quickly handle the abnormal situation, further improving the assembly efficiency and reducing the production cost. Therefore, realizing the monitoring of the workshop assembly process and the prediction of assembly precision based on digital twin technology and machine learning algorithms can effectively solve the problems faced by the current assembly workshop.

[0004] The prior art CN112947294A - A monitoring and simulation system for an automobile assembly workshop based on digital twin. This patent proposes a monitoring and simulation system for an automobile assembly workshop based on digital twin, including a static model establishment module for the assembly workshop, a two-way channel data acquisition module, a dynamic model establishment module for the assembly workshop, a production operation synchronization and correction module, and an entity control module for the assembly workshop. In the static model establishment module for the assembly workshop, a static model of the assembly workshop entities is initially established. The two-way channel data acquisition module is used to collect real-time operation data of the static model sub-items through a two-way communication channel. Combining the static model of the assembly workshop and the real-time operation data of the collected static model sub-items, a dynamic model of the assembly workshop is established. The synchronization and correction module is used to monitor the dynamic model of the assembly workshop, calculate data errors and correct them in a timely manner, synchronize the operation process of the assembly workshop and output a correction and synchronization signal. Finally, in the assembly workshop control module, the assembly workshop entities are controlled through the correction and synchronization signal. However, this method lacks effective processing of the measured data and it is difficult to ensure the accuracy of the data.

[0005] The prior art CN108171805A - A method for predicting assembly accuracy. According to each set of mating surfaces during assembly in the target model, geometric surfaces that meet the preset part design tolerances are respectively generated. Multiple simulation calculations are performed on the generated geometric surfaces to obtain the relative positions of each set of mating surfaces in the target model. According to the results of multiple simulation calculations, the assembly accuracy when at least two parts are assembled is obtained. However, this method only considers the influence of part surface topography and deformation, does not consider other random errors in the assembly process, has a large difference from the actual assembly process, and the calculation process is complex, making it difficult to quickly predict the assembly accuracy.

[0006] The existing patented technologies have the following problems:

[0007] (1) Traditional monitoring of the assembly workshop is mainly based on simple models and it is difficult to comprehensively and meticulously describe complex assembly processes.

[0008] (2) Traditional solutions lack effective processing of measured data. The noise in the data affects the reliability of the assembly process data, restricting the accuracy of assembly accuracy prediction and the credibility of monitoring data.

[0009] (3) Most traditional assembly accuracy prediction methods are based on the rigid body assumption or the first-order deformation theory, ignoring the influence of multi-dimensional coupling errors existing in the assembly process. The assembly accuracy prediction model has a large difference from the real assembly process, restricting the accuracy of assembly accuracy prediction.

[0010] (4) Traditional workshop monitoring and accuracy prediction methods only issue alarms when abnormal situations occur and do not provide reference solutions, which is not conducive to timely handling of abnormal situations. Summary of the Invention

[0011] Objective of the Invention: The present invention provides a method for monitoring and precision prediction of complex assembly processes based on digital twins, which realizes real-time monitoring of the assembly process and accurate perception of assembly precision, automatically pushes solution strategies, improves the assembly process, reduces production costs, and increases the success rate of the first assembly.

[0012] Technical Solution: A method for monitoring and precision prediction of an assembly process based on digital twins, comprising the following steps:

[0013] (1) Build a digital twin assembly workshop according to the real scenario of the assembly workshop;

[0014] (2) Establish an assembly behavior model according to the requirements of the assembly process document; the assembly behavior model adopts a nested fuzzy finite state machine model;

[0015] (3) Use the analytic hierarchy process-entropy weight method to determine the project data to be monitored in the digital twin assembly workshop;

[0016] (4) Optimize the project data to be monitored obtained in step (3) by using the particle swarm-SG data processing method; the particle swarm-SG data processing method optimizes the window size and polynomial order of the SG filtering algorithm by using the particle swarm algorithm;

[0017] (5) Build an assembly precision prediction model, including various error couplings and superpositions;

[0018] Improve the support vector machine model for predicting assembly deformation; the improved support vector machine model uses a hierarchical coarse and fine-grained grid search algorithm to optimize the regularization parameter C and the gamma parameter g;

[0019] When performing precision prediction, sample within the error distribution interval to simulate the random error in the assembly process;

[0020] (6) Use the Monte Carlo method to simulate the assembly process, set a sufficient number of simulation rounds, substitute the assembly deformation prediction value and the sampling value of the random error in step (5) into the assembly precision prediction model to obtain the assembly precision prediction value;

[0021] (7) For the precision non-conformance or abnormal situation in the assembly workshop, establish an assembly process optimization knowledge base in the monitoring system.

[0022] Further, the assembly behavior model adopts a nested fuzzy finite state machine model, takes the finite state machine model of the assembly process of sub-components in the equipment process as a state of the finite state machine model of the equipment process, sets the state switching condition as a set, and can realize state switching when one condition in the state switching condition set is satisfied.

[0023] Further, the analytic hierarchy process-entropy weight method first classifies the workshop data according to information categories, constructs a hierarchical structure model, and only retains the criterion layer; secondly, constructs a judgment matrix, uses the analytic hierarchy process to solve the importance degree of various types of information in the criterion layer, conducts sorting, and eliminates unimportant information categories; then, for various types of information under the retained important information categories, uses the entropy weight method to calculate weights and selects the project data that finally needs to be monitored.

[0024] Further, the specific steps of the particle swarm-SG data processing method are as follows:

[0025] Set the allowable fluctuation range of the data, and eliminate abnormal outliers; after eliminating the outliers, use the interpolation method to fill in the missing values; construct a model and design a fitness function through the particle swarm algorithm, take the window size and polynomial order as the optimization dimensions, and gradually iterate to find the parameter combination that optimizes the performance of the SG algorithm; after completing the parameter optimization, input the best window size and polynomial order into the SG algorithm. Within the window range, the SG algorithm fits the data according to the set polynomial order and outputs the fitted value of the window center point as the processed data value; through the movement of the window, gradually process the entire signal data.

[0026] Further, the fitness function comprehensively considers the mean square error, smoothness, and trend fitting degree of the filtered signal.

[0027] Further, the assembly accuracy prediction model:

[0028]

[0029] In the formula, ε is the assembly accuracy, δ 尺寸 is the dimensional error, δ 变形 is the deformation error generated by process parameters, δ 重复定位 is the repeated positioning error in the assembly process, δ 工装制造 is the tooling manufacturing error, δ 检测 is the detection error, represents the error coupling and superposition. The assembly deformation amount in step (5) is δ 变形 , and the random errors include δ 重复定位 , δ 工装制造 , δ 检测 .

[0030] Further, the improved support vector machine model uses a hierarchical coarse-fine granularity grid search algorithm to optimize the regularization parameter C and the gamma parameter g. The specific steps include:

[0031] 1) The search ranges of the parameters C and g are determined according to the maximum standard deviation of the input data, as shown in the following formula:

[0032]

[0033] where a is the base number, used to increment the parameter on a logarithmic scale, and σ max is the maximum standard deviation of the input data, b is the step size, representing the logarithmic value increased each time, which determines the granularity of parameter search, l is the number of increment steps, and lb is the increment for adjusting the search range;

[0034] 2) Divide the parameter optimization process into two stages: coarse-grained search and fine-grained search, and establish a set G of parameter combinations:

[0035] G = {(C 1 , g 1 ), …, (C 1 , g n ); (C 2 , g 1 ), …, (C 2 , g n ); …, (C m , g n )}

[0036] In the formula, (C m , g n ) is a possible parameter combination, m and n are positive integers. In the coarse-grained search stage, set the random sampling rate α, randomly select parameter combinations, and find the optimal parameter combination as follows:

[0037] G′ = {G[i] | i ∈ randperm(N, k)}

[0038] where k = max(1, α·N) is the number of samples, N = m × n is the total number of combinations, G′ is the parameter combination after random sampling, and randperm(N, k) is a function that randomly selects k indices from 1 to N;

[0039] Set a fine-grained search interval around the optimal parameter combination obtained from the coarse-grained search. The fine-grained search interval is further divided into n search intervals for fine-grained search. The values of C and g in the fine-grained search range are shown as follows:

[0040]

[0041] In the formula, C i and g i are the parameters for fine-grained search respectively, C * and g * are the optimal C and g parameters obtained from the coarse-grained search respectively, δ is the parameter reduction ratio, ξ is the parameter increase ratio, n is the number of subdivisions, C * × ξ - C * × δ represents the length of the subdivision interval of parameter C, and g * × ξ - g* ×δ represents the length of the subdivision interval of parameter g.

[0042] Furthermore, for the random errors in the simulation assembly process, based on the error statistical data, determine the distribution type and distribution interval of the errors, and sample within the error distribution interval when making accuracy predictions to simulate the random errors in the assembly process.

[0043] Furthermore, establish an optimization knowledge base for the assembly process in the monitoring system, number the possible abnormal situations under different states of different workstations, and store the corresponding solution strategies. When the monitoring system detects the corresponding abnormality, automatically query the solution strategy according to the abnormality number for the user's reference; the user can edit this knowledge base.

[0044] An assembly process monitoring and accuracy prediction system based on digital twin includes a digital twin assembly workshop, an assembly behavior model module, a data processing module, an assembly accuracy prediction model module, an assembly simulation module, and a knowledge base module;

[0045] The digital twin assembly workshop is built according to the real scene of the assembly workshop;

[0046] The assembly behavior model module establishes an assembly behavior model according to the requirements of the assembly process document; the assembly behavior model adopts a nested fuzzy finite state machine model;

[0047] The data processing module uses the analytic hierarchy process-entropy weight method to determine the project data to be monitored in the digital twin assembly workshop; optimize the obtained project data to be monitored by using the particle swarm-SG data processing method; the particle swarm-SG data processing method uses the particle swarm algorithm to optimize the window size and polynomial order of the SG filtering algorithm;

[0048] The assembly accuracy prediction model module constructs an assembly accuracy prediction model, which includes various error couplings and superpositions; improves the support vector machine model for predicting the assembly deformation amount; the improved support vector machine model uses a hierarchical coarse-grained grid search algorithm to optimize the regularization parameter C and the gamma parameter g; sample within the error distribution interval when making accuracy predictions to simulate the random errors in the assembly process;

[0049] The assembly simulation module uses the Monte Carlo method to simulate the assembly process, sets a sufficient number of simulation rounds, substitutes the predicted value of the assembly deformation amount and the sampled value of the random error into the assembly accuracy prediction model to obtain the assembly accuracy prediction value;

[0050] The knowledge base module establishes an optimization knowledge base for the assembly process in the monitoring system for the accuracy non-compliance or abnormal situations that occur in the assembly workshop.

[0051] Beneficial effects:

[0052] 1. The nested fuzzy finite state machine model proposed by the present invention can describe complex processes more comprehensively and meticulously compared with traditional models, and is applicable to the monitoring of complex processes.

[0053] 2. By comprehensively using the analytic hierarchy process and the entropy weight method to select monitoring items, it avoids the problems of excessive subjectivity, missing and redundant monitoring items caused by traditional manual selection.

[0054] 3. Through data processing, the authenticity and reliability of data can be significantly improved, further ensuring and enhancing the accuracy of monitoring and precision prediction.

[0055] 4. The support vector machine model is improved to fully consider data characteristics during training, accelerating the training speed and improving the prediction accuracy.

[0056] 5. When predicting the assembly accuracy, the influence of multi-dimensional coupling errors is considered, making the established accuracy prediction model more in line with the actual situation, overcoming the shortcomings of traditional accuracy prediction methods that only consider partial factors, and improving the accuracy of accuracy prediction.

[0057] 6. When the prediction accuracy exceeds the tolerance or other abnormal situations occur, based on the established assembly process optimization knowledge base, the solution strategy is automatically pushed to provide a reference for solving the abnormality. Description of the drawings

[0058] Figure 1 is the overall flow chart of the assembly workshop monitoring and assembly accuracy prediction method based on digital twin;

[0059] Figure 2 is the schematic diagram of the nested fuzzy finite state machine model established by the present invention;

[0060] Figure 3 is the flow chart for selecting monitoring items;

[0061] Figure 4 is the flow chart of the particle swarm - SG filtering algorithm;

[0062] Figure 5 is the data processing effect diagram;

[0063] Figure 6 is the assembly accuracy prediction flow chart. Detailed implementation manners

[0064] The technical solutions of the present invention will be further described below in conjunction with the drawings.

[0065] As Figure 1 shown, an assembly workshop monitoring and assembly accuracy prediction method based on digital twin includes the following steps:

[0066] (1) Based on enterprise research, construct a product model library and build a virtual assembly scenario

[0067] The product model library described in step (1) is used to store 3D models of the marine diesel engine cylinder head assembly workshop, including workshop buildings, equipment, production lines, assembly parts, etc. Deploy these 3D models in the virtual scenario according to the layout of the actual assembly workshop to construct a virtual assembly workshop that highly restores the assembly scenario, which is used to display the on-site status of the cylinder head assembly workshop. In addition, build a friendly user interface, use charts to display the assembly process data in real time, and realize the visual display of the measured data, so as to facilitate users to perceive the assembly process data in a timely manner.

[0068] (2) Establish an assembly behavior model

[0069] Deeply analyze the assembly process documents and establish an assembly behavior model according to the requirements of the assembly process documents. Considering the characteristics that the assembly process of complex assemblies is complex and divided into many stages such as sub-assembly and general assembly, traditional simple models are not applicable. Therefore, the present invention proposes a nested fuzzy finite state machine model for expressing the behavior of the assembly workshop. The specific steps are as follows:

[0070] (2.1) Determination of nested levels

[0071] According to the assembly process, the assembly behaviors that are executed separately can be used as an independent state machine model. When the behaviors are too complex, they can be appropriately split and expressed using a multi-layer state machine. Divide each independent assembly process into a finite number of assembly states according to the assembly steps.

[0072] (2.2) State machine nesting

[0073] According to the requirements of the assembly process card, nest the state machine model. For example, if an assembly process C requires assembling part A and part B, then use the finite state machine model representing the assembly process of part A as a state of the finite state machine model representing the assembly process C, so as to realize the nesting of the state machine model.

[0074] (2.3) Determination of state transition conditions

[0075] The assembly process is continuous and uninterrupted. The progress of the assembly process is essentially a continuous transition of multiple assembly states. To ensure the orderly and accurate transition between assembly states, it is necessary to set state transition conditions. The state transition conditions of complex assembly processes are intricate, while the transition conditions of traditional finite state machine models are single and difficult to meet the requirements of complex assembly processes. Therefore, the present invention improves the traditional state machine model and sets the transition conditions as a set. When one of the conditions in the transition condition set is met, the state transition can be realized.

[0076] The present invention improves the traditional finite state machine model, nests multiple fuzzy state machine models, and describes the assembly behavior comprehensively and multi-levelly, overcoming the disadvantages of poor adaptability of the traditional model and difficulty in describing complex assembly processes, and laying a foundation for realizing workshop twin monitoring.

[0077] Figure 2 Figure shows the model proposed by the present invention. In the figure, S i represents each state, S 0 represents the initial state, F represents the final state, S i，j represents the j-th sub-state in the i-th state, C m，n represents the transition condition from state m to state n. To improve the adaptability of the model to complex production processes, the present invention sets each state transition condition as a set, and δ 0 to δ k are the elements therein, all of which are transition conditions. When one of them is met, the state transition can be triggered.

[0078] As Figure 2 shown, based on the cylinder head assembly process document, the assembly process is split. According to the nested fuzzy finite state machine model proposed by the present invention, states and switching conditions are set. Through the nesting of multiple fuzzy state machines, the entire assembly process is comprehensively described, laying a foundation for workshop monitoring. When the state switching condition is met, the state switching is realized, the model is driven to update, and the virtual scene is guaranteed to be synchronized with the real scene.

[0079] (3) Selection of monitoring items

[0080] There is a large amount of data in the assembly workshop. If all data are collected and monitored, it will lead to too many detection devices, increase production costs, and some of the data have little impact on the assembly process, and it is meaningless to collect this data. Therefore, it is necessary to optimize the monitoring items. The traditional monitoring items are determined manually, with strong subjectivity and easy to cause the lack or redundancy of monitoring items. To overcome the above disadvantages, the present invention proposes a method for determining monitoring items by analytic hierarchy process-entropy weight method. The specific steps are as follows:

[0081] First, classify the workshop data according to information categories, construct a hierarchical structure model, and only retain the criterion layer; secondly, construct a judgment matrix, use the analytic hierarchy process to solve the importance degree of various information in the criterion layer, sort them, and eliminate unimportant information categories; then, for various information under the retained important information categories, calculate the weights using the entropy weight method, and select the final detection items.

[0082] First, for each information category, the analytic hierarchy process is used to screen the key categories. Then, for each index under the key categories, the entropy weight method is used to calculate the entropy weights. Finally, the detection items are determined. This serial combination method effectively reduces the computational complexity of the model and avoids overly subjective results, improving the efficiency on the basis of ensuring the scientific nature of the selected items.

[0083] As Figure 3 shown, the data in the cylinder head assembly workshop is classified according to information categories, and a hierarchical structure model is constructed, only retaining the criterion layer. Secondly, a judgment matrix is constructed, and the analytic hierarchy process is used to solve the importance degree of various information in the criterion layer, conduct sorting, and eliminate unimportant information categories. Then, for various information under the retained important information categories, the entropy weight method is used to calculate the weights, and the final detection information items are selected.

[0084] (4) Data processing

[0085] For the selected monitoring items, a data acquisition network is constructed for data acquisition. However, during the data acquisition process, it may be affected by other factors such as vibration, resulting in a large amount of noise in the measurement data. To improve the reliability of the data, the present invention proposes a particle swarm - SG data processing method, which uses the particle swarm algorithm to optimize the two parameters of the window size and polynomial order of the SG filtering algorithm, improves the performance of the SG filtering algorithm, and realizes data smoothing while filtering the data. The specific steps are as follows:

[0086] (4.1) Outlier rejection

[0087] Set the allowable fluctuation range of the data and reject the abnormal outliers;

[0088] (4.2) Missing value completion

[0089] After rejecting the outliers, use the interpolation method to complete the missing values;

[0090] (4.3) SG algorithm optimization and data processing

[0091] First, an optimization model is constructed through the particle swarm optimization algorithm and a fitness function is designed. With the window size and polynomial order as the optimization dimensions, the parameter combination that optimizes the performance of the SG algorithm is gradually found through iterative search. The fitness function comprehensively considers the mean square error, smoothness, and trend fitting degree of the filtered signal to ensure the accuracy and stability of the filtering result. Then, after the parameter optimization is completed, the optimal window size and polynomial order are input into the SG algorithm. Within the window range, the SG algorithm fits the data according to the set polynomial order and outputs the fitting value at the center point of the window as the processed data value. By moving the window, the entire signal data is gradually processed. The SG algorithm optimized based on PSO can significantly improve the data processing effect, achieving better smoothness and trend retention performance. The specific parameter optimization process is as follows:

[0092] 1) Design the fitness function

[0093] To ensure the accuracy of the filtered data, the ability to retain the original characteristics of the collected data, and improve the smoothness of the filtered signal, the present invention comprehensively considers three indicators, namely the mean square error (MSE), trend error, and signal smoothness, when designing the fitness function, and performs weighted synthesis so that these three indicators jointly determine the fitness function value. The constructed fitness function is shown in formula (1):

[0094] Fitness = α·MSE + β·trend_diff + γ·std_filtered(1)

[0095] In the formula, Fitness is the fitness function value, MSE is the mean square error, trend_diff is the trend error, std_filtered is the signal smoothness, and α, β, and γ are the weights of the three indicators of the mean square error, trend error, and signal smoothness, respectively.

[0096] The mean square error is used to measure the accuracy of the data processing result. Its calculation formula is as shown in formula (2):

[0097]

[0098] In the formula, x i is the original signal, is the filtered signal, and N is the total number of samples of the signal.

[0099] The trend error trend_diff is used to evaluate the trend fitting degree, and its calculation formula is as shown in formula (3):

[0100]

[0101] In the formula, T i is the extracted trend signal, is the filtered signal, and N is the number of samples.

[0102] Use the standard deviation of the first-order difference of the filtered signal as the signal smoothness index, and the calculation formula is as shown in Equation (4):

[0103]

[0104] In the formula is the first-order difference value, is the average value of the difference values, and M is the number of difference samples.

[0105] 1) Establish an optimization model

[0106] Based on the above analysis results, the smaller the fitness function, the better the filtering effect. The optimization parameters of the present invention are the window size and the polynomial order of the SG algorithm. Set the constraint conditions of the model and establish the optimization model as shown in Equation (5):

[0107]

[0108] In the formula, Fitness is the fitness function, w is the window size, w_min and w_max are the minimum and maximum values of its search range respectively, p is the polynomial order, p_min and p_max are the minimum and maximum values of its search range respectively. To ensure that the filter has a unique center point, w must be an odd number, and L is the signal length.

[0109] Use the particle swarm optimization algorithm to solve the model, and assign the best parameters to the SG filtering algorithm for real-time filtering processing. After data processing, the influence of noise can be weakened and the data reliability can be improved. Using the processed data for accuracy prediction can further improve the accuracy of accuracy prediction.

[0110] As Figure 4 shown, after outlier rejection and missing value filling, comprehensively considering the three indicators of mean square error, trend error, and signal smoothness, establish an optimization objective function through weighted synthesis, combine the variation range of parameters, establish an optimization model, use the particle swarm algorithm to solve, obtain the best window size and polynomial order, and assign these two parameters to the SG filtering algorithm. Within the window size range, fit the data points according to the specified polynomial order, output the filtered data, slide the window, and repeat this process until all data processing is completed. The data processing effect is as Figure 5 shown. Finally, the processed data is obtained, the influence of noise is weakened, and the data reliability is improved. On the one hand, the data is displayed, and on the other hand, it is convenient for the system to use the processed data for assembly accuracy prediction.

[0111] (5) Construction of the overall model for assembly accuracy prediction

[0112] The assembly process involves multiple aspects such as parts, equipment, assembly processes, and inspections. Therefore, assembly accuracy is often affected by multiple factors. Assembly parts are the objects of assembly, and their manufacturing errors and actual dimensions directly affect assembly accuracy; assembly equipment is the tool used in the assembly process, and its manufacturing errors and performance also have an important impact on assembly accuracy; forces and torques during the assembly process can cause deformation of the assembled body, affecting assembly accuracy and quality. Therefore, the rational selection of assembly process parameters also affects assembly accuracy; in terms of inspection, due to errors in inspection methods and tools, the measured assembly accuracy indicators will deviate from the true values. To comprehensively consider the combined effects of these factors, an assembly accuracy prediction model is established as shown in Equation (6):

[0113]

[0114] In the formula, ε is the assembly accuracy, δ 尺寸 is the dimensional error, δ 变形 is the deformation error caused by process parameters, δ 重复定位 is the repeat positioning error during the assembly process, δ 工装制造 is the fixture manufacturing error, δ 检测 is the inspection error, represents the error coupling and superposition. The assembly deformation amount described in step (5) is δ 变形 , and the random errors include δ 重复定位 , δ 工装制造 , δ 检测 .

[0115] The assembly process of complex assembled bodies such as marine diesel engine cylinder heads involves the combined effects of multiple factors, and it is difficult to guarantee their assembly accuracy. Assembly accuracy prediction is often required. Traditional accuracy prediction methods do not adequately consider the sources of errors. To improve the accuracy of assembly accuracy prediction, the present invention proposes an assembly accuracy prediction process as shown in Figure 6 . Based on a full analysis of the error sources, an assembly accuracy prediction model as shown in Equation (6) is constructed. The present invention comprehensively considers multiple aspects such as assembly deformation amount, part dimensional error, repeat positioning error, fixture manufacturing error, measurement error, etc.

[0116] (6) Determination of multi-dimensional error values

[0117] (6.1) Prediction of assembly deformation amount based on improved support vector machine

[0118] The deformation of parts during the assembly process is an important factor affecting assembly accuracy. The deformation amount of parts is jointly affected by multiple factors such as material properties and assembly process parameters, and it is difficult to calculate using mathematical formulas. Therefore, based on machine learning theory, the present invention improves the support vector machine model for predicting the assembly deformation amount. The specific improvement steps are as follows:

[0119] (6.1.1) Selection of Adaptive Parameter Search Range Based on the Maximum Standard Deviation of Input Samples

[0120] The prediction performance of the support vector machine model depends on the selection of key parameters C and g. The rationality of these parameters directly affects the training effect and prediction accuracy of the model. The traditional model parameter search range is specified manually, ignoring the characteristics of the data itself, which is prone to generating unreasonable parameter combinations and reducing the model training speed and prediction accuracy. To fully consider the distribution characteristics of the data itself, the parameter search range of the present invention is determined based on the maximum standard deviation of the input data, as shown in Equation (7):

[0121]

[0122] where a is the base number used to increment the parameter on a logarithmic scale, σ max is the maximum standard deviation of the input data, b is the step size, indicating the logarithm value increased each time, which determines the granularity of the parameter search, l is the number of increment steps, and lb is the increment for adjusting the search range to ensure coverage of a wider range of possible values.

[0123] (6.1.2) Parameter Optimization Based on Hierarchical Coarse-Fine Granularity Grid Search Algorithm

[0124] To balance the parameter search efficiency and model performance, the present invention improves the parameter optimization algorithm and uses a hierarchical coarse-fine granularity grid search algorithm for parameter optimization. The parameter optimization process is divided into two stages: coarse-grained search and fine-grained search, and a set of parameter combinations as shown in Equation (8) is established:

[0125] G = {(C 1 , g 1 ),…, (C 1 , g n ), (C 2 , g 1 ),…, (C 2 , g n ),…, (C m , g n )}(8)

[0126] In the formula, (C m , g n ) is a possible parameter combination. In the coarse-grained search stage, a random sampling rate α is set, and parameter combinations are randomly selected to find the optimal parameter combination, as shown in Equation (9).

[0127] G′ = {G[i]|i ∈ randperm(N, k)}(9)

[0128] where \(k = \max(1, \alpha\cdot N)\) is the number of samples, \(N = m\times n\) is the total number of combinations, \(\alpha\) is the random sampling ratio, \(G'\) is the parameter combination after random sampling, and \(randperm(N,k)\) is a function that randomly selects \(k\) indices from 1 to \(N\). Through random search, the computational load can be significantly reduced and the efficiency of parameter optimization can be improved.

[0129] Set a fine-grained search interval around the optimal parameter combination obtained from the coarse-grained search. The fine-grained search interval is further evenly divided into \(n\) search intervals for fine-grained search. The values of \(C\) and \(g\) in the fine-grained search range are shown in Equations (10) and (11):

[0130]

[0131] where \(C\) i and \(g\) i are the parameters for fine-grained search respectively, \(C\) * and \(g\) * are the optimal \(C\) and \(g\) parameters obtained from the coarse-grained search respectively, \(\delta\) is the parameter reduction ratio, \(\xi\) is the parameter increase ratio, \(n\) is the number of subdivisions, \(C\) * \(\times\xi - C\) * represents the length of the subdivision interval of parameter \(C\), and \(g\) * \(\times\xi - g\) * \(\times\delta\) represents the length of the subdivision interval of parameter \(g\). By further searching near the optimal parameters obtained after the coarse-grained search, a better parameter combination can be obtained.

[0132] The surrounding range is jointly determined by \(\delta\) (parameter reduction ratio) and \(\xi\) (parameter increase ratio), and the specific values can be modified based on the prediction accuracy of the support vector machine model. Here, the surrounding can be the left and right intervals (\(\delta\lt1,\xi\gt1\)), or only the left interval (\(\delta\lt\xi\lt1\)), or only the right interval (\(1\lt\delta\lt\xi\)). Here, through multiple experiments, the parameter search interval (the best fine-grained search interval) that maximizes the prediction accuracy of the support vector machine model can be determined, which may be different when this method is applied to other objects or problems.

[0133] By reasonably determining the parameter search interval and improving the parameter optimization algorithm, the present invention can significantly improve the prediction performance of the model and reduce the model training time.

[0134] The assembly deformation amount is an important factor affecting the assembly accuracy. Since it is jointly affected by process parameters, part sizes, part materials, etc. and is difficult to calculate using a mathematical formula, the present invention uses an improved support vector machine model to predict the assembly deformation amount. First, based on the maximum standard deviation of the input data, the search ranges of the hyperparameters C and g are determined. To find the optimal parameter combination, a coarse-to-fine granularity search method is used to optimize the parameter combination. First, all parameter combinations are placed in the set shown in Equation (8), the sampling ratio is set, and k parameter combinations are randomly selected to evaluate the MSE and R 2 value; then the optimal combination is selected, and according to Equations (10) and (11), the parameter search interval is further subdivided, and all parameter combinations within this interval are substituted into the prediction model to determine the best parameter combination, which is substituted into the support vector machine model to predict the assembly deformation amount.

[0135] (6.2) Determination of other error values

[0136] In addition to the deformation error, the assembly process of complex assemblies such as cylinder heads is also jointly affected by many other errors, but these errors are often random values, such as measurement errors, repeat positioning errors, etc., and are difficult to predict using specific formulas or models. To fully consider these effects on the assembly accuracy during the assembly process, the present invention determines the error distribution type and the error distribution interval based on statistical data. When performing accuracy prediction, random sampling is performed within the error distribution interval to simulate the random errors during the assembly process.

[0137] (7) Assembly accuracy prediction considering multi-dimensional coupling errors

[0138] To comprehensively consider the influence of multi-dimensional coupling errors on the assembly accuracy, the present invention uses the Monte Carlo method to perform multiple simulations on the assembly accuracy prediction model. Set a sufficient number of simulation rounds, substitute the predicted deformation amount value and the sampled values of other random errors into the assembly accuracy prediction model established in step (5), repeat this process until the predetermined number of simulation times is reached, and finally take the average value to obtain the final assembly accuracy prediction value. This method can overcome the influence of contingency.

[0139] (8) Exception handling

[0140] For the possible precision non - compliance or other abnormal situations that may occur in the assembly workshop, the present invention establishes an optimization knowledge base for the assembly process, stores the possible abnormal situations and assembly precision requirements that may occur during the assembly process, and numbers each abnormal type. To facilitate the handling of abnormalities, the present invention adds solution strategies to the optimization knowledge base for the assembly process and allows users to modify and improve them to meet different usage requirements. When the system detects an abnormality in the assembly process or the precision prediction result is out of tolerance, the system queries the corresponding fault number according to the assembly process and the abnormal type, automatically retrieves the solution strategy based on the fault number, and timely feeds back the solution suggestions to the system interface for the user to refer to. In addition, the user can edit this knowledge base to enrich the abnormal handling strategies and increase the system's abnormal handling ability.

[0141] The present invention discloses a method for monitoring the assembly workshop and predicting the assembly precision based on digital twin, and the steps are as follows: Based on the current situation of the enterprise assembly workshop, construct a product model library and build a virtual assembly workshop; On the basis of fully analyzing the assembly process, based on the nested fuzzy finite - state machine model, divide the assembly process into multiple states, and the switching condition between each state is a set. Through the nesting of multiple fuzzy finite - state machines, comprehensively, clearly and completely characterize the assembly process; Comprehensively utilize the analytic hierarchy process and the entropy weight method to scientifically and reasonably determine the monitoring items, and establish a data acquisition network for collecting real - time assembly data. Based on the measured data, propose a particle swarm (Particle Swarm Optimization, PSO) - SG (Savitzky - Golay) filtering data - processing method to simultaneously achieve data smoothing while removing the noise of the measured data, and improve the credibility of the data; Establish an assembly precision prediction model considering multi - dimensional coupling errors, optimize the parameter optimization process of the support vector machine model (Support Vector Machine, SVM) by improving the grid - search algorithm, and based on the Monte Carlo simulation, overcome the influence of accidental factors on the assembly precision prediction and improve the accuracy of the assembly precision prediction. Construct an optimization knowledge base for the assembly process. When the predicted precision does not meet the standard, the system automatically throws out the corresponding solution for the user to refer to.

[0142] Based on the above process, it is possible to realize the monitoring of the assembly process and the prediction of precision based on digital twin. Using digital twin technology, users can timely perceive the assembly state, process the measured data in real - time, predict the assembly precision, and handle it in a timely manner when an abnormality occurs, improve the assembly precision of marine diesel engine cylinder heads, reduce the assembly cost, and improve the one - time assembly success rate.

Claims

1. A method for monitoring and predicting assembly process based on digital twins, characterized in that: The following steps are involved: (1) Build a digital twin assembly workshop according to the real scene of the assembly workshop; (2) establishing an assembly behavior model according to the requirements of the assembly process document; the assembly behavior model adopts a nested fuzzy finite state machine model; (3) Use the analytic hierarchy process-entropy weight method to determine the project data that needs to be monitored in the digital twin assembly workshop; (4) optimizing the project data to be monitored obtained in step (3) by using a particle swarm-SG data processing method; the particle swarm-SG data processing method optimizes the window size and polynomial order of the SG filter algorithm by using a particle swarm algorithm; (5) Construct an assembly accuracy prediction model that includes multiple error coupling superpositions; Improved support vector machine model for assembly deformation prediction; The improved support vector machine model uses a hierarchical coarse-grained and fine-grained grid search algorithm to optimize the regularization parameter C and the gamma parameter g; When making accuracy predictions, samples are taken within the error distribution interval to simulate random errors in the assembly process; (6) using the Monte Carlo method to simulate the assembly process, setting a sufficient number of simulation rounds, substituting the predicted value of the assembly deformation and the sampled value of the random error in step (5) into the assembly accuracy prediction model to obtain the assembly accuracy prediction value; (7) For any abnormal situation in the assembly workshop where the accuracy does not meet the requirements, an assembly process optimization knowledge base is established in the monitoring system.

2. The assembly process monitoring and accuracy prediction method based on digital twin according to claim 1 is characterized in that: The assembly behavior model adopts a nested fuzzy finite state machine model, takes the finite state machine model of the assembly process of the subcomponents in the equipment process as a state of the finite state machine model of the equipment process, sets the state switching condition as a set, and realizes state switching when one condition in the state switching condition set is met.

3. The assembly process monitoring and accuracy prediction method based on digital twin according to claim 1 is characterized in that: The hierarchical analysis-entropy weight method first classifies the workshop data according to information categories, constructs a hierarchical model, and only retains the criterion layer; secondly, a judgment matrix is ​​constructed, and the hierarchical analysis method is used to solve the importance of various types of information in the criterion layer, sort them, and eliminate unimportant information categories; then, for various information under the retained important information categories, the entropy weight method is used to calculate the weights, and the project data that ultimately needs to be monitored is selected.

4. The assembly process monitoring and accuracy prediction method based on digital twin according to claim 1 is characterized in that: The specific steps of the particle swarm-SG data processing method are as follows: Set the permissible fluctuation range of the data and remove abnormal outliers; after removing the outliers, use the interpolation method to fill in the missing values; build the model and design the fitness function through the particle swarm algorithm, use the window size and polynomial order as the optimization dimensions, and gradually iterate to find the parameter combination that makes the SG algorithm perform best; after completing the parameter optimization, input the optimal window size and polynomial order into the SG algorithm, and within the window range, the SG algorithm fits the data according to the set polynomial order, and outputs the fitting value of the window center point as the processed data value; By moving the window, the entire signal data is processed step by step.

5. The assembly process monitoring and accuracy prediction method based on digital twin according to claim 1 is characterized in that: The fitness function comprehensively considers the mean square error, smoothness and trend fitting degree of the filtered signal.

6. The assembly process monitoring and accuracy prediction method based on digital twin according to claim 1 is characterized in that: The assembly accuracy prediction model: Where ε is the assembly accuracy, δ 尺寸 is the size error, δ 变形 is the deformation error caused by process parameters, δ 重复定位 is the repeated positioning error of the assembly process, δ 工装制造 is the tooling manufacturing error, δ 检测 To detect errors, represents the error coupling superposition; the assembly deformation in step (5) is δ 变形 , the random error includes δ 重复定位 , δ 工装制造 , δ 检测 .

7. The assembly process monitoring and accuracy prediction method based on digital twin according to claim 1 is characterized in that: The improved support vector machine model uses a hierarchical coarse-grained and fine-grained grid search algorithm to optimize the regularization parameter C and the gamma parameter g. The specific steps include: 1) The search range of parameters C and g is determined based on the maximum standard deviation of the input data, as shown in the following formula: Where a is the base used to increase the parameter on a logarithmic scale, σ max is the maximum standard deviation of the input data, b is the step size, which indicates the logarithmic value added each time, and determines the granularity of the parameter search, l is the number of incremental steps, and lb is the increment for adjusting the search range; 2) The parameter optimization process is divided into two stages: coarse-grained search and fine-grained search, and a set of parameter combinations G is established: G={(C1,g1),…,(C1,g n );(C2,g1),…,(C2,g n );…,(C m ,g n )} In the formula (C m ,g n ) is a possible parameter combination, m, n are positive integers. In the coarse-grained search stage, the random sampling rate α is set, and the parameter combination is randomly selected to find the optimal parameter combination, as shown in the following formula: G′={G[i]|i∈randperm(N,k)} Where k = max(1, α·N) is the number of samples, N = m×n is the total number of combinations, G′ is the parameter combination after random sampling, and randperm(N,k) is a function that randomly selects k indexes from 1 to N; A fine-grained search interval is set around the optimal parameter combination obtained by the coarse-grained search, and the fine-grained search interval is further divided into n search intervals for fine-grained search; the values ​​of C and g in the fine-grained search range are shown in the following formula: Where C i and g i They are the parameters of fine-grained search, C * and g * are the optimal C and g parameters obtained by coarse-grained search, δ is the parameter reduction ratio, ξ is the parameter growth ratio, n is the number of subdivisions, C * ×ξ-C * ×δ represents the length of the subdivision interval of parameter C, g * ×ξ-g * ×δ represents the length of the subdivision interval of parameter g.

8. The assembly process monitoring and accuracy prediction method based on digital twin according to claim 1 is characterized in that: The random errors in the simulated assembly process are determined based on error statistical data to determine the distribution type and distribution range of the errors. When performing accuracy prediction, sampling is performed within the error distribution range to simulate the random errors in the assembly process.

9. The assembly process monitoring and accuracy prediction method based on digital twin according to claim 1 is characterized in that: The assembly process optimization knowledge base is established in the monitoring system, the abnormal situations that may occur under different conditions of different workstations are numbered, and the corresponding solution strategies are stored. When the monitoring system detects the corresponding abnormality, the solution strategy is automatically queried according to the abnormality number for reference by the user; Users can edit the knowledge base.

10. An assembly process monitoring and accuracy prediction system based on digital twins uses an assembly process monitoring and accuracy prediction method based on digital twins as described in any one of claims 1 to 8, characterized in that: It includes digital twin assembly workshop, assembly behavior model module, data processing module, assembly accuracy prediction model module, assembly simulation module, and knowledge base module; The digital twin assembly workshop is a digital twin assembly workshop built according to the real scene of the assembly workshop; The assembly behavior model module establishes an assembly behavior model according to the requirements of the assembly process file; the assembly behavior model adopts a nested fuzzy finite state machine model; The data processing module uses the hierarchical analysis-entropy weight method to determine the project data that needs to be monitored in the digital twin assembly workshop; the acquired project data that needs to be monitored is optimized using the particle swarm-SG data processing method; the particle swarm-SG data processing method uses the particle swarm algorithm to optimize the window size and polynomial order of the SG filtering algorithm; The assembly accuracy prediction model module constructs an assembly accuracy prediction model, including multiple error coupling superpositions; an improved support vector machine model is used for assembly deformation prediction; The improved support vector machine model uses a hierarchical coarse-grained and fine-grained grid search algorithm to optimize the regularization parameter C and the gamma parameter g; when performing accuracy prediction, sampling is performed within the error distribution interval to simulate random errors in the assembly process; The assembly simulation module simulates the assembly process using the Monte Carlo method, sets a sufficient number of simulation rounds, substitutes the assembly deformation prediction value and the sampling value of the random error into the assembly accuracy prediction model, and obtains the assembly accuracy prediction value; The knowledge base module establishes an assembly process optimization knowledge base in the monitoring system for the accuracy that does not meet the requirements or abnormal situations that occur in the assembly workshop.

Images