An engineering problem solving method and system based on a fusion model

Abstract

The application relates to the field of data processing and provides an engineering problem solving method and system based on a fusion model, which comprises the following steps: performing data enhancement on real samples through physical constraint noise disturbance to obtain an enhanced sample set; performing parameter optimization on a physical constitutive model according to the enhanced sample set to obtain optimal constitutive model parameters; performing interpolation on the enhanced sample set through the physical constitutive model optimized by the optimal constitutive model parameters to obtain a virtual sample set; performing multi-source collaborative training on a neural network model according to the real samples, the enhanced sample set and the virtual sample set to obtain a hybrid training model; and performing prediction and solving on an engineering problem to be solved through the hybrid training model to obtain a solving result. The application realizes collaborative optimization of physical driving and data driving, improves the adaptability of the model to samples outside the data distribution under the premise of ensuring the credibility of generated data.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of data processing, and particularly relates to an engineering problem solving method and system based on a fusion model. BACKGROUND

[0002] With the rapid development of artificial intelligence technology, machine learning models have become the core tool for industrial intelligent transformation and are widely used in key fields such as performance prediction, parameter optimization, fault diagnosis, etc. Traditional machine learning methods establish the mapping relationship between input and output by mining the potential laws in massive data, showing strong non-linear modeling capability. In engineering fields such as equipment manufacturing, material research and development, chemical process, and biomedicine, researchers usually use algorithms such as support vector machine, random forest, and neural network to build prediction models for solving complex engineering optimization problems. Existing data-driven modeling methods have achieved remarkable results in a big data environment, effectively capturing the complex non-linear relationships between multiple variables and providing quantitative support for engineering decision-making. At the same time, the constitutive model based on physical mechanism has a deep theoretical foundation in the engineering field, which describes the internal laws of physical phenomena through mathematical equations, and has good interpretability and theoretical basis.

[0003] However, the existing technical solutions have significant limitations in actual engineering applications: traditional machine learning methods have strong dependence on data size and quality, and require a large number of labeled complete and balanced samples to support model training. However, in actual engineering scenarios, due to factors such as short research and development cycle, insufficient digitalization of equipment, high experimental cost, etc., the amount of available effective data is often only tens to hundreds of orders of magnitude. Data scarcity leads to the model being prone to overfitting, which is manifested as high accuracy in the training set but significant decline in actual generalization performance. On the other hand, existing small sample learning solutions rely on generative adversarial networks or transfer learning techniques. However, pure data-driven generation methods lack domain knowledge constraints and are prone to produce low-quality synthetic data that violates physical laws. Cross-domain transfer learning is limited by the difference in feature distribution between the source domain and the target domain, making it difficult to adapt to the specific needs of complex engineering problems. In addition, existing methods often separate data-driven modeling and physical mechanism modeling, failing to fully utilize domain knowledge to guide the data augmentation process, resulting in insufficient robustness and generalization ability of the model under small sample conditions. SUMMARY

[0004] In view of the deficiencies of the above prior art, the purpose of the application is to provide an engineering problem solving method and system based on a fusion model.

[0005] The application provides an engineering problem solving method based on a fusion model, comprising:

[0006] S1: data augmentation is performed on real samples by physical constraint noise disturbance to obtain an augmented sample set;

[0007] S2: parameter optimization is performed on the physical constitutive model according to the enhanced sample set, and optimal constitutive model parameters are obtained;

[0008] S3: virtual data is generated by the physical constitutive model optimized by the optimal constitutive model parameters, and virtual interpolation is performed on the virtual data, and a virtual sample set is obtained;

[0009] S4: a multi-source collaborative training is performed on a neural network model according to the real sample, the enhanced sample set and the virtual sample set, and a hybrid training model is obtained;

[0010] S5: a to-be-solved engineering problem is predicted and solved by the hybrid training model, and a solving result is obtained.

[0011] According to the engineering problem solving method based on the fusion model provided by the application, step S1 further comprises:

[0012] S11: difference noise parameters are set for features in the real sample according to feature statistical characteristics, and the real sample is disturbed and transformed by the difference noise parameters, and a disturbed sample is obtained;

[0013] S12: physical constraint conditions are added to the disturbed sample according to the domain law corresponding to the engineering problem, and a physically constrained sample is obtained;

[0014] S13: the physically constrained sample is adjusted in relevance, and an enhanced sample set is obtained.

[0015] According to the engineering problem solving method based on the fusion model provided by the application, the difference noise parameters in step S11 are adaptively set according to the engineering field physical mechanism corresponding to the to-be-solved engineering problem.

[0016] According to the engineering problem solving method based on the fusion model provided by the application, step S2 further comprises:

[0017] S21: a constitutive equation is trained according to the real sample and the enhanced sample set, and a physical constitutive model is obtained;

[0018] S22: global optimization is performed on the parameter constraint space of the physical constitutive model by a differential evolution method, and optimal constitutive model parameters are obtained.

[0019] According to the engineering problem solving method based on the fusion model provided by the application, the constitutive equation in step S21 is constructed based on the engineering field physical mechanism corresponding to the to-be-solved engineering problem.

[0020] According to the engineering problem solving method based on the fusion model provided by the application, step S3 further comprises:

[0021] S31: generating virtual data through the physical constitutive model optimized by the optimal constitutive model parameters, to obtain a basic virtual sample;

[0022] S32: supplementing the basic virtual sample by directional interpolation covering sparse areas of the real data, to obtain a supplemented virtual sample;

[0023] S33: performing physical constraint on the supplemented virtual sample, to obtain a virtual sample set.

[0024] According to the engineering problem solving method based on the fusion model provided by the application, the physical constraint condition used for physical constraint in step S33 is set according to the physical mechanism of the engineering field corresponding to the engineering problem to be solved.

[0025] According to the engineering problem solving method based on the fusion model provided by the application, step S4 further includes:

[0026] S41: assigning different weights to the real sample, the enhanced sample set and the virtual sample set and performing weighted fusion to obtain a weighted sample set;

[0027] S42: performing physical feature extraction on the weighted sample set to obtain an engineering feature set;

[0028] S43: performing deep learning on the engineering feature set to train a neural network model, to obtain a hybrid training model.

[0029] According to the engineering problem solving method based on the fusion model provided by the application, in step S41, the weight of the real sample is greater than the weight of the enhanced sample set, and the weight of the real sample is greater than the weight of the virtual sample set.

[0030] The application also provides an engineering problem solving system based on a fusion model, which is used to execute the engineering problem solving method based on the fusion model as described in any one of the above, and includes:

[0031] The enhancement module is used to perform data enhancement on the real sample by physical constraint noise disturbance, to obtain an enhanced sample set.

[0032] The optimization module is used to perform parameter optimization on the physical constitutive model according to the enhanced sample set, to obtain optimal constitutive model parameters.

[0033] The interpolation module is configured as the physical constitutive model optimized by the optimal constitutive model parameters, and is used to interpolate the enhanced sample set, to obtain a virtual sample set.

[0034] The training module is configured to perform multi-source collaborative training on the neural network model according to the real samples, the enhanced sample set and the virtual sample set, and obtain a hybrid training model.

[0035] The solving module is configured to use the hybrid training model obtained by the training module to perform predictive solving on an engineering problem to be solved, and obtain a solving result.

[0036] The engineering problem solving method and system based on the fusion model provided by the application can significantly expand the diversity and coverage of the training samples while maintaining the physical rationality of the data, ensure that the generated enhanced samples conform to the physical laws in the engineering field compared with the traditional random noise addition method, avoid the generation of invalid data that violates the physical constraints, thereby providing a high-quality basic data set for subsequent model training, and effectively alleviate the fundamental problem of insufficient data in small sample learning.

[0037] The application optimizes the constitutive model parameters of the enhanced sample set through the differential evolution algorithm, can globally search for the optimal parameter combination, avoids the defect that the traditional gradient descent algorithm is easy to fall into local optimization, fully utilizes the strong robustness and global convergence characteristics of the differential evolution algorithm in the multidimensional parameter space, ensures that the constitutive model can accurately capture the internal physical mechanism of the engineering system, lays a reliable theoretical foundation for virtual sample generation, and significantly improves the accuracy and stability of the model parameters.

[0038] Subsequently, the parameter space is generated for virtual sample generation according to the optimal constitutive model parameters, realizes targeted supplement of the sparse area covered by the real data, generates high-credibility virtual samples through the interpolation mechanism driven by physics, compared with the pure data interpolation or extrapolation method, can ensure the physical consistency and engineering rationality of the generated samples, effectively fills the data blank area in the parameter space, provides more comprehensive and balanced training data distribution for the model, and greatly improves the prediction ability and adaptability of the model to unknown working conditions.

[0039] The application also cooperatively trains the multi-source data through a weighted fusion strategy, fully gives the complementary advantages of different data sources, wherein the real samples provide the most reliable benchmark information, the expanded samples enhance the diversity of data, and the virtual samples fill the data gaps, the three cooperate to form a complete training data ecology, through differentiated weight distribution, ensures the hierarchical management of data quality, avoids the negative impact of low-quality data on model performance, and maximizes the use of limited real data value. In addition, the application also constructs the main control parameter density, dimensionless ratio feature and other composite features with clear engineering significance, converts the original input features into a more physically meaningful representation form, this feature engineering strategy not only reduces the dimension complexity of the feature space, but also enhances the relevance between the features and the target variables, so that the neural network can more effectively learn and generalize the engineering laws, significantly improves the convergence speed and prediction accuracy of the model.

[0040] Overall, the application solves the engineering problem through a hybrid model of fusion training, has strong robustness and high reliability, and can provide stable and reliable prediction results in actual engineering applications, providing a solid technical support for engineering decision-making. BRIEF DESCRIPTION OF DRAWINGS

[0041] The accompanying drawings are included to provide a further understanding of the application and are incorporated in and constitute a part of this specification, illustrate embodiments of the application and together with the description serve to explain the principles of the application. In the drawings:

[0042] Figure 1 A flowchart of an engineering problem solving method based on a fusion model provided by the embodiment of the application;

[0043] Figure 2 A structure schematic diagram of an engineering problem solving system based on a fusion model provided by the embodiment of the application.

[0044] Reference signs:

[0045] 100, enhancement module; 200, optimization module; 300, interpolation module; 400, training module; 500, solving module. DETAILED DESCRIPTION

[0046] In the following, the technical solutions of the present application will be described clearly and completely in connection with the drawings, which are obviously a part of the embodiments of the present application, but not all the embodiments. It should be understood that the description is only exemplary, and is not intended to limit the scope of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative labor should be within the scope of protection of the present application.

[0047] In addition, in the following description, the description of well-known structures and techniques is omitted to avoid unnecessary confusion of the concepts disclosed in the present application.

[0048] In the description of the present application, it should be noted that, unless otherwise explicitly specified and limited, the orientation or positional relationship indicated by the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer" and the like is based on the orientation or positional relationship shown in the drawings, and is only for the convenience of describing the present application and simplifying the description, and therefore cannot be understood as indicating or implying that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as limiting the present application. In addition, the terms "first", "second", "third" are only for description purposes and cannot be understood as indicating or implying relative importance. The terms "mounting", "connection", "connection" should be understood broadly, for example, it can be fixed connection, or detachable connection, or integral connection; it can be mechanical connection, or electrical connection; it can be direct connection, or indirect connection through intermediate medium, or internal communication of two elements. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0049] The exemplary embodiments will be described in detail hereinbelow, examples of which are shown in the drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The implementations described in the following exemplary embodiments do not represent all implementations consistent with the present application. Instead, they are merely examples of methods and systems consistent with some aspects of the present application as detailed in the appended claims.

[0050] The embodiments of the present application will be described in connection with the drawings.

[0051] As Figure 1 shown, the present application provides a fusion model-based engineering problem solving method, comprising:

[0052] S1: data augmentation is performed on real samples by physical constraint noise disturbance to obtain an augmented sample set.

[0053] The application carries out data enhancement of real samples in step S1, and after physical constraint and correlation adjustment, a single real sample generates a plurality of enhanced samples conforming to the engineering physical law, the obtained enhanced samples not only maintain the statistical characteristics of the original data, but also expand the coverage range of the parameter space, thereby providing sufficient training data basis for subsequent constitutive model training.

[0054] The step S1 further comprises:

[0055] S11: setting a differential noise parameter for a feature in the real sample according to a feature statistical characteristic, and performing perturbation transformation on the real sample through the differential noise parameter to obtain a perturbed sample.

[0056] In step S11, the noise type and the added strength are first determined according to the physical mechanism of the corresponding engineering field of the to-be-solved engineering problem, that is, according to the variation range and sensitivity of each parameter in the actual engineering, and then the perturbation is realized by adding random noise conforming to the normal distribution to the original feature value. For example, in the welding engineering problem, a Gaussian noise with a mean value of 0 and a standard deviation of 1.5 is added to the welding current value of each real sample, and the original welding current is 75A. After the noise perturbation, the welding current value is , wherein represents a normal distribution random number with a mean value of 0 and a variance of 2.25, and the probability of the perturbed current value falling within the range of 72.5A to 77.5A is about 68%. Finally, a plurality of perturbed samples are generated from a single real sample by simultaneously perturbing all features, thereby expanding the distribution range of the original data.

[0057] The differential noise parameter in step S11 is adaptively set according to the physical mechanism of the corresponding engineering field of the to-be-solved engineering problem, and the differential noise parameter comprises a process parameter noise amplitude, a geometric parameter noise amplitude, a temperature parameter noise amplitude and a time parameter noise amplitude.

[0058] In the high-temperature alloy welding crack prediction scene, the process parameter noise amplitude is set as a Gaussian noise of positive and negative amperes for the welding current, because the welding current directly affects the heat input, and its change has a significant impact on the welding quality. The geometric parameter noise amplitude is the noise set for the base material thickness, because the slight change of the geometric size will affect the heat conduction path. The temperature parameter noise amplitude is the noise set for the preheating temperature, because the temperature parameter affects the thermal expansion and contraction behavior of the material. The time parameter noise amplitude is the noise set for the single pass time length, because the time parameter is directly related to the duration of the thermal cycle process.

[0059] S12: adding a physical constraint condition to the perturbed sample according to the field rule corresponding to the engineering problem to obtain a physically constrained sample.

[0060] Further, the application adds physical constraint conditions to the perturbed samples according to the field rules corresponding to the engineering problem in step S12. The physical constraint conditions refer to constraint relationships between parameters established based on engineering physical principles to ensure that the generated perturbed samples are physically reasonable and feasible.

[0061] For example, in welding engineering, the physical constraint conditions include geometric constraints such as the weld depth cannot exceed the thickness of the base material, process matching constraints between the welding wire diameter and the welding current, energy conservation constraints, etc. The application of constraint conditions is realized through conditional judgment and numerical clipping in the data processing process. In the specific data processing process, first, the relationship between the perturbed weld depth and the thickness of the base material is checked. When the weld depth exceeds the thickness of the base material, the weld depth value is clipped to the thickness of the base material. For example, when the perturbed weld depth is 5.2 mm and the thickness of the base material is 4.8 mm, the weld depth is adjusted to 4.8 mm. Then, the matching relationship between the welding wire diameter and the welding current is checked. According to the welding process specification, when the welding wire diameter is less than 1.2 mm, the welding current should not exceed 70 A. When it is detected that the welding wire diameter is 1.0 mm and the welding current is 85 A, the welding current is adjusted to 70 A. Finally, the energy conservation constraint is applied. The ratio of heat input energy to heat dissipation energy is calculated. When the ratio exceeds the physically reasonable range, the sample is removed.

[0062] S13: Correlation adjustment is performed on the physical constraint samples to obtain an enhanced sample set.

[0063] In step S13, the application performs correlation adjustment on the physical constraint samples. Correlation adjustment refers to the linkage correction of sample characteristic values according to the mutual influence relationship between engineering parameters to ensure the coordination and consistency between multiple parameters. For example, in the welding process, there is a positive correlation between the welding current and the welding wire diameter, and there is an inverse correlation between the cooling time and the heat input. The greater the heat input, the longer the cooling time required to reach a certain temperature. Therefore, when the welding wire diameter changes from the standard value of 1.4 mm to 1.8 mm, the welding current increases by 8 A from the reference value, i.e. new current value = original current value + (1.8-1.4) x 20, where 20 is an empirical coefficient. Then, the effect of heat input on cooling time is calculated. When the heat input increases, the cooling time is adjusted according to the cooling factor. The adjusted cooling time = original cooling time x cooling factor. Through this correlation adjustment, the enhanced samples generated maintain reasonable mutual relationships between physical parameters.

[0064] S2: According to the enhanced sample set, the physical constitutive model is parameter optimized to obtain the optimal constitutive model parameters.

[0065] Step S2 further includes:

[0066] S21: training a constitutive equation according to the real samples and the enhanced sample set, and obtaining a physical constitutive model.

[0067] Further, in step S21, the training of the constitutive equation is first performed. In the training process of the constitutive equation, the feature data in the enhanced sample set is substituted into each component term of the constitutive equation, and the error between the calculation result of the constitutive equation and the sample label value is minimized by adjusting the parameter value. The data processing logic first extracts the feature values in the enhanced sample set, then substitutes these feature values into the corresponding constitutive equation terms for calculation, then calculates the value of the influence factor term, and obtains the contribution value of the influence factor term as the value raised to the power of α, and simultaneously calculates the values of other terms and combines them into the complete output value of the constitutive equation.

[0068] In step S21, the constitutive equation is constructed based on the physical mechanism of the engineering field corresponding to the engineering problem to be solved, and includes an influence factor term, a coupling effect term, an environmental condition term, and a material property term.

[0069] In step S21, the constitutive equation is trained according to the real samples and the enhanced sample set to obtain a physical constitutive model. The constitutive equation is a mathematical expression constructed based on the physical mechanism of the engineering field corresponding to the engineering problem to be solved, and is used to describe the internal physical relationship between parameters in the engineering system. The constitutive equation includes four components: an influence factor term, a coupling effect term, an environmental condition term, and a material property term. The influence factor term describes the direct influence of main process parameters on the target variable, the coupling effect term reflects the interaction relationship between multiple parameters, the environmental condition term describes the influence of external environment on the engineering process, and the material property term reflects the influence of inherent properties of materials on engineering behavior.

[0070] S22: performing global optimization on the parameter constraint space of the physical constitutive model by a differential evolution method to obtain optimal constitutive model parameters.

[0071] Further, the differential evolution method is a kind of random search algorithm based on population, which finds the global optimal solution in the parameter space through differential mutation and crossover operation between individuals. Specifically, first, an initial population is randomly generated in the parameter constraint space, each individual representing a combination of constitutive model parameters. The parameter constraint space is set with boundary ranges according to physical meaning and engineering experience. Then, through the differential evolution algorithm, mutation operation, crossover operation and selection operation are performed, and finally the individual with the optimal fitness is selected and reserved. The fitness function is defined as the accuracy between the predicted value of the constitutive equation and the real label value, i.e., the global optimal parameter combination is obtained.

[0072] S3: interpolating the enhanced sample set by the physical constitutive model optimized by the optimal constitutive model parameters to obtain a virtual sample set.

[0073] Further, step S3 is a virtual sample generation stage, using the parameter-optimized physical constitutive model, generating a virtual sample set through interpolation technology, aiming to use the generalization ability of the physical constitutive model to generate new virtual data points in the existing sample space, thereby expanding the training data set and improving the generalization performance of the subsequent model.

[0074] Further, step S3 further comprises:

[0075] S31: Interpolating the enhanced sample set through the physical constitutive model optimized by the optimal constitutive model parameters to obtain basic virtual samples.

[0076] Further, in step S31, the optimal constitutive model parameters obtained by globally optimizing the parameters of the physical constitutive model through the differential evolution method in step S2 are first used to drive the physical constitutive model, and then the subsequent virtual sample generation is performed using the constitutive model. The specific process is to input the optimal constitutive model parameters obtained in step S2 + the enhanced sample set in step S1, and then perform interpolation operation on the enhanced sample set using the optimized physical constitutive model. In the interpolation operation, interpolation is performed between the data points of the enhanced sample set, and intermediate data points conforming to the physical law are finally generated by using the continuity and smoothness of the physical constitutive model, and the basic virtual samples are output.

[0077] S32: Directly interpolating and supplementing the real data covering sparse areas in the basic virtual samples to obtain supplemented virtual samples.

[0078] Further, the real data covering sparse areas refers to areas with low real sample density in the multi-dimensional parameter space. The parameter combinations of the area are relatively rare in actual engineering, but they are needed due to the model generalization performance. Therefore, the present application directly interpolates and supplements in step S32 by identifying the sparse areas of the parameter space and generating virtual samples to fill the gaps. Specifically, the parameter space is first divided into multiple sub-regions, the number density of real samples in each sub-region is calculated, the sparse areas with sample density lower than the average density threshold are identified, and then the number of generated virtual samples in the sparse areas is increased. Specifically, the probability distribution of random sampling is adjusted, the sampling weight of the parameter values corresponding to the sparse areas is increased, so that the generated virtual samples cover these sparse areas more.

[0079] S33: Physically constraining the supplemented virtual samples to obtain a virtual sample set.

[0080] In step S33, the physical constraint conditions for physical constraint are set according to the physical mechanism of the engineering field corresponding to the to-be-solved engineering problem, including geometric compatibility constraint, process matching constraint and energy balance constraint.

[0081] The step S33 obtains the virtual sample set by physical constraints on the supplementary virtual sample, the physical constraint conditions including three types of geometric compatibility constraints, process matching constraints and energy balance constraints, the geometric compatibility constraints ensuring the physical rationality between geometric parameters, the process matching constraints ensuring the coordinated matching between process parameters, and the energy balance constraints ensuring the physical balance between energy input and heat dissipation. Through the physical constraints, it can be ensured that the virtual sample does not exceed the reasonable range, and finally the virtual sample set meeting the physical constraints is generated, thereby providing sufficient diversified training data for subsequent neural network training.

[0082] S4: According to the real sample, the enhanced sample set and the virtual sample set, the neural network model is trained in a multi-source collaborative manner to obtain a hybrid training model.

[0083] The multi-source collaborative training process in the step S4 firstly performs weighted fusion on the data of different sources through a differentiated weight allocation mechanism. The real sample is the most reliable data source, directly derived from actual engineering measurement and thus having the highest reliability. The enhanced sample is still disturbed by artificial disturbance characteristics although it is subjected to physical constraints. The virtual sample is generated by a constitutive model and has certain theoretical derivation properties. Therefore, the final weighted mechanism needs to ensure that the real data occupies a dominant position in the model training. Then, the neural network is trained in a multi-source collaborative manner according to the weighted data set to obtain a hybrid training model.

[0084] The step S4 further includes:

[0085] S41: The real sample, the enhanced sample set and the virtual sample set are allocated with differentiated weights and are subjected to weighted fusion to obtain a weighted sample set.

[0086] In the step S41, the weight of the real sample is greater than the weight of the enhanced sample set, and the weight of the real sample is greater than the weight of the virtual sample set.

[0087] Specifically, when the weights are allocated, the weight of the real sample is set to 5.0 because these data are directly derived from the measurement results of actual welding experiments and contain real physical processes and material properties. The weight of the enhanced sample set is set to 1.0 because the physical accuracy is relatively reduced although the statistical characteristics of the original data are maintained. The weight of the virtual sample set is also set to 1.0 because it lacks actual verification although it is generated by a constitutive model and meets the physical constraints. In the weighted fusion process, the loss function contribution degree of each sample is adjusted according to the weight, that is, the prediction error of the real sample will produce a gradient update amplitude 5 times that of the enhanced sample in the back propagation, so as to make the model fit the distribution characteristics of the real data more. At the same time, the enhanced sample and the virtual sample are used to expand the decision boundary and fill the data sparse area.

[0088] S42: Physical feature extraction is performed on the weighted sample set to obtain an engineering feature set.

[0089] The physical feature extraction process in step S42 converts the original engineering parameters into engineering features with stronger physical meaning. For example, in the welding engineering problem, the extraction of engineering features includes the calculation of heat input density characteristics, which are obtained by multiplying the welding current, welding voltage, and square root of the weld spot area and weld spot depth, and then dividing by the product of the welding wire cross-sectional area and the wire feed speed. The engineering features are calculated by physical formulas, have stronger physical interpretation and prediction ability, and can better reflect the physical mechanism and crack formation law in the welding process.

[0090] S43: Deep learning is performed on the engineering feature set to train a neural network model, obtaining a hybrid training model.

[0091] Further, the neural network model adopts a multi-layer neural network structure, including an input layer, two hidden layers, and an output layer. The input layer receives 15 engineering features. The first hidden layer contains 64 neurons and uses the ReLU activation function, with an L2 regularization coefficient of 0.01 and batch normalization processing. The Dropout ratio is set to 0.3 to prevent overfitting. The second hidden layer contains 32 neurons and also uses the ReLU activation function, with an L2 regularization coefficient of 0.01 and batch normalization. The Dropout ratio is set to 0.2. The output layer uses a single neuron and the Sigmoid activation function for binary classification prediction. The training process uses the Adam optimizer.

[0092] The final hybrid training model is a deep neural network trained by multiple sources of data. The model can accurately predict the physical parameters or required results of engineering problems, such as whether a crack will occur in the high-temperature alloy welding process under given welding process parameters. In this prediction scenario, the model receives 9 engineering input parameters including welding current, single pass duration, base material thickness, weld spot depth, weld spot area, preheating temperature, interlayer temperature, cooling time, and welding wire diameter. Through internal physical feature extraction and deep learning processing, it outputs a probability value between 0 and 1 representing the likelihood of crack occurrence. When the output value is greater than 0.5, it is determined that a crack will occur, and when it is less than or equal to 0.5, it is determined that a crack will not occur.

[0093] The mixed training model obtained by the application can process complex nonlinear mapping relationships under small sample conditions, and overcome the problem of overfitting of traditional machine learning models in the case of data scarcity by fusing physical mechanisms and data-driven methods. The model integrates physical constraints and engineering knowledge, so that the prediction result not only has statistical significance but also conforms to the physical law. For example, when the input parameters show that the welding current is too high and the cooling time is too short, the model will give a high crack risk prediction based on the heat accumulation effect and material properties. The obtained prediction ability is based on the deep understanding of the welding physical process rather than simple data fitting.

[0094] S5: predicting and solving the engineering problem by using the mixed training model to obtain a solution result.

[0095] Further, step S5 uses the mixed training model obtained by the foregoing training to solve the engineering problem. The solving process of the engineering problem is divided into two modes of prediction solving and optimization solving. In the prediction solving mode, the engineer inputs specific welding process parameters, the model calculates engineering features through the feature engineering module, and then inputs these features into the trained neural network for forward propagation calculation, and finally outputs the probability of the phenomenon to be predicted, such as crack probability. Subsequently, it can be judged that the process parameters are not suitable and need to be adjusted. In the optimization solving mode, the model is combined with an optimization algorithm to find optimal process parameters that meet specific constraint conditions, for example, the engineer sets the objective function as minimizing the crack probability while meeting the welding quality requirements.

[0096] The mixed training model can also perform sensitivity analysis and parameter influence evaluation. By fixing other parameters and changing the value of a single parameter, the change trend of the model output is observed, so as to identify the key parameter that has the greatest influence on the formation of the predicted phenomenon. The real-time prediction capability of the model enables it to be integrated into the control system of the welding equipment. During the welding process, the crack risk is dynamically evaluated according to the real-time monitored parameters, and the process parameters are automatically adjusted or a warning signal is issued when the prediction probability exceeds a preset threshold, realizing the transformation from offline analysis to online control, and greatly improving the stability and controllability of the welding quality.

[0097] As shown in Figure 2 The application also provides an engineering problem solving system based on a fusion model, which comprises:

[0098] The enhancement module 100 is used for performing data enhancement on the real samples by physical constraint noise disturbance to obtain an enhanced sample set;

[0099] The optimization module 200 is used for performing parameter optimization on the physical constitutive model according to the enhanced sample set to obtain optimal constitutive model parameters;

[0100] An interpolation module 300 configured as the physical constitutive model optimized by the optimal constitutive model parameters, for interpolating the enhanced sample set to obtain a virtual sample set;

[0101] A training module 400 for performing multi-source collaborative training on a neural network model according to the real samples, the enhanced sample set and the virtual sample set, to obtain a hybrid training model;

[0102] A solving module 500 configured as the hybrid training model trained by the training module 400, for predicting and solving a to-be-solved engineering problem to obtain a solving result.

[0103] The device embodiments described above are merely illustrative, wherein the units described as separate components can or can not be physically separated, and the components displayed as units can or can not be physical units, i.e., they can be located in one place or distributed on multiple network units. Part or all of the modules can be selected to achieve the purpose of the embodiment according to actual needs. Those skilled in the art can understand and implement it without creative labor.

[0104] From the above description of the embodiments, those skilled in the art can clearly understand that the embodiments can be realized by means of software plus necessary general hardware platforms, and of course, can also be realized by hardware. Based on such understanding, the above technical solutions can be embodied in the form of software products, which can be stored in a computer readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and include a number of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute the methods described in each embodiment or some parts of the embodiments.

[0105] In one specific embodiment, the present application is implemented by using a K439B high-temperature alloy welding crack data set, wherein the initial real samples are only 32, and include the following key features: welding current , single pass duration , base material thickness , weld depth , weld area , preheating temperature , interlayer temperature , duration of cooling to below 500℃ , wire diameter , and the final target variable is to judge whether welding cracks occur.

[0106] In a specific implementation, the data augmentation phase of the present application expands 32 real samples to 32,000 augmented samples, trains the physical constitutive model on the augmented data, and achieves an accuracy of 86%. Based on the optimal constitutive model, 32,032 virtual samples are generated, and the final neural network model achieves an accuracy of 89% on the test set and 83% on the original data.

[0107] The present application provides an engineering problem solving method and system based on a fusion model, proposes a hybrid training framework that fuses domain mechanisms and data augmentation, uses data augmentation techniques to expand sample diversity, and simultaneously introduces an engineering constitutive model to construct a virtual data generator that meets physical constraints for directional interpolation completion in the sparse interval of real data. This method realizes the collaborative optimization of physical driving and data driving, significantly improves the adaptability of the model to samples outside the data distribution under the premise of ensuring the credibility of the generated data, and provides an implementable solution for the construction of high-robustness machine learning models under small-sample conditions.

[0108] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the embodiments of the present application, and not to limit them. Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can modify the technical solutions recorded in the foregoing embodiments, or make equivalent substitutions for some technical features; and these modifications or substitutions do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application. Any changes or substitutions that can be easily thought of by those skilled in the art within the scope of the disclosed technology should be covered within the protection scope of the present application.

Claims

1. A method for solving engineering problems based on a fusion model, characterized in that, include: S1: Data augmentation of real samples is performed by physically constraining noise perturbation to obtain an augmented sample set; Step S1 further includes: S11: Set differential noise parameters for the features in the real sample according to the feature statistical characteristics, and perform perturbation transformation on the real sample through the differential noise parameters to obtain the perturbed sample; S12: Add physical constraints to the disturbance sample according to the domain rules corresponding to the engineering problem to obtain a physical constraint sample; S13: The physical constraint samples are relevance adjusted to obtain an enhanced sample set; This includes: determining the noise type and intensity based on the variation range and sensitivity of each parameter in actual engineering, and then adding random noise that follows a normal distribution to the original feature values. Specifically, Gaussian noise is added to the welding current value of each real sample, and the noise perturbation causes a single real sample to generate multiple perturbed samples. The physical constraints include the geometric constraint that the weld depth cannot exceed the thickness of the base material, the process matching constraint between the welding wire diameter and the welding current, and the energy conservation constraint. S2: Based on the enhanced sample set, optimize the parameters of the physical constitutive model to obtain the optimal constitutive model parameters; S3: Interpolate the enhanced sample set using the physical constitutive model optimized by the optimal constitutive model parameters to obtain a virtual sample set; S4: Based on the real samples, the enhanced sample set, and the virtual sample set, perform multi-source collaborative training on the neural network model to obtain a hybrid training model; The hybrid training model accurately predicts the physical parameters or required results of engineering problems, including determining whether cracks will occur during the welding of high-temperature alloys by extracting internal physical features and deep learning under given welding process parameters. The welding process parameters include welding current, single pass duration, base metal thickness, weld depth, weld area, preheating temperature, interpass temperature, cooling time, and welding wire diameter. S5: The hybrid training model is used to predict and solve the engineering problem to be solved, and the solution result is obtained. The prediction solution includes prediction solution and optimization solution; In the predictive solution mode, the process includes: inputting welding process parameters, the hybrid training model calculating engineering features through the feature engineering module, inputting the engineering features into the trained neural network for forward propagation calculation, and finally outputting the probability of the phenomenon to be predicted. In the optimization solution mode, the objective function is set to minimize the probability of crack occurrence while meeting the welding quality requirements.

2. The method for solving engineering problems based on a fusion model according to claim 1, characterized in that, The differentiated noise parameters in step S11 are adaptively set according to the physical mechanism of the engineering field corresponding to the engineering problem to be solved.

3. A method for solving engineering problems based on a fusion model, characterized in that, include: S1: Data augmentation of real samples is performed by physically constraining noise perturbation to obtain an augmented sample set; This includes: determining the noise type and intensity based on the variation range and sensitivity of each parameter in actual engineering, and then adding random noise that follows a normal distribution to the original feature values. Specifically, Gaussian noise is added to the welding current value of each real sample, and the noise perturbation causes a single real sample to generate multiple perturbed samples. Physical constraints are added to the disturbance samples according to the domain rules corresponding to the engineering problem to obtain physically constrained samples. The physical constraints include the geometric constraint that the weld depth cannot exceed the thickness of the base material, the process matching constraint between the welding wire diameter and the welding current, and the energy conservation constraint. S2: Based on the enhanced sample set, optimize the parameters of the physical constitutive model to obtain the optimal constitutive model parameters; Step S2 further includes: S21: Train constitutive equations based on the real samples and the enhanced sample set to obtain a physical constitutive model; S22: The optimal constitutive model parameters are obtained by globally optimizing the parameter constraint space of the physical constitutive model using the differential evolution method. S3: Interpolate the enhanced sample set using the physical constitutive model optimized by the optimal constitutive model parameters to obtain a virtual sample set; S4: Based on the real samples, the enhanced sample set, and the virtual sample set, perform multi-source collaborative training on the neural network model to obtain a hybrid training model; The hybrid training model accurately predicts the physical parameters or required results of engineering problems, including determining whether cracks will occur during the welding of high-temperature alloys by extracting internal physical features and deep learning under given welding process parameters. The welding process parameters include welding current, single pass duration, base metal thickness, weld depth, weld area, preheating temperature, interpass temperature, cooling time, and welding wire diameter. S5: The hybrid training model is used to predict and solve the engineering problem to be solved, and the solution result is obtained. The prediction solution includes prediction solution and optimization solution; In the prediction solution mode, the following is included: the hybrid training model calculates engineering features through the feature engineering module, inputs the engineering features into the trained neural network for forward propagation calculation, and finally outputs the probability of the phenomenon to be predicted; In the optimization solution mode, the objective function is set to minimize the probability of crack occurrence while meeting the welding quality requirements.

4. The method for solving engineering problems based on a fusion model according to claim 3, characterized in that, The constitutive equation in step S21 is constructed based on the physical mechanism of the engineering field corresponding to the engineering problem to be solved.

5. A method for solving engineering problems based on a fusion model, characterized in that, include: S1: Data augmentation of real samples is performed by physically constraining noise perturbation to obtain an augmented sample set; This includes: determining the noise type and intensity based on the variation range and sensitivity of each parameter in actual engineering, and then adding random noise that follows a normal distribution to the original feature values. Specifically, Gaussian noise is added to the welding current value of each real sample, and the noise perturbation causes a single real sample to generate multiple perturbed samples. Physical constraints are added to the disturbance samples according to the domain rules corresponding to the engineering problem to obtain physically constrained samples. The physical constraints include the geometric constraint that the weld depth cannot exceed the thickness of the base material, the process matching constraint between the welding wire diameter and the welding current, and the energy conservation constraint. S2: Based on the enhanced sample set, optimize the parameters of the physical constitutive model to obtain the optimal constitutive model parameters; S3: Interpolate the enhanced sample set using the physical constitutive model optimized by the optimal constitutive model parameters to obtain a virtual sample set; Step S3 further includes: S31: Interpolate the enhanced sample set using the physical constitutive model optimized by the optimal constitutive model parameters to obtain basic virtual samples; S32: Perform directional interpolation to supplement the sparsely covered areas of real data in the basic virtual sample to obtain supplementary virtual samples; S33: Apply physical constraints to the supplementary virtual samples to obtain a virtual sample set; S4: Based on the real samples, the enhanced sample set, and the virtual sample set, perform multi-source collaborative training on the neural network model to obtain a hybrid training model; The hybrid training model accurately predicts the physical parameters or required results of engineering problems, including determining whether cracks will occur during the welding of high-temperature alloys by extracting internal physical features and deep learning under given welding process parameters. The welding process parameters include welding current, single pass duration, base metal thickness, weld depth, weld area, preheating temperature, interpass temperature, cooling time, and welding wire diameter. S5: The hybrid training model is used to predict and solve the engineering problem to be solved, and the solution result is obtained. The prediction solution includes prediction solution and optimization solution; In the predictive solution mode, the process includes: inputting welding process parameters, the hybrid training model calculating engineering features through the feature engineering module, inputting the engineering features into the trained neural network for forward propagation calculation, and finally outputting the probability of the phenomenon to be predicted. In the optimization solution mode, the objective function is set to minimize the probability of crack occurrence while meeting the welding quality requirements.

6. The method for solving engineering problems based on a fusion model according to claim 5, characterized in that, In step S33, the physical constraints used for physical constraints are set according to the physical mechanism of the engineering field corresponding to the engineering problem to be solved.

7. A method for solving engineering problems based on a fusion model, characterized in that, include: S1: Data augmentation of real samples is performed by physically constraining noise perturbation to obtain an augmented sample set; This includes: determining the noise type and intensity based on the variation range and sensitivity of each parameter in actual engineering, and then adding random noise that follows a normal distribution to the original feature values. Specifically, Gaussian noise is added to the welding current value of each real sample, and the noise perturbation causes a single real sample to generate multiple perturbed samples. Physical constraints are added to the disturbance samples according to the domain rules corresponding to the engineering problem to obtain physically constrained samples. The physical constraints include the geometric constraint that the weld depth cannot exceed the thickness of the base material, the process matching constraint between the welding wire diameter and the welding current, and the energy conservation constraint. S2: Based on the enhanced sample set, optimize the parameters of the physical constitutive model to obtain the optimal constitutive model parameters; S3: Interpolate the enhanced sample set using the physical constitutive model optimized by the optimal constitutive model parameters to obtain a virtual sample set; S4: Based on the real samples, the enhanced sample set, and the virtual sample set, perform multi-source collaborative training on the neural network model to obtain a hybrid training model; Step S4 further includes: S41: Assign differentiated weights to the real samples, augmented sample set, and virtual sample set, and perform weighted fusion to obtain a weighted sample set; S42: Extract physical features from the weighted sample set to obtain an engineering feature set; S43: Perform deep learning on the engineering feature set to train a neural network model and obtain a hybrid training model; The hybrid training model accurately predicts the physical parameters or required results of engineering problems, including determining whether cracks will occur during the welding of high-temperature alloys by extracting internal physical features and deep learning under given welding process parameters. The welding process parameters include welding current, single pass duration, base metal thickness, weld depth, weld area, preheating temperature, interpass temperature, cooling time, and welding wire diameter. S5: The hybrid training model is used to predict and solve the engineering problem to be solved, and the solution result is obtained. The prediction solution includes prediction solution and optimization solution; In the predictive solution mode, the process includes: inputting welding process parameters, the hybrid training model calculating engineering features through the feature engineering module, inputting the engineering features into the trained neural network for forward propagation calculation, and finally outputting the probability of the phenomenon to be predicted. In the optimization solution mode, the objective function is set to minimize the probability of crack occurrence while meeting the welding quality requirements.

8. The method for solving engineering problems based on a fusion model according to claim 7, characterized in that, In step S41, the weight of the real sample is greater than the weight of the augmented sample set, and the weight of the real sample is greater than the weight of the virtual sample set.

9. A system for solving engineering problems based on a fusion model, used to execute a method for solving engineering problems based on a fusion model as described in any one of claims 1 to 8, characterized in that, include: Augmentation module: Used to augment real samples by physically constraining noise perturbations, resulting in an augmented sample set; Optimization module: used to optimize the parameters of the physical constitutive model based on the enhanced sample set, and obtain the optimal constitutive model parameters; An interpolation module, configured as the physical constitutive model after the optimal constitutive model parameters are optimized, is used to interpolate the enhanced sample set to obtain a virtual sample set; Training module: used to perform multi-source collaborative training on the neural network model based on the real samples, the augmented sample set, and the virtual sample set to obtain a hybrid training model; The solution module is configured to use the hybrid training model obtained by the training module to predict and solve the engineering problem to be solved, and obtain the solution result.

Images