Carriage riveting structure performance collaborative optimization model training method and system based on multi-task learning

Abstract

The present application relates to the technical field of model training, and discloses a riveted structure performance collaborative optimization model training method and system based on multi-task learning, which comprises the following steps: step 1, obtaining sample parameters, generating a structured input vector and a performance supervision label; step 2, obtaining a riveting coupling state vector based on the structured input vector; step 3, constructing a riveting coupling state graph and extracting shared features; step 4, obtaining task features and performance prediction values; step 5, constructing a basic loss, a coupling consistency loss and a total training target function; step 6, identifying task conflicts and updating model parameters until convergence; and step 7, obtaining a performance collaborative optimization model and determining a collaborative optimization evaluation value. The present application realizes joint prediction and collaborative optimization evaluation of multiple performances of a riveted structure of a carriage.

Description

Technical Field

[0001] This invention belongs to the field of model training technology, specifically relating to a training method and system for a collaborative optimization model of the performance of a car body riveting structure based on multi-task learning. Background Technology

[0002] Riveted structures for train carriages are widely used in the manufacturing of rail vehicles and other car bodies. Their design typically affects overall load-bearing capacity, connection durability, structural quality, and assembly adaptability. Because riveted connections are related to the connection relationships between the main car body components and are influenced by factors such as rivet arrangement, plate layer combination, hole fit, assembly deviations, and clamping conditions, the same structural design often corresponds to multiple interrelated performance indicators. In practical design, optimizing only one performance aspect can easily lead to a decline in other performance aspects, making it difficult to meet the comprehensive requirements of the train carriage structure for load-bearing capacity, durability, lightweighting, and assembly consistency.

[0003] In existing technologies, the performance analysis of riveted structures in train carriages often employs finite element analysis, empirical rule verification, or the establishment of surrogate models for individual indicators. These methods typically treat overall load-bearing capacity, connection durability, structural quality, and assembly sensitivity separately, lacking a unified modeling of the relationships between multiple performance parameters. This can easily lead to numerous design iterations, long analysis cycles, and difficulty in simultaneously addressing the synergistic optimization requirements of multiple performance parameters. Summary of the Invention

[0004] This invention provides a training method and system for a collaborative optimization model of the performance of a car body riveting structure based on multi-task learning. It solves the technical problems in related technologies where multiple performance aspects of the car body riveting structure are usually modeled separately, making it difficult to uniformly learn the coupling information between structural relationships and connection states. Furthermore, the multi-task training process is prone to mutual interference between tasks, resulting in insufficient collaborative optimization effect and training stability.

[0005] This invention provides a training method for a collaborative optimization model of the performance of a car body riveting structure based on multi-task learning, comprising the following steps:

[0006] Step 1: Obtain the original design and process parameters of multiple car body riveting structure samples to obtain structured input vectors and performance supervision labels; wherein, the structured input vectors include component layer information, riveting layer information and process constraint layer information;

[0007] Step 2: Based on the component layer information, riveting layer information and process constraint layer information in the structured input vector, perform combined calculations to obtain the riveting coupling state vector;

[0008] Step 3: Based on the structured input vector, the riveting coupling state vector, and the basic data of the geometric and connection relationships of the riveting structure of the carriage, construct the riveting coupling state diagram and extract the shared features;

[0009] Step 4: Based on the shared features and the riveted coupling state vector, obtain the task features corresponding to each performance task, and obtain the performance prediction value based on each task feature;

[0010] Step 5: Calculate the basic loss based on the performance prediction value and the performance supervision label, calculate the coupling consistency loss based on the preset correlation between the riveted coupling state vector and the performance prediction value, and construct the overall training objective function;

[0011] Step 6: Calculate the gradient based on the basic loss corresponding to each performance task, identify task conflicts and determine the intensity of task conflicts, dynamically adjust the task weights according to the intensity of task conflicts, and update the parameters of the shared encoding network, task-specific decoupling module and prediction head based on the adjusted task weights until the overall training objective function converges.

[0012] Step 7: After the overall training objective function converges, the collaborative optimization model of the car body riveting structure performance is obtained, and the collaborative optimization evaluation value is determined based on the performance prediction value corresponding to the candidate car body riveting structure scheme.

[0013] This invention also provides a training system for a collaborative optimization model of the performance of riveting structures in train carriages based on multi-task learning, including:

[0014] The sample construction module obtains the original design and process parameters of multiple car body riveting structure samples, and obtains structured input vectors and performance supervision labels; wherein, the structured input vectors include component layer information, riveting layer information and process constraint layer information;

[0015] The state generation module performs combined calculations based on the component layer information, riveting layer information, and process constraint layer information in the structured input vector to obtain the riveting coupling state vector;

[0016] The graph encoding module constructs a riveting coupling state graph and extracts shared features based on the structured input vector, the riveting coupling state vector, and the basic data of the geometric and connection relationships of the riveting structure of the carriage.

[0017] The task prediction module obtains the task features corresponding to each performance task based on shared features and riveted coupling state vectors, and obtains the performance prediction value based on each task feature.

[0018] The objective construction module calculates the basic loss based on the performance prediction value and the performance supervision label, calculates the coupling consistency loss based on the preset correlation between the riveted coupling state vector and the performance prediction value, and constructs the overall training objective function;

[0019] The conflict control module calculates the gradient based on the base loss corresponding to each performance task, identifies task conflicts and determines the intensity of task conflicts, dynamically adjusts the task weights according to the intensity of task conflicts, and updates the parameters of the shared encoding network, task-specific decoupling module and prediction head based on the adjusted task weights until the overall training objective function converges.

[0020] The optimization evaluation module obtains a collaborative optimization model for the performance of the riveting structure of the carriage after the overall training objective function converges, and determines the collaborative optimization evaluation value based on the performance prediction value corresponding to the candidate carriage riveting structure scheme.

[0021] The beneficial effects of this invention are as follows: By hierarchically organizing, constructing state variables, modeling graph structures, and jointly training multiple tasks on the riveted car body structure samples, this invention can simultaneously consider multiple performance parameters such as overall load-bearing margin, connection durability damage, structural quality, and assembly tolerance sensitivity within the same model, avoiding the inefficiency and fragmented results caused by analyzing each performance parameter separately in existing technologies. Furthermore, by introducing shared feature extraction, task-specific decoupling, coupling consistency constraints, and task conflict control mechanisms, this invention enables the model to learn structural relationships and connection states while also considering the correlations and differences between different performance tasks, thereby improving the stability, consistency, and overall reliability of model training and prediction results. The resulting collaborative optimization model can also be directly used for candidate scheme evaluation, which is beneficial for improving the efficiency of riveted car body structure design and optimization. Attached Figure Description

[0022] Figure 1 This is a flowchart of the training method for the collaborative optimization model of the performance of the riveting structure of the carriage based on multi-task learning according to the present invention. Detailed Implementation

[0023] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.

[0024] like Figure 1 As shown, the training method for the collaborative optimization model of the performance of the riveting structure of the carriage based on multi-task learning includes the following steps:

[0025] Step 1: Obtain the original design and process parameters of multiple car body riveting structure samples to obtain structured input vectors and performance supervision labels; wherein, the structured input vectors include component layer information, riveting layer information and process constraint layer information;

[0026] Step 2: Based on the component layer information, riveting layer information and process constraint layer information in the structured input vector, perform combined calculations to obtain the riveting coupling state vector;

[0027] Step 3: Based on the structured input vector, the riveting coupling state vector, and the basic data of the geometric and connection relationships of the riveting structure of the carriage, construct the riveting coupling state diagram and extract the shared features;

[0028] Step 4: Based on the shared features and the riveted coupling state vector, obtain the task features corresponding to each performance task, and obtain the performance prediction value based on each task feature;

[0029] Step 5: Calculate the basic loss based on the performance prediction value and the performance supervision label, calculate the coupling consistency loss based on the preset correlation between the riveted coupling state vector and the performance prediction value, and construct the overall training objective function;

[0030] Step 6: Calculate the gradient based on the basic loss corresponding to each performance task, identify task conflicts and determine the intensity of task conflicts, dynamically adjust the task weights according to the intensity of task conflicts, and update the parameters of the shared encoding network, task-specific decoupling module and prediction head based on the adjusted task weights until the overall training objective function converges.

[0031] Step 7: After the overall training objective function converges, the collaborative optimization model of the car body riveting structure performance is obtained, and the collaborative optimization evaluation value is determined based on the performance prediction value corresponding to the candidate car body riveting structure scheme.

[0032] In one embodiment of the present invention, to ensure that the training of the subsequent collaborative optimization model for the riveting structure performance of the carriage is based on unified, computable, and correspondable data, the original design and process parameters of multiple carriage riveting structure samples are first obtained, and structured input vectors and performance supervision labels are generated accordingly. Here, a carriage riveting structure sample refers to a data unit that corresponds to a specific carriage riveting structure scheme. This data unit includes structural design parameters, assembly process parameters, and performance results corresponding to that structural scheme. The structured input vector refers to input data arranged according to a unified field order, used as sample input for model training; the performance supervision label refers to the target output data corresponding to this input, used to constrain the model to learn the mapping relationship between different performance parameters such as overall load-bearing margin, connection durability damage, structural quality, and assembly tolerance sensitivity.

[0033] Specifically, in step 11, the original design and process parameters of multiple car body riveting structure samples are collected according to sample numbers and divided into component layer information, riveting layer information, and process constraint layer information according to the object type represented by the parameters. The sample number is used to identify all parameters and performance data under the same car body riveting structure sample so that the subsequent input data and label data can maintain a correspondence. The component layer information is used to reflect the basic attributes of the main car body components, including plate thickness, beam size, material type, component spatial location, and geometric contour of the connection area. Among them, plate thickness and beam size are used to describe the geometric scale of the main components, material type is used to reflect the material properties of the components, and component spatial location and geometric contour of the connection area are used to reflect the distribution of the connection area in the overall structure of the car body.

[0034] The riveting layer information reflects the design attributes of the riveting connection itself, including rivet diameter, rivet spacing, edge distance, arrangement, number of riveting layers, and riveting area distribution. Specifically, rivet diameter, rivet spacing, and edge distance reflect the basic dimensional relationships of the riveting connection; the arrangement reflects the form of rivets in the connection area; and the number of riveting layers and riveting area distribution reflect the interlayer relationships and spatial distribution of the connection. The process constraint layer information reflects the influence of manufacturing and assembly conditions on the riveting structure, including hole fit grade, clamping conditions, assembly deviations, tolerance boundaries, and process limitations. Specifically, hole fit grade reflects the fit between the hole and the rivet; clamping conditions reflect the tightness of the connection area during assembly; assembly deviations and tolerance boundaries reflect the allowable error range in actual manufacturing and assembly; and process limitations reflect the constraints imposed by the processing method, assembly sequence, or manufacturing boundaries.

[0035] In step 12, the component layer information, riveting layer information, and process constraint layer information are processed uniformly and spliced ​​in a fixed order to obtain a structured input vector corresponding to the sample number. This uniform processing refers to converting sample parameters from different sources, types, and formats into a standardized data form suitable for a unified input model. For numerical information, units are standardized to eliminate numerical differences caused by different units of measurement in different samples; for categorical information, category coding is performed according to preset category numbers to convert discontinuous attributes such as material category, arrangement method, and hole fit level into a processable data form; for sequential information, field sorting is performed according to preset field order to ensure consistent arrangement of similar data in different samples; and for missing information, completion or removal is performed according to preset completion rules to reduce interference from abnormal samples in the training process.

[0036] After unified processing, the information is concatenated in the order of component layer information, riveting layer information, and process constraint layer information to obtain a structured input vector. The positions of each field in this structured input vector are predetermined, and different samples have the same technical meaning at the same field position, thus ensuring that the model can establish stable input recognition relationships for similar information during training. In other words, if a fixed position is used to represent rivet spacing, this meaning remains unchanged in all samples, so that the impact on related performance can be uniformly learned in subsequent training.

[0037] In step 13, the overall load-bearing margin, connection durability damage, structural quality, and assembly tolerance sensitivity corresponding to the sample number are obtained and organized into performance supervision labels that correspond one-to-one with the structured input vector in a unified order. Then, the structured input vector corresponding to each sample number is paired with the performance supervision label corresponding to the same sample number to obtain the training dataset. The overall load-bearing margin is used to reflect the load-bearing margin of the riveted structure of the carriage under the target load condition; the connection durability damage is used to reflect the degree of damage to the riveted connection under long-term cyclic load or service conditions; the structural quality is used to reflect the quality index of the corresponding structural scheme; and the assembly tolerance sensitivity is used to reflect the sensitivity of the structural performance to changes in assembly errors or manufacturing deviations.

[0038] The one-to-one correspondence mentioned here means that the structured input vector and performance supervision label under the same sample number must originate from the same riveting structure scheme and the same data standard, avoiding mismatches where the input comes from one scheme but the label comes from another. The training dataset refers to a collection of data consisting of structured input vectors from multiple samples and their corresponding performance supervision labels, used for input-output mapping learning in the subsequent model training phase. By constructing this training dataset, the model can simultaneously learn relevant information for multiple tasks such as load-bearing capacity, durability, quality, and assembly sensitivity within a unified sample system.

[0039] Through the above implementation process, this embodiment completes the standardized organization of the car body riveting structure samples from the original design and process parameters to the training dataset. On the one hand, through the hierarchical collection and unified processing of component layer information, riveting layer information, and process constraint layer information, the input data can more clearly reflect the geometric attributes, connection attributes, and assembly constraint attributes of the car body riveting structure, facilitating subsequent model learning of the correlation between information from different sources. On the other hand, by pairing structured input vectors with multi-class performance supervision labels, a data foundation suitable for multi-task training is established, enabling subsequent models to jointly learn around overall load-bearing margin, connection durability damage, structural quality, and assembly tolerance sensitivity.

[0040] In one embodiment of the present invention, after constructing the structured input vector, the component layer information, riveting layer information, and process constraint layer information in the structured input vector are further combined and calculated to obtain the riveting coupling state vector. Here, the riveting coupling state vector refers to vector data composed of multiple state indicators reflecting the load transfer state, local damage tendency, and mass redundancy relationship in the riveting structure of the carriage. Compared to directly using the original geometric parameters and process parameters, this method does not simply splice the original data, but rather extracts the combined relationships between multiple parameters, enabling subsequent model training to learn around the state changes formed by the combined effects of the parameters, rather than relying solely on the numerical changes of a single field.

[0041] Specifically, in step 21, the structured input vector is first decomposed according to the field order to obtain component layer information, riveting layer information, and process constraint layer information. Based on the component layer information, the plate thickness ratio correction term, effective connection bandwidth, and number of effective load-bearing paths are determined. The plate thickness ratio correction term reflects the influence of different plate thickness combinations on the load transfer state of the connection area; the effective connection bandwidth reflects the effective width of the connection area actually involved in load transfer; and the number of effective load-bearing paths reflects the number of effective paths through which the load can be distributed and transferred in the riveting structure of the carriage. Based on the riveting layer information, the riveting path density, connection path dispersion, average rivet diameter, and number of rivets on the same path are determined. The riveting path density reflects the tightness of the connection path distribution, the connection path dispersion reflects the spatial dispersion of the connection paths, the average rivet diameter reflects the average level of rivet size in the connection area, and the number of rivets on the same path reflects the number of rivets on the same load transfer path. Based on the process constraint layer information, the hole edge geometry sensitivity term, assembly deviation term, and clamping retention term are determined. Among them, the hole edge geometry sensitivity term reflects the influence of the local geometry of the hole edge on the stress sensitivity, the assembly deviation term reflects the influence of assembly errors on the connection state, and the clamping retention term reflects the ability of the connection area to maintain a clamped state after assembly. In addition, the additional mass of the connection is determined based on the component layer information and the riveting layer information to reflect the mass increment brought about by the riveting connection and its local overlapping area.

[0042] In step 22, the product of the riveting path density and the plate thickness ratio correction term is used as the first calculated value, and the sum of the connection path dispersion and a constant is used as the second calculated value. The ratio of the first calculated value to the second calculated value is determined as the load concentration index, which can be expressed as: ,in, Indicates the first The transmission concentration index of each sample; Indicates the first The riveting path density of each sample; Indicates the first The plate thickness ratio correction term for each sample; Indicates the first The discreteness of the connection paths for each sample; the constant 1 in the denominator is used to avoid the denominator being zero and to keep the index calculation stable. The load concentration index is used to reflect the degree of load concentration on the riveting connection paths. Generally speaking, when the riveting paths are denser, the plate thickness combination is more favorable for local load transfer, and the connection path distribution is more concentrated, the index is higher, indicating that the load is more likely to concentrate in a specific connection path. Adding the discreteness of the connection paths to the constant 1 and using it as the denominator helps to avoid the denominator being zero and keeps the calculation process stable.

[0043] In the same step, the product of the average rivet diameter and the number of rivets along the same path is used as the third calculated value, and the ratio of the effective connection bandwidth to the third calculated value is determined as the bypass load tendency index. The bypass load tendency index can be expressed as: ,in, Indicates the first Bypass load tendency index for each sample; Indicates the first The effective connection bandwidth of each sample; Indicates the first The mean diameter of the rivets in each sample; Indicates the first The number of rivets along the same path for each sample. The bypass load tendency index is used to reflect the degree to which loads are likely to bypass the main connection path and be redistributed in the adjacent area. When the effective connection bandwidth is large and the connection density on the same path is relatively limited, the load is more likely to be distributed and transferred to the surrounding area. Therefore, this index can be used to represent the load redistribution trend in the connection area. This index does not directly represent the strength result, but is used to provide load flow state information for subsequent models.

[0044] In step 23, the product of the hole edge geometric sensitivity term and the assembly deviation term is taken as the fourth calculated value, and the sum of the clamping and holding term and constant 1 is taken as the fifth calculated value. The ratio of the fourth calculated value to the fifth calculated value is determined as the local damage sensitivity index. The local damage sensitivity coefficient can be expressed as: ,in, Indicates the first Local damage sensitivity index of each sample; Indicates the first The hole edge geometry sensitivity of each sample; Indicates the first Assembly deviation item for each sample; Indicates the first The compression retention term is used for each sample; the constant 1 in the denominator is used to avoid the denominator being zero. The local damage sensitivity index is used to reflect the tendency of the connection area to generate local damage under the combined effects of the local geometry of the hole edge and assembly deviations. If the local state of the hole edge is more likely to cause stress concentration and the assembly deviation is large, while the compression retention capacity is weak, then the index is high, indicating that the connection area is more prone to local damage. Setting the compression retention term on the denominator side can reflect the inhibitory effect of the compression state on the tendency of local damage.

[0045] Simultaneously, the ratio of the added mass to the number of effective carrying paths is defined as the mass redundancy coupling index. This index reflects the correspondence between the added mass input and the expansion of carrying paths. When the added mass increases significantly while the increase in effective carrying paths is limited, it indicates a high mass cost per unit carrying path; conversely, it indicates that a smaller mass increment can achieve a greater expansion of carrying paths. This index reflects the coupling relationship between mass and redundancy, rather than simply representing the magnitude of mass.

[0046] After obtaining the load concentration index, bypass load tendency index, local damage sensitivity index, and mass redundancy coupling index, these indices are combined in a fixed order to obtain a riveting coupling state vector that corresponds one-to-one with the structured input vector. The fixed order here means that the arrangement of each state index in all samples remains consistent to ensure that each position has a definite meaning during subsequent training; the one-to-one correspondence means that each structured input vector generates a riveting coupling state vector to maintain the consistency of the sample data.

[0047] Through the above implementation process, this embodiment transforms the original component parameters, connection parameters, and process parameters into state data that reflects the internal coupling relationship of the riveted structure of the carriage. This method can consolidate scattered inputs into state indicators that are more suitable for model learning, reduce the model's direct dependence on a single original parameter, improve the consistency of input data and training stability, and thus provide a more suitable data foundation for the joint training of multiple performance tasks.

[0048] In one embodiment of the present invention, after obtaining the structured input vector and the riveting coupling state vector, a riveting coupling state diagram is further constructed based on the geometric and connection relationship data of the riveting structure of the carriage, and shared features are extracted on this basis. Here, the geometric and connection relationship data refers to the basic data used to describe the main components of the carriage, the riveting connection areas, and their spatial connection relationships, including the relative positional relationships between components, the distribution relationship of connection areas, and the possible path relationships of load transmission. The riveting coupling state diagram refers to the data form in which the component objects, connection objects, and their interrelationships in the riveting structure of the carriage are organized into a graph structure, used to incorporate the originally scattered structural parameters, connection parameters, and state indicators into a unified data expression framework.

[0049] Specifically, in step 31, based on the geometric and connection relationship data of the riveting structure of the carriage, the node set and edge set corresponding to the riveting coupling state diagram are determined. The node set includes component nodes and connection nodes, where component nodes represent plates, beams, or other load-bearing components in the main body of the carriage, and connection nodes represent riveting connection areas or local connection units. By simultaneously setting component nodes and connection nodes, the relationship between the main components and connection areas can be reflected in the same diagram structure, rather than treating the entire structure as a single object. The edge set includes adjacent edges, riveting constraint edges, and path load transfer edges. Adjacent edges represent the geometric adjacency relationship between nodes, reflecting the spatial proximity of different components or connection areas; riveting constraint edges represent the constraint relationship formed by the riveting connection, reflecting the connection effect between the component and the connection area; path load transfer edges represent the possible load transfer paths between the component and the connection area, reflecting the load transfer relationship in the structure.

[0050] After determining the node set and edge set, a node attribute matrix corresponding to the node set is further constructed based on the structured input vector and the riveting coupling state vector. Here, the node attribute matrix refers to the attribute data set formed by organizing the input information corresponding to each node according to the node order in the node set, used to provide a computable feature description for each node in the graph. Specifically, the structured input vector mainly provides component layer information, riveting layer information, and process constraint layer information, while the riveting coupling state vector mainly provides state information such as load concentration index, bypass load tendency index, local damage sensitivity index, and mass redundancy coupling index. By using both types of data as node attributes, the nodes in the graph not only possess basic attributes at the geometric and connection levels but also state information extracted from previous steps, enabling subsequent networks to simultaneously learn structural distribution relationships and state association relationships. For example, for two connection nodes located in the same sidewall connection area, if their corresponding local damage sensitivity indices are different, the subsequent shared coding network can use this to distinguish their differences in role within the overall structure.

[0051] In step 32, for any two nodes connected in the edge set, the edge weight is determined based on geometric adjacency strength and assembly discontinuity penalty term. Here, edge weight refers to a quantitative parameter used to measure the tightness of the connection or the strength of the relationship between two nodes. Geometric adjacency strength reflects the proximity of two nodes in terms of spatial location, boundary contact, or connection distance; the closer the nodes are and the clearer the basis of their interaction, the higher their geometric adjacency strength is usually. The assembly discontinuity penalty term reflects the degree of connection weakening caused by factors such as assembly deviations, uneven connections, local misalignments, or process discontinuities; the larger the penalty term, the worse the continuity of the connection relationship in the actual structure. By using the aforementioned geometric adjacency strength as the numerator and the sum of the assembly discontinuity penalty term and a constant as the denominator to determine the edge weight, the stability of the edge weight calculation can be maintained while allowing assembly discontinuities to correct the edge relationship. In other words, when two nodes are geometrically close but the assembly discontinuity is significant, the corresponding edge weight will be reduced, thus more closely reflecting the actual connection state.

[0052] In step 33, a riveting coupling state graph is constructed based on the node set, edge set, node attribute matrix, and edge weights. This riveting coupling state graph is then input into a shared encoding network to obtain shared features that correspond one-to-one with the structured input vector and the riveting coupling state vector. Here, the shared encoding network refers to a common network component set at the front end of the model for unified feature extraction of the riveting coupling state graph. Its input is graph structure data composed of the node set, edge set, node attribute matrix, and edge weights, and its output is shared features that can be used by multiple subsequent performance tasks. The shared encoding network can be a neural network, preferably a neural network suitable for processing graph structure data, such as a graph neural network, graph convolutional network, graph attention network, or other network structures capable of jointly encoding node attributes, connections between nodes, and edge weights.

[0053] The shared features here refer to the intermediate feature results output by the shared coding network, used to represent the distribution of the main components of the carriage, the arrangement of riveted connection areas, the strength of connection relationships between nodes, the load transfer path connections, and the state association information reflected by the riveting coupling state vector. The shared features do not directly correspond to a single performance prediction result, but rather serve as the basic input for multiple subsequent performance tasks, such as overall load-bearing margin, connection durability damage, structural quality, and assembly tolerance sensitivity. In other words, the shared coding network first extracts the structural and connection relationships in the same sample uniformly, and then different subsequent tasks extract information relevant to their respective tasks based on these shared features.

[0054] For example, for the riveted connection area between the side wall of the carriage and the underframe, the shared coding network can combine the corresponding component nodes, connection nodes, adjacent edges, path load edges, and load concentration index and local damage sensitivity index in the node attributes to uniformly encode the connection status and load connection of the area, forming a shared feature that can be called by different performance tasks in the future.

[0055] Through the above implementation process, this embodiment transforms the components, connection areas, and their relationships in the riveted structure of the carriage into unified graph structure data and extracts shared features. This method can organize scattered input information into a data form with relationships, enabling subsequent models to learn the relationship information and transmission connections between nodes. At the same time, it provides a unified source of basic features for different performance tasks, thereby improving the consistency, stability, and overall relationship expression ability of the collaborative optimization model for the performance of the riveted structure of the carriage.

[0056] In one embodiment of the present invention, after obtaining the shared features and the riveting coupling state vector, task features corresponding to each performance task are further obtained based on the shared features and the riveting coupling state vector, and performance prediction values ​​are obtained based on each task feature. Here, task features refer to specific feature information extracted from the shared features and the riveting coupling state vector for a specific performance task, reflecting the structural relationships, connection states, or process influences that the task is more concerned with. Performance prediction values ​​refer to the prediction results output by the model for the corresponding performance task, representing the predicted performance of the riveted structure sample of the carriage in terms of overall load-bearing margin, connection durability damage, structural quality, and assembly tolerance sensitivity. Through this step, the common features obtained in the previous stage can be further converted into specific features for different tasks, thereby meeting the need for joint prediction of multiple performance parameters.

[0057] Specifically, in step 41, task-specific decoupling modules and prediction heads are established, corresponding to overall load-bearing margin, connection durability damage, structural quality, and assembly tolerance sensitivity, respectively. The task-specific decoupling module is a network component placed after shared features to further extract task-related features for a specific performance task. Its function is to further decompose shared features used by multiple performance tasks into task features suitable for the current task. Here, decoupling refers to further distinguishing the differences in the focus of different performance tasks on input information while retaining prior shared information. The prediction head is the output network component corresponding to each task-specific decoupling module, used to convert task features into corresponding performance prediction results. The task-specific decoupling modules and prediction heads are set up one by one in the order of overall load-bearing margin, connection durability damage, structural quality, and assembly tolerance sensitivity to ensure that each performance task has an independent feature extraction path and output path.

[0058] In this embodiment, the task-specific decoupling module corresponding to the overall load-bearing margin is mainly used to extract information related to the overall load-bearing capacity, load-bearing path distribution, and structural continuity; the task-specific decoupling module corresponding to connection durability damage is mainly used to extract information related to local damage tendency, stress concentration in the connection area, and the impact of process deviations; the task-specific decoupling module corresponding to structural quality is mainly used to extract information related to component dimensions, connection-added mass, and redundant configuration; and the task-specific decoupling module corresponding to assembly tolerance sensitivity is mainly used to extract information related to assembly deviations, hole fit levels, and changes in clamping state. By setting the above-mentioned task-specific decoupling modules separately, the information mixing caused by different performance tasks directly sharing the same output path can be avoided, thereby enabling each task to form task features more suitable for its own prediction target based on shared features.

[0059] In step 42, the shared features and riveting coupling state vectors are input into the task-specific decoupling modules respectively to obtain task features corresponding to overall load margin, connection durability damage, structural quality, and assembly tolerance sensitivity, respectively. Then, each task feature is input into its corresponding prediction head to obtain the performance prediction values ​​corresponding to the four performance tasks mentioned above. Here, "input separately" means that for each performance task, the shared features and riveting coupling state vectors of the same sample are simultaneously sent to the task-specific decoupling module corresponding to that task for processing, rather than using different decoupling modules across tasks. By simultaneously introducing shared features and riveting coupling state vectors, the task-specific decoupling modules can utilize both the common relationship information extracted in the preceding shared encoding stage and the state index information contained in the riveting coupling state vectors, thereby forming task features that take into account both common information and task differences.

[0060] The shared features provide fundamental relationship information that multiple performance tasks commonly depend on, such as component distribution relationships, connection strength, and load transfer path connections. The riveting coupling state vector provides state information related to load transfer status, local damage sensitivity, and mass redundancy. The task-specific decoupling module, after jointly processing the two types of information, outputs the task features corresponding to the task. The prediction head then maps these task features to specific performance prediction values. The prediction head can be an output layer or a multi-layer output structure in a neural network, and its output format can be set to scalar values ​​or vector values ​​depending on the task type. In this embodiment, all four performance tasks can use corresponding numerical prediction outputs.

[0061] For example, in a connection durability damage task, the task-specific decoupling module can further extract damage-related information of the connection region by combining the local damage sensitivity index and the bypass load tendency index, based on the connection relationships reflected by the shared features. The corresponding prediction head then outputs the connection durability damage prediction value for that sample. Similarly, in a structural quality task, the task-specific decoupling module can further focus on relevant state information such as connection-added mass and the number of effective load-bearing paths, based on the shared features, and then the corresponding prediction head outputs the structural quality prediction value.

[0062] In step 43, the performance prediction values ​​corresponding to the overall load-bearing margin, connection durability damage, structural quality, and assembly tolerance sensitivity are integrated in the order of overall load-bearing margin, connection durability damage, structural quality, and assembly tolerance sensitivity, resulting in a set of performance prediction values ​​that correspond one-to-one with the shared features and riveting coupling state vectors. This set of performance prediction values ​​refers to the result data set composed of prediction results from multiple performance tasks in a predetermined order, used as the input basis for subsequent loss calculation and model training updates. Integrating in a fixed order ensures that the position and meaning of each performance prediction value in the result set remain consistent, thus facilitating subsequent steps to calculate the loss of each task and the overall training objective according to unified rules.

[0063] Through the above implementation process, this embodiment sets up task-specific decoupling modules and prediction heads on the basis of shared features, enabling the model to output multiple performance prediction results for the same sample simultaneously. This approach not only preserves the structural relationship information shared by different tasks, but also enhances the specificity of feature extraction for each task, reduces information mixing between tasks, and thus provides a stable predictive output foundation for subsequent loss calculation, parameter updates, and performance co-optimization model training.

[0064] In one embodiment of the present invention, after obtaining the performance prediction values ​​corresponding to each performance task, a basic loss is further calculated based on the performance prediction values ​​and performance supervision labels, and a coupling consistency loss is calculated based on the preset correlation between the riveting coupling state vector and the performance prediction values, thereby constructing the overall training objective function. Here, the basic loss refers to the loss value formed by the deviation between the prediction results of each performance task and the corresponding supervision label, used to reflect the degree of fit of the model to each task objective. The coupling consistency loss refers to the constraint loss constructed based on the correlation between each state index and the performance result in the riveting structure of the carriage, used to ensure that the model output not only approaches the supervision label but also conforms to the state correlation logic within the riveting structure. The overall training objective function is used to comprehensively evaluate the overall error level of the model in the current training state and serves as the basis for subsequent parameter updates.

[0065] Specifically, in step 51, the average squared deviation between the predicted performance values ​​and the performance supervision labels corresponding to overall load-bearing margin, connection durability damage, structural quality, and assembly tolerance sensitivity is calculated to obtain the basic loss for each of the four performance tasks. The average squared deviation here refers to the result of averaging the squared differences between the predicted values ​​and the corresponding true labels for each sample in the same task. This method allows samples with larger prediction deviations to have higher weights in the loss calculation, thus prompting the model to pay more attention to the task outputs with larger deviations during training. Subsequently, the basic losses corresponding to overall load-bearing margin, connection durability damage, structural quality, and assembly tolerance sensitivity are weighted and integrated to obtain the basic task loss. This weighted integration refers to merging the losses of different tasks according to pre-set or subsequently adjustable task weights to form a basic loss result representing the overall fit of multiple tasks. This processing allows the model to consider multiple performance objectives simultaneously during the same training process, rather than optimizing only a single performance.

[0066] In step 52, a first consistency constraint loss is constructed based on the predicted performance value corresponding to the structural quality, the load concentration index, and the preset correlation coefficient and state coefficient. Specifically, the product of the predicted performance value corresponding to the structural quality and the first correlation coefficient is calculated first, and then the product of the load concentration index and the state coefficient is calculated. The sum of these two products is used as the constraint benchmark. Then, the difference between the predicted performance value corresponding to the connection durability damage and the constraint benchmark is determined as the first consistency constraint loss after non-negative constraint processing. The preset correlation relationship here refers to the constraint logic pre-set according to the relationship between mass, load concentration state, and durability damage in the riveted structure of the carriage. Non-negative constraint processing means that when the difference is less than zero, it is set to zero, and when the difference is greater than zero, its original value is retained, so that only the part that violates the constraint will generate a loss. In this way, the model output can satisfy the constraint relationship that, under a certain structural quality and load concentration state, the predicted value of connection durability damage should not deviate from the preset benchmark without constraint. That is to say, this consistency constraint is not a replacement for the supervision label, but rather adds a reasonable constraint on the prediction result at the structural state level.

[0067] In step 53, the second consistency constraint loss and the third consistency constraint loss are constructed respectively. For the second consistency constraint loss, the sum of the mass redundancy coupling index and the constant is first used as the denominator, and then the ratio of the second correlation coefficient to the denominator is calculated as the first constraint benchmark. Then, the difference between the change in the performance prediction value corresponding to the overall load-bearing margin and the first constraint benchmark is processed by non-negativity constraint and determined as the second consistency constraint loss. Here, the change in the performance prediction value corresponding to the overall load-bearing margin refers to the predicted increment or relative change of the current sample in the direction of the load-bearing margin, which is used to reflect the changing trend of load-bearing performance with the change of structural redundancy relationship. This constraint is used to limit the predicted range of the load-bearing performance improvement when the mass redundancy relationship changes, so that the model output is more consistent with the correspondence between mass increase and load-bearing path expansion in structural design.

[0068] For the third consistency constraint loss, the product of the local damage sensitivity index and the third correlation coefficient is first used as the second constraint benchmark. Then, the difference between this second constraint benchmark and the performance prediction value corresponding to the assembly tolerance sensitivity is processed by non-negativity constraint and determined as the third consistency constraint loss. This constraint is used to reflect the correlation between the local damage sensitivity state and the assembly tolerance sensitivity result. When the local damage sensitivity index is high, the assembly tolerance sensitivity prediction value output by the model should also maintain a corresponding relationship; otherwise, a constraint loss will occur. By setting this consistency constraint, the model can consider the impact of local state changes on the assembly sensitivity result during training, rather than relying solely on the surface correspondence in the data.

[0069] For example, in two training samples, if their structural masses are similar, but one sample has a higher load concentration index, its predicted connection durability damage value should generally not decrease arbitrarily, deviating from this state difference. Similarly, when the mass redundancy coupling index is large, the predicted increment of the overall load-bearing margin should generally not increase indefinitely. These constraints are not intended to arbitrarily specify specific performance results, but rather to limit the model output to maintain a trend consistent with changes in the structural state.

[0070] In step 54, the first consistency constraint loss, the second consistency constraint loss, and the third consistency constraint loss are weighted and integrated to obtain the coupling consistency loss. Then, the basic task loss is added to the coupling consistency loss to obtain the overall training objective function. This weighted integration refers to combining different consistency constraint losses according to preset weights to reflect the importance of different constraint relationships to model training. The overall training objective function is the comprehensive optimization objective of the model training phase, which includes both the requirement for fitting the supervision label and the constraint requirements for the internal state relationships of the riveting structure. After this construction, model training no longer only pursues numerical closeness between the predicted performance values ​​and the supervision label, but also requires that the predicted results satisfy the correlation logic related to the riveting coupling state vector.

[0071] Through the above implementation process, this embodiment incorporates both the basic task loss and the coupling consistency loss into the overall training objective function. This ensures that model training considers both the prediction accuracy of multiple performance tasks and the correlation and coordination between different state indicators and performance results. This reduces the likelihood of the model relying solely on sample fitting while ignoring structural state relationships, improving the consistency and stability of training results. Consequently, it provides a more reliable objective function foundation for subsequent training of models for task conflict control and collaborative optimization of carriage riveting structure performance.

[0072] In one embodiment of the present invention, after constructing the overall training objective function, gradients are further calculated based on the base losses corresponding to each performance task. Task conflicts are identified and their intensity is determined. Task weights are dynamically adjusted according to the conflict intensity, and the parameters of the shared encoding network, task-specific decoupling modules, and prediction heads are updated based on the adjusted weights until the overall training objective function converges. Here, the task gradient refers to the direction and magnitude of change of the base loss corresponding to a certain performance task relative to the shared encoding network parameters, reflecting the trend of the task's influence on the shared parameter updates. Task conflict refers to the inconsistent update directions proposed by different performance tasks for the shared encoding network parameters, which may cause the optimization process of one task to weaken the learning effect of another task. The task conflict intensity is used to quantify the degree of conflict between a certain performance task and other tasks.

[0073] Specifically, in step 61, the parameters of the shared coding network are differentiated based on the basic loss corresponding to the overall load-bearing margin, the basic loss corresponding to the connection durability impairment, the basic loss corresponding to the structural quality, and the basic loss corresponding to the assembly tolerance sensitivity, to obtain the task gradients corresponding to each performance task. Differentiation of the shared coding network parameters is performed here because the shared coding network is a common feature extraction part used by multiple performance tasks, and its parameter update direction best reflects whether there is a relationship of mutual benefit or mutual interference between tasks. Then, for any two performance tasks, the ratio of their inner product to their norm product is calculated to obtain the task conflict coefficient between any two performance tasks. This task conflict coefficient reflects the degree of consistency between the gradient directions of the two tasks. When the coefficient is greater than zero, it indicates that the two tasks have relatively consistent update directions in the current training state; when the coefficient is near zero, it indicates that the update directions of the two tasks are basically unrelated; when the coefficient is less than zero, it indicates that the update directions of the two tasks are opposite, which is identified as a task conflict. In this way, whether different performance tasks interfere with each other can be transformed into a calculable quantitative result.

[0074] In step 62, based on the task conflict coefficient between any two performance tasks, the negative task conflict coefficients between each performance task and the remaining performance tasks are summarized to obtain the task conflict intensity corresponding to each performance task. This summarization refers to, for a given performance task, counting all task conflict coefficients less than zero between it and all other performance tasks, and forming the conflict intensity result corresponding to that task. A higher task conflict intensity indicates a higher degree of inconsistency in updates between that task and other tasks in the current training round, making it more prone to mutual interference during shared parameter updates. Subsequently, the task weights corresponding to each performance task are dynamically adjusted based on the task conflict intensity. This dynamic adjustment means that the task weights are no longer fixed, but change with the changes in the conflict state of each task during training, ensuring that tasks with strong task conflicts or significant changes in training state receive weight allocations that match their current state during parameter updates. In this way, the training process can respond to changes in the relationships between different tasks, thereby reducing the task imbalance problem that easily occurs under a fixed weight model.

[0075] In step 63, for performance tasks with conflicting tasks, the corresponding task gradients are projected and corrected to obtain corrected task gradients. Projection correction here refers to adjusting the components in a task gradient that conflict with the gradients of conflicting tasks, thereby reducing mutual interference between different tasks on the update of shared coding network parameters. This process reduces the negative impact of a task on other tasks without completely discarding its learning information. Subsequently, the corrected task gradients are weighted and integrated based on the adjusted task weights to obtain aggregated gradients. Aggregated gradients refer to the basis for updating shared parameters by combining the corrected gradients of multiple performance tasks with the current task weights. Then, the shared coding network parameters are updated based on the aggregated gradients; simultaneously, the task-specific decoupling module parameters and prediction head parameters for each performance task are updated based on the adjusted task weights and the overall training objective function. In other words, the shared coding network parameters are mainly updated based on the aggregated gradients to ensure that the common feature extraction part takes into account the relationships between multiple tasks; while the task-specific decoupling module parameters and prediction head parameters are updated based on the overall training objective function and the corresponding task weights to ensure continuous optimization of the prediction capabilities of each task.

[0076] The shared encoding network parameters, task-specific decoupling module parameters, and prediction head parameters here correspond to the common feature extraction part, task feature extraction part, and task output part in the model, respectively. By differentiating the update methods of different parameter parts, the shared part and the task-specific part can play different roles in training: the shared part focuses more on multi-task coordination, while the task-specific part focuses more on the prediction accuracy of individual performance metrics.

[0077] In step 64, the overall training objective function is recalculated based on the updated shared coding network parameters, the task-specific decoupling module parameters corresponding to each performance task, and the prediction head parameters. If the change in the overall training objective function within a consecutive preset training epochs is less than a preset threshold, the update stops. If the change in the overall training objective function within a consecutive preset training epochs is greater than or equal to the preset threshold, the process returns to step 61 to continue execution. Convergence of the overall training objective function here means that, within a certain number of consecutive training epochs, the change in the overall training objective function remains within the preset threshold, indicating that the model parameter updates have stabilized. The preset training epochs and preset threshold can be set according to the actual sample size, training accuracy requirements, and computational resources. This iterative judgment mechanism avoids premature training termination and also avoids ineffective updates when the objective function has basically stabilized.

[0078] Through the above implementation process, this embodiment combines task gradient calculation, task conflict identification, conflict intensity quantification, dynamic adjustment of task weights, gradient projection correction, and parameter iterative update, enabling the model to dynamically adjust the parameter update strategy based on the actual relationship between multiple performance tasks during training. This reduces mutual interference between different performance tasks during shared parameter updates, improving the training stability of the common feature extraction part. Furthermore, it allows each task to retain an update path suitable for its own objective while sharing learning, thus providing support for the formation of stable training results for the subsequent collaborative optimization model of the carriage riveting structure performance.

[0079] In one embodiment of the present invention, after the overall training objective function converges, a collaborative optimization model for the riveting structure performance of the carriage is further obtained, and a collaborative optimization evaluation value is determined based on the performance prediction values ​​corresponding to the candidate carriage riveting structure schemes. Here, the collaborative optimization model for the carriage riveting structure performance refers to a model that, after completing parameter training, can receive the original design and process parameters of the candidate carriage riveting structure schemes and output multiple performance prediction results and a comprehensive evaluation result. The collaborative optimization evaluation value refers to the evaluation result calculated based on multiple performance prediction values, used to reflect the comprehensive balance of the candidate carriage riveting structure schemes in terms of overall load-bearing capacity, connection durability, structural quality, and assembly tolerance adaptability.

[0080] Specifically, in step 71, after the overall training objective function converges, the shared coding network parameters, the task-specific decoupling module parameters corresponding to each performance task, and the prediction head parameters corresponding to each performance task are read. The generation link of the structured input vector, the riveting coupling state vector, the shared features, and the performance prediction value set is then solidified to obtain the collaborative optimization model for the riveting structure performance of the carriage. Solidifying the generation link means fixing the input organization method, state quantity calculation method, graph structure processing method, shared feature extraction method, and multi-task prediction method determined in the training phase, so that subsequent candidate solutions can be processed in the inference phase according to the same order and rules as in the training phase. The resulting collaborative optimization model not only retains the parameter results obtained from training but also retains the complete processing flow from the original design and process parameters to the final performance prediction results, thus ensuring consistency in data processing between the training and application phases.

[0081] The shared encoding network parameters here refer to the parameter results obtained after the common feature extraction part has converged during training; the task-specific decoupling module parameters refer to the parameter results obtained after the task feature extraction part corresponding to each performance task has converged during training; and the prediction head parameters refer to the parameter results obtained after the output part of each performance task has converged during training. By solidifying the above parameters together with the preceding processing steps, the model can directly output multiple performance prediction results without retraining when faced with new candidate solutions.

[0082] In step 72, the original design and process parameters of the candidate car body riveting structure schemes are obtained, and processed sequentially according to the generation order of the structured input vector, riveting coupling state vector, shared features, and performance prediction value set. This yields the performance prediction values ​​corresponding to the overall load-bearing margin, connection durability damage, structural quality, and assembly tolerance sensitivity of the candidate car body riveting structure schemes. Here, the candidate car body riveting structure scheme refers to the car body riveting structure design schemes to be compared, screened, or optimized; it can be any one of the multiple alternative schemes generated during the design phase. Sequential processing means that the original parameters of the candidate schemes are first organized into a structured input vector according to the aforementioned rules, then a riveting coupling state vector is generated, shared features are extracted based on this, and the corresponding performance prediction values ​​are output through the task-specific decoupling module and prediction head. This processing order is consistent with the training phase to ensure that the input data of the candidate schemes undergo the same processing logic before participating in performance evaluation.

[0083] In step 73, the predicted performance values ​​corresponding to overall load-bearing margin, connection durability damage, structural quality, and assembly tolerance sensitivity are normalized. Then, based on the normalized predicted performance values ​​and their corresponding output evaluation weights, a weighted integration is performed to obtain the collaboratively optimized evaluation value. The normalization process here refers to mapping different performance prediction values ​​to a unified numerical range to eliminate differences in dimensions and numerical scales between different indicators, facilitating subsequent comprehensive evaluation. The output evaluation weights are parameters pre-set according to actual design requirements, performance priorities, or application scenarios, indicating the degree of influence of each performance component in the comprehensive evaluation. For example, in scenarios where lightweight design is prioritized, the evaluation weight corresponding to structural quality can be appropriately increased; in scenarios where long-term service stability is prioritized, the evaluation weight corresponding to connection durability damage can be appropriately increased.

[0084] The weighted and integrated collaborative optimization evaluation value can be used as a basis for ranking, screening, or further optimizing candidate car body riveting structure schemes. This collaborative optimization evaluation value is not a direct replacement of a single performance index, but rather a comprehensive reflection of the balance between multiple performance aspects of the candidate scheme. In other words, even if a candidate scheme is not optimal in a single performance aspect, it can still obtain a high collaborative optimization evaluation value as long as it has a more reasonable comprehensive matching relationship between overall load-bearing capacity, connection durability, structural quality, and assembly tolerance sensitivity.

[0085] Through the above implementation process, this embodiment solidifies the trained parameter results and processing links into a collaborative optimization model for the performance of the riveted car body structure. This model is then used to perform multiple performance predictions and comprehensive evaluations of candidate riveted car body structure schemes. On the one hand, this allows the model to directly serve the candidate scheme selection and design decision-making process, improving the efficiency of riveted car body structure optimization applications. On the other hand, by uniformly normalizing and weighting the multiple performance prediction values, the applicability of candidate schemes can be evaluated from a comprehensive perspective, avoiding judgments based solely on a single performance indicator. This provides an executable evaluation basis for the collaborative optimization of riveted car body structure schemes.

[0086] This embodiment provides a training system for a collaborative optimization model of the performance of riveted car body structures based on multi-task learning, including:

[0087] The sample construction module obtains the original design and process parameters of multiple car body riveting structure samples, and obtains structured input vectors and performance supervision labels; wherein, the structured input vectors include component layer information, riveting layer information and process constraint layer information;

[0088] The state generation module performs combined calculations based on the component layer information, riveting layer information, and process constraint layer information in the structured input vector to obtain the riveting coupling state vector;

[0089] The graph encoding module constructs a riveting coupling state graph and extracts shared features based on the structured input vector, the riveting coupling state vector, and the basic data of the geometric and connection relationships of the riveting structure of the carriage.

[0090] The task prediction module obtains the task features corresponding to each performance task based on shared features and riveted coupling state vectors, and obtains the performance prediction value based on each task feature.

[0091] The objective construction module calculates the basic loss based on the performance prediction value and the performance supervision label, calculates the coupling consistency loss based on the preset correlation between the riveted coupling state vector and the performance prediction value, and constructs the overall training objective function;

[0092] The conflict control module calculates the gradient based on the base loss corresponding to each performance task, identifies task conflicts and determines the intensity of task conflicts, dynamically adjusts the task weights according to the intensity of task conflicts, and updates the parameters of the shared encoding network, task-specific decoupling module and prediction head based on the adjusted task weights until the overall training objective function converges.

[0093] The optimization evaluation module obtains a collaborative optimization model for the performance of the riveting structure of the carriage after the overall training objective function converges, and determines the collaborative optimization evaluation value based on the performance prediction value corresponding to the candidate carriage riveting structure scheme.

[0094] It should be noted that the interval and threshold sizes are set for ease of comparison. The size of the threshold depends on the amount of sample data and the base number set by those skilled in the art for each set of sample data, as long as it does not affect the proportional relationship between the parameter and the quantized value. Furthermore, the above formulas are all dimensionless calculations, and the formulas are derived from software simulations using a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0095] The embodiments of the present invention have been described above, but the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of the present embodiments, all of which are within the protection scope of the present embodiments.

Claims

1. A training method for a collaborative optimization model of the performance of a car body riveting structure based on multi-task learning, characterized in that, Includes the following steps: Step 1: Obtain the original design and process parameters of multiple car body riveting structure samples to obtain structured input vectors and performance supervision labels; wherein, the structured input vectors include component layer information, riveting layer information and process constraint layer information; Step 2: Based on the component layer information, riveting layer information and process constraint layer information in the structured input vector, perform combined calculations to obtain the riveting coupling state vector; Step 3: Based on the structured input vector, the riveting coupling state vector, and the basic data on the geometric and connection relationships of the riveting structure of the carriage, construct the riveting coupling state diagram and extract shared features, including: Step 31: Based on the basic data of the geometric and connection relationships of the riveting structure of the carriage, determine the node set and edge set corresponding to the riveting coupling state diagram; wherein, the node set includes component nodes representing the main components of the carriage and connection nodes representing the riveting connection area, and the edge set includes adjacent edges representing geometric adjacency relationships, riveting constraint edges representing riveting connection relationships, and path transmission edges representing load transmission path relationships; then, construct the node attribute matrix corresponding to the node set based on the structured input vector and the riveting coupling state vector; Step 32: For any two nodes in the edge set that have a connection relationship, determine the edge weight based on the geometric adjacency strength and the assembly discontinuity penalty term; wherein, the geometric adjacency strength is used as the numerator, the sum of the assembly discontinuity penalty term and the constant is used as the denominator, and the ratio of the numerator to the denominator is determined as the edge weight. Step 33: Construct a riveting coupling state graph based on the node set, edge set, node attribute matrix and edge weight, and input the riveting coupling state graph into the shared coding network to obtain shared features that correspond one-to-one with the structured input vector and the riveting coupling state vector; Step 4: Based on the shared features and the riveted coupling state vector, obtain the task features corresponding to each performance task, and obtain the performance prediction value based on each task feature; Step 5: Calculate the basic loss for each performance task based on the performance prediction value and the performance supervision label, calculate the coupling consistency loss based on the preset correlation between the riveted coupling state vector and the performance prediction value, and construct the overall training objective function. Step 6: Calculate the gradient based on the basic loss corresponding to each performance task, identify task conflicts and determine the intensity of task conflicts, dynamically adjust the task weights according to the intensity of task conflicts, and update the parameters of the shared encoding network, task-specific decoupling module and prediction head based on the adjusted task weights until the overall training objective function converges. Step 7: After the overall training objective function converges, the collaborative optimization model of the car body riveting structure performance is obtained, and the collaborative optimization evaluation value is determined based on the performance prediction value corresponding to the candidate car body riveting structure scheme.

2. The training method for the collaborative optimization model of carriage riveting structure performance based on multi-task learning according to claim 1, characterized in that, Obtain the original design and process parameters of multiple car body riveting structure samples to obtain structured input vectors and performance supervision labels, including: Step 11: Collect the original design and process parameters of multiple car body riveting structure samples according to the sample number, and divide them according to the main car body components, riveting connections and process constraints represented by the parameters to obtain component layer information, riveting layer information and process constraint layer information corresponding to the sample number; wherein, the component layer information includes plate thickness, beam size, material type, component spatial position and geometric contour of connection area; the riveting layer information includes rivet diameter, rivet spacing, edge distance, arrangement method, number of riveting layers and distribution of riveting area; the process constraint layer information includes hole fit level, clamping conditions, assembly deviation, tolerance boundary and process restriction conditions; Step 12: Perform unified processing on the component layer information, riveting layer information, and process constraint layer information; the unified processing includes: unifying the unit for numerical information, performing category coding according to the preset category number for categorical information, performing field sorting according to the preset field order for sequential information, and performing completion or removal of missing information according to the preset completion rules; then concatenate the component layer information, riveting layer information, and process constraint layer information in the order to obtain the structured input vector corresponding to the sample number; Step 13: Obtain the overall load margin, connection durability damage, structural quality, and assembly tolerance sensitivity corresponding to the sample number, and organize them into performance supervision labels that correspond one-to-one with the structured input vector in the order of overall load margin, connection durability damage, structural quality, and assembly tolerance sensitivity; then pair the structured input vector corresponding to each sample number with the performance supervision label corresponding to the same sample number to obtain the training dataset.

3. The training method for the collaborative optimization model of carriage riveting structure performance based on multi-task learning according to claim 1, characterized in that, Based on the component layer information, riveting layer information, and process constraint layer information in the structured input vector, a combined calculation is performed to obtain the riveting coupling state vector, including: Step 21: Based on the structured input vector decomposition, obtain component layer information, riveting layer information, and process constraint layer information. Based on the component layer information, determine the plate thickness ratio correction term, effective connection bandwidth, and effective load-bearing path number. Based on the riveting layer information, determine the riveting path density, connection path dispersion, rivet diameter mean, and number of rivets on the same path. Based on the process constraint layer information, determine the hole edge geometry sensitivity term, assembly deviation term, and clamping retention term. Finally, determine the connection additional mass based on the component layer information and the riveting layer information. Step 22: The product of the riveting path density and the plate thickness ratio correction term is used as the first calculated value; the sum of the connection path dispersion and constant 1 is used as the second calculated value; the ratio of the first calculated value to the second calculated value is determined as the load concentration index; the product of the average rivet diameter and the number of rivets in the same path is used as the third calculated value; the ratio of the effective connection bandwidth to the third calculated value is determined as the bypass load tendency index. Step 23: The product of the hole edge geometric sensitivity term and the assembly deviation term is taken as the fourth calculated value, and the sum of the clamping and holding term and constant 1 is taken as the fifth calculated value. The ratio of the fourth calculated value to the fifth calculated value is determined as the local damage sensitivity index. The ratio of the connection additional mass to the number of effective load-bearing paths is determined as the mass redundancy coupling index. Then, the load concentration index, bypass load tendency index, local damage sensitivity index and mass redundancy coupling index are combined in a fixed order to obtain the riveting coupling state vector that corresponds one-to-one with the structured input vector.

4. The training method for the collaborative optimization model of carriage riveting structure performance based on multi-task learning according to claim 1, characterized in that, Based on shared features and riveted coupling state vectors, task features corresponding to each performance task are obtained, and performance prediction values ​​are obtained based on each task feature, including: Step 41: Establish task-specific decoupling modules and prediction heads corresponding to overall load-bearing margin, connection durability damage, structural quality, and assembly tolerance sensitivity, respectively; wherein, the task-specific decoupling modules and prediction heads are set one by one in the order of overall load-bearing margin, connection durability damage, structural quality, and assembly tolerance sensitivity. Step 42: Input the shared features and riveting coupling state vector into the decoupling module of each task respectively to obtain the task features corresponding to the overall load margin, connection durability damage, structural quality and assembly tolerance sensitivity respectively; then input each task feature into the corresponding prediction head to obtain the performance prediction value corresponding to the overall load margin, connection durability damage, structural quality and assembly tolerance sensitivity respectively. Step 43: Integrate the performance prediction values ​​corresponding to the overall load-bearing margin, the performance prediction values ​​corresponding to the connection durability damage, the performance prediction values ​​corresponding to the structural quality, and the performance prediction values ​​corresponding to the assembly tolerance sensitivity in the order of overall load-bearing margin, connection durability damage, structural quality, and assembly tolerance sensitivity to obtain a set of performance prediction values ​​that correspond one-to-one with the shared features and the riveting coupling state vector.

5. The training method for the collaborative optimization model of carriage riveting structure performance based on multi-task learning according to claim 1, characterized in that, The basic loss for each performance task is calculated based on the predicted performance value and the performance supervision label. The coupling consistency loss is calculated based on the pre-defined correlation between the riveted coupling state vector and the predicted performance value. Finally, the overall training objective function is constructed, including: Step 51: Calculate the average squared deviation between the predicted performance values ​​and the performance monitoring labels corresponding to the overall load-bearing margin, connection durability damage, structural quality, and assembly tolerance sensitivity, respectively, to obtain the basic loss corresponding to the overall load-bearing margin, the basic loss corresponding to the connection durability damage, the basic loss corresponding to the structural quality, and the basic loss corresponding to the assembly tolerance sensitivity; then, perform weighted integration of the above basic losses to obtain the basic task loss. Step 52: Calculate the product of the performance prediction value corresponding to the structural quality and the first correlation coefficient, and the product of the load concentration index and the state coefficient. Use the sum of the two products as the constraint benchmark. After processing the performance prediction value corresponding to the connection durability damage and the constraint benchmark with non-negative constraints, determine the first consistency constraint loss. Step 53: Using the sum of the mass redundancy coupling index and constant 1 as the denominator, calculate the ratio of the second correlation coefficient to the denominator as the first constraint benchmark. The difference between the performance prediction value change corresponding to the overall load-bearing margin and the first constraint benchmark is determined as the second consistency constraint loss after non-negative constraint processing. The product of the local damage sensitivity index and the third correlation coefficient is used as the second constraint benchmark. The difference between the second constraint benchmark and the performance prediction value corresponding to the assembly tolerance sensitivity is determined as the third consistency constraint loss after non-negative constraint processing. Step 54: Weighted integration of the first consistency constraint loss, the second consistency constraint loss, and the third consistency constraint loss to obtain the coupling consistency loss; then add the basic task loss to the coupling consistency loss to obtain the overall training objective function.

6. The training method for the collaborative optimization model of carriage riveting structure performance based on multi-task learning according to claim 1, characterized in that, The gradient is calculated based on the base loss corresponding to each performance task. Task conflicts are identified and their intensity is determined. Task weights are dynamically adjusted according to the conflict intensity. The parameters of the shared encoder network, task-specific decoupling module, and prediction head are updated based on the adjusted task weights until the overall training objective function converges. This process includes: Step 61: Based on the basic loss corresponding to the overall load-bearing margin, the basic loss corresponding to the connection durability damage, the basic loss corresponding to the structural quality, and the basic loss corresponding to the assembly tolerance sensitivity, the parameters of the shared coding network are differentiated to obtain the task gradient corresponding to each performance task; then, for any two performance tasks, the ratio of their inner product to their norm product is calculated to obtain the task conflict coefficient between any two performance tasks, and a task conflict is identified when the task conflict coefficient is less than zero. Step 62: Based on the task conflict coefficient between any two performance tasks, summarize the negative task conflict coefficients between each performance task and the other performance tasks to obtain the task conflict intensity corresponding to each performance task; then dynamically adjust the task weight corresponding to each performance task based on the task conflict intensity corresponding to each performance task. Step 63: For performance tasks with task conflicts, project and correct the corresponding task gradients to obtain corrected task gradients; then, weight and integrate the corrected task gradients based on the adjusted task weights to obtain aggregated gradients; and update the shared coding network parameters based on the aggregated gradients. For each performance task, update the task-specific decoupling module parameters and prediction head parameters corresponding to that performance task based on the adjusted task weights and the overall training objective function. Step 64: Recalculate the overall training objective function based on the updated shared coding network parameters, the task-specific decoupling module parameters corresponding to each performance task, and the prediction head parameters; if the change in the overall training objective function within consecutive preset training rounds is less than the preset threshold, then stop updating; if the change in the overall training objective function within consecutive preset training rounds is greater than or equal to the preset threshold, then return to step 61 to continue execution.

7. The training method for the collaborative optimization model of carriage riveting structure performance based on multi-task learning according to claim 1, characterized in that, After the overall training objective function converges, a collaborative optimization model for the riveting structure performance of the carriage is obtained. Based on the performance prediction values ​​corresponding to the candidate carriage riveting structure schemes, the collaborative optimization evaluation value is determined, including: Step 71: After the overall training objective function converges, read the shared coding network parameters, the task-specific decoupling module parameters corresponding to each performance task, and the prediction head parameters corresponding to each performance task, and solidify the generation link of the structured input vector, the riveting coupling state vector, the shared features, and the performance prediction value set to obtain the collaborative optimization model of the riveting structure performance of the carriage. Step 72: Obtain the original design and process parameters of the candidate car body riveting structure scheme, and process them in the order of generating structured input vector, riveting coupling state vector, shared features and performance prediction value set to obtain the overall load margin, connection durability damage, structural quality and assembly tolerance sensitivity corresponding to the candidate car body riveting structure scheme. Step 73: Normalize the performance prediction values ​​corresponding to overall load-bearing margin, connection durability damage, structural quality, and assembly tolerance sensitivity respectively; then, based on the normalized performance prediction values ​​corresponding to overall load-bearing margin, connection durability damage, structural quality, and assembly tolerance sensitivity, and the corresponding output evaluation weights, perform weighted integration to obtain the collaborative optimization evaluation value.

8. A training system for a collaborative optimization model of carriage riveting structure performance based on multi-task learning, characterized in that, The training method for the collaborative optimization model of the performance of the riveting structure of the carriage based on multi-task learning as described in any one of claims 1-7 includes: The sample construction module obtains the original design and process parameters of multiple car body riveting structure samples, and obtains structured input vectors and performance supervision labels; wherein, the structured input vectors include component layer information, riveting layer information and process constraint layer information; The state generation module performs combined calculations based on the component layer information, riveting layer information, and process constraint layer information in the structured input vector to obtain the riveting coupling state vector; The graph encoding module constructs a riveting coupling state graph and extracts shared features based on the structured input vector, the riveting coupling state vector, and the basic data of the geometric and connection relationships of the riveting structure of the carriage. The task prediction module obtains the task features corresponding to each performance task based on shared features and riveted coupling state vectors, and obtains the performance prediction value based on each task feature. The objective construction module calculates the basic loss based on the performance prediction value and the performance supervision label, calculates the coupling consistency loss based on the preset correlation between the riveted coupling state vector and the performance prediction value, and constructs the overall training objective function; The conflict control module calculates the gradient based on the base loss corresponding to each performance task, identifies task conflicts and determines the intensity of task conflicts, dynamically adjusts the task weights according to the intensity of task conflicts, and updates the parameters of the shared encoding network, task-specific decoupling module and prediction head based on the adjusted task weights until the overall training objective function converges. The optimization evaluation module obtains a collaborative optimization model for the performance of the riveting structure of the carriage after the overall training objective function converges, and determines the collaborative optimization evaluation value based on the performance prediction value corresponding to the candidate carriage riveting structure scheme.

Images