A method and device for multi-step quality prediction of a plastic forming process based on gradient boosting decision trees

Abstract

The application discloses a plastic forming process multi-step quality prediction method and equipment based on a gradient boosting decision tree, and relates to the technical field of ring rolling. The application constructs a size parameter prediction model representing a nonlinear mapping relationship from a pre-rolling state (i.e. an input state sequence) to future geometric size evolution based on a gradient boosting decision tree model, realizes dynamic prediction of sizes at future time steps by using the size parameter prediction model, and introduces coding of a target prediction step, so that the size parameter prediction model can distinguish prediction tasks at different future time steps and improve prediction accuracy. The application realizes real-time, accurate and stable dynamic prediction of the geometric size of a ring during rolling.

Description

Technical Field

[0001] This application relates to the field of ring rolling technology, and in particular to a multi-step quality prediction method and equipment for plastic forming process based on gradient boosting decision tree. Background Technology

[0002] Ring rolling is a key plastic forming process in the manufacture of rotating parts such as large bearing rings, flange rings, aerospace rings, and wind power and energy equipment rings. This process is typically carried out under high temperature and large deformation conditions, where the material undergoes continuous diameter expansion, wall thickness reduction, and height changes under the combined action of multiple actuators, including the main roll, mandrel, tapered roll, and guide roll. For ring-rolled products, outer diameter, height, and wall thickness are the most direct and critical quality indicators. Their accuracy not only determines whether the part meets the final tolerance requirements but also directly affects subsequent machining allowances, material utilization, scrap rate, and production costs. Therefore, how to perform real-time, accurate, and stable dynamic prediction of the ring's geometric dimensions during the rolling process is a key issue in achieving online quality control and digital closed-loop regulation of the ring rolling process. Summary of the Invention

[0003] The purpose of this application is to provide a multi-step quality prediction method and equipment for plastic forming process based on gradient boosting decision tree, so as to realize real-time, accurate and stable dynamic prediction of the geometric dimensions of ring parts during rolling.

[0004] To achieve the above objectives, this application provides the following solution.

[0005] In a first aspect, this application provides a multi-step quality prediction method for plastic forming processes based on gradient boosting decision trees, including: The input state sequence is obtained based on the input window; the input state sequence includes time step t and the ring rolling state vector of L-1 time steps before time step t; the ring rolling state vector includes process state variables, equipment operation variables and geometric dimension variables; where t is the prediction starting point and L is the length of the input window; The target prediction step is encoded to construct the input state sequence as the structured input feature of the target prediction step; the target prediction step is any time step within the prediction window, and the prediction window includes K time steps after time step t, where K is the length of the prediction window; The structured input features are input into the size parameter prediction model to obtain the predicted value of the size parameter for the target prediction step; the size parameter prediction model is obtained by training a gradient boosting decision tree model.

[0006] Optionally, the structured input features are: ; in, This represents the structured input features at the h-th time step within the prediction window. Represents the input state sequence. This represents the encoding vector at the h-th time step within the prediction window. Indicates the flattening operation. This indicates a feature splicing operation.

[0007] Optionally, the loss function used in training the gradient boosting decision tree model is: ; in, For loss function, Represents the actual value of the dimensional parameter Between and predicted values The residual loss function is given by N, where N is the number of training samples in a single batch, and i represents the i-th training sample. This represents the total number of regression trees. Indicates the first The number of leaf nodes in a tree. This represents the penalty coefficient for the number of leaf nodes, used to limit excessive complexity in the tree structure. This represents the regularization coefficient for the leaf node weights. Indicates the first The first of the trees The prediction weights of each leaf node.

[0008] Optionally, the ratio of the length of the input window to the length of the prediction window is 4:1.

[0009] Optionally, the size parameter prediction model includes: an outer diameter prediction model, a height prediction model, and a wall thickness prediction model; The dimensional parameters include outer diameter, height, and wall thickness.

[0010] Optionally, the outer diameter prediction model is also used to analyze the importance of each variable in the structured input features relative to the outer diameter prediction, and to determine the key variables whose contribution to the outer diameter prediction is greater than the first contribution threshold, as the key variables corresponding to the outer diameter. The height prediction model is also used to analyze the importance of each variable in the structured input features relative to the height prediction, and to determine the key variables whose contribution to the height prediction is greater than the second contribution threshold, as the key variables corresponding to the height. The wall thickness prediction model is also used to analyze the importance of each variable in the structured input features relative to the wall thickness prediction, and to identify key variables whose contribution to the wall thickness prediction is greater than the third contribution threshold, as the key variables corresponding to the wall thickness.

[0011] Optionally, the structured input features are input into the size parameter prediction model to obtain the predicted size parameters for the target prediction step, and the method further includes: Calculate the prediction deviation between the predicted values ​​of the size parameters of the target prediction step and the target values; A quality warning is triggered when the predicted deviation of the dimensional parameters is outside the preset range.

[0012] Optionally, the prediction deviation of the dimensional parameters not being within the preset range means that at least one of the prediction deviation of the outer diameter, the prediction deviation of the height, and the prediction deviation of the wall thickness is not within the corresponding preset range.

[0013] Optionally, a quality warning is triggered when the predicted deviation of the dimensional parameters is outside the preset range, followed by: The size parameters whose prediction deviation is outside the corresponding preset range are identified as the target size parameters. Identify the causes of quality influencing factors based on the key variables corresponding to the target size parameters.

[0014] Secondly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described multi-step quality prediction method for plastic forming process based on gradient boosting decision tree.

[0015] According to the specific embodiments provided in this application, this application has the following technical effects.

[0016] This application provides a multi-step quality prediction method and device for plastic forming process based on gradient boosting decision tree. This application constructs a size parameter prediction model based on gradient boosting decision tree model to represent the nonlinear mapping relationship from the pre-rolling state (i.e., input state sequence) to the future geometric dimension evolution. This size parameter prediction model is used to realize dynamic prediction of dimensions in multiple future time steps. The encoding of the target prediction step is introduced to enable the size parameter prediction model to distinguish the prediction tasks of different future time steps and improve the prediction accuracy. This application realizes real-time, accurate and stable dynamic prediction of ring geometry during rolling process. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1This is a flowchart illustrating a multi-step quality prediction method for plastic forming process based on gradient boosting decision tree, provided as an embodiment of this application.

[0019] Figure 2 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0020] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0022] In related ring forging technologies, some patents mainly focus on pre-rolling process design and blank size determination. For example, the patent "A Method for Determining the Dimensions of a Ring Forging in the Radial and Axial Directions" (CN101829686A) calculates the blank height, outer diameter, inner diameter, and wall thickness under different deformation modes using parameters such as the outer diameter, inner diameter, height, wall thickness, mandrel diameter, rolling ratio, and radial and axial deformation distribution ratio of the ring forging. This improves blank design efficiency, shortens production cycles, and reduces costs. While this approach is significant for pre-rolling process preparation and blank design, its technical focus lies in pre-rolling dimension design and parameter selection. It cannot yet predict the future evolution of outer diameter, height, and wall thickness online based on real-time data collected during rolling, including temperature, rolling force, roll position, speed, and historical dimensional status.

[0023] In terms of process equipment and forming paths, existing technologies focus more on the structure of ring rolling equipment, rolling paths, and the forming control of ring parts made of specific materials. For example, "Ring rolling process and apparatus" (WO2016142661A1) discloses a ring rolling process and apparatus, which applies radial pressure to the inner and outer surfaces of the ring workpiece through forming rolls and core rolls to achieve the forming of ring parts with specific geometries. "A fine-grained rolling forming method for large high-alloy steel ring parts" (CN102489639B) designs the blanking, heating, forging, billet making, and rolling forming processes of large high-alloy steel ring parts, and divides the rolling process into stages such as pre-rolling, main rolling, and rounding and sizing, and improves the grain structure and product performance through temperature and rolling speed control. The above solutions can improve the forming capacity, microstructure and material utilization under specific process conditions, but they mainly solve the problems of equipment operation, process system or material microstructure control, and have not formed a multi-step prediction model for ring rolling dimensions based on multi-source process data.

[0024] In the field of predictive control for rolling processes, some patents have attempted to introduce predictive concepts into process control. For example, the paper "A Shape Model Predictive Control Method" (CN103357669A) proposes using static and dynamic prediction models to obtain the shape components and then performing control compensation based on the prediction results to reduce feedback control lag. This type of method demonstrates that predictive models have practical value in online control of the rolling process, but its application is primarily to strip shape control, dealing with the shape components in continuous strip rolling, rather than the three-dimensional geometric coupling evolution between outer diameter, height, and wall thickness in axial rolling of ring-shaped pieces. Ring rolling is characterized by closed-loop geometry, simultaneous radial expansion, axial reduction, and wall thickness thinning; its dimensional state is influenced by both historical geometric states and equipment conditions, making it impossible to directly apply predictive models from strip shape control.

[0025] With the development of industrial big data and machine learning technologies, patents related to manufacturing quality prediction are shifting from single-process parameter control to multi-source data modeling. For example, the "Product Quality Control Method Based on Process Network Model and Machine Learning Algorithm" (CN113469241A) combines a complex process network model with machine learning algorithms to mine key process factors and their interactions, achieving quality control in complex product processing. The "A Die-casting Defect Prediction and Diagnosis System" (CN113486457A) obtains process parameters for die casting, uses a deep learning model to output defect prediction results, and filters process parameters through correlation analysis to improve defect diagnosis efficiency. These solutions reflect the application trend of machine learning in manufacturing quality control, but their prediction objects are mostly quality defects, final inspection indicators, or process factor contribution analysis. They typically do not involve multi-step dynamic regression prediction of continuous geometric dimensions during ring rolling, nor do they fully utilize the time dependency between historical dimensional states and real-time equipment states.

[0026] In the field of metal processing quality prediction, existing solutions combine finite element simulation with machine learning algorithms. For example, the patent "A Method for Predicting Decarburization Depth of Medium Carbon Steel Based on Finite Element and XGBoost Algorithms" (CN117494531B) establishes the billet temperature field using an Abaqus finite element model, combines production process data, production plan data, and decarburization quality data, and then utilizes eXtreme Gradient Boosting (XGBoost) to construct a decarburization depth prediction model. This solution demonstrates that tree models such as XGBoost can be used for metal processing quality prediction and can complement the finite element model. However, this patent focuses on the decarburization depth of medium carbon steel, a surface quality indicator. Its input-output logic, physical objects, and prediction targets differ from the dynamic evolution of the outer diameter, height, and wall thickness in ring rolling. Furthermore, its technical approach still relies heavily on finite element model preprocessing and does not directly establish a lightweight, interpretable, and online-deployable multi-step dimensional prediction framework for multi-sensor time-series data from the ring rolling process.

[0027] Therefore, it is evident that related technologies have covered areas such as ring rolling blank design, ring rolling process and equipment control, rolling process predictive control, general manufacturing quality machine learning modeling, and material quality prediction combining finite element method (FEM) and XGBoost. However, significant technological gaps remain between these solutions and online dimensional prediction in ring rolling. Ring rolling field data is characterized by short sequences, high dimensionality, non-steady states, strong noise, and limited sample size. Furthermore, outer diameter, height, and wall thickness are subject to physical constraints such as volume conservation, radial expansion, and axial compression coupling. Simply relying on finite element simulation is insufficient to meet real-time requirements, while using complex deep time-series models can easily lead to overfitting, prediction smoothing, error accumulation, or generalization instability on small-scale structured industrial data. Preliminary validation results based on real industrial multi-sensor data show that, in the scenario of short-sequence structured data for ring rolling, the gradient boosting decision tree model can better adapt to the nonlinear coupling relationship between multi-source process variables and historical dimensional states, and exhibits high prediction accuracy, stability, and interpretability in the prediction of outer diameter, height, and wall thickness.

[0028] Therefore, this application addresses the issue that related ring rolling technologies tend to focus on pre-rolling design, process control, and equipment structure optimization, while lacking online multi-step dimensional prediction capabilities. It proposes a prediction method that integrates multi-source process variables and historical geometric states for the radial-axial ring rolling process, namely, a multi-step quality prediction method for the plastic forming process based on gradient boosting decision trees. This method takes temperature, radial force, axial force, rolling time, spindle speed, mandrel position, main roll position, feed and discharge platform positions, ring growth rate, and historical outer diameter, height, and wall thickness as inputs to directly predict changes in outer diameter, height, and wall thickness over multiple future time steps. This achieves high-precision, low-latency, and interpretable dynamic tracking of ring rolling geometric quality without relying on costly 3D finite element real-time simulation. This method overcomes the shortcomings of related ring rolling patents, which emphasize pre-rolling design and process control, and related quality prediction patents, which lack 3D dimensional time-series modeling of ring rolling. It provides a new technical path for real-time quality monitoring, process parameter adjustment, and digital closed-loop control of ring rolling.

[0029] In one exemplary embodiment, a multi-step quality prediction method for plastic forming processes based on gradient boosting decision trees is provided, such as... Figure 1 As shown, it includes the following steps 101-103.

[0030] Step 101: Obtain the input state sequence based on the input window; the input state sequence includes time step t and the ring rolling state vector of L-1 time steps before time step t; the ring rolling state vector includes process state variables, equipment operation variables and geometric dimension variables; where t is the prediction starting point and L is the length of the input window; Step 102: Introduce the encoding of the target prediction step, and construct the input state sequence as the structured input feature of the target prediction step; the target prediction step is any time step within the prediction window, and the prediction window includes K time steps after time step t, where K is the length of the prediction window; Step 103: Input the structured input features into the size parameter prediction model to obtain the predicted value of the size parameter of the target prediction step; the size parameter prediction model is obtained by training a gradient boosting decision tree model.

[0031] Steps 101-103 above are used to predict the dimensional parameters for several future time steps during the radial and axial rolling process of the ring, based on the real-time collected multi-source process status, equipment operating status and historical geometric dimension information. Steps 101-103 above can realize real-time, accurate and stable dynamic prediction of the geometric dimensions of the ring during the rolling process.

[0032] In another exemplary embodiment, the above-described dimensional parameter prediction model includes: an outer diameter prediction model, a height prediction model, and a wall thickness prediction model; the above-described dimensional parameters include outer diameter, height, and wall thickness.

[0033] In another exemplary embodiment, the method unifies sensor data, roll motion data, and historical dimensional states during the ring rolling process into a structured time-series sample. Through a sliding window, multi-step prediction step-size encoding, and a regularized gradient boosting tree regression model, it establishes a nonlinear mapping relationship from the current rolling state to future geometric dimensional evolution—a dimensional parameter prediction model. This provides a predictive basis for dimensional quality monitoring, process adjustment, and digital closed-loop control during the ring rolling process. The construction of this dimensional parameter prediction model includes: S1. Collect and construct the status data of the ring rolling process.

[0034] S2. Construct multi-step prediction samples based on sliding window.

[0035] S3. Construct structured input features suitable for tree models.

[0036] S4. Train the gradient boosting decision tree model.

[0037] In another exemplary embodiment, step S1 is implemented as follows: During the ring rolling process, process state variables, equipment operation variables, and geometric dimensional variables are collected at each time step through equipment sensors, control systems, and online measurement devices. Process state variables include temperature, radial force, axial force, rolling time, and spindle speed percentage; equipment operation variables include mandrel position, main roll position, feed table position, discharge table position, and ring growth rate; geometric dimensional variables include the ring's outer diameter, height, and wall thickness.

[0038] To facilitate subsequent modeling, the first The rolling state vector of the ring at each time step is represented as: in, Indicates the first The complete state vector at each time step; Indicates the first The combination of process state variables and equipment operation variables collected at each time step includes temperature, radial force, axial force, rolling time, spindle speed, mandrel position, main roll position, feed table position, discharge table position, and ring growth rate, etc. , , They represent the first The outer diameter, height, and wall thickness are measured at each time step. This representation allows the model input to include not only real-time process parameters but also historical geometric states, enabling the predictive model to learn the coupling relationship between process variations and dimensional evolution during ring rolling.

[0039] For each individual ring piece, its complete rolling process is organized into a sample sequence in chronological order. Subsequent training and testing are performed by dividing the data into ring pieces to avoid different sliding windows of the same ring piece appearing simultaneously in both the training and testing sets, thereby reducing the risk of data leakage and improving the reliability of model evaluation results.

[0040] In another exemplary embodiment, step S2 is implemented as follows: The ring rolling process exhibits a clear temporal continuity; the current outer diameter, height, and wall thickness are not only affected by the current process parameters but also closely related to the geometric state of the previous stage. Therefore, this application employs a sliding window approach to construct training samples. Specifically, at a certain time step... As the starting point for prediction, a continuous forward selection is made. Each time step is used as an input window to predict the future. The outer diameter, height, and wall thickness at each time step. The history input window is represented as follows: in, Indicates time step The input state sequence is used to predict the starting point; to This represents a continuous rolling state vector of ring parts within the input window; This indicates the length of the input window. This window is able to preserve short-term state changes, dimensional continuity, and cumulative process effects during the rolling process.

[0041] The future prediction target is the future sequence of three geometric dimensions: outer diameter, height, and wall thickness. For the future... At each time step, the prediction target can be expressed as: ,in, The range of values ​​is to K represents the prediction window length. In specific implementations, it is preferable to use an input window length and a prediction window length that are... The proportion, for example , , , Window configuration. This setting can retain sufficient historical state information while avoiding excessively long input sequences that lead to high feature dimensionality and noise accumulation, thus improving the stability of multi-step prediction.

[0042] In another exemplary embodiment, step S3 is implemented as follows: Since gradient boosting decision trees are well-suited for processing structured tabular data, this invention expands the multi-time-step state vectors within the historical sliding window into one-dimensional features and introduces prediction step size encoding, enabling the model to distinguish prediction tasks at different future time steps. The structured input features for the h-th time step within the prediction window are constructed as follows: in, This represents the structured input features at the h-th time step within the prediction window. Represents the input state sequence. This represents the encoding vector at the h-th time step within the prediction window. This represents the flattening operation, used to expand the history state window into a one-dimensional vector. This indicates a feature splicing operation.

[0043] The preferred encoding method described above is one-hot encoding, which maintains an orthogonal relationship between different prediction steps and prevents the model from simply interpreting the prediction step size as a continuous numerical value, thereby more stably distinguishing between short-term predictions and predictions with longer step sizes.

[0044] During model training, three regression models can be established separately for outer diameter, height, and wall thickness, or a three-target prediction module can be formed by parallel training. Since the dimensions and variation ranges of outer diameter, height, and wall thickness are different, modeling them separately helps to reduce the impact of target scale differences on model training and improve the prediction accuracy of targets of various sizes.

[0045] In another exemplary embodiment, step S4 is implemented as follows.

[0046] A gradient boosting decision tree is used to establish a nonlinear mapping relationship between input features and geometric dimensions. For any dimension parameter, the model is formed by progressively stacking multiple regression trees, and its prediction form is as follows: in, This represents the predicted value of the size parameter, where i represents the i-th training sample; Indicates the first The structured input features of each training sample; Indicates the number of regression trees; Indicates the first A regression tree. This additive model combines multiple weak predictors into a strong predictor by fitting the residuals of the previous round in each round, making it suitable for characterizing the nonlinear relationships between multiple source process variables, equipment operating status, and dimensional status during ring rolling.

[0047] To avoid overfitting the model to local sensor noise or special operating conditions of individual components, this embodiment incorporates a tree structure complexity constraint into the loss function. The model optimization objective can be expressed as: In this formula, This represents the loss function, which is the overall optimization objective of the model. Indicates the number of training samples in a single batch; Represents the actual value of the dimensional parameter Between and predicted values The residual loss function; Indicates the first The number of leaf nodes in a tree; This represents the penalty coefficient for the number of leaf nodes, used to limit excessive complexity in the tree structure; This represents the regularization coefficient for the leaf node weights; Indicates the first The first of the trees The prediction weights of each leaf node.

[0048] By introducing the above regularization term, the model can improve its fitting ability while suppressing excessively deep tree structures and excessively large leaf node weights, thereby enhancing the model's generalization ability under different rings and different rolling stages.

[0049] In another exemplary embodiment, sample subsampling, feature subsampling, and learning rate shrinkage mechanisms can be employed during the specific training process. Sample subsampling is used to reduce the impact of a single abnormal rolling segment on the model, feature subsampling is used to reduce the model's over-reliance on a single sensing variable, and learning rate shrinkage is used to control the contribution of each newly added regression tree to the final prediction result. Through these mechanisms, the model can maintain good stability and robustness under multi-source short-sequence industrial data conditions.

[0050] In another exemplary embodiment, steps 101-103 described above constitute the implementation process of the actual online prediction phase.

[0051] After training, in the actual online prediction phase, the rolling state vectors of the ring piece for the L-1 time steps prior to the current time step are obtained to form the input state sequence. Structured input features are constructed as described above and input into the outer diameter prediction model, height prediction model, and wall thickness prediction model respectively, to obtain the predicted values ​​of the dimensional parameters for each time step within the prediction window. For the h-th time step within the prediction window, the output result is: in, This represents the predicted geometric dimension vector at the h-th time step within the prediction window; , , These represent the predicted outer diameter, height, and wall thickness, respectively. By arranging the prediction results at different time steps within the prediction window in chronological order, the dimensional evolution trajectory of the ring component over a future period can be formed, providing a basis for process engineers to judge dimensional change trends and detect out-of-tolerance risks in advance.

[0052] When a closer approximation of online operation scenarios is required, a recursive prediction method can be adopted, where the predicted future dimensions are used as historical dimension inputs for subsequent prediction windows. This method can simulate the continuous prediction process when the actual future dimensions in the production environment have not yet been measured, giving the model continuous rolling prediction capabilities, making it suitable for online monitoring and closed-loop control.

[0053] In another exemplary embodiment, the above method also includes quality warnings and process adjustment support based on predicted deviations.

[0054] To achieve online quality control during the ring rolling process, this application compares the predicted values ​​of dimensional parameters with the actual values ​​or process tolerances to calculate the predicted deviation. When any predicted deviation of the outer diameter, height, or wall thickness exceeds the corresponding set threshold, a quality warning can be triggered, indicating that adjustments may be needed to the mandrel feed, main roll position, axial reduction, rolling speed, or other process parameters.

[0055] The prediction bias at the h-th time step within the prediction window can be expressed as: in, This represents the prediction bias vector at the h-th time step within the prediction window; This represents the predicted geometric dimension vector at the h-th time step within the prediction window; This represents the target dimension vector at the h-th time step within the prediction window. By analyzing this prediction bias, it's possible to anticipate potential risks such as out-of-tolerance outer diameter, height deviation, or insufficient wall thickness in the ring component.

[0056] Meanwhile, the gradient boosting decision tree model can provide feature importance analysis to identify key variables that contribute significantly to the prediction of outer diameter, height, and wall thickness. For example, the importance of variables such as historical outer diameter, height, wall thickness, rolling time, rolling force, roll position, and ring growth rate can help explain the causes of dimensional variations. This explanation helps process engineers understand the main influencing factors of the prediction results, providing a reference for subsequent process optimization and parameter adjustment.

[0057] In another exemplary embodiment, the above-described outer diameter prediction model is also used to analyze the importance of each variable in the structured input features relative to the outer diameter prediction, and to determine the key variables whose contribution to the outer diameter prediction is greater than a first contribution threshold, as the key variables corresponding to the outer diameter. The height prediction model described above is also used to analyze the importance of each variable in the structured input features relative to height prediction, and to identify key variables whose contribution to height prediction is greater than the second contribution threshold, as key variables corresponding to height. The aforementioned wall thickness prediction model is also used to analyze the importance of each variable in the structured input features relative to the wall thickness prediction, and to identify the key variables whose contribution to the wall thickness prediction is greater than the third contribution threshold, as the key variables corresponding to the wall thickness.

[0058] In another exemplary embodiment, the first contribution threshold, the second contribution threshold, and the third contribution threshold described above may be the same or different.

[0059] In another exemplary embodiment, step 103 above is followed by: Calculate the prediction deviation between the predicted values ​​of the size parameters of the target prediction step and the target values; A quality warning is triggered when the predicted deviation of the dimensional parameters is outside the preset range. When a quality warning is triggered, the size parameter whose predicted deviation is outside the corresponding preset range is identified as the target size parameter; the cause of the quality impact is determined based on the key variables corresponding to the target size parameter.

[0060] In another exemplary embodiment, the prediction deviation of the above-mentioned dimensional parameters not being within the preset range means that at least one of the prediction deviation of the outer diameter, the prediction deviation of the height, and the prediction deviation of the wall thickness is not within the corresponding preset range.

[0061] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 2As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and databases. The internal memory provides the environment for the operating system and computer programs stored in the non-volatile storage media. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When the computer program is executed by the processor, it implements a multi-step quality prediction method for plastic forming processes based on gradient boosting decision trees.

[0062] Those skilled in the art will understand that Figure 2 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0063] In one exemplary embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0064] According to the specific embodiments provided in this application, this application establishes a prediction method for the radial-axial ring rolling process through multi-source process data acquisition, sliding window sample construction, prediction step size encoding, regularized gradient boosting decision tree modeling, multi-step size prediction, and deviation early warning analysis. Compared with methods relying on finite element simulation, this scheme has higher computational efficiency and is suitable for online deployment. Compared with general deep time series models, this scheme is more suitable for short-sequence, high-dimensional, and small-sample structured industrial data scenarios, and also has high prediction accuracy, stability, and interpretability, as detailed below.

[0065] Multi-source process state variables, equipment operation variables, and historical outer diameter, height, and wall thickness are uniformly constructed into a sliding window time series sample. Combined with prediction step size coding, dynamic prediction of the outer diameter, height, and wall thickness of the ring component for multiple future time steps is achieved.

[0066] A regularized gradient boosting decision tree model is used to model short-sequence, high-dimensional, and unsteady-state industrial data of ring rolling, avoiding the problems of insufficient real-time performance of traditional finite element simulation and the generalization instability of complex deep time series models on small sample structured data.

[0067] The application method of monitoring dimensional deviations and providing early warnings based on prediction results involves identifying out-of-tolerance risks in advance based on the predicted deviations between the outer diameter, height, and wall thickness and the target dimensions, and providing a basis for adjusting process parameters such as core roll feed, main roll position, axial reduction, and rolling speed.

[0068] In another exemplary embodiment, the following verification results and descriptions were provided to illustrate the technical effects of the present application's technical solution.

[0069] Regarding the technical effectiveness of the above-mentioned solution in significantly improving the accuracy and stability of ring rolling dimension prediction: In industrial data verification based on 40 independent ring rolling sequences and a total of 5822 multi-sensor samples, the gradient boosting decision tree model used in this application embodiment achieved near-perfect fitting results on the three targets of outer diameter, height, and wall thickness, with an average MAE of 0.41 and a Macro-R... 2 The value is 0.99; where R represents the outer diameter, height, and wall thickness. 2 The results were approximately 1.00, 0.98, and 1.00, respectively. These results are significantly better than the depth sequence model and the traditional statistical baseline, indicating that the proposed scheme can more effectively handle structured, short-sequence, and unsteady-state data from the ring rolling field.

[0070] Regarding the technical effect of the above-mentioned technical solution in reducing the instability caused by the recursive prediction error of multi-step prediction: In the recursive prediction mode of this application embodiment, the model's average MAE only increases from 0.41 to 0.52, Macro-R 2 The accuracy remains at 0.99, indicating that even under online inference conditions where the actual future dimensions are not directly available, the model can still maintain stable trajectory predictions. In contrast, some deep learning models exhibit significantly amplified errors in recursive predictions, making it difficult to meet the reliability requirements of closed-loop control.

[0071] Regarding the technical effect of the above-mentioned technical solution having good generalization ability across ring components: the embodiments of this application are based on grouped 5-fold cross-validation, and the average MAE of the Gradient Boosting Decision Tree (GBDT) model is 0.29±0.06, Macro-R 2 The result is 1.00±0.00, which is significantly better than the large fluctuations in the performance of deep models under different data partitions. This effect comes from the synergistic effect of regularized tree ensemble, random subsampling of samples / features, and prediction step encoding mechanism, which enables the model to stably learn the nonlinear mapping between device operating variables, historical size states, and future geometric evolution under limited sample conditions.

[0072] Regarding the technical effects of the above-mentioned technical solution in improving industrial deployment efficiency: Compared with finite element simulation, the inference speed after training in this application embodiment is fast, and it can be deployed in production line monitoring systems or host computers for real-time dimensional tracking; compared with deep models, this solution has low training cost, intuitive parameter adjustment, and strong interpretability. The model feature importance results are consistent with the ring rolling plastic deformation mechanism, and can identify the contributions of key variables such as historical geometric state and rolling time to outer diameter expansion, height adjustment, and wall thickness reduction, providing a basis for process optimization, anomaly diagnosis, and control strategy correction.

[0073] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0074] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0075] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0076] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0077] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A multi-step quality prediction method for plastic forming process based on gradient boosting decision tree, characterized in that, include: The input state sequence is obtained based on the input window; the input state sequence includes time step t and the ring rolling state vector of L-1 time steps before time step t; the ring rolling state vector includes process state variables, equipment operation variables and geometric dimension variables; where t is the prediction starting point and L is the length of the input window; The target prediction step is encoded to construct the input state sequence as the structured input feature of the target prediction step; the target prediction step is any time step within the prediction window, and the prediction window includes K time steps after time step t, where K is the length of the prediction window; The structured input features are input into the size parameter prediction model to obtain the predicted value of the size parameter for the target prediction step; the size parameter prediction model is obtained by training a gradient boosting decision tree model.

2. The multi-step quality prediction method for plastic forming process based on gradient boosting decision tree according to claim 1, characterized in that, The structured input features are: ； in, This represents the structured input features at the h-th time step within the prediction window. Represents the input state sequence. This represents the encoding vector at the h-th time step within the prediction window. Indicates the flattening operation. This indicates a feature splicing operation.

3. The multi-step quality prediction method for plastic forming process based on gradient boosting decision tree according to claim 1, characterized in that, The loss function used in training the gradient boosting decision tree model is: ； in, For loss function, Represents the actual value of the dimensional parameter Compared with the predicted value The residual loss function between them, where N is the number of training samples in a single batch, and i represents the i-th training sample. This represents the total number of regression trees. Indicates the first The number of leaf nodes in a tree. This represents the penalty coefficient for the number of leaf nodes. This represents the regularization coefficient for the leaf node weights. Indicates the first The first of the trees The prediction weights of each leaf node.

4. The multi-step quality prediction method for plastic forming process based on gradient boosting decision tree according to claim 1, characterized in that, The ratio of the length of the input window to the length of the prediction window is 4:

1.

5. The multi-step quality prediction method for plastic forming process based on gradient boosting decision tree according to claim 1, characterized in that, The dimensional parameter prediction model includes: outer diameter prediction model, height prediction model, and wall thickness prediction model; The dimensional parameters include outer diameter, height, and wall thickness.

6. The multi-step quality prediction method for plastic forming process based on gradient boosting decision tree according to claim 5, characterized in that, The outer diameter prediction model is also used to analyze the importance of each variable in the structured input features relative to the outer diameter prediction, and to determine the key variables whose contribution to the outer diameter prediction is greater than the first contribution threshold, as the key variables corresponding to the outer diameter. The height prediction model is also used to analyze the importance of each variable in the structured input features relative to the height prediction, and to determine the key variables whose contribution to the height prediction is greater than the second contribution threshold, as the key variables corresponding to the height. The wall thickness prediction model is also used to analyze the importance of each variable in the structured input features relative to the wall thickness prediction, and to identify key variables whose contribution to the wall thickness prediction is greater than the third contribution threshold, as the key variables corresponding to the wall thickness.

7. The multi-step quality prediction method for plastic forming process based on gradient boosting decision tree according to claim 6, characterized in that, The structured input features are input into the size parameter prediction model to obtain the predicted size parameter values ​​for the target prediction step, followed by: Calculate the prediction deviation between the predicted values ​​of the size parameters of the target prediction step and the target values; A quality warning is triggered when the predicted deviation of the dimensional parameters is outside the preset range.

8. The multi-step quality prediction method for plastic forming process based on gradient boosting decision tree according to claim 7, characterized in that, The predicted deviation of the dimensional parameters is not within the preset range, which means that at least one of the predicted deviations of the outer diameter, height, and wall thickness is not within the corresponding preset range.

9. The multi-step quality prediction method for plastic forming process based on gradient boosting decision tree according to claim 7, characterized in that, A quality warning is triggered when the predicted deviation of the dimensional parameters is outside the preset range, followed by: The size parameters whose prediction deviation is outside the corresponding preset range are identified as the target size parameters. Identify the causes of quality influencing factors based on the key variables corresponding to the target size parameters.

10. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the multi-step quality prediction method for plastic forming process based on gradient boosting decision tree as described in any one of claims 1-9.

Images