A shot peening process for improving the surface hardness of precision long shafts

Abstract

This invention provides a shot peening strengthening process for improving the surface hardness of precision long shaft parts, belonging to the field of precision machining technology. The invention involves pre-treating the surface of the precision long shaft part, establishing a discrete mesh model of the long shaft and collecting initial residual stress distribution data, activating an online dynamic load compensation support system to control the axial runout within a threshold range, and driving the spray gun to perform continuous shot peening along a spiral trajectory based on the process timing parameters output by the physical information-constrained spatiotemporal jump graph attention model. The invention also utilizes a multi-channel acoustic emission signal sequence to output real-time predicted values ​​of the residual stress field, inputting these predicted values ​​into a shot peening strengthening process game-theoretic control model to drive a follow-up reverse hydraulic straightening device to actively suppress bending deformation and perform supplementary shot peening on areas that do not meet the standards. This solves the technical problem of difficulty in real-time suppression of bending deformation caused by uneven residual stress field distribution during the shot peening strengthening process of precision long shaft parts.

Description

Technical Field

[0001] This invention belongs to the field of precision machining technology, and specifically relates to a shot peening strengthening process for improving the surface hardness of precision long shaft parts. Background Technology

[0002] Precision long shaft components are widely used in aerospace, precision machine tools, and marine transmission. Their surface hardness and residual stress distribution directly affect the fatigue life and service reliability of the components. Shot peening is the main process for improving the surface hardness of long shaft components and introducing compressive residual stress. Traditional shot peening processes rely on manual experience to set process parameters such as air pressure, spray gun speed, and rotation speed, and judge whether the strengthening effect meets the standards through post-process inspection.

[0003] However, for precision long shafts with a large length-to-diameter ratio, the uneven distribution of shot peening impact force along the axial and circumferential directions can induce asymmetric residual stress accumulation inside the shaft, thereby triggering bending deformation. Traditional processes rely on static support devices and post-processing straightening, which cannot detect the real-time evolution of bending deformation. Straightening operations often lag behind the development of deformation, resulting in inaccurate application of straightening force at the wrong location and timing, making it difficult to fundamentally suppress deformation.

[0004] In existing technologies, the spatial distribution of residual stress fields during shot peening cannot be perceived in real time. Therefore, there is a lack of immediate feedback for adjusting process parameters and applying straightening force. The localized stress asymmetry caused by the shot peening gun trajectory continues to accumulate, eventually leading to excessive bending deformation of the shaft. Furthermore, once deformation occurs, the straightening process cannot completely eliminate the internal residual asymmetric stress. In other words, existing technologies suffer from the technical problem of uneven residual stress field distribution during shot peening of precision long shafts, making it difficult to suppress bending deformation of the shaft in real time. Summary of the Invention

[0005] In view of this, the present invention provides a shot peening strengthening process for improving the surface hardness of precision long shaft parts, which can solve the technical problem in the prior art that the uneven distribution of residual stress field during the shot peening strengthening process of precision long shaft parts makes it difficult to suppress the bending deformation of the shaft parts in real time.

[0006] This invention is achieved as follows: This invention provides a shot peening strengthening process for improving the surface hardness of precision long shaft parts, comprising the following steps:

[0007] The surface of the precision long shaft is pretreated, the coordinates of three-dimensional measuring points along the entire length of the long shaft are measured, a discrete mesh model of the long shaft is established, and the initial residual stress distribution data of the long shaft is collected as the reference input for subsequent process parameter control.

[0008] The online dynamic load compensation support system is activated, and multiple sets of follow-up hydraulic support units are arranged along the long axis. Sensors collect axial runout and bending strain in real time, and the support stiffness of each support unit is dynamically adjusted to control the axial runout within the axial runout threshold range.

[0009] Based on the process timing parameters output by the physical information-constrained spatiotemporal jump graph attention model, the air pressure, spray gun moving speed and long shaft rotation speed are set, and the shot peening gun is driven to continuously shot peening the long shaft surface along the spiral trajectory.

[0010] During the shot peening process, a multi-channel acoustic emission sensor collects the acoustic emission signal sequence in real time. The physical information constrained spatiotemporal jump graph attention model outputs the residual stress field prediction value in real time based on the acoustic emission signal sequence and the three-dimensional measurement point coordinates. The residual stress field prediction value is input into the shot peening process game control model to drive the follow-up reverse hydraulic straightening device to actively suppress bending deformation.

[0011] After shot peening, a dynamic spatial octree partitioning load balancing parallel scheduling algorithm is activated to perform parallel post-processing on the simulation results of the full-field discrete element method, verify the symmetry of the residual stress distribution on the long axis surface, and carry out supplementary shot peening on the areas that do not meet the standards.

[0012] The three-dimensional measuring point coordinates refer to the set of spatial coordinate points uniformly distributed along the entire length of the long axis, and the measuring point spacing is the measuring point spacing value, which is used to describe the spatial geometry of the long axis in the processing state.

[0013] The long-axis discrete mesh model uses three-dimensional measurement point coordinates as nodes to divide the long-axis surface into several continuous triangular or quadrilateral mesh units, which are used to support the spatial modeling input of the graph attention network.

[0014] The initial residual stress distribution data is obtained by X-ray diffraction or blind hole method on multiple sets of long-axis specimens. The measurement points correspond to discrete grid nodes, and the measurement results are used as the initial stress labels of the model training dataset after statistical analysis.

[0015] The online dynamic load compensation support system consists of multiple sets of hydraulic support units arranged along the axial direction. Each set of support units is equipped with a displacement sensor and a pressure sensor. The support stiffness adjustment range is the support stiffness value, and the support stiffness adjustment response time does not exceed the response time threshold.

[0016] The spiral trajectory refers to the spray gun moving at a constant speed along the long axis while the long axis rotates at a constant speed around the axis. The spray gun coverage path unfolds into a spiral shape on the surface of the long axis, and the spiral pitch is determined by the spray gun moving speed and the long axis rotation speed.

[0017] The input layer of the physical information constrained spatiotemporal jump graph attention model receives the coordinates of the three-dimensional measurement points along the full length of the major axis, process timing parameters, and multi-channel acoustic emission signal sequences. The three types of inputs are mapped to the feature space through independent linear projection layers. The spatial modeling module uses a graph attention network, and the temporal modeling module uses a temporal convolutional network.

[0018] In the graph attention network, the attention weights are modulated by the Euclidean distance between nodes and the residual stress gradient. The graph attention network is stacked in multiple layers, and the output of each layer is normalized to output node features.

[0019] The dilation coefficient of the temporal convolutional network increases layer by layer. The temporal convolutional network captures the evolution of the acoustic emission signal sequence on the time axis. The spatiotemporal jump connection mechanism establishes direct additive jump connections between the bottom and top outputs of the graph attention network and between the bottom and top layers of the temporal convolutional network.

[0020] In the physical constraint fusion module of the physical information constrained spatiotemporal jump graph attention model, the constitutive equation of metal elastoplastic mechanics and the one-dimensional heat conduction partial differential equation are embedded as regularization terms into the loss function, and the regularization coefficient is determined by cross-validation experiments.

[0021] The following reverse hydraulic straightening device can apply a reverse straightening force in real time according to the direction of bending deformation. The straightening force range is the straightening force value. The position of the force application point moves synchronously with the position of the spray gun to ensure that the straightening force always acts on the section with the maximum bending deformation.

[0022] The shot peening strengthening process game control model consists of a two-layer game structure, with an upper-layer model that aims to maximize stress symmetry and a lower-layer model that aims to minimize straightening force. The two objectives are related through a coupling stiffness coefficient.

[0023] The dynamic spatial octree partitioning load balancing parallel scheduling algorithm dynamically reconstructs the spatial octree based on the local density of the projectile in the simulation space, divides the high-density collision region into finer subdomains, merges the sparse region into larger subdomains, and uses an asynchronous message communication mechanism to eliminate waiting between nodes.

[0024] The acoustic emission signal sequence is acquired by a multi-channel piezoelectric sensor arranged on the long axis or a fixture. The sampling frequency is the sampling frequency value, and the number of channels is the channel value, which is used to reflect the spatial location and impact intensity of the projectile impact event.

[0025] Wherein, the axial runout threshold is 0.01–0.05 mm; the measuring point spacing is 20–50 mm; and the support stiffness is... ~ The response time threshold is 20ms; the straightening force value is 500-5000N; the sampling frequency value is 1-5. The channel value is 4–8; the helical pitch ranges from 3 to 15 mm; the regularization coefficient ranges from 0.01 to 0.1; and the coupling stiffness coefficient ranges from 100 to 500. .

[0026] This invention constructs a physical information-constrained spatiotemporal jump graph attention model, takes a multi-channel acoustic emission signal sequence and three-dimensional measurement point coordinates as input, outputs the predicted value of residual stress field in real time, and then drives the shot peening strengthening process game control model to coordinate the spray gun trajectory and straightening force, thus solving the technical problem that the uneven distribution of residual stress field makes it difficult to suppress bending deformation in real time.

[0027] This invention combines the ability of graph attention networks to model spatial topology with the ability of temporal convolutional networks to capture the temporal evolution of acoustic emission signals. This allows the model to proactively trigger the straightening device based on the predicted stress field value before bending deformation occurs, intervening in the accumulation process of asymmetric stress from the source, rather than passively implementing straightening after deformation occurs. This ensures that the timing and location of the straightening force application are always synchronized with the deformation trend.

[0028] In summary, the present invention solves the technical problem mentioned in the background art of the difficulty in real-time suppression of bending deformation of precision long shafts due to uneven distribution of residual stress field during shot peening. Attached Figure Description

[0029] Figure 1 This is a flowchart of the method of the present invention.

[0030] Figure 2 This is a schematic diagram of the discrete mesh model of the entire length of the optical rod axis and the coordinate distribution of the three-dimensional measurement points.

[0031] Figure 3 This is a schematic diagram of the spatial distribution of residual stress across the entire field at a typical moment of shot peening.

[0032] Figure 4 A line graph showing the axial distribution of the hardened layer depth at each test section. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0034] like Figure 1 The diagram shows a flowchart of a shot peening process for improving the surface hardness of precision long shaft parts, provided by the present invention. This method includes the following steps:

[0035] S01. Perform surface pretreatment on the precision long shaft component, measure the coordinates of three-dimensional measuring points along the entire length of the long shaft, establish a discrete mesh model of the long shaft, and collect the initial residual stress distribution data of the long shaft as the reference input for subsequent process parameter control;

[0036] S02. Start the online dynamic load compensation support system, arrange multiple sets of follow-up hydraulic support units along the long axis, and use sensors to collect axial runout and bending strain in real time, dynamically adjust the support stiffness of each support unit, and control the axial runout within the range of 0.01 to 0.05 mm.

[0037] S03. Based on the process timing parameters output by the attention model of the spatiotemporal jump diagram constrained by physical information, set the air pressure to 0.3-0.6MPa, the spray gun moving speed to 50-200mm / s, and the long shaft rotation speed to 30-120r / min, and drive the shot peening gun to continuously shot peening the long shaft surface along the spiral trajectory.

[0038] S04. During the shot peening process, the multi-channel acoustic emission sensor collects the acoustic emission signal sequence in real time. The physical information constrained spatiotemporal jump graph attention model outputs the predicted value of the residual stress field in real time based on the acoustic emission signal sequence and the coordinates of the three-dimensional measuring point. The predicted value of the residual stress field is then input into the shot peening process game control model to drive the follow-up reverse hydraulic straightening device to actively suppress bending deformation.

[0039] S05. After shot peening, the dynamic spatial octree partitioning load balancing parallel scheduling algorithm is started to perform parallel post-processing on the simulation results of the whole field discrete element method, verify the symmetry of the residual stress distribution on the long axis surface, and carry out supplementary peening for the substandard areas. The supplementary peening pressure is 0.3 to 0.5 MPa.

[0040] S06. Optionally, it also includes: performing surface hardness and straightness tests on the long shaft parts that have undergone shot peening. The criteria for passing the surface hardness test are that the depth of the hardened layer is not less than 0.2 to 0.5 mm and the straightness error is not more than 0.05 to 0.1 mm / m. After passing the tests, the parts are subjected to rust prevention and sealing treatment.

[0041] The three-dimensional measuring point coordinates refer to a set of spatial coordinate points evenly distributed along the entire length of the long axis, used to describe the spatial geometry of the long axis in the processing state, with a measuring point spacing of 20-50mm.

[0042] The long-axis discrete mesh model refers to a geometrical numerical representation that divides the long-axis surface into several continuous triangular or quadrilateral mesh units using three-dimensional measurement point coordinates as nodes, and is used to support the spatial modeling input of graph attention networks.

[0043] The initial residual stress distribution data is obtained by X-ray diffraction or blind hole method on multiple sets of long-axis specimens. The measurement points correspond to discrete grid nodes, and the measurement results are used as the initial stress labels of the model training dataset after statistical analysis.

[0044] The online dynamic load compensation support system consists of 3 to 8 sets of hydraulic support units arranged axially. Each support unit is equipped with a displacement sensor and a pressure sensor, and the support stiffness adjustment range is [range missing]. ~ The response time for stiffness adjustment is no more than 20ms.

[0045] The aforementioned follow-up reverse hydraulic straightening device refers to a hydraulic actuator that can apply a reverse straightening force in real time according to the direction of bending deformation. The straightening force ranges from 500 to 5000 N, and the position of the force application point moves synchronously with the position of the spray gun to ensure that the straightening force always acts on the section with the maximum bending deformation.

[0046] The axial runout control threshold of 0.01 to 0.05 mm is obtained by the following method: shot peening tests are conducted on long shaft specimens with different length-to-diameter ratios (range of 50 to 200) under different support stiffnesses. The runout and the final straightness error of each group are recorded. The mapping relationship between runout and straightness error is determined by regression analysis. With the straightness error not exceeding 0.05 mm / m as a constraint, the upper limit of runout is calculated.

[0047] The spiral trajectory refers to the spray gun moving at a constant speed along the long axis while the long axis rotates at a constant speed around the axis. The spray gun coverage path unfolds into a spiral shape on the surface of the long axis. The pitch of the spiral is determined by the spray gun moving speed and the rotation speed of the long axis, and the pitch range is 3 to 15 mm.

[0048] The long shaft component can be a smooth shaft, etc.

[0049] The physical information-constrained spatiotemporal jump graph attention model is based on deep learning. Its specific structure is as follows: The model input layer receives the coordinates of three-dimensional measurement points along the entire major axis, process timing parameters (air pressure, spray gun movement speed, and rotational speed), and a multi-channel acoustic emission signal sequence. These three types of inputs are mapped to a 128-dimensional feature space through independent linear projection layers. The spatial modeling module uses a graph attention network, with discrete grid nodes as graph nodes and adjacency relationships between nodes as graph edges. The multi-head attention mechanism has 8 heads, each with a 16-dimensional feature dimension. The attention weights are modulated by the Euclidean distance between nodes and the residual stress gradient. The graph attention network is stacked in three layers, and the output of each layer is normalized to produce 128-dimensional node features. The temporal modeling module uses a temporal convolutional network with a kernel size of 3 and a dilation coefficient of [missing information]. The architecture features four dilated convolutional layers, each with 128 output channels, progressively increasing in complexity. The temporal convolutional network captures the evolution of acoustic emission signal sequences along the time axis. A spatiotemporal jump connection mechanism establishes a direct additive jump connection between the outputs of the first and third layer graph attention networks, and simultaneously between the first and fourth layers of the temporal convolutional network. This allows initial features of residual stress at lower levels to be directly fused with features at higher levels, avoiding deep gradient vanishing. The physical constraint fusion module embeds the constitutive equations of metal elastoplastic mechanics and the one-dimensional partial differential equation of heat conduction as regularization terms into the loss function. The regularization coefficient... The range is 0.01–0.1, determined by cross-validation experiments; the output layer consists of two parallel fully connected layers, which output the predicted residual stress field and hardening layer depth for each node on the long axis surface, respectively, with the output dimension consistent with the number of discrete grid nodes; the weight matrix between neurons is stored sparsely, with the sparsity rate controlled by the attention weight threshold, ranging from 0.05 to 0.15; the feature tensors between layers are stored in half-precision floating-point (FP16) format to compress memory usage; the data loop uses an asynchronous preloading mechanism with a preloading queue depth of 4, and memory allocation... The memory allocation is statically divided into a 40% allocation for the graph attention network layer and the 40% allocation for the temporal convolutional network layer, with a 20% allocation for the skip connection cache. The graph structure adjacency matrix is ​​stored in a compressed sparse row format. In CUDA stream allocation, the forward propagation of the graph attention network and the forward propagation of the temporal convolutional network are assigned to two independent CUDA streams for parallel execution. Shared memory allocation for CUDA is used to store the intermediate matrix of multi-head attention within the same graph attention network layer. In hierarchical allocation, the calculation of physical constraint regularization gradients is assigned to independent computational levels to avoid aliasing with prediction error gradients. (Spray gun movement speed...) With long shaft speed The model output is adjusted by an adaptive learning rate function. After mapping, update to the motion controller, function The definition will be explained later.

[0050] The steps for establishing the training dataset for the physical information-constrained spatiotemporal jump graph attention model specifically include: collecting experimental data on shot peening strengthening processes for precision long shaft components of different materials and with different aspect ratios (range 50–200); recording process timing parameters, three-dimensional measurement point coordinates, acoustic emission signal sequences, and X-ray diffraction residual stress and microhardness measurements after shot peening for each experimental group; performing wavelet packet decomposition on the acoustic emission signal sequences to extract the frequency band from 50–500. The feature components are used to convert the coordinates of the three-dimensional measurement points into a discrete mesh adjacency matrix. The residual stress measurement value and the hardened layer depth measurement value are used as supervision labels. The dataset size is no less than 500 experimental samples, which are divided into training set, validation set and test set in a ratio of 7:2:1.

[0051] The specific steps for training the physical information-constrained spatiotemporal jump graph attention model include: using the Adam optimizer with an initial learning rate of... The weight decay coefficient is Batch size is 16; total training epochs are 200; the loss function is a weighted sum of the prediction mean square error term and the physical constraint regularization term, with regularization coefficients... The model parameters are determined by cross-validation within the range of 0.01 to 0.1; the prediction error is evaluated on the validation set every 10 rounds, and early stopping is triggered when the validation set error does not decrease for 20 consecutive rounds; the model parameters are saved based on the weights corresponding to the minimum validation set error.

[0052] The technical effects of the physical information-constrained spatiotemporal jump graph attention model are as follows: the graph attention network establishes spatial topological relationships between discrete grid nodes on the long axis surface, enabling the model to perceive the spatial gradient distribution of residual stress in the radial and axial directions of the long axis; the temporal convolutional network captures the dynamic evolution of acoustic emission signals in the processing sequence, incorporating the time accumulation effect of impact events into the prediction; the spatiotemporal jump connection mechanism preserves the underlying geometry and initial stress characteristics, preventing the deep network from forgetting the initial constraint information; the physical partial differential equation constraint enables the model to maintain stress tensor balance and energy conservation under industrial sparse data conditions, significantly improving the generalization ability and predictive physical rationality of unseen working conditions.

[0053] Wherein, the adaptive learning rate adjustment function The motion controller updates its step size based on the current residual stress prediction error and acoustic emission signal energy, adjusting the spray gun's moving speed in conjunction with the current residual stress prediction error. Inputs include the root mean square error of the residual stress in the current validation set. (Unit: MPa) and normalized acoustic emission signal energy (Dimensionless, ranging from 0 to 1), the output is the learning rate adjustment coefficient. (Dimensionless); the function is defined as follows: ,in This is the dimensionless adjustment coefficient, with units of . The range of values ​​is Obtained through experimental calibration; when When the learning rate remains unchanged, the learning rate remains at its current value; when When the learning rate is multiplied by a coefficient of 0.9, it undergoes a slight decay; when At that time, the learning rate is multiplied by a coefficient of 0.7 for moderate decay; when When the time comes, the learning rate is reset to the initial learning rate. To escape local minima The value is obtained in the following way: Record the values ​​in 5-10 sets of verification experiments under different working conditions. and The typical distribution range, with Falling into under most operating conditions Using the interval as the target, we work backwards. Within a reasonable range, take the median value.

[0054] The shot peening strengthening process game-theoretic control model is used to optimize the coordination between the spray gun trajectory and the straightening force. An upper-level model aiming to maximize stress symmetry and a lower-level model aiming to minimize the straightening force constitute a two-layer game structure; the upper-level objective function... The symmetry exponent used to maximize the residual stress distribution on the major axis surface is used as input, which includes the predicted residual stress field output from the physically constrained spatiotemporal jump graph attention model. (Unit: MPa) and the pitch of the spray gun spiral trajectory (Unit: mm), the objective function is: ,in The mean value of the predicted residual stress field (in MPa) is given, with the constraint being the pitch. mm and the residual stress difference between adjacent nodes does not exceed 50MPa; lower-level objective function To minimize the total straightening force applied by the follow-up reverse hydraulic straightening device, the input includes the straightening force. (Unit: N) and axial runout (Unit: mm), objective function is: ,in Coupled stiffness coefficient (unit) ), the constraints are mm and N; the coupling terms of the two objective functions are This coupling term appears simultaneously in both the upper-level constraints and the lower-level target, reflecting the physical mechanism by which the stress asymmetry caused by the spray gun trajectory drives an increase in the lower-level straightening force. The range of values ​​is The straightness error and residual stress symmetry index were obtained by conducting orthogonal experiments under different combinations of axial runout (0.01–0.05 mm) and straightening force (500–5000 N), recording the straightness error and residual stress symmetry index under each combination, and then determining the straightness error through least squares fitting. .

[0055] The dynamic spatial octree partitioning load balancing parallel scheduling algorithm refers to dynamically reconstructing the spatial octree based on the local density of the projectiles in the simulation space, dividing the high-density collision area into finer subdomains and merging the sparse area into larger subdomains, so that each parallel computing node bears an equivalent computing load, and using an asynchronous message communication mechanism to eliminate waiting between nodes; the subdomain partitioning threshold (i.e., the maximum number of projectiles in a single subdomain) ranges from 500 to 2000, which is determined by minimizing the variance of the computing time of each node in the pre-experiment.

[0056] The acoustic emission signal sequence refers to the time-domain signal of elastic waves acquired by a multi-channel piezoelectric sensor arranged on the long axis or a fixture, with a sampling frequency of 1 to 5. The number of channels is 4 to 8, which are used to reflect the spatial location and impact intensity of a projectile impact event.

[0057] The method for detecting the depth of the hardened layer involves uniformly cutting 3 to 5 cross sections along the long axis. For each cross section, a micro Vickers hardness tester is used to measure the hardness distribution radially from the surface to the inside. The depth corresponding to the point where the hardness value drops to 110% of the matrix hardness is taken as the depth of the hardened layer.

[0058] The specific implementation of step S01 is as follows: First, the surface of the precision long shaft component is cleaned and the oxide layer is removed to ensure consistency between the surface state of subsequent measurements and shot peening. Then, three-dimensional measuring points are evenly distributed along the entire length of the long shaft at intervals of 20-50 mm. A coordinate measuring instrument is used to acquire the spatial coordinates of each measuring point, forming a set of three-dimensional measuring point coordinates to describe the spatial geometry of the long shaft under machining conditions. Using these coordinate points as nodes, the topological relationships between adjacent nodes are encoded into triangular or quadrilateral mesh elements to establish a discrete mesh model of the long shaft. This model is stored in the form of an adjacency matrix, using a compressed sparse row format to save storage space. Initial residual stress distribution data is obtained by measuring multiple sets of long shaft samples using X-ray diffraction or blind hole methods. The measurement points correspond one-to-one with the discrete mesh nodes. After statistical analysis, the measurement results are used as the initial stress labels for the training dataset of the physical information constrained spatiotemporal jump graph attention model, providing a benchmark input for subsequent process parameter control.

[0059] The specific implementation of step S02 is as follows: 3 to 8 sets of follow-up hydraulic support units are evenly arranged along the long axis. Each set of units is equipped with a displacement sensor and a pressure sensor. The sensors collect the axial runout and bending strain of the current section in real time with a response period of no more than 20ms. The controller calculates the support stiffness adjustment of each section based on the collected data, and the adjustment range is... ~ The axial runout was controlled within a threshold range of 0.01–0.05 mm. The axial runout threshold was determined as follows: shot peening tests were conducted on long-axis specimens with different aspect ratios (50–200) under different support stiffnesses. The runout and the final straightness error of each group were recorded. The mapping relationship between the runout and the straightness error was established through regression analysis. The upper limit of the runout was calculated by using the straightness error not exceeding 0.05 mm / m as a constraint.

[0060] The specific implementation of step S03 is as follows: the process timing parameters are determined by the physical information-constrained spatiotemporal jump graph attention model based on the current residual stress field prediction value and acoustic emission signal energy, through an adaptive learning rate adjustment function. The mapped output is sent to the motion controller. The spray gun air pressure setting range is 0.3–0.6 MPa, the spray gun moving speed is 50–200 mm / s, and the long axis rotation speed is 30–120 r / min. While the spray gun moves uniformly along the long axis, the long axis rotates uniformly around the axis. The spray gun coverage path unfolds as a spiral on the surface of the long axis. The spiral pitch is determined by both the spray gun moving speed and the long axis rotation speed, and the pitch range is 3–15 mm. Adaptive learning rate adjustment function. The input is the root mean square error of the residual stress in the current validation set. With normalized acoustic emission signal energy The output is the learning rate adjustment coefficient. The function is defined as ,in accordance with The interval in which the motion controller updates its step size is adjusted in stages.

[0061] The specific implementation of step S04 is as follows: a multi-channel acoustic emission sensor is arranged on a long shaft or fixture, and the sampling frequency is 1 to 5. The system has 4–8 channels and acquires the time-domain signal sequence of elastic waves generated by projectile impact in real time. The acoustic emission signal sequence is then extracted from the wavelet packet decomposition to obtain 50–500... After the frequency band feature components are input, along with the three-dimensional measurement point coordinates and process timing parameters, they are input into a physical information-constrained spatiotemporal jump graph attention model. The graph attention network of this model uses discrete grid nodes as graph nodes and adjacency relationships between nodes as graph edges. The attention weights are jointly modulated by the Euclidean distance between nodes and the residual stress gradient, stacked in three layers to capture the spatial gradient distribution of residual stress; the temporal convolutional network dilation coefficient is calculated according to... Layer-by-layer increments capture the temporal evolution of acoustic emission signals; a spatiotemporal jump connection mechanism establishes direct additive jump connections between lower and higher layers to prevent the loss of initial constraint information in deep networks; the physical constraint fusion module embeds the constitutive equations of metal elastoplastic mechanics and the one-dimensional partial differential equation of heat conduction into the loss function, with regularization coefficients... The range is 0.01 to 0.1, determined by cross-validation. The model outputs the predicted residual stress field values ​​for each grid node in real time. These predicted values ​​are input into the shot peening strengthening process game control model. The upper-level objective function adjusts the pitch based on the principle of maximizing stress symmetry, while the lower-level objective function controls the axial runout based on the principle of minimizing straightening force. The coupling stiffness coefficient... The range is 100 to 500 The two-layer target collaborative drive follow-up reverse hydraulic straightening device actively suppresses bending deformation.

[0062] The specific implementation of step S05 is as follows: After shot peening, the spatial octree is dynamically reconstructed based on the local density of the projectiles in the simulation space. The high-density collision zone is divided into finer subdomains, and the sparse zone is merged into a larger subdomain. This allows each parallel computing node to bear an equivalent computational load. The maximum number of projectiles in the subdomain is within the range of 500 to 2000, determined by minimizing the variance of the computation time of each node in the pre-experiment. An asynchronous message communication mechanism is used to eliminate waiting between nodes, and parallel post-processing of the simulation results of the full-field discrete element method is completed to verify the symmetry of the residual stress distribution on the long axis surface. For areas that do not meet the requirements, the additional shot pitch is calculated based on the upper-level objective function. The additional shot pressure is set to 0.3 to 0.5 MPa, and local additional shot is performed until the stress symmetry of the entire field meets the requirements.

[0063] It should be noted that the key technologies of this invention include: explicitly modeling the spatial topological relationship of discrete grid nodes along the long axis using a graph attention network, enabling the spatial gradient distribution of residual stress to be perceived, thus overcoming the limitation of traditional point-by-point measurement which can only obtain local isolated readings; capturing the cumulative evolution of acoustic emission signals along the time axis using a temporal convolutional network, allowing the temporal cumulative effect of impact events to be included in the prediction, rather than relying solely on the instantaneous signal amplitude; and coordinating the spray gun trajectory and straightening force using a two-layer game structure, enabling the stress symmetry target and the straightening force minimization target to be synergistically optimized within the same framework, avoiding mutual interference between the two. The synergistic effect of the three key technologies is that spatial modeling provides information on the location and degree of stress asymmetry, temporal modeling provides information on the trend of stress evolution, and game control transforms these two types of information into specific spray gun trajectory adjustment and straightening force commands, forming a complete closed loop of perception, prediction, and intervention, so that bending deformation is suppressed at the nascent stage, rather than being passively dealt with after deformation has formed.

[0064] It should be noted that during shot peening of precision long shafts with large length-to-diameter ratios, when there is a slight imbalance in the distribution of shot impact energy across different cross-sections along the shaft axis, the magnitude and direction of the accumulated residual compressive stress at each cross-section will exhibit a gradual deviation. This deviation is difficult to detect by macroscopic measuring instruments in the early stages of processing. However, as the shot peening process progresses, the asymmetric residual stress continuously accumulates on the shaft cross-section, eventually forming a bending moment sufficient to induce bending deformation within the shaft. This results in the shaft's straightness error exceeding tolerance after the strengthening process. At this point, the bending deformation has penetrated deep into the shaft, and simply relying on external force for straightening cannot eliminate the internal residual asymmetric stress. The shaft often springs back after unloading, making it difficult to maintain the straightening effect. The reason for the above technical problems is that the spatial distribution of shot peening impact force is affected by multiple factors such as the spray gun trajectory, air pressure fluctuations, and shaft rotation speed. A slight drift in any single parameter will produce significant cross-sectional stress asymmetry after axial integration. Traditional processes lack real-time sensing methods for the spatial distribution of the residual stress field and cannot issue intervention signals in the early stages of asymmetric trend formation. Therefore, passive straightening can only be performed after deformation has occurred. The usual solution to the aforementioned technical problems is to add an intermediate inspection process, which involves periodically stopping the machine during shot peening to measure residual stress at a designated cross-section using an X-ray diffractometer. Process parameters are then manually adjusted based on the measurement results before continuing processing. However, this method has two fundamental drawbacks: first, the stoppage inspection disrupts the sequential continuity of the shot peening process, and the repositioning of the spray gun after each stop introduces stress discontinuities at the junctions; second, X-ray diffraction measurements can only obtain local stress values ​​at a designated cross-section and cannot reconstruct the overall stress distribution. Adjusting overall process parameters based on local readings suffers from severe spatial sampling deficiencies, and adjustment decisions rely entirely on operator experience, lacking a systematic basis. This invention effectively solves this technical problem. The physical information-constrained spatiotemporal jump graph attention model uses multi-channel acoustic emission signal sequences as a real-time sensing means. The graph attention network encodes the spatial topological relationship of the discrete grid nodes in the entire field into a graph structure, enabling the model to continuously output the predicted value of the residual stress field in the entire field with a millisecond-level response cycle without stopping the machine or interfering with the processing. When the model detects a stress asymmetry trend in a certain area, the shot peening strengthening process game control model immediately adjusts the pitch of the shot peening gun and drives the follow-up reverse hydraulic straightening device to apply a reverse straightening force at the section with the maximum deformation. This cancels the accumulation of asymmetric stress before the bending moment of the section exceeds the elastic limit, thereby ensuring that the shaft remains within the allowable straightness range throughout the strengthening process. This fundamentally solves the defect of traditional stop-and-detection methods that cannot achieve continuous real-time intervention.

[0065] Specifically, the principle of this invention is:

[0066] The fundamental reason why this invention can solve the above-mentioned technical problems is that it constructs a closed-loop feedback control system with real-time prediction of residual stress field as the core, so that the adjustment of process parameters and the application of straightening force can be effectively intervened before the bending deformation develops significantly.

[0067] First, the physical information-constrained spatiotemporal jump graph attention model uses a graph attention network as the backbone of spatial modeling. It explicitly encodes the spatial topological relationships of discrete grid nodes on the long-axis surface into a graph structure. The attention weights between nodes are jointly modulated by Euclidean distance and residual stress gradient, enabling the model to perceive the gradient distribution of residual stress in the radial and axial directions, rather than relying solely on independent readings from local measurement points. The fundamental difference between this spatial modeling approach and traditional point-by-point measurement methods lies in its ability to capture the spatial propagation path of asymmetric stress accumulation, thus issuing an early warning signal before local stress anomalies evolve into macroscopic bending deformation.

[0068] Secondly, the temporal convolutional network expands the receptive field layer by layer through dilated convolution, incorporating the cumulative effect of projectile impact events over time into the prediction. This enables the model to distinguish between short-term impact disturbances and persistent stress accumulation trends, avoiding misjudgments caused by instantaneous noise. The spatiotemporal jump connection mechanism preserves the underlying geometric features and initial residual stress information, preventing deep networks from forgetting initial constraints and ensuring that the predicted values ​​always use the initial stress state as a reference throughout the entire reinforcement process.

[0069] Furthermore, the physical constraint fusion module embeds the constitutive equations of metal elastoplastic mechanics and the partial differential equations of heat conduction into the loss function, so that the model prediction results mathematically satisfy the stress tensor balance and energy conservation conditions. Thus, even under industrial sparse data conditions, the model can still output physically reasonable residual stress field predictions, rather than relying solely on empirical extrapolation from data fitting.

[0070] Finally, the shot peening strengthening process game-theoretic control model coordinates the shot blasting trajectory and straightening force using a two-layer game structure. The upper layer adjusts the pitch with the goal of maximizing stress symmetry, while the lower layer controls axial runout with the goal of minimizing straightening force. The two objectives are linked through a coupling stiffness coefficient, enabling the local stress asymmetry caused by the shot blasting trajectory to drive the straightening force adjustment in real time, forming a collaborative closed loop between the shot peening process and the straightening device. It is this system, based on real-time prediction and using proactive intervention, that enables the present invention to suppress bending deformation before it accumulates beyond tolerance, fundamentally solving the technical problem of difficulty in real-time suppression of shaft bending deformation caused by uneven residual stress field distribution.

[0071] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0072] The specific implementation method of step S01 is as follows.

[0073] After surface pretreatment to remove oxide layers and contaminants from the precision long shaft, three-dimensional measuring points are evenly distributed along the entire length of the shaft, with a spacing of 20–50 mm between the measuring points. A coordinate measuring machine is used to obtain the spatial coordinates of each measuring point, forming a coordinate point set. ,in The total number of measurement points. , , The first The coordinate components of each measuring point in three-dimensional space are given, in mm. Using the set of coordinate points as nodes, the surface along the major axis is divided into several continuous triangular or quadrilateral mesh elements, forming a discrete mesh model. The mesh adjacency matrix is... Defined as:

[0074] ;

[0075] In the formula, For the first The coordinate vectors of each node, in mm. For nodes With nodes The Euclidean distance between them is in mm. This is the adjacent distance threshold, typically set to 1.5 times the distance between measuring points, in mm. For the adjacency matrix, the first... Line 1 Column elements, with values ​​of 0 or 1. Adjacency matrix. A compressed sparse row format is used for storage to save memory. Initial residual stress distribution data are acquired by X-ray diffraction or blind hole method at measurement points corresponding to grid nodes on multiple sets of long-axis specimens. The measurement results are statistically analyzed to form an initial stress label vector. , of which Each component For the first The initial residual stress values ​​of each node, in MPa, are used as the baseline input for subsequent model training.

[0076] The specific implementation method of step S02 is as follows.

[0077] Three to eight sets of hydraulic support units are arranged axially along the long axis, each set equipped with a displacement sensor and a pressure sensor, and the support stiffness adjustment range is [range missing]. ~ The response time is no more than 20ms. The sensor collects the axial runout of each cross-section in real time. (Unit: mm) and bending strain, the control target is to The thickness should be controlled within the range of 0.01 to 0.05 mm. For supporting unit numbering, , This represents the total number of support units. The axial runout control threshold is determined by regression analysis, with a straightness error not exceeding 0.05 mm / m as a constraint, establishing a mapping relationship between runout and straightness error:

[0078] ;

[0079] In the formula, This refers to straightness error, expressed in mm / m. , , The regression coefficients were obtained by fitting orthogonal experimental data of specimens with different aspect ratios (50–200) using the least squares method. The unit is mm / m. Units are , Units are This represents the regression error term, in mm / m. Using mm / m as a constraint, the reverse calculation... The upper limit is the control threshold of 0.01 to 0.05 mm.

[0080] The specific implementation method of step S03 is as follows.

[0081] Based on the process timing parameters output by the physical information-constrained spatiotemporal jump graph attention model, the air pressure is set to 0.3–0.6 MPa, and the spray gun moving speed is... The speed ranges from 50 to 200 mm / s, with the long shaft rotating at a speed of 100 mm / s. The speed is 30–120 r / min. The spray gun moves at a constant speed along the axial direction while its long axis rotates at a constant speed. The coverage path unfolds as a spiral on the surface of the long axis, with a pitch of… Determined by the following formula:

[0082] ;

[0083] In the formula, The pitch of the helix is ​​expressed in mm / r. The spray gun movement speed is expressed in mm / s. The speed is the long shaft speed, expressed in r / s, with a pitch range of 3–15 mm / r.

[0084] The specific implementation method of step S04 is as follows.

[0085] A multi-channel acoustic emission sensor (sampling frequency of 1–5 MHz, number of channels of 4–8) acquires acoustic emission time-domain signal sequences in real time. Feature components in the 50–500 kHz frequency band are extracted via wavelet packet decomposition to form a feature matrix. ,in The number of sensor channels (4-8). The number of time sampling points is denoted as . The model's input layer receives the coordinates of three-dimensional measurement points, process timing parameters, and acoustic emission signal sequences. These three types of inputs are mapped to a 128-dimensional feature space via independent linear projection layers. The projection transformation is expressed as follows:

[0086] ;

[0087] ;

[0088] ;

[0089] In the formula, The original feature vector of coordinates is formed by flattening and concatenating the coordinates of all nodes. This is the original feature vector of process parameters, containing three components: air pressure, spray gun moving speed, and rotation speed. Acoustic emission characteristic matrix Flattened vector , , These are the corresponding projection weight matrices. , , These are the corresponding bias vectors. , , These are the projected 128-dimensional feature vectors.

[0090] The spatial modeling module uses a graph attention network with 8 multi-head attention heads, each with 16-dimensional features. Nodes in a layered graph attention network For nodes attention weights The stress is modulated by the Euclidean distance between nodes and the residual stress gradient, and the calculation formula is as follows:

[0091] ;

[0092] In the formula, For the first The layer can learn attention vectors with a dimension of 256 (the dimension of the concatenation of two 128-dimensional features). For the first Layer nodes eigenvectors Represents vector concatenation operation This is the distance modulation coefficient, in units of The gradient modulation coefficient is expressed in units of 1000 ppm. For nodes With nodes The magnitude of the residual stress gradient between them is determined by Calculated, unit is For nodes The set of neighboring nodes This is the neighborhood node index, which corresponds to the support unit number in step S02. The meanings differ, and should be distinguished by context. For dimensionless attention weights, satisfying The function is a linear rectified function with leakage, defined as outputting the original value when the input value is greater than 0, and outputting the original value multiplied by a leakage coefficient when the input value is less than or equal to 0. The empirical value of the leakage coefficient is 0.2. The attention weight threshold ranges from 0.05 to 0.15. Weights below the threshold are reset to zero to achieve sparse storage. All terms in the exponential function are dimensionless. and All are dimensionless quantities.

[0093] The formula for feature aggregation at each layer of a graph attention network is:

[0094] ;

[0095] In the formula, For the first Layer nodes The output feature vector As the activation function, use For the first The learnable aggregated weight matrix of each layer. The graph attention network has 3 stacked layers. The output of each layer is normalized to output 128-dimensional node features. The spatiotemporal skip connection establishes a direct additive connection between the outputs of the 1st and 3rd layers:

[0096] ;

[0097] In the formula, and These are the output feature vectors of layers 1 and 3 of the graph attention network, respectively. This is the feature vector after skip fusion.

[0098] The temporal modeling module uses a temporal convolutional network with a kernel size of 3 and a dilation coefficient of 1. The number of dilated convolutional layers increases progressively, with a total of 4 layers, each with 128 output channels. The formula for the output of layer dilated convolution is:

[0099] ;

[0100] In the formula, For temporal convolutional networks Layer at time The output feature vector For the first Layer convolution kernel, kernel size is 3 Indicates the coefficient of thermal expansion is The dilated convolution operation, For the first The output features of the layer, when Input time is For the first Layer bias vector This is a linear rectified activation function, defined as outputting the original value when the input value is greater than 0, and outputting 0 otherwise. A skip connection is established between layer 1 and layer 4.

[0101] ;

[0102] In the formula, and These are the output feature vectors of the 1st and 4th layers of the temporal convolutional network, respectively. This is the feature vector after skip fusion.

[0103] The physical constraint fusion module embeds the constitutive equations of metal elastoplastic mechanics and the one-dimensional partial differential equation of heat conduction as regularization terms into the loss function. The prediction mean square error term is:

[0104] ;

[0105] In the formula, For the first The predicted residual stress values ​​for each node are in MPa. For the first Residual stress monitoring labels for each node, in MPa. For the first The predicted hardened layer depth for each node, in mm. For the first Supervision labels for the hardened layer depth of each node, in mm. This is the dimensionless adjustment coefficient, with units of . Experience value This unifies the two dimensions into The physical constraint regularization term is:

[0106] ;

[0107] In the formula, For the first Predicted shear stress values ​​for each node, in MPa. For the first Temperature prediction values ​​for each node, in Kelvin (K). This is the thermal diffusivity, in units of... Time, in seconds Units are , Units are The first dimension is Units are , Units are The dimension of the second term in parentheses is , This is the dimensionless adjustment coefficient, with units of . Right now The empirical value is taken as the calibration value that makes the two orders of magnitude comparable, so that Overall dimensions are The total loss function is:

[0108] ;

[0109] In the formula, Units are Units are This is the regularization coefficient, in units of... The range is 0.01 to 0.1. Determined by cross-validation, making Dimensions are unified as The output layer consists of two parallel fully connected layers that output the predicted residual stress field vectors for each node. (Unit: MPa) and vector of predicted hardened layer depth (Unit: mm), dimension is consistent with the number of discrete grid nodes.

[0110] Model output and Adaptive learning rate adjustment function The motion controller is updated after processing; the function is defined as follows:

[0111] ;

[0112] In the formula, The learning rate adjustment coefficient is dimensionless. The root mean square error of residual stress in the current validation set, in MPa. The normalized acoustic emission signal energy is dimensionless and ranges from 0 to 1. This is the dimensionless adjustment coefficient, with units of . The range of values ​​is ~ Recorded from 5 to 10 sets of verification experiments under different working conditions and The typical distribution range, with Falling into under most operating conditions The interval is obtained by reverse calculation of the target, so that Maintain dimensionless. When The learning rate remains constant when Multiply the learning rate by 0.9, when Multiply the learning rate by 0.7, when Learning rate reset .

[0113] The shot peening enhancement process game regulation model adopts a two-layer game structure, with the upper-layer objective function being:

[0114] ;

[0115] In the formula, The average value of the predicted residual stress field is expressed in MPa. Units are Right now The denominator unit is Units are The objective is to maximize the symmetric optimization; the constraint is as follows. mm / r and the residual stress difference between adjacent nodes does not exceed 50 MPa. The lower-level objective function is:

[0116] ;

[0117] In the formula, The unit is N (unit: N). For the first The axial runout at the support unit, in mm, is the same as in step S02. Consistent in meaning This is the coupling stiffness coefficient, in units of... The value range is 100 to 500. The determination was made by least-squares fitting through orthogonal experiments with different combinations of axial runout and different straightening forces. The unit is N, multiplied by a dimension adjustment factor of 1mm / 1N to get mm. The overall dimension is unified to N, and the second term is revised to... The unit is N, meaning the lower-level objective function is modified as follows:

[0118] ;

[0119] The constraints are mm and N, the follow-up reverse hydraulic straightening device is based on the output of the lower-level model. Active alignment is implemented, with the point of application of alignment force moving synchronously with the position of the spray gun.

[0120] The specific implementation method of step S05 is as follows.

[0121] After shot peening, a dynamic spatial octree partitioning load-balancing parallel scheduling algorithm is initiated to perform parallel post-processing of the full-field discrete element method simulation results. This algorithm dynamically reconstructs the spatial octree based on the local density of the projectile in the simulation space, dividing high-density collision regions into finer subdomains and merging sparse regions into larger subdomains. The subdomain partitioning objective is:

[0122] ;

[0123] In the formula, Total number of parallel computing nodes For the first The computation time of each parallel computing node, in seconds. The average computation time for each node is expressed in seconds. An octree subdomain partitioning scheme was adopted; the subdomain partitioning threshold (i.e., the maximum number of projectiles within a single subdomain) ranged from 500 to 2000, determined by solving the aforementioned variance minimization problem in preliminary experiments. An asynchronous message communication mechanism was used to eliminate waiting between nodes, verify the symmetry of the residual stress distribution on the long axis surface, and perform supplementary spraying on areas that did not meet the standards, with a supplementary spray pressure of 0.3–0.5 MPa.

[0124] The specific implementation method of step S06 is as follows.

[0125] For long shaft parts that have undergone shot peening, 3 to 5 sections are uniformly cut along the axial direction. For each section, the hardness distribution is measured radially from the surface to the inside using a micro Vickers hardness tester. The depth corresponding to the point where the hardness value drops to 110% of the matrix hardness is taken as the hardened layer depth. The qualified criteria are that the hardened layer depth is not less than 0.2 to 0.5 mm and the straightness error does not exceed 0.05 to 0.1 mm / m. After passing the test, rust prevention and sealing treatment is carried out.

[0126] To better understand and implement this invention, a specific application scenario of this invention is provided below as Example 2: This example uses a precision transmission guide shaft as the object. The material is 42CrMo alloy steel, with a nominal diameter of 40mm, a total length of 5000mm, and an aspect ratio of 125, which is a typical large aspect ratio precision long shaft component. Before processing, the surface of the guide shaft is cleaned and the oxide layer is removed to ensure that the surface condition meets the requirements of the shot peening strengthening process.

[0127] Measurement points were evenly distributed along the entire length of the smooth rod axis at 30mm intervals, totaling 168 three-dimensional measurement points. The spatial coordinates of each measurement point were acquired using a coordinate measuring machine, forming a three-dimensional measurement point coordinate set. The adjacency matrix was stored in a compressed sparse row format to establish a discrete mesh model of the smooth rod axis. Figure 2 As shown. The initial residual stress distribution data were obtained by X-ray diffraction on 6 groups of samples from the same batch. The measurement points corresponded to the discrete grid nodes, and the statistical mean was used as the initial stress label for the physical information constrained spatiotemporal jump graph attention model.

[0128] Five sets of follow-up hydraulic support units are evenly arranged along the axis of the polished rod. Each unit is equipped with a displacement sensor and a pressure sensor, with a response period of 15ms and a support stiffness adjustment range of [missing information]. ~ After the system is started, each sensor continuously collects axial runout and bending strain. The controller adjusts the support stiffness in real time based on the regression analysis mapping relationship to keep the axial runout of the entire shaft within 0.03mm.

[0129] The physical information-constrained spatiotemporal jump graph attention model, based on the current predicted residual stress field and acoustic emission signal energy, adjusts the learning rate using an adaptive learning rate function. Output process timing parameters: spray gun air pressure is set to 0.45 MPa, spray gun moving speed to 120 mm / s, guide shaft speed to 60 r / min, and helical pitch to 8 mm. A 4-channel piezoelectric sensor with 2... Acoustic emission signal sequences are acquired at sampling frequency, and 50–500 MHz are extracted after wavelet packet decomposition. Frequency band feature components, along with 3D measurement point coordinates and process timing parameters, are input into the model. The model's graphical attention network consists of 3 stacked layers, with 8 multi-head attention heads, each with 16-dimensional features; the temporal convolutional network's dilation coefficient is calculated according to... Increasing layer by layer; physical constraint regularization coefficient The value is set to 0.05. During this shot peening process, the model continuously outputs the predicted value of the global residual stress field with a period of 15ms. The spatial distribution of the residual stress field at a typical moment is shown below. Figure 3 As shown.

[0130] The shot peening strengthening process game-theoretic control model receives real-time predictions of the residual stress field. The upper objective function fine-tunes the pitch based on the principle of maximizing stress symmetry, with an adjustment range of 6–10 mm. The lower objective function drives the follow-up reverse hydraulic straightening device based on the principle of minimizing straightening force, with a coupling stiffness coefficient. Take 200 The straightening force was controlled within the range of 800–2200 N. During the entire strengthening process, the model triggered straightening intervention 12 times, with each intervention lasting approximately 200 ms. The straightness error of the shaft was consistently maintained within 0.04 mm / m.

[0131] After shot peening, a dynamic spatial octree partitioning load balancing parallel scheduling algorithm was initiated, with the maximum number of projectiles in the subdomain set to 1200. An asynchronous message communication mechanism was used to perform parallel post-processing of the full-field discrete element method simulation results, verifying the symmetry of the residual stress distribution across the entire shaft surface. Verification revealed two nodes in the middle section of the shaft with residual stress differences exceeding 50 MPa, classifying them as substandard areas. These areas underwent supplementary shot peening with an injection pressure of 0.4 MPa. After supplementary peening, the full-field symmetry was satisfied. Surface hardness testing was performed on the shot-peened shaft. Four sections were uniformly cut along the axial direction, and the hardness distribution on each section was measured radially from the surface inwards using a micro Vickers hardness tester. The depth corresponding to the point where the hardness value decreased to 110% of the matrix hardness was taken as the hardened layer depth. The test results are shown in Table 1.

[0132] Table 1 Summary of Hardened Layer Depth and Straightness Errors at Various Test Sections of the Smooth Rod Axis

[0133]

[0134] The depth of the hardened layer on all cross sections is not less than 0.2 mm, and the straightness error does not exceed 0.1 mm / m. After passing the inspection, the smooth shaft is subjected to anti-rust sealing treatment.

[0135] like Figure 3 As shown, the spatial distribution diagram of the residual stress field intuitively presents the magnitude and gradient distribution of the residual compressive stress at each grid node during the shot peening process. It can be seen that after active intervention, the symmetry of the residual stress distribution in the circumferential and axial directions of the shaft is significantly improved, and local stress concentration areas are promptly identified and eliminated through compensated shot peening.

[0136] Compared to traditional stop-and-test methods and post-operative straightening, this invention achieves the following technological advancements: Traditional methods rely on local X-ray diffraction sampling, which can only obtain isolated stress readings of discrete sections and cannot reconstruct the spatial distribution of the entire field. In contrast, this invention uses a graph attention network to explicitly encode the spatial topological relationship of the grid nodes across the entire field, enabling the continuous distribution of the residual stress field to be reconstructed in real time, fundamentally breaking through the spatial resolution limitations of local sampling. Traditional methods suffer from temporal interruptions between stop-and-test and resumption of work. In contrast, this invention uses acoustic emission signals as a continuous sensing method, with the model continuously outputting predicted values ​​at millisecond intervals, ensuring the shot peening process is uninterrupted and guaranteeing the continuity and controllability of stress field accumulation. Traditional straightening methods apply external forces after deformation occurs, and the straightening effect is difficult to maintain in the long term due to the elastic rebound of the shaft components. In contrast, this invention uses an artificial intelligence model to detect the stress asymmetry trend in advance and applies a reverse straightening force before the bending moment of the section exceeds the elastic limit, allowing the suppression effect to intervene in the nascent stage of deformation and fundamentally avoiding the rebound problem.

[0137] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A shot peening strengthening process for improving the surface hardness of precision long shaft parts, characterized in that, Includes the following steps: The surface of the precision long shaft is pretreated, the coordinates of three-dimensional measuring points along the entire length of the long shaft are measured, a discrete mesh model of the long shaft is established, and the initial residual stress distribution data of the long shaft is collected as the reference input for subsequent process parameter control. The online dynamic load compensation support system is activated, and multiple sets of follow-up hydraulic support units are arranged along the long axis. Sensors collect axial runout and bending strain in real time, and the support stiffness of each support unit is dynamically adjusted to control the axial runout within the axial runout threshold range. Based on the process timing parameters output by the physical information-constrained spatiotemporal jump graph attention model, the air pressure, spray gun moving speed and long shaft rotation speed are set, and the shot peening gun is driven to continuously shot peening the long shaft surface along the spiral trajectory. During the shot peening process, a multi-channel acoustic emission sensor collects the acoustic emission signal sequence in real time. The physical information constrained spatiotemporal jump graph attention model outputs the residual stress field prediction value in real time based on the acoustic emission signal sequence and the three-dimensional measurement point coordinates. The residual stress field prediction value is input into the shot peening process game control model to drive the follow-up reverse hydraulic straightening device to actively suppress bending deformation. After shot peening, a dynamic spatial octree partitioning load balancing parallel scheduling algorithm is activated to perform parallel post-processing on the simulation results of the full-field discrete element method, verify the symmetry of the residual stress distribution on the long axis surface, and carry out supplementary shot peening on the areas that do not meet the standards.

2. The shot peening strengthening process for improving the surface hardness of precision long shaft parts according to claim 1, characterized in that, The three-dimensional measuring point coordinates refer to the set of spatial coordinate points uniformly distributed along the entire length of the long axis, and the measuring point spacing is the measuring point spacing value, which is used to describe the spatial geometry of the long axis in the processing state.

3. The shot peening strengthening process for improving the surface hardness of precision long shaft parts according to claim 2, characterized in that, The long-axis discrete mesh model uses the coordinates of three-dimensional measurement points as nodes to divide the long-axis surface into several continuous triangular or quadrilateral mesh units, which are used to support the spatial modeling input of the graph attention network.

4. The shot peening strengthening process for improving the surface hardness of precision long shaft parts according to claim 3, characterized in that, The initial residual stress distribution data were obtained by X-ray diffraction or blind hole method on multiple sets of long-axis specimens. The measurement points corresponded to the discrete grid nodes, and the measurement results were statistically analyzed and used as the initial stress labels for the model training dataset.

5. The shot peening strengthening process for improving the surface hardness of precision long shaft parts according to claim 4, characterized in that, The online dynamic load compensation support system consists of multiple sets of hydraulic support units arranged along the axial direction. Each set of support units is equipped with a displacement sensor and a pressure sensor. The support stiffness adjustment range is the support stiffness value, and the support stiffness adjustment response time does not exceed the response time threshold.

6. The shot peening strengthening process for improving the surface hardness of precision long shaft parts according to claim 5, characterized in that, The spiral trajectory refers to the spray gun moving at a constant speed along the long axis while the long axis rotates at a constant speed around the axis. The spray gun coverage path unfolds into a spiral shape on the surface of the long axis, and the spiral pitch is determined by the spray gun moving speed and the long axis rotation speed.

7. The shot peening strengthening process for improving the surface hardness of precision long shaft parts according to claim 6, characterized in that, The input layer of the physical information constrained spatiotemporal jump graph attention model receives the coordinates of the full-length three-dimensional measurement points along the major axis, process timing parameters, and multi-channel acoustic emission signal sequences. The three types of inputs are mapped to the feature space through independent linear projection layers. The spatial modeling module uses a graph attention network, and the temporal modeling module uses a temporal convolutional network.

8. The shot peening strengthening process for improving the surface hardness of precision long shaft parts according to claim 7, characterized in that, In the graph attention network, the attention weights are modulated by the Euclidean distance between nodes and the residual stress gradient. The graph attention network is stacked in multiple layers, and the output of each layer is normalized to output node features.

9. The shot peening strengthening process for improving the surface hardness of precision long shaft parts according to claim 8, characterized in that, The dilation coefficient of the temporal convolutional network increases layer by layer. The temporal convolutional network captures the evolution of the acoustic emission signal sequence on the time axis. The spatiotemporal jump connection mechanism establishes direct additive jump connections between the bottom and top outputs of the graph attention network and between the bottom and top layers of the temporal convolutional network.

10. The shot peening strengthening process for improving the surface hardness of precision long shaft parts according to claim 9, characterized in that, The physical constraint fusion module of the physical information constrained spatiotemporal jump graph attention model embeds the metal elastoplastic mechanical constitutive equation and the one-dimensional heat conduction partial differential equation as regularization terms into the loss function, and the regularization coefficient is determined by cross-validation experiments.

Images