Optimization design method and system for forming process of lightweight caliper piston

Abstract

The application discloses an optimization design method and system for a forming process of a lightweight caliper piston, which comprises the following steps: determining an upper limit threshold of a grain size of a fillet area according to a target piston fatigue life requirement, generating a pass parameter sequence in a feasible range of a multi-pass forming parameter, and constructing a training sample set through forming simulation and grain evolution calculation; a grain prediction network is established in a recursive manner, and the grain refinement amount and the growth amount are constrained and predicted through a forming refinement channel and a residence growth channel in sequence; the trained grain prediction network is used as a prediction tool, a gradient search is performed on the pass parameter sequence under the constraints of a total thinning amount and equipment capacity, and a minimum final grain size is taken as a target to obtain an optimized forming process scheme. The application can improve the optimization efficiency of the multi-pass forming process parameter and enhance the consistency between the grain size prediction and the actual forming rule.

Description

Technical Field

[0001] This application relates to the field of process optimization and control technology, specifically to an optimization design method and system for the molding process of a lightweight caliper piston. Background Technology

[0002] With increasing demands for lightweight automotive components, lightweight materials such as aluminum alloys are increasingly being used in braking system parts. Caliper pistons endure repeated hydraulic loads during operation, and the transition fillet area between the cup bottom and sidewalls is prone to high local stress. Therefore, in addition to dimensional accuracy and forming integrity, the microstructure after forming is also a crucial factor affecting its service performance. For thin-walled aluminum alloy cup-shaped pistons, plastic forming processes such as reverse extrusion and spinning are often employed, with material flow and wall thickness controlled through multiple deformation passes. During hot forming, aluminum alloys simultaneously undergo work hardening, dynamic recovery, dynamic recrystallization, and inter-pass grain growth, among other microstructural evolution behaviors. Grain size is influenced by parameters such as reduction rate, forming temperature, deformation rate, and residence time, and the material state formed in the previous pass continues to affect subsequent passes.

[0003] Current process design typically combines empirical parameters, finite element thermoforming simulations, and grain evolution models to calculate and compare different pass schemes. This method can obtain relatively detailed results on strain, temperature, and microstructure evolution; however, the computational load for a single high-fidelity simulation is large, and its efficiency is low when performing repeated calculations in a multi-pass, multi-parameter continuous space. In recent years, data-driven surrogate models have begun to be used to assist in optimizing metal forming parameters. However, ordinary prediction models mainly rely on sample fitting, and when the training samples are limited, it is difficult to guarantee that the grain evolution results for each pass consistently conform to the microstructure changes in the actual forming process. Therefore, how to balance the physical rationality of grain evolution prediction, computational efficiency, and multi-pass process parameter optimization capabilities remains a problem to be solved in the design of lightweight caliper piston forming processes. Summary of the Invention

[0004] In view of this, this disclosure provides an optimized design method and system for the molding process of lightweight caliper pistons, which can alleviate, reduce or even eliminate at least the above-mentioned problems.

[0005] According to one aspect of this application, an optimized design method for the molding process of a lightweight caliper piston is provided, comprising: Based on the fatigue life requirements of the target piston, determine the upper limit threshold of the grain size in the rounded corner region; Within the feasible range of multi-pass forming parameters, multiple sets of pass parameter sequences are sampled and generated. Forming simulation and grain evolution calculation are run on each set of pass parameter sequences to obtain the grain size of the rounded corner area after each pass and form a training sample set. A pass sequence grain prediction network is constructed and trained with the training sample set. The pass sequence grain prediction network is recursively applied step by step. Each time step corresponds to a pass and passes through the forming and refining channel and the dwell and growth channel in sequence. The forming and refining channel outputs a non-negative refining amount and subtracts it from the current grain size. The dwell and growth channel outputs a non-negative growth amount and adds it to the current grain size. Using the trained pass sequence grain prediction network as a prediction tool, the pass parameter sequence as an optimization variable, and the final grain size output by the network as the optimization target, under the constraints of total thinning amount and equipment capability, the pass parameter sequence that minimizes the final grain size is obtained through gradient search, which is then used as the optimized forming process scheme.

[0006] Optionally, the forming and refining channel and the dwell and growth channel each include a sub-network, and the output of each sub-network is processed by a non-negative activation function to obtain the refining amount and the growth amount, respectively; the non-negative activation function structurally constrains the output to always be non-negative, so that the grain size only decreases and does not increase in the forming stage and only increases and does not decrease in the dwell stage.

[0007] Optionally, the internal state of the pass sequence grain prediction network includes a grain size scalar and an auxiliary material state vector; the auxiliary material state vector participates in the input calculation of the forming refinement channel and the dwell growth channel, and is updated sequentially by the two channels in each time step.

[0008] Optionally, the input to the forming refinement channel is a concatenation of the grain size scalar from the previous time step, the auxiliary material state vector from the previous time step, and the forming parameters of the current pass; the input to the dwell growth channel is a concatenation of the updated grain size scalar from the forming refinement channel, the updated auxiliary material state vector, and the dwell parameters of the current pass.

[0009] Optionally, when training the pass sequence grain prediction network, the loss value is composed of a weighted sum of the final grain size prediction error term and the intermediate pass grain size prediction error term; the intermediate pass grain size prediction error term is calculated as the deviation between the predicted grain size after each pass and the corresponding simulated grain size in the training sample set.

[0010] Optionally, the loss value further includes a monotonicity regularization term, which penalizes the deviation of the grain size being greater than the input grain size after the output of the forming refinement channel and the deviation of the grain size being less than the input grain size after the output of the dwell growth channel.

[0011] Optionally, the parameters for each pass in the pass parameter sequence include reduction rate, forming speed, forming temperature, inter-pass dwell time, and dwell temperature; before inputting the pass sequence grain prediction network, each parameter is subjected to minimum-maximum normalization processing according to the upper and lower bounds of its sampling range.

[0012] Optionally, the gradient search adopts the projected gradient descent method, randomly generating multiple initial points from the parameter space and searching independently. After each gradient update, each parameter is projected back to the feasible range, and the reduction rate of each pass is adjusted so that their product satisfies the total thinning constraint. The result with the smallest final grain size that satisfies all constraints is taken as the optimized forming process scheme.

[0013] Optionally, the sub-networks in the forming and refining channel and the dwelling and growing channel are each two-layer fully connected networks. The first layer uses a modified linear unit activation function, and the second layer has no activation function. The output of the second layer is split into a scalar component and a vector component. The scalar component is processed by the non-negative activation function to obtain the corresponding refining amount or growing amount. The vector component is processed by the bounded activation function and used as the update increment of the auxiliary material state vector.

[0014] According to another aspect, an optimized design system for the molding process of a lightweight caliper piston is provided, comprising: The grain size threshold determination module is used to determine the upper limit threshold of the grain size in the rounded corner region based on the fatigue life requirements of the target piston. The training sample construction module is used to sample and generate multiple sets of pass parameter sequences within the feasible range of multi-pass molding parameters, run molding simulation and grain evolution calculation on each set of pass parameter sequences, obtain the grain size of the rounded corner area after each pass, and form a training sample set. The grain prediction network construction and training module is used to construct a pass sequence grain prediction network and train it with the training sample set. The pass sequence grain prediction network is recursively calculated step by step, with each time step corresponding to a pass and passing through the forming and refining channel and the dwell and growing channel in sequence. The forming and refining channel outputs a non-negative refining amount and subtracts it from the current grain size. The dwell and growing channel outputs a non-negative growth amount and adds it to the current grain size. The molding process optimization module uses the trained pass sequence grain prediction network as a prediction tool, the pass parameter sequence as an optimization variable, and the final grain size output by the network as the optimization target. Under the constraints of total thinning amount and equipment capability, it uses gradient search to find the pass parameter sequence that minimizes the final grain size, which is then used as the optimized molding process scheme.

[0015] According to the method in this application, addressing the problem that the grain size in the rounded corner region is affected by the coupling of parameters in consecutive passes during the multi-pass forming process of lightweight caliper pistons, and the computational burden of traditional simulation optimization is large, the method embeds forming refinement and inter-pass dwell growth into the pass sequence grain prediction network. By constraining the direction of change of refinement and dwell growth through non-negative activation, and combining this with the auxiliary material state to recursively extrapolate between adjacent passes, the prediction process is made consistent with the actual grain evolution law in forming. Simultaneously, using the trained prediction network as a differentiable surrogate model, under the constraints of total thinning amount and equipment capacity, the method directly uses the gradient of the final grain size with respect to the pass parameters to search for a forming scheme. The optimization results are verified through forming simulation and grain evolution calculation. When the prediction deviation exceeds a set threshold, samples are added and retraining is performed. This reduces the burden of repeated multi-parameter simulation calculations while improving the physical rationality and engineering feasibility of the obtained forming scheme, enabling the final grain size in the rounded corner region to be specifically controlled according to fatigue life requirements. Attached Figure Description

[0016] This application will be more fully understood by referring to the following detailed description of specific embodiments in conjunction with the accompanying drawings, wherein the same reference numerals in the drawings refer to the same elements, wherein: Figure 1 A flowchart illustrating an optimized design method for the molding process of a lightweight caliper piston according to some embodiments of the present disclosure is shown.

[0017] Figure 2 The diagram schematically illustrates the cross-channel recursive relationship and dual-channel structure of the prediction network.

[0018] Figure 3 A schematic block diagram of an optimized design system for the molding process of a lightweight caliper piston according to some embodiments of the present disclosure is shown.

[0019] Figure 4 A schematic diagram of the structure of an electronic device according to some embodiments of the present disclosure is shown. Detailed Implementation

[0020] Several embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings to enable those skilled in the art to understand and implement the disclosure. However, the present disclosure may be embodied in many different forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided herein to make the disclosure comprehensive and complete, and to fully convey the scope of the disclosure to those skilled in the art. These embodiments are illustrative and not limiting of the disclosure.

[0021] It will be understood that although the terms first, second, third, etc., may be used herein to describe various elements, steps, and / or parts, these elements, steps, and / or parts should not be limited by these terms. These terms are used only to distinguish one element, step, or part from another. Therefore, the first element, step, or part discussed below may be referred to as the second element, step, or part without departing from the teachings of this disclosure.

[0022] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit this disclosure. As used herein, the singular forms “a,” “an,” and “the” are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms “comprising” and / or “including” as used in this specification specify the presence of a feature, integral, step, operation, element, and / or component, but do not exclude the presence of one or more other features, integrals, steps, operations, elements, components, and / or groups thereof. As used herein, the term “and / or” includes any and all combinations of one or more of the associated listed items.

[0023] Unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure pertains. It will be further understood that terms such as those defined in commonly used dictionaries should be interpreted as having meanings consistent with their meanings in the relevant field and / or the context of this specification, and will not be interpreted in an idealized or overly formal sense unless expressly defined herein.

[0024] It should be noted that, provided there is no conflict, features in different embodiments can be combined and used arbitrarily.

[0025] This embodiment provides an optimized design method for the forming process of lightweight caliper pistons, applicable to multi-pass reverse extrusion or spinning forming scenarios for thin-walled cup-shaped aluminum alloy caliper pistons. Typical applications include 6061 or 7075 aluminum alloy brake caliper pistons with an outer diameter of 30 to 60 mm and a wall thickness of 2 to 4 mm. Implementing this method typically involves: a computational platform capable of running finite element thermoforming simulations and grain evolution calculations; calibrated constitutive model parameters and recrystallization kinetic parameters for the target aluminum alloy; and clearly defined geometric dimensions, wall thickness requirements, and fatigue life indices for the target piston. The method ultimately outputs a set of optimized multi-pass forming parameter sequences, minimizing the final grain size of the rounded section (hereinafter referred to as the rounded corner region) where the piston cup bottom transitions to the sidewall while meeting geometric accuracy requirements.

[0026] The implementation process of the method described in this application will be described in detail below with reference to specific embodiments. It should be noted that this embodiment is only used to explain this application and is not intended to limit the scope of protection of this application. Conventional adjustments or substitutions of each step by those skilled in the art without departing from the concept of this application should be included in the scope of protection of this application.

[0027] Please see Figure 1 , Figure 1 A flowchart illustrating an optimized design method for the molding process of a lightweight caliper piston according to some embodiments of the present disclosure is shown.

[0028] In step S1, the upper limit threshold of the grain size in the rounded corner region is determined based on the fatigue life requirements of the target piston. In one embodiment, the caliper piston repeatedly bears braking hydraulic loads during use, and the rounded corner region is a stress concentration location where fatigue cracks preferentially initiate. A Hall-Petch relationship exists between the fatigue limit and grain size of aluminum alloys.

[0029] in The fatigue limit is expressed in MPa. The stress is the lattice friction stress, in MPa. The Hall-Petch coefficient is the fatigue coefficient, in MPa·μm. ; The values ​​represent the average grain size in μm. These material constants were obtained from material handbooks or calibrated using standard fatigue tests. Based on the stress amplitude requirement corresponding to the target fatigue life, the upper limit threshold for the grain size in the fillet region is obtained through inverse kinematics. For example, for a 6061-T6 aluminum alloy piston that needs to withstand 1 million cycles of loading, this threshold can be determined to be 25 μm.

[0030] In step S2, multiple sets of pass parameter sequences are sampled and generated, and forming simulation and grain evolution calculations are run to form a training sample set. Within the feasible range of multiple pass forming parameters, multiple sets of pass parameter sequences are sampled and generated. Forming simulation and grain evolution calculations are run on each set of pass parameter sequences to obtain the grain size of the rounded corner area after each pass, thus forming a training sample set.

[0031] In some embodiments, a finite element thermodynamic simulation model of the target piston reverse extrusion molding is established, and a grain evolution calculation subroutine is embedded in the simulation solution. After each simulation increment step, this subroutine updates the grain size based on the strain, strain rate, and temperature history of the elements. A dynamic recrystallization model is used in the molding stage: recrystallization is triggered when the local cumulative strain exceeds the critical strain, and the recrystallized grain size is determined by the Zener-Hollomon parameters. Decide:

[0032]

[0033] in The equivalent rate of change is expressed in seconds. ; The activation energy for thermal deformation is expressed in J / mol. The gas constant is taken as 8.314 J / (mol·K); This is absolute temperature, in Kelvin (K). For 6061 aluminum alloy, For example, take 145 kJ / mol. For example, take 650 μm·s , For example, the value of 0.19 was obtained through regression analysis of thermal simulation compression tests under different temperature and strain rate conditions.

[0034] In one embodiment, the inter-pass dwell stage employs a static grain growth model:

[0035] in Grain size at the start of the dwell time, in μm; The dwell time is expressed in seconds (s). The residence temperature is expressed in Kelvin (K). The activation energy for grain growth is expressed in J / mol. The velocity constant is expressed in μm. / s; The growth index. For 6061 aluminum alloy, For example, take 115 kJ / mol. For example, take 1.2 × 10 μm / s, For example, take 2.5.

[0036] Within the engineering feasibility of the pass / fail scheme, Latin hypercube sampling was used to generate multiple sets of pass / fail parameter sequences as sample points. Latin hypercube sampling is a hierarchical random sampling method that divides the value range of each parameter into several layers and then randomly selects a value from each layer, ensuring uniform sample distribution in the parameter space. The sampling range included: 3 to 6 passes, single-pass reduction rate of 10% to 40%, molding temperature of 573K to 723K, molding speed of 5 to 50 mm / s, and dwell time between passes of 5 to 60 s. A total of, for example, 300 sets of valid samples were generated. The simulation was run sequentially for each sample, and the average grain size of representative units in the fillet region was extracted after each pass. Each sample was labeled with the grain size sequence of the fillet region after each pass, and all samples constituted the training sample set.

[0037] In step S3, a pass sequence grain prediction network is constructed and trained using a training sample set. The pass sequence grain prediction network is constructed and trained using the training sample set. The network iterates step-by-step, with each time step corresponding to a pass and sequentially passing through a forming and refining channel and a dwell and growth channel. The forming and refining channel outputs a non-negative refining amount, which is subtracted from the current grain size. The dwell and growth channel outputs a non-negative growth amount, which is added to the current grain size.

[0038] According to the embodiments of this disclosure, the aforementioned co-simulation takes several hours per run, and searching for the optimal pass order in the multidimensional continuous parameter space requires hundreds of simulation calls, making the computational cost unacceptable in engineering. Therefore, a dedicated recurrent prediction network is constructed as a fast proxy model. The difference between this network and a general recurrent network lies in that it encodes the physical laws governing grain refinement during aluminum alloy hot forming and grain growth during dwell time into the network architecture through structural constraints. Even with limited training samples, it can still output predictions that conform to physical laws for unseen pass order combinations.

[0039] Specifically, the training sample set is divided into a training set and a validation set in an 8:2 ratio. For samples with different number of passes, the sequence length is padded to the maximum number of passes. (For example, take 6), set the padding position parameter value to 0 and mark it with a mask; ignore the loss calculation for padding positions during training. The parameters for each pass in each sample group include: indentation rate. Molding speed Molding temperature Stop time between tracks , residence temperature There are a total of 5 parameters. These 5 parameters are then subjected to minimum-maximum normalization based on their respective sampling ranges (upper and lower bounds). according to Mapped to the 0-1 range. Initial grain size. The grain size is normalized in the same way as the grain size labeled in each pass. After normalization, all values ​​are dimensionless, eliminating dimensional and numerical scale differences between different physical quantities.

[0040] The input format for each sample group after preprocessing is: dimension A two-dimensional tensor, with an additional scalar The annotation output is the length. A one-dimensional vector, where each component is the normalized grain size of each pass.

[0041] In one embodiment, the pass sequence grain prediction network treats the multi-pass forming process as a temporal process, recursively calculating step-by-step, with each time step corresponding to one pass. The network's internal state is determined by the grain size scalar. and auxiliary material state vector It consists of two parts. The dimension of the auxiliary material state vector is set to, for example, 15, to characterize potential microscopic information in the current material, such as dislocation density and subgrain structure, that influence subsequent recrystallization behavior but cannot be directly observed. Initial time. Take the normalized initial grain size , Initialize as a vector of all zeros.

[0042] The correlation between the pass parameter sequence and the grain size in the fillet region is that the deformation conditions of each pass determine the degree of triggering dynamic recrystallization and the fineness of the recrystallized grains, while the dwell conditions determine the magnitude of static grain growth. These conditions accumulate gradually along the pass sequence, and the material state at the end of the previous pass constitutes the initial conditions for the next pass, forming a causal dependency in the time series.

[0043] Within each time step, the update of the grain size state sequentially goes through two calculation stages: the forming refinement channel and the dwell growth channel. Please refer to [link / reference]. Figure 2 , Figure 2 The diagram schematically illustrates the cross-channel recursive relationship and dual-channel structure of the prediction network, as shown below. Figure 2 As shown, the pass-sequence grain prediction network iterates through each pass in turn, with each pass sequentially passing through a grain refinement channel and a dwell-growth channel. For the i-th pass, the grain refinement channel concatenates the grain size d(i-1) of the previous pass, the auxiliary material state h(i-1), and the reduction rate, forming speed, and forming temperature of the current pass. After processing by two fully connected layers, the output is split into scalar and vector components. The scalar component is processed by the Softplus function to obtain the non-negative refinement amount Δd. - This yields the grain size dF(i) after molding. The vector components are then processed by the tanh function to obtain the auxiliary state increment and update it to hF(i). Subsequently, the dwell growth channel concatenates dF(i) and hF(i) with dwell time and dwell temperature, and calculates the non-negative growth mass Δd through the corresponding sub-network. + The auxiliary state increment is used to obtain the final grain size d(i) and auxiliary material state h(i) of the current pass. During the forming stage, the grain size only decreases and does not increase, while during the dwell stage, the grain size only increases and does not decrease. The output state of the current pass continues to serve as the input for the next pass until the recursion of all effective passes is completed.

[0044] Specifically, in one embodiment, the grain refinement channel simulates the grain refinement effect of dynamic recrystallization during the current forming pass. The channel input is the grain size scalar from the previous time step. Auxiliary material state vector of the previous time step (Dimension 15) and current stage forming parameters (Dimension 3) concatenation, the total input dimension is 19.

[0045] The subnetwork within the channel is a two-layer fully connected network. Layer 1 has an input dimension of 19 and an output dimension of 32, with the activation function being a Corrected Linear Unit (ReLU), which takes the larger value between zero and itself. Layer 2 has an input dimension of 32 and an output dimension of 16, with no activation function. The 16-dimensional output of Layer 2 is split into two parts: the first component is a scalar. The remaining 15 components form a vector. .

[0046] right Applying the Softplus function yields the refinement amount. :

[0047] The output of the Softplus function is always positive, thus structurally ensuring the refinement amount. The grain size has been updated to:

[0048] in This represents the lower bound of the grain size in normalized coordinates, corresponding to an actual grain size of approximately 5 μm; for example, 0.05 in normalized coordinates. Because... , This holds true throughout, meaning that the grain size only decreases and never increases during the forming stage.

[0049] The auxiliary material state vector is updated as follows:

[0050] tanh is the hyperbolic tangent function, which is a bounded activation function. It constrains the update increment to between -1 and 1 to prevent the auxiliary state values ​​from diverging.

[0051] For example, suppose The first component of the output of the second layer of the subnetwork ,but After normalization, this value corresponds to a relatively large amount of actual refinement. It reaches the lower bound. After actual training convergence, the network learns... Typically between -2 and 2, corresponding to The range of 0.13 to 2.13 reasonably reflects the level of detail required for a single pass.

[0052] In one embodiment, the dwell-growth channel simulates the effect of static grain growth during the dwell period after a pass ends. The channel input is the grain size scalar after the forming refinement channel update. Updated auxiliary material state vector (Dimension 15) and current track dwell parameters (Dimension 2) concatenation, the total input dimension is 18.

[0053] The subnetworks within the channel are also two-layer fully connected networks: Layer 1 has an input dimension of 18 and an output dimension of 32, with the activation function being a modified linear unit; Layer 2 has an input dimension of 32 and an output dimension of 16, with no activation function. The output of Layer 2 is split in the same way as the shaping and refining channel: the first component is a scalar. The remaining 15 components form a vector. .

[0054] right Applying the Softplus function yields a large amount of data. :

[0055] Similarly, Guaranteed by the network structure. Grain size updated to:

[0056] because , This always holds true, meaning the grain size only increases and never decreases during the dwell phase. The auxiliary material state vector is updated as follows:

[0057] After sequential calculations of the forming and refining channels and the dwell and lengthening channels, the first... Step output That is, the first The predicted grain size (normalized value) of the fillet region after each pass. After the network completes the recursion for all valid passes, the output of the last pass is... The final predicted grain size is then denormalized to the actual physical quantity (unit: μm). Denormalization is performed according to... Calculation, where and These represent the upper and lower bounds for observing grain size in the sample set.

[0058] The mechanism of the Softplus activation function in the dual-channel structure lies in the fact that during the hot deformation process of aluminum alloy, dynamic recrystallization only refines or maintains the grain size, without coarsening it; while static grain growth during the stationary phase only enlarges or maintains the grain size, without refining it. By ensuring these two monotonic laws at the network structure level through Softplus, the network will not make predictions that violate physical laws when the training samples are limited.

[0059] In some embodiments, the loss value used for training is a weighted sum of three parts: the final grain size prediction error term, the intermediate pass grain size prediction error term, and the monotonicity regularization term.

[0060] Final grain size prediction error term The mean square of the difference between the predicted value and the labeled value for the last pass of each sample within the batch:

[0061] in This represents the batch sample size. Let j be the predicted value of the j-th sample at the last valid pass. All values ​​correspond to the simulation annotations and are normalized dimensionless quantities. The intermediate pass grain size prediction error term... The intermediate grain size of each pass is used as an additional monitoring signal:

[0062] This approach utilizes simulated, pass-by-pass grain data to constrain the network to learn the correct evolution path, rather than simply fitting the final result. With a limited sample size, intermediate supervision significantly reduces the network's degrees of freedom in fitting the final value, thereby improving the generalization accuracy for unseen pass combinations.

[0063] Monotonicity regularization term Penalize minor monotonic deviations that may occur in numerical calculations:

[0064] in Let be the output value of the i-th sample in the shaping and refining channel. The ReLU function outputs positive values ​​as is and zero for non-positive values, and only generates a positive penalty value when monotonicity is violated.

[0065] Among them, the weighting coefficient The method for determining this is to perform a test on the validation set. Experiment with each value in increments of 0.1 within the range of 0.1 to 1.0, keeping other training configurations constant, and select the value that makes the validation set... The lowest value; in the above 6061 aluminum alloy piston scenario, the calibration result is, for example, 0.5. Its function is to assist in regularization, and its value is much smaller than that of the main loss term, fixed at 0.1.

[0066] Optionally, the optimizer is Adam, with an initial learning rate of 0.001. The learning rate adjustment strategy is cosine annealing, where the learning rate smoothly decays from the initial value according to a cosine function curve, decreasing to 0.01 times the initial value at the end of the total training epochs. The training epochs are set to 500, and the batch size is set to 32. L2 regularization is applied to the network weights to suppress overfitting, and the regularization coefficient is determined by performing a grid search on the validation set loss, with a search range of 1×10⁻⁶. Up to 1×10 In the above scenario, for example, 1×10 During training, the validation set loss is monitored. The training is terminated early when the validation set loss no longer decreases after 50 consecutive rounds, and the network parameters with the lowest validation set loss are taken as the final model.

[0067] The computational logic of the training and inference phases is completely consistent: the inputs are both a sequence of pass parameters and the initial grain size, which are recursively processed through the forming and refining channels and the dwell and growing channels in sequence according to time steps, and the output is the grain size of each pass. The only difference between the two phases is the use of the gradient: in the training phase, the gradient of the loss with respect to the network weights is used to update the weights; in the inference phase (i.e., in the subsequent optimization steps), the gradient of the final grain size with respect to the input parameters is used to search for the optimal pass scheme.

[0068] Optionally, after training is completed, check the prediction results on the validation set: if the maximum relative error exceeds 15%, simulation samples can be added to the parameter region of the error set and then retrained.

[0069] In step S4, the trained pass sequence grain prediction network is used as a prediction tool to perform gradient search to obtain an optimization scheme. Using the trained pass sequence grain prediction network as the prediction tool, the pass parameter sequence as the optimization variable, and the final grain size output by the network as the optimization objective, under the constraints of total thinning amount and equipment capability, the pass parameter sequence that minimizes the final grain size is obtained through gradient search, and this is taken as the optimized forming process scheme.

[0070] In one embodiment, after the network is trained, the network is used as a fast prediction tool to search for the pass scheme that minimizes the final grain size in the pass parameter space.

[0071] Specifically, given the number of paths Under these conditions, the pass parameter sequence is used as the optimization variable, and the final grain size output by the pass sequence grain prediction network is used. To optimize the objective, we need to make The minimum combination of parameters that satisfies both types of constraints.

[0072] Total thinning constraint: The product of the reduction rates of each pass equals the required wall thickness ratio as designed.

[0073] in This represents the initial wall thickness of the billet. The target final wall thickness is given in mm.

[0074] Equipment capacity constraints: The forming force in each pass shall not exceed the rated capacity of the press. The forming force is estimated using a simplified formula:

[0075] in For the first The rheological stress under the pass conditions, in MPa, is calculated by the Sellars-Tegart constitutive model based on temperature, strain, and strain rate. The cross-sectional area of ​​the punch is in mm. ; This is the friction correction factor, ranging from 1.5 to 2.5, determined based on the lubrication method. Average strain. average strain rate , For the first Height of the deformation zone for each pass, in mm.

[0076] The range constraints for each parameter are consistent with those during sampling: , mm / s, K, s.

[0077] In one embodiment, since the fully connected layers, ReLU, Softplus, and max operations in the network are all differentiable or have subgradients, the final grain size... The gradient of the input trace parameter sequence can be directly calculated via backpropagation. Specifically, the projective gradient descent method is used. For example, 50 initial points are randomly generated from the parameter space, and gradient descent iterations are performed independently for each initial point, with an iteration count of, for example, 200 steps. After each update, two projection operations are performed: first, the parameters are truncated to a feasible range, and then the reduction rate of each trace is adjusted proportionally to restore the product to its original value. Specifically, for each Multiply by a uniform correction factor to make the product equal to the target value. The step size is determined using Armijo line search, that is, try the step size along the gradient direction. If the target value does not decrease sufficiently, the step size is halved and the trial is repeated. The initial trial step size is, for example, 0.01.

[0078] After 50 sets of searches are completed, select the set that satisfies all constraints and The minimum result is taken as the optimized molding process scheme. In some embodiments, the optimized comparison can be performed separately for different number of passes, and the overall optimal number of passes and its parameter sequence can be selected.

[0079] For example, for a 6061 aluminum alloy piston with an outer diameter of 50 mm, a target wall thickness of 3 mm, and a cup depth of 35 mm, the initial billet wall thickness is 12 mm, and the initial grain size is 80 μm. An optimized four-pass scheme was obtained: Pass 1: 32% reduction, temperature 693 K, speed 28 mm / s, dwell time 18 s; Pass 2: 28% reduction, temperature 673 K, speed 22 mm / s, dwell time 12 s; Pass 3: 22% reduction, temperature 653 K, speed 18 mm / s, dwell time 8 s; Pass 4: 15% reduction, temperature 633 K, speed 12 mm / s, dwell time 5 s. The network predicts the final grain size in the rounded corner region to be 19 μm, which is less than the upper threshold of 25 μm. This scheme shows a trend of progressively decreasing temperature and speed, and decreasing reduction: the high temperature and high pressure in the early stage fully trigger recrystallization to refine the grains, while the low temperature and low pressure in the later stage, along with a shortened dwell time, suppresses static growth.

[0080] In some embodiments, the optimized molding process scheme is substituted into molding simulation and grain evolution calculation for verification. The simulated value of the final grain size in the rounded corner region is extracted, and the relative deviation between it and the predicted value of the pass sequence grain prediction network is calculated. A threshold is set, for example, 10%. When the relative deviation does not exceed the set threshold, the current scheme is accepted as the final optimization result. When the relative deviation exceeds the set threshold, the simulation result of this verification is added to the training sample set as a new sample, the pass sequence grain prediction network is retrained, and gradient search is performed again until the deviation converges. In the above example, the grain size obtained by simulation verification is 21 μm, and the relative deviation from the predicted value of 19 μm is about 10.5%. After one round of supplementary training and re-optimization, the deviation is reduced to 7%, which meets the set threshold requirement. The specific value of the set threshold is determined according to the engineering accuracy requirements: if the machining allowance of subsequent processes is sufficient to accommodate microstructure fluctuations, it can be appropriately relaxed to 15%; if the piston is directly assembled without cutting, it is tightened to 5%.

[0081] This method transforms the multi-pass molding process design from the traditional experience-based simulation verification mode to a sequential optimization mode guided by the target grain size, allowing microstructure requirements to be considered during the pass design stage. The dual-channel network structure with physical monotonicity constraints maintains the physical rationality of the prediction under limited training samples, and the loss function supervised by intermediate passes improves the prediction accuracy of the progressive evolution path. The gradient optimization search based on this network can traverse the parameter space that traditional methods require hundreds of simulations to cover in minutes, shortening the process development cycle of new piston models from weeks to days. At the same time, the optimization scheme takes into account both geometric thinning requirements and grain refinement goals in the rounded corner areas.

[0082] Figure 3 A schematic block diagram illustrating an optimized design system for the molding process of a lightweight caliper piston according to some embodiments of the present disclosure is shown. Figure 3 As shown, the system includes: The grain size threshold determination module 301 is used to determine the upper limit threshold of the grain size in the rounded corner region based on the fatigue life requirements of the target piston. The training sample construction module 302 is used to sample and generate multiple sets of pass parameter sequences within the feasible range of multi-pass forming parameters, run forming simulation and grain evolution calculation on each set of pass parameter sequences, obtain the grain size of the rounded corner area after each pass, and form a training sample set. The grain prediction network construction and training module 303 is used to construct a pass sequence grain prediction network and train it with the training sample set. The pass sequence grain prediction network is recursively calculated step by step, with each time step corresponding to a pass and passing through the forming and refining channel and the dwell and growing channel in sequence. The forming and refining channel outputs a non-negative refining amount and subtracts it from the current grain size. The dwell and growing channel outputs a non-negative growth amount and adds it to the current grain size. The molding process optimization module 304 is used to use the trained pass sequence grain prediction network as a prediction tool, the pass parameter sequence as an optimization variable, and the final grain size output by the network as an optimization target. Under the constraints of total thinning amount and equipment capability, it uses gradient search to find the pass parameter sequence that minimizes the final grain size, and uses it as the optimized molding process scheme.

[0083] The specific implementation method and technical effects of this system can be referred to the description of the method above, and will not be repeated here.

[0084] Figure 4 A schematic block diagram of a computing device capable of implementing some embodiments of the present disclosure is shown. In some embodiments, the computing device may include various different types of devices, such as computing devices, computers, client devices, system-on-a-chip, and / or any other suitable computing devices.

[0085] The computing device may include at least one processor 1302, memory 1304, multiple communication interfaces 1306, display device 1308, other input / output (I / O) devices 1310, and one or more mass storage devices 1313 that are capable of communicating with each other, such as via system bus 1311 or other suitable means.

[0086] Processor 1302 may be a single processing unit or multiple processing units, and all processing units may include a single or multiple computing units or multiple cores. Processor 1302 may be implemented as one or more microprocessors, microcomputers, microcontrollers, digital signal processors, central processing units, state machines, logic circuits, and / or any device that manipulates signals based on operating instructions. Among other capabilities, processor 1302 may be configured to acquire and execute computer-readable instructions stored in memory 1304, mass storage device 1313, or other computer-readable media, such as program code of operating system 1316, program code of application program 1318, program code of other program 1320, etc., to implement the methods provided in the embodiments of this disclosure.

[0087] Memory 1304 and mass storage device 1313 are examples of computer storage media for storing instructions that are executed by processor 1302 to perform the various functions described above. For example, memory 1304 may generally include both volatile and non-volatile memory (e.g., RAM, ROM, etc.). Furthermore, mass storage device 1313 may generally include hard disk drives, solid-state drives, removable media, including external and removable drives, memory cards, flash memory, floppy disks, optical disks (e.g., CDs, DVDs), storage arrays, network-attached storage, storage area networks, etc. Both memory 1304 and mass storage device 1313 may be collectively referred to herein as memory or computer storage media, and may be non-transitory media capable of storing computer-readable, processor-executable program instructions as computer program code, which may be executed by processor 1302 as a specific machine configured to perform the operations and functions described in the examples herein.

[0088] Multiple program modules can be stored on mass storage device 1313. These programs include operating system 1316, one or more application programs 1318, other programs 1320, and program data 1322, and they can be loaded into memory 1304 for execution. Examples of such application programs or program modules may include, for example, computer program logic (e.g., computer program code or instructions) for implementing the methods provided herein. Moreover, these program modules can be distributed across different physical locations to achieve their respective functions.

[0089] Although Figure 4The modules 1313, 1318, 1320, and 1322, or portions thereof, are illustrated as being stored in memory 1304 of a computing device; however, modules 1313, 1318, 1320, and 1322 may be implemented using any form of computer-readable medium accessible by a computing device. As used herein, “computer-readable medium” includes at least two types of computer-readable media: computer storage media and communication media.

[0090] Computer storage media includes volatile and non-volatile, removable and non-removable media implemented by any method or technology for storing information such as computer-readable instructions, data structures, program modules, or other data. Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technologies, CD-ROM, DVD, or other optical storage devices, magnetic cartridges, magnetic tapes, disk storage devices or other magnetic storage devices, or any other non-transfer medium that can be used to store information for access by computing devices.

[0091] In contrast, communication media can embody computer-readable instructions, data structures, program modules, or other data within modulated data signals such as carrier waves or other transmission mechanisms. Computer storage media as defined herein do not include communication media.

[0092] The computing device may also include one or more communication interfaces 1306 for exchanging data with other devices, such as via a network, direct connection, etc. Communication interface 1306 can facilitate communication across various network and protocol types, including wired networks (e.g., LAN, cable, etc.) and wireless networks (e.g., WLAN, cellular, satellite, etc.), the Internet, etc. Communication interface 1306 can also provide communication with external storage devices (not shown), such as storage arrays, network-attached storage, storage area networks, etc.

[0093] In some examples, a display device 1308, such as a monitor, may be included for displaying information and images. Other I / O devices 1310 may be devices that receive various inputs from a user and provide various outputs to the user, and may include touch input devices, gesture input devices, cameras, keyboards, remote controls, mice, printers, audio input / output devices, and so on.

[0094] This application provides a computer-readable storage medium storing computer-readable instructions thereon, which, when executed, implement the above-described method.

[0095] This application provides a computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computing device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computing device to perform the methods provided in the various optional implementations described above.

[0096] In the description of this specification, the terms "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to a specific feature, structure, material, or characteristic described in connection with that embodiment or example that is included in at least one embodiment or example of this disclosure. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0097] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing custom logic functions or processes, and the scope of preferred embodiments of this disclosure includes additional implementations in which functions may be performed not in the order shown or discussed (including substantially simultaneously or in reverse order depending on the functions involved), as will be understood by those skilled in the art to which embodiments of this disclosure pertain.

[0098] Furthermore, the functional units in the various embodiments of this disclosure can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0099] In the above embodiments, the descriptions of each embodiment have different focuses. Parts not described in detail in a certain embodiment can be referred to in the relevant descriptions of other embodiments. The above descriptions are merely preferred embodiments of this application and explanations of the technical principles used. Those skilled in the art should understand that the scope of the invention involved in this application is not limited to the technical solutions formed by specific combinations of the above technical features, but should also cover other technical solutions formed by arbitrary combinations of the above technical features or their equivalent features without departing from the inventive concept.

Claims

1. An optimized design method for the molding process of a lightweight caliper piston, characterized in that, include: Based on the fatigue life requirements of the target piston, determine the upper limit threshold of the grain size in the rounded corner region; Within the feasible range of multi-pass forming parameters, multiple sets of pass parameter sequences are sampled and generated. Forming simulation and grain evolution calculation are run on each set of pass parameter sequences to obtain the grain size of the rounded corner area after each pass and form a training sample set. A pass sequence grain prediction network is constructed and trained with the training sample set. The pass sequence grain prediction network is recursively applied step by step. Each time step corresponds to a pass and passes through the forming and refining channel and the dwell and growth channel in sequence. The forming and refining channel outputs a non-negative refining amount and subtracts it from the current grain size. The dwell and growth channel outputs a non-negative growth amount and adds it to the current grain size. Using the trained pass sequence grain prediction network as a prediction tool, the pass parameter sequence as an optimization variable, and the final grain size output by the network as the optimization target, under the constraints of total thinning amount and equipment capability, the pass parameter sequence that minimizes the final grain size is obtained through gradient search, which is then used as the optimized forming process scheme.

2. The method according to claim 1, characterized in that, The forming and refining channel and the dwell and growth channel each contain a sub-network. The output of each sub-network is processed by a non-negative activation function to obtain the refining amount and the growth amount, respectively. The non-negative activation function structurally constrains the output to always be non-negative, so that the grain size only decreases and does not increase during the forming stage and only increases and does not decrease during the dwell stage.

3. The method according to claim 1, characterized in that, The internal state of the sequence grain prediction network includes a grain size scalar and an auxiliary material state vector; the auxiliary material state vector participates in the input calculation of the forming refinement channel and the dwell growth channel, and is updated sequentially by the two channels in each time step.

4. The method according to claim 3, characterized in that, The input to the forming refinement channel is a concatenation of the grain size scalar from the previous time step, the auxiliary material state vector from the previous time step, and the forming parameters of the current pass; the input to the dwell growth channel is a concatenation of the updated grain size scalar from the forming refinement channel, the updated auxiliary material state vector, and the dwell parameters of the current pass.

5. The method according to claim 1, characterized in that, When training the pass sequence grain prediction network, the loss value is composed of the weighted sum of the final grain size prediction error term and the intermediate pass grain size prediction error term; the intermediate pass grain size prediction error term is calculated as the deviation between the predicted grain size after each pass and the corresponding simulated grain size in the training sample set.

6. The method according to claim 5, characterized in that, The loss value also includes a monotonicity regularization term, which penalizes the deviation of the grain size being greater than the input grain size after the output of the forming refinement channel and the deviation of the grain size being less than the input grain size after the output of the dwell growth channel.

7. The method according to claim 1, characterized in that, The parameters for each pass in the pass parameter sequence include reduction rate, forming speed, forming temperature, inter-pass dwell time, and dwell temperature. Before inputting the pass sequence into the grain prediction network, each parameter is subjected to minimum-maximum normalization processing according to the upper and lower bounds of its sampling range.

8. The method according to claim 1, characterized in that, The gradient search adopts the projected gradient descent method, randomly generating multiple initial points from the parameter space and searching independently. After each gradient update, each parameter is projected back to the feasible range, and the reduction rate of each pass is adjusted so that their product satisfies the total thinning constraint. The result with the smallest final grain size that satisfies all constraints is taken as the optimized forming process scheme.

9. The method according to claim 2, characterized in that, The sub-networks in the forming and refining channel and the dwelling and growing channel are each two-layer fully connected networks. The first layer uses a modified linear unit activation function, and the second layer has no activation function. The output of the second layer is split into a scalar component and a vector component. The scalar component is processed by the non-negative activation function to obtain the corresponding refining amount or growing amount. The vector component is processed by the bounded activation function and used as the update increment of the auxiliary material state vector.

10. An optimized design system for the molding process of a lightweight caliper piston, characterized in that, include: The grain size threshold determination module is used to determine the upper limit threshold of the grain size in the rounded corner region based on the fatigue life requirements of the target piston. The training sample construction module is used to sample and generate multiple sets of pass parameter sequences within the feasible range of multi-pass molding parameters, run molding simulation and grain evolution calculation on each set of pass parameter sequences, obtain the grain size of the rounded corner area after each pass, and form a training sample set. The grain prediction network construction and training module is used to construct a pass sequence grain prediction network and train it with the training sample set. The pass sequence grain prediction network is recursively calculated step by step, with each time step corresponding to a pass and passing through the forming and refining channel and the dwell and growing channel in sequence. The forming and refining channel outputs a non-negative refining amount and subtracts it from the current grain size. The dwell and growing channel outputs a non-negative growth amount and adds it to the current grain size. The molding process optimization module uses the trained pass sequence grain prediction network as a prediction tool, the pass parameter sequence as an optimization variable, and the final grain size output by the network as the optimization target. Under the constraints of total thinning amount and equipment capability, it uses gradient search to find the pass parameter sequence that minimizes the final grain size, which is then used as the optimized molding process scheme.

Images