Titanium alloy milling deformation control method based on dynamic stiffness evolution and reinforcement learning

Abstract

The application discloses a titanium alloy milling deformation control method based on dynamic stiffness evolution and reinforcement learning, and belongs to the technical field of numerical control machining and deformation control of aviation titanium alloy integral structural parts. The method comprises the following steps: establishing a part finite element model, discretely voxelizing a region to be removed, extracting dynamic stiffness evolution characteristics of each machining step, and constructing a stiffness description vector of a current machining state; constructing a machining deformation proxy model to realize fast prediction of a machining process deformation; modeling the machining process as a Markov decision process, comprehensively representing maximum deformation, stiffness weakening, machining smoothness and machining efficiency through a multi-objective reward function, and autonomously searching for an optimal asymmetric layered milling sequence and parameter sequence satisfying manufacturing constraints by using a near-optimal strategy optimization algorithm. The application can significantly reduce the machining warping of titanium alloy integral component parts and improve the dimensional accuracy, surface quality and process reliability without increasing additional processes such as grooving.

Description

Technical Field

[0001] This invention relates to a method for controlling deformation during milling of titanium alloys, specifically a method for controlling deformation during milling of titanium alloys based on dynamic stiffness evolution and reinforcement learning, belonging to the field of CNC machining and deformation control technology for integral titanium alloy structural components for aerospace applications. Background Technology

[0002] Integral titanium alloy structural components for aerospace applications are typically formed by CNC milling of thick plate blanks, resulting in a high material removal rate of 80%–95%. During this process, the original residual stress balance within the blank is continuously disrupted, leading to warping, twisting, or localized collapse of the workpiece, severely impacting subsequent assembly positioning accuracy and service reliability. This problem is particularly pronounced for complex components with thin base plates, dense ribs, or multiple coupled cavities.

[0003] Existing control methods mainly include symmetrical layer milling, natural or artificial aging, clamping compensation, stress relief grooves, and machining sequence adjustments based on the experience of process engineers. While these methods can play a certain role in suppressing stress in specific scenarios, they generally have the following shortcomings: First, they are difficult to reflect the time-varying and nonlinear weakening law of stiffness changes with geometric shape during the layer-by-layer material removal process of complex thin-walled parts; second, they are difficult to simultaneously consider multiple factors such as residual stress release, cutting load, machining continuity, and tool accessibility; third, for multi-cavity, asymmetrical structures, or multi-region alternating material removal tasks, traditional empirical methods are difficult to obtain the globally optimal material removal sequence in a holistic sense.

[0004] In particular, existing solutions typically treat workpiece stiffness as a static quantity or only assess deformation trends once on the final configuration, failing to establish an online optimization mechanism for the strong coupling relationship between "current remaining material state—next material removal action—subsequent deformation risks." Therefore, a deformation control method is needed that can dynamically perceive the evolution of structural stiffness and stress redistribution throughout the entire processing process, and search for the optimal processing sequence and parameters through intelligent decision-making methods. Summary of the Invention

[0005] This invention aims to address the shortcomings of existing methods for controlling deformation during the machining of integral titanium alloy components. These methods struggle to accurately track the dynamic stiffness evolution during machining, comprehensively quantify the influence of time-varying residual stress, and lack a global dynamic optimization mechanism for the cutting sequence. A new deformation control method for titanium alloy milling based on dynamic stiffness evolution and reinforcement learning is proposed. This method accurately extracts the dynamic stiffness evolution characteristics of each machining step by performing voxel-based layering and partitioning of the cutting region and stepwise material removal simulation. A surrogate model integrating the topological encoding of remaining material, residual stress, and stiffness vectors is constructed for deformation prediction. Furthermore, the entire machining process is reconstructed as a Markov decision process, relying on a proximal policy optimization algorithm and a multi-objective reward function for global optimization. This achieves accurate feature modeling and dynamic decision optimization for the entire machining process of complex components, ultimately obtaining an optimal machining strategy with smaller deformation and more stable processes.

[0006] To achieve the above objectives, this invention provides a method for controlling deformation during milling of titanium alloys based on dynamic stiffness evolution and reinforcement learning, comprising the following steps:

[0007] Step 1: Establish a finite element model of the part, divide the area to be removed into layers and partitions to form a set of voxelized cutting feature blocks;

[0008] Step 2: Based on the set of voxelized cutting feature blocks, perform stepwise material removal simulation processing, extract the dynamic stiffness evolution features of each machining step, and construct the stiffness description vector of the current machining state.

[0009] Step 3: Construct a processing deformation proxy model, taking the remaining material topology code, residual stress embedding vector, stiffness description vector and candidate action code as input, and output the deformation prediction for the next processing step;

[0010] Step 4: Model the processing process as a Markov decision process, design a multi-objective reward function based on the deformation prediction, and use the near-end policy optimization algorithm to train the policy and obtain the optimal processing policy.

[0011] Furthermore, the dynamic stiffness evolution characteristics include global flexibility index, minimum characteristic stiffness, peak displacement at key observation points, local deformation gradient, and modal stiffness vector.

[0012] Furthermore, in step 3, the processing deformation proxy model includes a dual-channel 3D convolutional encoder and a fully connected regression head. The training process of the processing deformation proxy model adopts a design experiment combined with active learning to generate training samples and iteratively update the model.

[0013] Further, step 3 includes:

[0014] Receive the remaining material topology code, residual stress embedding vector, stiffness description vector, and candidate action code;

[0015] The remaining material topology encoding and residual stress embedding vector are respectively input into the first and second branches of the dual-channel 3D convolutional encoder for processing to obtain the first feature and the second feature.

[0016] The first feature, the second feature, the stiffness description vector, and the candidate action code are concatenated to obtain the fused feature;

[0017] The fused features are input into multiple cascaded fully connected layers in the fully connected regression head for processing, and the deformation prediction or risk label for the next processing step is output.

[0018] Further, step 4 includes:

[0019] Construct the state space and action space, and set manufacturing constraints;

[0020] The state space includes a remaining material occupancy matrix and a stiffness description vector; the action space is defined as the feature block number to be removed in the next machining step and its associated process parameter combination; the manufacturing constraints include minimum remaining wall thickness, maximum cantilever length, and tool reachability.

[0021] Define a multi-objective reward function and train it using a proximal policy optimization algorithm to output an optimized processing policy.

[0022] Furthermore, the multi-objective reward function is expressed as:

[0023] ;

[0024] in, This represents the predicted deformation amount for the current processing step. The difference between the predicted deformation amount of the current processing step and the previous processing step is calculated as 0 for the previous state of the first processing step. This is the normalized equivalent stiffness index; This is the stiffness safety threshold; For the cost of time; This is a continuous penalty item; To terminate the reward; For multi-objective reward functions; The weight of the largest deformation term; The weight of the deformation increment term; This represents the weight of the stiffness safety term; This represents the maximum difference between the stiffness safety threshold and the normalized equivalent stiffness index and 0. Weighting of the processing time item; This represents the weight of the penalty term for process continuity.

[0025] Furthermore, the titanium alloy milling deformation control method also includes: converting the optimal machining strategy into CAM toolpath planning input, generating a CNC program, and performing interference and stability checks before actual machining, ultimately completing the process implementation and verification.

[0026] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0027] 1. This invention elevates machining deformation control from static process planning relying solely on experience to dynamic optimization decision-making throughout the entire process, significantly improving the global optimality of machining strategies for complex structural parts. The invention rigorously reconstructs the multi-step cutting process into a Markov decision process, utilizing a proximal strategy optimization algorithm to globally optimize the action sequence over a large time domain throughout the entire machining cycle. This mechanism overcomes the limitation of traditional single-step greedy algorithms, which are prone to getting trapped in local optima, fundamentally achieving dynamic spatiotemporal matching between the residual stress release trajectory and the workpiece's macroscopic deformation resistance.

[0028] 2. By explicitly introducing dynamic stiffness evolution and residual stress characteristics updated with each machining step, unreasonable process routes can be avoided in advance before the structural stiffness drops sharply, reducing the probability of instability and warping. With a high proportion of material removal, the stiffness matrix of thin-walled parts exhibits strong nonlinear decay. This invention explicitly constructs a stiffness description vector containing characteristic stiffness and modal vectors in the state space through eigenvalue decomposition, and integrates the residual stress embedding vector extracted by a multi-branch network. This enables the agent to physically perceive the underlying mechanical coupling mechanism of material removal—stress release—stiffness weakening, and actively adjust cutting parameters or shift the machining area before approaching the critical point of stiffness instability.

[0029] 3. By employing a surrogate model to replace large-scale repetitive full finite element calculations, optimization time is significantly shortened, improving the algorithm's engineering feasibility. A 3D convolutional neural network is used to extract the three-dimensional spatial features of voxelized topology and stress field, and combined with an active learning closed-loop mechanism, targeted high-precision finite element verification is added for high-uncertainty risk boundary samples (such as stiffness abrupt change regions). This data-driven and physical prior-based approach, while ensuring micrometer-level deformation prediction accuracy, compresses the computational cost of single-step physical response evaluation from minutes to milliseconds in conventional nonlinear iterative solutions, effectively overcoming the barrier between massive trial-and-error in reinforcement learning and the high computational cost of finite element methods.

[0030] 4. The reward function comprehensively considers maximum deformation, stiffness safety margin, machining efficiency, and process continuity, ensuring that the optimization result balances geometric accuracy and manufacturing feasibility. Its scientific basis lies in the fact that purely pursuing deformation minimization easily leads to unrealistic cutting commands such as frequent tool jumps and extremely low feed rates. The multi-objective reward function constructed in this invention not only directly penalizes warpage displacement but also imposes soft constraints on the system that conform to the dynamic stability of machining by introducing a normalized stiffness safety threshold term and a process continuity penalty term. This ensures that the output optimal strategy meets the practical requirements of smooth tool transition and avoidance of cutting chatter.

[0031] 5. Without adding extra processes such as stress relief grooves, smaller final deformation and higher dimensional consistency can be achieved, reducing material waste and processing costs. Based on the asymmetric alternating material removal strategy autonomously generated by reinforcement learning, it essentially adaptively counteracts the bending moment generated by the initial residual stress release in both spatial and temporal dimensions; that is, it utilizes the higher residual stiffness of the uncut area to resist the local stress release deformation of the current cutting area, thereby eliminating the reliance on high-cost processes such as artificial aging or grooving in terms of mechanical mechanism. Attached Figure Description

[0032] Figure 1 This is the overall flowchart of the method of the present invention;

[0033] Figure 2 This is a schematic diagram showing the voxelization of the area to be removed in a titanium alloy integral component according to the depth of the feature area hierarchy;

[0034] Figure 3 It is a graph showing the evolution of residual stress redistribution and equivalent stiffness index under different processing steps;

[0035] Figure 4 This is a schematic diagram of the network structure of the processing deformation proxy model;

[0036] Figure 5 It is a cumulative reward convergence curve during the training process of the agent;

[0037] Figure 6 This is a comparison diagram of deformation cloud maps under traditional symmetrical layered milling and the optimized asymmetrical alternating material removal strategy of this invention;

[0038] Figure 7 It is a comparison chart of the actual measurement results of the invention. Detailed Implementation

[0039] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, and the advantages of the present invention in the above and / or other aspects will become clearer.

[0040] This embodiment provides a method for controlling deformation during titanium alloy milling based on dynamic stiffness evolution and reinforcement learning. The overall process of this method is as follows: Figure 1 As shown, the specific steps are as follows:

[0041] Step 1: Establish the finite element model of the part. Create a three-dimensional solid model of the titanium alloy integral component blank in finite element software such as ABAQUS or ANASYS. and the target part 3D solid model And the material region to be removed is precisely defined through spatial Boolean subtraction. ;

[0042] Specifically, in this embodiment, a 7075-T7451 integral component blank with dimensions of 300mm×150mm×30mm is selected, and the final base plate thickness of the target part is set to 3mm, the rib height to 20mm, and the local rib width to 8mm~12mm. The above precise geometric dimensions define the absolute spatial boundary of the material removal simulation and provide a basis for subsequent mesh size allocation and process constraints.

[0043] The initial residual stress distribution of the blank is obtained using the contour method, blind hole method, X-ray diffraction method, or ultrasonic inversion method; a three-dimensional residual stress field is formed using spline interpolation, Kriging interpolation, etc. Establish a database of parameters for anisotropic elastoplastic materials based on their grades.

[0044] In this embodiment, the residual stress in the thickness direction is measured using the contour method. The maximum compressive stress on the surface is approximately -150 MPa, and the maximum tensile stress in the core is approximately 80 MPa. The three-dimensional residual stress field is reconstructed based on spline interpolation and mapped (assigned) to the mesh elements (or nodes) of the finite element model. The finite element model uses a 2.5 mm mesh in the base plate area and a 5 mm mesh in the area away from the processing area. Twelve key observation points are set to monitor deformation.

[0045] Step 2: Divide the area to be removed into layers and sections according to cavity area, rib area, depth level, feed direction or process features, forming a set of voxelized cutting feature blocks. This set not only transforms the continuous removal volume into discrete mesh elements for finite element simulation, but also directly constitutes the discrete action space for the reinforcement learning agent to optimize the processing sequence.

[0046] In specific implementation, such as Figure 2As shown, the entire region to be removed is divided into 8 main feature regions (AH) and 5 depth levels (1-5), forming a total of 40 voxelized cutting feature blocks. This is used to match the actual process of CNC milling feature-driven tool movement and to accurately isolate the nonlinear residual stress gradient in the thickness direction. At the same time, the total number of action candidates is controlled to 40, which effectively avoids the "curse of dimensionality" in the reinforcement learning exploration process and ensures the usability of the algorithm while ensuring the accuracy of deformation control.

[0047] Step 3: Establish a discrete model of the blank in the finite element platform and define the three-dimensional residual stress field. Map to elements or nodes, and set clamping boundary conditions, reference cutting loads, and measurement observation point sets. .

[0048] The discrete model of the blank is a mathematical mesh representation of the aforementioned continuous three-dimensional solid geometry. The reconstructed three-dimensional residual stress field is accurately mapped to the integration points or nodes of the mesh elements in this discrete model, giving the finite element solution system a realistic initial internal stress state. Simultaneously, strict boundary and load parameters are set: the translational degrees of freedom of the corresponding nodes in the base plate support area are fully constrained to eliminate rigid body displacements (and the set of finite element elements within this constrained area is defined as the support mesh), and an equivalent uniformly distributed surface load is applied as a reference cutting load based on cutting parameters, thereby highly replicating the effect of real workshop fixtures on the stiffness of the part and the cutting force disturbance.

[0049] Step 4: Simulate gradual material removal using unit activation / deactivation techniques. In the... After removing the feature block set in each processing step, the remaining structure is solved by static or thermo-mechanical coupled equilibrium to obtain the displacement field. Stress field .

[0050] The specific implementation process of the element activation / deactivation technique is as follows: For the finite element elements within the selected feature block spatial envelope, the system introduces a minimum attenuation coefficient into their material stiffness matrix to cause their load-bearing capacity to fail, thereby simulating physical material removal and avoiding global mesh remapping. After removing the feature block, the remaining structure is solved using a nonlinear static equilibrium method based on the Newton-Raphson iterative method. The obtained displacement field is used to directly extract the current maximum deformation, constituting the core penalty index of the reinforcement learning model; the obtained stress field reflects the redistribution of residual stress caused by material removal in real time. This physical stress field will serve as the basic data for extracting and generating residual stress embedding vectors in subsequent steps, which are then input into the processing deformation proxy model to assess the potential deformation risk of the next action.

[0051] Step 5: Extract dynamic stiffness evolution features for each processing step. The extraction process is as follows: Based on the global stiffness matrix updated in each step (this matrix is ​​inherited from Step 4 and extracted by the finite element solver after completing the nonlinear static equilibrium solution, recording the evolution of the structure's resistance to deformation in real time), the minimum characteristic stiffness and modal stiffness vector are calculated through eigenvalue decomposition; at the same time, the local deformation gradient and global compliance index are calculated by the displacement response difference of the support mesh under the reference load in Step 3.

[0052] The fundamental technical purpose of extracting and constructing this stiffness description vector is to explicitly inject it into the state space of reinforcement learning as a physical early warning indicator. This breaks through the limitation of traditional processes treating workpiece stiffness as a static constant, enabling intelligent decision-making systems to keenly capture the critical point of nonlinear, precipitous stiffness weakening under high material removal rates. The evolution curves of residual stress redistribution and equivalent stiffness index under different processing steps are shown below. Figure 3 As shown, this allows for proactive avoidance of high-risk material removal actions that could easily cause severe warping in the next trajectory planning step.

[0053] Stiffness description vector that constitutes the current processing state Among them, the global flexibility index Minimum characteristic stiffness Peak displacement of key observation points (This peak value represents the extreme values ​​of the node displacements at the 12 key observation points preset in step one within the displacement field of step four), local deformation gradient. and modal stiffness vector .

[0054] Step Six: Construct a processing deformation proxy model Topological coding of remaining materials Residual stress embedding vector Stiffness description vector Candidate action coding As input, the output is the deformation prediction or risk label for the next processing step. A "designed experiment + active learning" approach is used to generate training samples and iteratively update the model, specifically including:

[0055] S01: 2000 sets of random machining sequences and parameter combinations were generated using Latin hypercube sampling. The random machining sequence refers to different removal order routes set for the aforementioned 40 voxelized cutting feature blocks; the parameter combination refers to the set of discrete process parameters such as cutting depth, lateral cutting width, and feed rate matched when performing each removal action. Offline finite element simulations were performed on these 2000 schemes to extract the corresponding stiffness evolution and deformation results, forming the basic database for the initial training of the surrogate model, corresponding to... Figure 4 The initial training sample library is shown.

[0056] S02: Active learning phase, targeted addition of high-risk samples and model iteration. After completing the initial training of the surrogate model using the basic samples, the model's prediction uncertainty (such as output variance) for the unknown state space is evaluated, and 300 sets of high-risk samples are targeted and added.

[0057] High-risk samples specifically include: combinations of processing states where the surrogate model's prediction uncertainty is significantly higher than a preset threshold (a knowledge blind spot), and extreme process combinations where the prediction results are on the verge of a sharp drop in stiffness or warping deformation exceeding tolerance. These high-risk samples are then re-imported into the finite element platform for high-precision solving to obtain the true physical response, and appended to the training sample library for iterative updates of the model parameters. This mechanism significantly improves the prediction accuracy of the surrogate model in critical process boundary regions with minimal additional finite element computation cost.

[0058] S03: The surrogate model is implemented based on a hybrid architecture of a dual-channel 3D convolutional encoder and a fully connected regression head. Specifically, the model input receives the original three-dimensional matrix of residual material topology and three-dimensional residual stress field data; the dual-channel 3D convolutional encoder processes the two original inputs respectively, extracts spatial topological features and stress distribution features, and transforms them into one-dimensional residual material topology encoding and residual stress embedding vector.

[0059] like Figure 4 As shown, the model processing flow is divided into a feature extraction layer, a fusion layer, and a decision prediction layer. The feature extraction layer uses a first-path 3D convolutional structure with three consecutive "convolution-pooling-activation" operators to process the voxel occupancy matrix. The convolutional kernel size is 3×3×3, aiming to capture the connectivity and supporting topology of the remaining structure. The second-path 3D convolutional structure simultaneously processes the stress tensor components mapped to the voxel mesh, extracting the spatial gradient features of the residual stress. The fusion layer flattens the outputs of the two encoders, and the resulting topology encoding and stress embedding vector are concatenated with the stiffness description vector and candidate action encoding of the current step to form a global feature vector. The decision prediction layer sequentially inputs the fused feature vector into a fully connected regression head, where it undergoes nonlinear mapping through fully connected layers with widths of 256, 128, 64, and 32, respectively.

[0060] In terms of training mechanism, the model relies on a closed-loop iterative strategy combining experimental design and active learning. Based on the initial Latin hypercube sampling, it selectively supplements high-risk samples with high uncertainty for finite element verification and updating, thereby ensuring the prediction accuracy of key process areas while controlling computational costs. In particular, a multi-dimensional stiffness description vector covering global flexibility, minimum characteristic stiffness, and local deformation gradient is explicitly injected as a core physical prior in this architecture. This precisely quantifies the nonlinear and time-varying stiffness weakening law during the layer-by-layer material removal process of complex components, enabling the decision system to accurately anchor the strong coupling relationship between "remaining state - material removal action - subsequent deformation risk" and achieve proactive avoidance of high-risk processing steps before the structural stiffness drops sharply.

[0061] To ensure that the machining deformation proxy model can accurately map the nonlinear strong coupling relationship between the physical state and the process action, this invention clarifies the model's data interaction logic: at the input end, the remaining material topology encoding is derived from the current set of voxelized cutting feature blocks. Real-time state generation is achieved by assigning a value of 1 to unremoved regions and a value of 0 to removed regions, constructing a binary voxel occupancy matrix; the residual stress embedding vector is obtained by solving the instantaneous stress field in step four. The model resamples to a voxel grid identical to the topological matrix, and extracts the average value of the stress components within each voxel as a feature tensor. Candidate action encoding uses one-hot encoding to represent the feature block number to be removed in the next processing step, and then concatenates this with a normalized sequence of process parameters. At the output, the model generates the deformation prediction. Directly input into the multi-objective reward function in step eight. In this process, the core negative reward item guides the agent to search for the action path with the least deformation; risk labels, such as Boolean values ​​or classification probabilities, are used to determine whether the current action combination violates manufacturing constraints or causes instantaneous stiffness collapse. In turn, the action masking mechanism directly shields high-risk processing spaces in the training of the near-end policy optimization algorithm, effectively avoiding the risk of workpiece scrap.

[0062] S04: After training is completed, the average relative error of the maximum deformation prediction on the validation set is controlled within 5%.

[0063] Step 7: Establish a reinforcement learning environment model, abstracting the processing process into a Markov decision process. The state space includes the current remaining material occupancy matrix, the residual stress embedding vector corresponding to the current processing step, and the stiffness description vector from Step 5.

[0064] The remaining material occupancy matrix, used to describe the instantaneous topology of the workpiece, is a binary voxel occupancy matrix generated from the real-time state of the voxelized cutting feature block set (1 for unremoved material and 0 for removed material). Based on this, the state can be further incorporating clamping position changes (e.g., support point coordinates), tool wear state (e.g., flank wear value), spindle power constraints, and ambient temperature factors to achieve a digital twin representation of complex and variable engineering scenarios. Its motion space... Defined as the feature block number to be removed in the next machining step and its associated process parameter combination, including the depth of cut. Lateral cut width Feed rate Spindle speed .

[0065] Step 8: Design a multi-objective reward function and train the policy. The constructed reward function takes the following form:

[0066] ;

[0067] In the formula, This represents the predicted deformation value for the current step. The deformation increment is the difference between the predicted deformation of the current processing step and the previous processing step. The deformation of the previous state in the first processing step is counted as 0. This is the normalized equivalent stiffness index; This is the stiffness safety threshold; For the cost of time; This is a continuous penalty item; To terminate the reward; For multi-objective reward functions; The weight of the largest deformation term; The weight of the deformation increment term; This represents the weight of the stiffness safety term; This represents the maximum difference between the stiffness safety threshold and the normalized equivalent stiffness index and 0. Weighting of the processing time item; Weighting of the process continuity penalty term;

[0068] Furthermore, minimum remaining wall thickness of 2.5 mm, maximum cantilever length of 45 mm, and tool accessibility are set as manufacturing constraints; a near-end strategy optimization algorithm is used for training, with a discount factor of 0.99, a shear coefficient of 0.2, and a learning rate of 1×10⁻⁶. -4 The maximum number of training rounds is set to 5000.

[0069] Furthermore, the multi-objective reward function used in this embodiment is:

[0070] ;

[0071] When the maximum deformation of the final part is less than 0.15mm, the system provides a termination reward. The cumulative reward convergence curve during the agent training process is shown in the figure below. Figure 5 As shown.

[0072] Step Nine: Addressing the limitations of traditional empirical methods such as symmetrical layer milling in handling time-varying stiffness weakening and the inability to optimize the global material removal sequence, this invention combines the dynamic state monitoring of a Kistler force gauge with the high-precision topography detection of a Hexagon coordinate measuring machine to conduct a comparative experiment. The optimal machining strategy is then implemented. The process is converted into CAM toolpath planning and CNC program generation. After digital simulation and machine tool interference verification are completed using platforms such as Siemens NX CAM and VERICUT, the final practical verification is carried out using a high-precision five-axis machining center.

[0073] like Figure 6 As shown, the traditional symmetrical layered milling process, due to its failure to dynamically track the stiffness weakening during the material removal process, resulted in a maximum bottom surface deformation of 0.452 mm, as predicted and measured by the baseline process. Furthermore, combined with... Figure 7 The actual measurement comparison results further verified the engineering applicability of this innovative method. In comparison, after adopting the optimization strategy of this invention, the maximum deformation was significantly reduced to 0.128 mm, a reduction of 71.6%. This significant deformation suppression effect is mainly due to the fact that this method explicitly introduces the dynamic stiffness evolution characteristics into the reinforcement learning state space and constructs a multi-objective reward function. This mechanism breaks through the local optima limitation of traditional algorithms. In the global optimization process, it not only imposes a severe penalty on the maximum warping displacement, but also introduces a stiffness safety threshold soft constraint to drive the agent to actively adjust the cutting path before the structural stiffness drops sharply, thereby outputting a globally optimal machining sequence that takes into account geometric accuracy, machining stability, and manufacturing continuity.

[0074] The asymmetric alternating material removal strategy independently generated in this invention cleverly utilizes the high residual stiffness of the uncut region in terms of physical mechanism, achieving adaptive hedging against the bending moment generated by the release of local residual stress in the current cutting zone. This dynamic internal stress balance mechanism based on the spatiotemporal dimension enables complex titanium alloy components to achieve better final morphological accuracy and dimensional consistency without adding additional high-cost processes such as stress relief grooves or artificial aging, fundamentally reducing the scrap rate and material loss.

[0075] This invention provides a method for controlling deformation during titanium alloy milling based on dynamic stiffness evolution and reinforcement learning. Many methods and approaches exist for implementing this technical solution; the above description is merely a preferred embodiment of the invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention. All components not explicitly stated in this embodiment can be implemented using existing technologies.

Claims

1. A method for controlling deformation during milling of titanium alloys based on dynamic stiffness evolution and reinforcement learning, characterized in that, Includes the following steps: Step 1: Establish a finite element model of the part, divide the area to be removed into layers and partitions to form a set of voxelized cutting feature blocks; Step 2: Based on the set of voxelized cutting feature blocks, perform stepwise material removal simulation processing, extract the dynamic stiffness evolution features of each machining step, and construct the stiffness description vector of the current machining state. Step 3: Construct a processing deformation proxy model, taking the remaining material topology code, residual stress embedding vector, stiffness description vector and candidate action code as input, and output the deformation prediction for the next processing step; The processing deformation proxy model includes a dual-channel 3D convolutional encoder and a fully connected regression head; Step 3 includes: Receive the remaining material topology code, residual stress embedding vector, stiffness description vector, and candidate action code; The remaining material topology encoding and residual stress embedding vector are respectively input into the first and second branches of the dual-channel 3D convolutional encoder for processing to obtain the first feature and the second feature. The first feature, the second feature, the stiffness description vector, and the candidate action code are concatenated to obtain the fused feature; The fused features are input into multiple cascaded fully connected layers in the fully connected regression head for processing, and the deformation prediction or risk label for the next processing step is output. Step 4: Model the processing process as a Markov decision process, design a multi-objective reward function based on the deformation prediction, and use the proximal policy optimization algorithm to train the policy and obtain the optimal processing policy; Step 4 includes: Construct the state space and action space, and set manufacturing constraints; The state space includes a remaining material occupancy matrix and a stiffness description vector; the action space is defined as the feature block number to be removed in the next machining step and its associated process parameter combination; the manufacturing constraints include minimum remaining wall thickness, maximum cantilever length, and tool reachability. Define a multi-objective reward function and train it using a proximal policy optimization algorithm to output the optimal processing policy; The multi-objective reward function Represented as: ； in, This represents the predicted deformation amount for the current processing step. This is the difference between the predicted deformation amount of the current processing step and the previous processing step; This is the normalized equivalent stiffness index; This is the stiffness safety threshold; For the cost of time; This is a continuous penalty item; To terminate the reward; The weight of the largest deformation term; The weight of the deformation increment term; This represents the weight of the stiffness safety term; Represents the maximum value function; Weighting of the processing time item; This represents the weight of the penalty term for process continuity.

2. The titanium alloy milling deformation control method based on dynamic stiffness evolution and reinforcement learning according to claim 1, characterized in that, The dynamic stiffness evolution characteristics include global flexibility index, minimum characteristic stiffness, peak displacement at key observation points, local deformation gradient, and modal stiffness vector.

3. The titanium alloy milling deformation control method based on dynamic stiffness evolution and reinforcement learning according to claim 1, characterized in that, In step 3, the training process of the processing deformation proxy model adopts a combination of experimental design and active learning to generate training samples and iteratively update the model.

4. The titanium alloy milling deformation control method based on dynamic stiffness evolution and reinforcement learning according to claim 1, characterized in that, Also includes: The optimal machining strategy is converted into CAM toolpath planning input, a CNC program is generated, and interference and stability checks are performed before actual machining, ultimately completing the process implementation and verification.

5. An electronic device, characterized in that, It includes a processor and a memory, the memory storing program code that, when executed by the processor, causes the processor to perform the steps of the titanium alloy milling deformation control method based on dynamic stiffness evolution and reinforcement learning as described in any one of claims 1 to 4.

6. A storage medium, characterized in that, The device stores a computer program or instructions that, when run on a computer, execute the steps of the titanium alloy milling deformation control method based on dynamic stiffness evolution and reinforcement learning as described in any one of claims 1 to 4.

Images