Data-mechanism fusion thin-walled part machining deformation control method and system

Abstract

The present application belongs to the technical field related to thin-walled part machining, in order to solve the problem of low utilization rate of existing multi-source heterogeneous physical data and unclear mechanism, a data-mechanism fusion thin-walled part machining deformation control method and system is proposed, the bayesian optimization algorithm is used to screen high sensitive process parameters; based on the reinforcement learning algorithm, combined with high sensitive process parameter combination, low-dimensional multi-source physical feature vector and full amount process parameter value, the full amount process parameter adjustment candidate solution is obtained; the double objective function is constructed with the minimum residual stress unevenness and deformation as the target, the double objective function is solved based on the full amount process parameter adjustment candidate solution through the non-dominated sorting genetic algorithm, the optimal process parameter combination is obtained to realize the control of thin-walled part machining, and the problem of low utilization rate of multi-source heterogeneous physical data and unclear mechanism of traditional method is solved.

Description

Technical Field

[0001] This invention belongs to the technical field of thin-walled part processing, and in particular relates to a data-mechanism fusion method and system for controlling deformation during thin-walled part processing. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Thin-walled parts, such as aero-engine blades and integral wing panels, are widely used in aerospace, automotive manufacturing, and electronic equipment industries due to their lightweight and high strength. However, during milling, these thin-walled parts are highly susceptible to deformation due to their poor structural rigidity, thin wall thickness, and compact structure.

[0004] Thin-walled parts are often made of difficult-to-machine materials such as nickel-based alloys and titanium alloys. Traditional machining methods lead to problems such as decreased dimensional accuracy and surface quality, seriously affecting the assembly accuracy and service performance of the workpiece. Milling is prone to problems such as high cutting forces and severe tool wear. Although ultrasonic vibration-assisted milling technology has advantages over traditional milling, such as lower cutting forces and better surface quality, the evolution mechanism of residual stress under the multi-physics coupling effect in ultrasonic vibration-assisted milling is still unclear, and the process parameters involved in ultrasonic vibration-assisted technology are complex and lack theoretical guidance.

[0005] Current methods for controlling machining deformation mainly rely on traditional experience-based trial cutting, single-mechanism models, and suffer from insufficient data utilization and poor adaptability. With the development of intelligent sensing technology and artificial intelligence, data acquisition has become more convenient. Compared to single-source modeling methods, multi-source information fusion can provide multiple information sources, enhancing information complementarity and fault tolerance, and achieving comprehensive perception and cross-validation of the machining state. However, purely data-driven models rely on large amounts of historical data, have weak physical interpretability, limited generalization ability, and are difficult to accurately control machining deformation. Summary of the Invention

[0006] To overcome the shortcomings of the prior art, this invention provides a data-mechanism fusion method and system for controlling deformation during the processing of thin-walled parts. This method solves the problems of low utilization rate of multi-source heterogeneous physical data and uncertain mechanism in traditional methods, and is particularly suitable for high-precision processing of complex thin-walled structural parts in aerospace.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] In a first aspect, the present invention provides a data-mechanism fusion-based method for controlling deformation during the processing of thin-walled parts, comprising:

[0009] The full range of process parameter values ​​and synchronously acquired multi-source heterogeneous physical data during the processing of thin-walled parts are obtained, and the multi-source heterogeneous physical data is preprocessed to obtain low-dimensional multi-source physical feature vectors.

[0010] A Bayesian optimization algorithm is used to screen highly sensitive process parameters and generate high-risk process parameter combinations.

[0011] Based on reinforcement learning algorithms, and combining highly sensitive process parameter combinations, low-dimensional multi-source physical feature vectors, and full process parameter values, candidate solutions for adjusting full process parameters are obtained.

[0012] A dual objective function is constructed with the goals of minimizing residual stress nonuniformity and deformation. Candidate solutions are adjusted based on the process parameters, and the dual objective function is solved using a non-dominated sorting genetic algorithm to obtain the optimal combination of process parameters. The processing of thin-walled parts is then controlled based on the optimal combination of process parameters. In the optimization solution using the non-dominated sorting genetic algorithm, the corresponding residual stress and deformation are predicted using a trained physical information neural network model based on each set of process parameter values ​​in the population and the corresponding low-dimensional multi-source physical feature vector.

[0013] Secondly, the present invention provides a data-mechanism fusion-based deformation control system for thin-walled part processing, comprising:

[0014] The acquisition module is configured to: acquire all process parameter values ​​during the thin-walled part processing and simultaneously collect multi-source heterogeneous physical data, and preprocess the multi-source heterogeneous physical data to obtain low-dimensional multi-source physical feature vectors;

[0015] The screening module is configured to use a Bayesian optimization algorithm to screen highly sensitive process parameters and generate high-risk process parameter combinations.

[0016] The adjustment module is configured to: obtain candidate solutions for adjusting all process parameters based on reinforcement learning algorithms, combining highly sensitive process parameter combinations, low-dimensional multi-source physical feature vectors, and all process parameter values;

[0017] The optimization module is configured to: construct a dual objective function with the goals of minimizing residual stress nonuniformity and deformation; adjust candidate solutions based on the process parameters; solve the dual objective function using a non-dominated sorting genetic algorithm to obtain the optimal combination of process parameters; and control the processing of thin-walled parts based on the optimal combination of process parameters. Specifically, in the optimization solution using the non-dominated sorting genetic algorithm, the corresponding residual stress and deformation are predicted using a trained physical information neural network model based on each set of process parameter values ​​in the population and the corresponding low-dimensional multi-source physical feature vector.

[0018] Thirdly, the present invention provides an electronic device including a memory and a processor, and computer instructions stored in the memory and executable on the processor, wherein the computer instructions, when executed by the processor, perform the method described in the first aspect.

[0019] Fourthly, the present invention provides a computer-readable storage medium for storing computer instructions, which, when executed by a processor, perform the method described in the first aspect.

[0020] The above one or more technical solutions have the following beneficial effects:

[0021] In this invention, a mechanism-data dual-driven model is established by combining physical information neural networks and Bayesian-reinforcement learning hybrid optimization strategies to improve the reliability and interpretability of the model. The Pareto optimal process parameters that satisfy the residual stress gradient and deformation are generated by the NSGA-III algorithm, thereby realizing the optimization of residual stress distribution and active suppression of deformation.

[0022] The present invention solves the problems of low utilization rate and unclear mechanism of multi-source heterogeneous physical data in traditional methods, and is particularly suitable for high-precision machining of complex thin-walled structural parts in aerospace.

[0023] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0024] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0025] Figure 1 This is a flowchart of the deformation control method for thin-walled parts in Embodiment 1 of the present invention;

[0026] Figure 2 This is a block diagram of the deformation control method for thin-walled parts in Embodiment 1 of the present invention;

[0027] Figure 3 This is a structural diagram of the thin-walled part processing device in Embodiment 1 of the present invention;

[0028] Figure 4 This is a schematic diagram of the multi-sensor information synchronous acquisition device in Embodiment 1 of the present invention;

[0029] Figure 5 This is a schematic diagram of the physical information neural network model structure in Embodiment 1 of the present invention;

[0030] Figure 6 This is a flowchart of the NSGA-III algorithm in Embodiment 1 of the present invention;

[0031] In the figure, 1. Magnetic base, 2. Infrared thermal imager, 3. Tool holder, 4. Accelerometer, 5. Force gauge, 6. Spindle, 7. Universal bracket, 8. Worktable, 9. Laser displacement sensor, 10. CNC milling machine, 11. Data collector, 12. Thin-walled part. Detailed Implementation

[0032] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0033] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0034] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0035] Example 1

[0036] This embodiment discloses a data-mechanism fusion method for controlling deformation during the machining of thin-walled parts, including:

[0037] The full range of process parameters during the thin-walled part processing is obtained and multi-source heterogeneous physical data is collected simultaneously. The multi-source heterogeneous physical data is preprocessed to obtain low-dimensional multi-source physical feature vectors.

[0038] A Bayesian optimization algorithm is used to screen highly sensitive process parameters and generate high-risk process parameter combinations. Based on a reinforcement learning algorithm, the candidate solutions for adjusting all process parameters are obtained by combining highly sensitive process parameters, multi-source physical feature vectors, and all process parameter values.

[0039] A dual objective function is constructed with the goals of minimizing residual stress nonuniformity and deformation. Candidate solutions are adjusted based on the process parameters, and the dual objective function is solved using a non-dominated sorting genetic algorithm to obtain the optimal combination of process parameters. The processing of thin-walled parts is then controlled based on the optimal combination of process parameters. In the optimization solution using the non-dominated sorting genetic algorithm, the corresponding residual stress and deformation are predicted using a trained physical information neural network model based on each set of process parameter values ​​in the population and the corresponding multi-source physical feature vectors.

[0040] In this embodiment, the experiment was set as the machining object of A7075-T6 aerospace aluminum alloy thin-walled part with dimensions of 100×100×50 mm and initial residual stress of -150MPa. An ultra-precision milling machine was used, equipped with an ultrasonic vibration spindle, and the cutting tool was a carbide end mill with a diameter of 6mm, 4 flutes, a rake angle of 10°, and a helix angle of 30°.

[0041] First, the thin-walled part processing apparatus involved in this embodiment will be described by way of example, such as Figure 3 As shown, the thin-walled part machining device includes a magnetic base 1, an infrared thermal imager 2, a tool holder 3, an accelerometer 4, a force gauge 5, a spindle 6, a universal support 7, a worktable 8, a laser displacement sensor 9, and a CNC milling machine 10; a Kistler 9257B triaxial piezoelectric force sensor is selected and installed below the workpiece fixture to measure the dynamic milling force in the X, Y, and Z axes. F x , F y , F z The sampling frequency was 10kHz; a PCB 352C33 accelerometer was selected and installed on the tool holder to measure vibration signals at a sampling frequency of 20kHz; an infrared thermal imager of FLIR A655sc was used to monitor the temperature field distribution in the processing area at a sampling frequency of 100Hz; a laser displacement sensor of Keyence LK-G500 was used, vertically aligned with the area to be measured on the thin-walled part, to measure three-dimensional deformation at a sampling frequency of 1kHz; the thin-walled part 12 was processed, and after processing, an X-ray diffractometer was used to perform offline measurements on the workpiece to obtain residual stress measurement values.

[0042] like Figure 4 As shown, all sensor signals are synchronously acquired through the same chassis and data collector 11. The time alignment of the above sensors is achieved by hardware trigger signal, and the synchronization error is ≤1μs. The force sensor signal is converted into a voltage signal by a charge amplifier. The vibration signal is connected to the acquisition card after being filtered by anti-aliasing. The infrared thermal imager data is transmitted through a gigabit network port.

[0043] The following is combined with Figures 1-2 The method for controlling deformation during the processing of thin-walled parts proposed in this embodiment will be described in detail:

[0044] Three key process parameters were used as experimental parameters: ultrasonic vibration parameters such as amplitude, frequency and mode; milling parameters such as spindle speed, feed rate and depth of cut; and thin-walled part structural parameters such as wall thickness. Each parameter was set to three levels, and a multi-level combined experiment was designed using the orthogonal array method.

[0045] As shown in Table 1, the experimental parameters and levels set in this experiment were determined using the L27(3) orthogonal array. 13A total of 27 experiments were conducted, with each experiment repeated 3 times. The experimental order was randomized to eliminate the influence of machine tool thermal deformation.

[0046] Table 1: Experimental parameters and levels

[0047]

[0048] For these 27 experimental combinations, multi-source heterogeneous physical data were collected simultaneously, and the maximum deformation after processing was directly measured by a laser displacement sensor. At the same time, residual stress was measured by an X-ray diffractometer, forming a basic mapping relationship between process parameters, sensor data, and deformation / residual stress.

[0049] The number of experimental groups, process parameters, sensor signals, and residual stress (deformation) are integrated into a four-dimensional feature tensor. ,in:

[0050]

[0051] in, Indicates the number of experimental groups; Indicates the dimension of process parameters; This indicates the dimensions of the sensing signal, namely vibration signal, temperature signal, and milling force; This represents the dimensions of the quality response, namely residual stress and deformation.

[0052] Feature extraction and dimensionality reduction fusion are performed on multi-source heterogeneous physical data composed of vibration signals, temperature signals, and milling forces. Specifically:

[0053] Step 11: Time-frequency domain alignment.

[0054] Using the milling force signal as a reference, the vibration signal is downsampled to 10 kHz by an 8th-order elliptic low-pass filter with a cutoff frequency of 5 kHz; the temperature signal is upsampled to 10 kHz by cubic spline interpolation, and FIR low-pass filtering is used to eliminate high-frequency noise, thus achieving time-frequency domain alignment of the data.

[0055] Step 12: Feature extraction.

[0056] For the aligned timing data, i.e., milling force ( F x , F y as well as F z The system extracts time-domain statistical features (such as maximum value, peak value, standard deviation, root mean square, skewness, peak index, impulse index, etc.) and frequency-domain features (such as wavelet packet energy entropy, energy ratio of each frequency band, etc.) from factors such as temperature and vibration, to form a high-dimensional feature set.

[0057] Step 13: Feature dimensionality reduction and fusion.

[0058] After feature extraction, all extracted high-dimensional feature sets are dimensionality reduced using Kernel Principal Component Analysis (KPCA). Radial basis functions are selected as the kernel function for KPCA, and the top K principal components with a cumulative contribution rate of over 95% are retained to form the final low-dimensional multi-source physical feature vector. G .

[0059] Step 14: Feature prediction model training.

[0060] Based on the 27 sets of sample datasets obtained from orthogonal experimental data, a random forest regressor was used to construct a model from process parameters. To low-dimensional multi-source physical feature vectors G The mapping model, i.e. This mapping model establishes a reliable surrogate model by learning the complex nonlinear relationship between process parameters and low-dimensional multi-source physical feature vectors. It can accurately obtain the corresponding multi-source physical features by inputting any combination of process parameters.

[0061] like Figure 5 The diagram shown is the structure of the Physical Information Neural Network (PINN) model in this embodiment. The input to the PINN model is a low-dimensional multi-source physical feature vector extracted from the collected data. G The output is the residual stress field, along with the corresponding process parameter combinations. The hidden layer configuration includes: LSTM temporal coding layer, CNN spatial feature extraction layer, and physical constraint embedding layer.

[0062] In this embodiment, the total loss function of the PINN network model includes a physical constraint loss function and a data-driven loss function.

[0063] Among them, the physical constraint loss function for:

[0064]

[0065] in, This represents the predicted residual stress, where t represents time. This represents the Johnson-Cook constitutive model. Describing the L2 norm, This represents the current plastic strain rate. This indicates the current absolute temperature.

[0066] The Johnson-Cook constitutive equation is:

[0067]

[0068] in, Indicates the initial yield stress. Indicates the strain hardening coefficient. Indicates the strain hardening index. Represents equivalent plastic strain. Represents the strain rate sensitivity coefficient. This represents the current plastic strain rate. Indicates the reference strain rate. Indicates the temperature softening index. This indicates the current absolute temperature.

[0069] Data-driven loss function It can be represented as:

[0070]

[0071] in, This represents the predicted residual stress field. This represents the measured residual stress field. This represents the predicted deformation. This represents the measured deformation. represents the weighting coefficient, and N represents the sample data in the training batch.

[0072] The total loss function is:

[0073] = +

[0074] in, These are the weighting coefficients of the physical constraint loss function.

[0075] In this embodiment, the total loss function of the physical information neural network model is formed by combining the data-driven loss function and the physical loss function. The data-driven loss function mainly focuses on the physical information neural network's ability to fit the measured data, ensuring the accuracy of the physical information neural network on known data; while the physical loss function constrains the model through physical laws, ensuring that the prediction results of the physical information neural network conform to physical laws, thereby improving the reliability and interpretability of the physical information neural network.

[0076] The experimental dataset was divided into an 80% training set, a 15% validation set, and a 5% test set. The AdamW optimizer (learning rate 3e-4, β1=0.9, β2=0.999) was used in conjunction with a cosine annealing strategy (initial period T0=50, multiple T...). mult =2) Optimize and set up early stopping mechanism and gradient clipping to stabilize the training process.

[0077] Then, the physical constraint fine-tuning phase begins. Initially, the feature extraction layer is fixed and only the physically constrained nodes are trained for 50 rounds. During the full-network joint training phase, the training is performed according to λ... phy=0.01*1.1 (k / 10) The weight coefficients of the physical loss function are dynamically adjusted, where k is the number of training epochs and the upper limit is λ. phy =0.1, and increases by 10% every 10 rounds until the target weight value is reached. This ultimately ensures that the validation set error meets the requirements.

[0078] By implementing INT8 quantization using TensorRT and calibrating the dynamic range using 50 sets of classic operating condition data, and by using FP16 and FP32 precision protection for the LSTM output layer and physical constraint layer respectively, the inference speed is increased by 3 times and the model size is compressed.

[0079] In this embodiment, the global sensitivity of process parameters to deformation is analyzed based on the Bayesian optimization algorithm (EI acquisition function), and key control factors are screened. Specifically:

[0080] To quantify the impact of each process parameter on the deformation of thin-walled parts, first, the types of parameters with different effects on deformation are identified, and then the normalized sensitivity index of the process parameters to deformation is calculated.

[0081]

[0082] in, Indicates the first i Sensitivity to normalization of process parameters Indicates the first i Process parameters such as amplitude and feed rate. This represents the maximum deformation under the full combination of process parameters.

[0083] Screening sensitivity Highly sensitive process parameters such as amplitude and feed rate reduce the optimization dimensionality.

[0084] By using the Expected Improvement (EI) function as the acquisition function for Bayesian optimization, 20 high-risk parameter combinations were selected based on the EI function, and these combinations were used as a set of potential initial exploration points.

[0085] Based on Bayesian optimization to identify highly sensitive process parameters such as amplitude and feed rate, the DDPG algorithm is used for real-time optimization of process parameters.

[0086] The state space is:

[0087]

[0088] in, as well as This indicates the highly sensitive process parameters at time t, namely amplitude and feed rate. Indicates the standard deviation of residual stress distribution; Indicates the remaining tool life; This represents the maximum deformation at time t-1; the superscript T indicates transpose. This represents the low-dimensional multi-source physical feature vector at time t.

[0089] The motion space employs a layered structure design: sensitive parameters allow for adjustment within a ±30% working range, with a step size factor of 0.8, enabling rapid exploration of optimal parameter values ​​over a wider range; secondary sensitive parameters, such as frequency... f , depth of cut Limiting the range to ±15% with a step size factor of 0.4 assists in optimization to a certain extent while avoiding excessive adjustments that could interfere with the core optimization objective. Non-sensitive parameters such as vibration mode M and wall thickness h are only allowed to be fine-tuned by ±5% with a step size factor of 0.1, which reduces the waste of computational resources caused by ineffective adjustments and ensures that the machining process conforms to the boundary constraints of the part's structural design and process mode.

[0090] The DDPG algorithm is used to adjust process parameters in real time for local optimization. The reward function is designed as follows:

[0091]

[0092] in, The sensitivity coefficient of amplitude to deformation. Indicates the remaining tool life. This is the sensitivity coefficient of the feed rate to the standard deviation of residual stress. The standard deviation of the residual stress distribution at the current moment is represented in MPa; 0.1 represents the tool life regularization coefficient. The learning rate is set to 0.0001, and the feature tensor is collected every 50 steps to update the state. The DDPG output action is the adjustment amount of all process parameters, gradually approaching the deformation minimization target.

[0093] The two stages work together to obtain the optimal adjustment strategy for the machining process parameters of thin-walled parts and the corresponding combination of process parameters. This provides high-quality input for subsequent NSGA-III multi-objective optimization.

[0094] Establish a model that includes residual stress nonuniformity (std(σ)) and maximum deformation (σ). Bi-objective function:

[0095]

[0096] In the formula, Indicates the process parameters to be optimized. Indicates the standard deviation of residual stress distribution. This indicates the maximum deformation of a thin-walled component.

[0097] Based on the full set of candidate solutions for process parameters output during the reinforcement learning phase, a high-quality combination of process parameters is selected from them. This serves as the initial population for the non-dominated genetic algorithm. During each evolutionary process, for each set of process parameters... The evaluations were conducted sequentially, starting with the feature prediction model established in step 14. Predicting multi-source physical feature vectors G Then, the process parameters are combined with the predicted features, and a pre-trained physical information neural network model is invoked to predict the residual stress and maximum deformation. Finally, the residual stress non-uniformity (std(σ)) and maximum deformation (σ) corresponding to each set of process parameters in the population are predicted. Batch prediction and evaluation are performed.

[0098] like Figure 6 As shown, the specific process of optimizing the candidate solutions for all process parameters output during the reinforcement learning stage using a non-dominated genetic algorithm is as follows:

[0099] Step 21: Initialize the population.

[0100] From the full set of candidate solutions for adjusting process parameters output by reinforcement learning, a preset number of process parameter combinations, such as 100, are randomly selected to form the initial population.

[0101] Step 22: Assess population fitness.

[0102] For each set of process parameters in the population, a corresponding low-dimensional multi-source physical feature vector is first generated using a random forest regressor.

[0103] The process parameters and low-dimensional multi-source physical feature vectors are input into the trained physical information neural network (PINN) to predict the residual stress non-uniformity and maximum deformation corresponding to the set of parameters, which are used as fitness evaluation indicators.

[0104] Step 23: Quick Non-Dominated Sort.

[0105] Based on the dual objective function, the non-dominated hierarchy is divided among all individuals in the population:

[0106] Individuals for which no other individual can simultaneously outperform it in both objectives are classified as the first frontier;

[0107] Individuals dominated by individuals in the first frontier, but not by other individuals, are classified as the second frontier.

[0108] This process continues until all individuals have completed the hierarchical division and the target advantages of each parameter combination are clearly defined.

[0109] Step 24: Calculate the crowding distance.

[0110] For each individual in the frontier layer, calculate its crowding distance to measure the sparsity of the individual in the frontier layer: the greater the distance, the fewer high-quality parameter combinations around the individual, and retaining it can improve population diversity;

[0111] The distance calculation method is as follows: sort each target dimension (residual stress, deformation) separately, and take the sum of the differences between adjacent individuals in that dimension to avoid the algorithm getting trapped in local optima.

[0112] Step 25: Select the parent individual.

[0113] Parental selection is performed by combining frontier hierarchy and crowding distance: individuals with higher frontier hierarchy are given priority.

[0114] If individuals are in the same frontal layer, select those with greater crowding distance to ensure that the parent population has both "high quality" and "diversity," providing a good foundation for subsequent genetic manipulation.

[0115] Step 26: Genetic manipulation (crossover and mutation).

[0116] Crossover: The simulated binary crossover (SBX) strategy is used to reconfigure the parameters of the selected parent individuals.

[0117] Mutation: For the offspring individuals after crossover, a certain process parameter is randomly adjusted according to a preset mutation probability (such as 0.1). For example, the amplitude is mutated from 10μm to 10.5μm to expand the parameter search range and avoid premature convergence of the algorithm.

[0118] Step 27: Generate offspring population.

[0119] Through the above crossover and mutation operations, a combination of offspring process parameters with the same number as the initial population is generated, forming the offspring population.

[0120] Step 28: Merge and secondary sorting and filtering.

[0121] The parent population and the offspring population are merged to form a mixed population that is twice the size of the initial population.

[0122] Repeat steps 23 and 24 to re-divide the frontier hierarchy of the mixed population and calculate the crowding distance.

[0123] Step 29: Elite Selection.

[0124] Individuals are selected from the mixed population, and those with high frontier levels and high crowding distances are selected in turn until the number is restored to the initial population size to form the next generation of population, thus achieving "elite preservation" and ensuring that the overall quality of each generation of population does not decline.

[0125] Iteration termination check: Check if the current iteration count has reached the preset maximum value (e.g., 200 iterations).

[0126] If not achieved, return to step 22 and repeat the "evaluation-sorting-selection-genetics" process;

[0127] If the target is reached, stop iterating and proceed to the result output stage.

[0128] After the iteration terminates, the first frontier individual in the final population is extracted to form the Pareto optimal solution set.

[0129] The parameter combination that meets the actual processing requirements is selected from the solution set and used as the final optimal process parameters for thin-walled part processing.

[0130] The final optimal solution is a high-quality solution set that satisfies the following conditions.

[0131] Residual stress gradient ≤ 40 MPa / mm;

[0132] Deformation ≤ 0.05 mm (for a wall thickness of 1 mm).

[0133] This embodiment constructs a multi-sensor synchronous monitoring platform to collect multi-source heterogeneous physical signals such as cutting force, vibration, and temperature in real time during the ultrasonic milling of thin-walled parts. Based on multi-factor orthogonal experiments, a four-dimensional mapping database of experimental group number, process parameters, sensor signals, and residual stress is constructed. Combining a neural network of physical information with a Bayesian-reinforcement learning hybrid optimization strategy, a mechanism-data dual-driven model is established to quantify the cross-scale correlation between multi-source data and deformation. The NSGA-III algorithm is used to generate Pareto optimal process parameters that satisfy the residual stress gradient and deformation, thereby achieving optimization of residual stress distribution and active suppression of deformation. This embodiment solves the problems of low utilization rate of multi-source heterogeneous physical data and unclear mechanism in traditional methods, and is particularly suitable for high-precision machining of complex thin-walled structural parts in aerospace applications.

[0134] Example 2

[0135] The purpose of this embodiment is to provide a deformation control system for thin-walled part processing, including:

[0136] The acquisition module is configured to: acquire all process parameter values ​​during the thin-walled part processing and simultaneously collect multi-source heterogeneous physical data, and preprocess the multi-source heterogeneous physical data to obtain low-dimensional multi-source physical feature vectors;

[0137] The screening module is configured to use a Bayesian optimization algorithm to screen highly sensitive process parameters and generate high-risk process parameter combinations.

[0138] The adjustment module is configured to: obtain candidate solutions for adjusting all process parameters based on reinforcement learning algorithms, combining highly sensitive process parameter combinations, low-dimensional multi-source physical feature vectors, and all process parameter values;

[0139] The optimization module is configured to: construct a dual objective function with the goals of minimizing residual stress nonuniformity and deformation; adjust candidate solutions based on the process parameters; solve the dual objective function using a non-dominated sorting genetic algorithm to obtain the optimal combination of process parameters; and control the processing of thin-walled parts based on the optimal combination of process parameters. Specifically, in the optimization solution using the non-dominated sorting genetic algorithm, the corresponding residual stress and deformation are predicted using a trained physical information neural network model based on each set of process parameter values ​​in the population and the corresponding low-dimensional multi-source physical feature vector.

[0140] In further embodiments, the following is also provided:

[0141] An electronic device includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor. When executed by the processor, the computer instructions perform the method described in Embodiment 1. For brevity, further details are omitted here.

[0142] It should be understood that in this embodiment, the processor can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc.

[0143] Memory may include read-only memory and random access memory, and provides instructions and data to the processor. A portion of memory may also include non-volatile random access memory. For example, memory may also store information about the device type.

[0144] A computer-readable storage medium for storing computer instructions, which, when executed by a processor, perform the method described in Embodiment 1.

[0145] The method in Embodiment 1 can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor. The software modules can reside in readily available storage media in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory; the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, a detailed description is not provided here.

[0146] A computer program product includes a computer program that, when executed by a processor, implements the method described in Embodiment 1.

[0147] The present invention also provides at least one computer program product tangibly stored on a non-transitory computer-readable storage medium. The computer program product includes computer-executable instructions, such as instructions included in program modules, which execute in a device on a target real or virtual processor to perform the processes / methods described above. Typically, program modules include routines, programs, libraries, objects, classes, components, data structures, etc., that perform specific tasks or implement specific abstract data types. In various embodiments, the functionality of program modules can be combined or divided among program modules as needed. The machine-executable instructions for the program modules can execute within a local or distributed device. In a distributed device, the program modules can reside in both local and remote storage media.

[0148] The computer program code used to implement the methods of the present invention may be written in one or more programming languages. This computer program code may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the computer or other programmable data processing device, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a computer, partially on a computer, as a stand-alone software package, partially on a computer and partially on a remote computer, or entirely on a remote computer or server.

[0149] In the context of this invention, computer program code or related data may be carried by any suitable carrier to enable a device, apparatus, or processor to perform the various processes and operations described above. Examples of carriers include signals, computer-readable media, and the like. Examples of signals may include electrical, optical, radio, sound, or other forms of propagation signals, such as carrier waves, infrared signals, etc.

[0150] Those skilled in the art will recognize that the units and algorithm steps described in conjunction with the embodiments herein can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0151] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A data-mechanism fusion method for controlling deformation during the machining of thin-walled parts, characterized in that, include: The full range of process parameters during the thin-walled part processing is obtained and multi-source heterogeneous physical data is collected simultaneously. The multi-source heterogeneous physical data is preprocessed to obtain low-dimensional multi-source physical feature vectors. A Bayesian optimization algorithm is used to screen highly sensitive process parameters and generate high-risk process parameter combinations. Based on reinforcement learning algorithms, and combining highly sensitive process parameter combinations, low-dimensional multi-source physical feature vectors, and full process parameter values, candidate solutions for adjusting full process parameters are obtained. A dual objective function is constructed with the goals of minimizing residual stress nonuniformity and deformation. Candidate solutions are adjusted based on the process parameters, and the dual objective function is solved using a non-dominated sorting genetic algorithm to obtain the optimal combination of process parameters. The processing of thin-walled parts is then controlled based on the optimal combination of process parameters. In the optimization solution using the non-dominated sorting genetic algorithm, the corresponding residual stress and deformation are predicted using a trained physical information neural network model based on each set of process parameter values ​​in the population and the corresponding low-dimensional multi-source physical feature vector.

2. The data-mechanism fusion method for controlling deformation during thin-walled part processing as described in claim 1, characterized in that, The preprocessing of multi-source heterogeneous physical data specifically involves: time-frequency domain alignment of the multi-source heterogeneous physical data, extraction of time-domain statistical features and frequency-domain features, and dimensionality reduction of the extracted features.

3. The data-mechanism fusion method for controlling deformation during thin-walled part processing as described in claim 1, characterized in that, The loss function of the physical information neural network model is: ； ； ； in, The physical constraint loss function, For data-driven loss functions, This represents the predicted residual stress, and t represents time. This represents the Johnson-Cook constitutive model. Represents the L2 norm. This represents the predicted residual stress field. This represents the measured residual stress field. This represents the predicted deformation. This represents the measured deformation. Indicates the weighting coefficient. The weighting coefficients represent the physical constraint loss function.

4. The data-mechanism fusion method for controlling deformation during thin-walled part processing as described in claim 1, characterized in that, The construction of the training set for the physical information neural network model is as follows: ultrasonic vibration parameters, milling parameters, and structural parameters are used as experimental parameters. Each experimental parameter is set with three levels. Multi-level combined experiments are designed using the orthogonal array method. Multi-source heterogeneous data are collected for each group of experiments and residual stress is measured to construct a preliminary training sample set.

5. The data-mechanism fusion method for controlling deformation during thin-walled part processing as described in claim 1, characterized in that, A Bayesian optimization algorithm is used to screen highly sensitive process parameters and generate high-risk process parameter combinations, specifically: The global sensitivity of different process parameters to deformation is calculated based on the normalized sensitivity index, and the highly sensitive process parameters are determined. The expected improvement function is used as the acquisition function to generate high-risk process parameter combinations based on highly sensitive process parameters.

6. The data-mechanism fusion method for controlling deformation during thin-walled part processing as described in claim 1, characterized in that, Based on reinforcement learning algorithms, and combining highly sensitive process parameter combinations, low-dimensional multi-source physical feature vectors, and all process parameter values, candidate solutions for adjusting all process parameters are obtained, specifically: The current high-sensitivity process parameter value, the current low-dimensional multi-source physical feature vector, the standard deviation of residual stress distribution, the remaining tool life, and the maximum deformation of the previous moment are used as the state space. The adjustment range and step size are designed in layers according to the sensitivity of process parameters, and the adjustment amount of process parameters is taken as the action. A reward function is designed based on the maximum deformation and the standard deviation of residual stress. The system is iteratively trained to generate candidate solutions for adjusting all process parameters.

7. The data-mechanism fusion method for controlling deformation during thin-walled part processing as described in claim 1, characterized in that, The solution is obtained by non-dominated sorting genetic algorithm. The solution set is divided into different Pareto front levels by fast non-dominated sorting and crowding distance is calculated. After merging the parent and offspring populations by adopting an elite retention strategy, individuals with high front level and large crowding distance are selected to enter the next generation.

8. A data-mechanism fusion-based deformation control system for thin-walled part machining, characterized in that, include: The acquisition module is configured to: acquire all process parameter values ​​during the thin-walled part processing and simultaneously collect multi-source heterogeneous physical data, and preprocess the multi-source heterogeneous physical data to obtain low-dimensional multi-source physical feature vectors; The screening module is configured to use a Bayesian optimization algorithm to screen highly sensitive process parameters and generate high-risk process parameter combinations. The adjustment module is configured to: obtain candidate solutions for adjusting all process parameters based on reinforcement learning algorithms, combining highly sensitive process parameter combinations, low-dimensional multi-source physical feature vectors, and all process parameter values; The optimization module is configured to: construct a dual objective function with the goals of minimizing residual stress nonuniformity and deformation; adjust candidate solutions based on the process parameters; solve the dual objective function using a non-dominated sorting genetic algorithm to obtain the optimal combination of process parameters; and control the processing of thin-walled parts based on the optimal combination of process parameters. Specifically, in the optimization solution using the non-dominated sorting genetic algorithm, the corresponding residual stress and deformation are predicted using a trained physical information neural network model based on each set of process parameter values ​​in the population and the corresponding low-dimensional multi-source physical feature vector.

9. An electronic device, characterized in that, It includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor, which, when executed by the processor, perform the method according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, Used to store computer instructions, which, when executed by a processor, perform the method described in any one of claims 1-7.

Images