A high-precision intelligent machining method and system for mold blanks

Abstract

This invention relates to the field of mold blank processing technology, specifically to a high-precision intelligent processing method and system for mold blanks, comprising: applying ultrasonic guided waves to the mold blank and constructing a target stress distribution field inside the mold blank based on the feedback signal returned by the ultrasonic guided waves; using preset principal stresses and target deformation amounts in the preset processing process as optimization targets, performing processing path planning on the target stress distribution field to generate stress processing data including target cutting parameters and target stress release paths; and controlling the processing equipment to perform processing operations on the mold blank according to the stress processing data, thereby solving the technical problem of low processing accuracy and stability caused by deformation induced by stress release after mold blank processing.

Description

Technical Field

[0001] This invention relates to the field of mold blank processing technology, specifically to a high-precision intelligent processing method and system for mold blanks. Background Technology

[0002] The precision of mold blank machining plays a crucial role in ensuring the dimensional stability, service life, and molding quality of the final mold product, and is a core aspect of precision manufacturing.

[0003] Currently, the main way to improve the machining accuracy of mold blanks is to perform residual stress detection on the mold blanks after machining, and then use subsequent processes such as heat treatment, vibration aging or compensation machining to eliminate stress and correct dimensions based on the detection results. This is a post-processing mode of "machining-measurement-remachining". It can be seen that it cannot actively predict and control the dynamic evolution and redistribution of internal stress caused by material removal during the machining process. This makes it difficult to suppress the deformation caused by stress release in advance during the machining stage, which seriously restricts the further improvement of machining accuracy and the fundamental guarantee of stability. Summary of the Invention

[0004] To address the technical problem of low machining accuracy and stability caused by deformation resulting from stress release after mold blank machining, this application provides a high-precision intelligent machining method and system for mold blanks.

[0005] The high-precision intelligent machining method and system for mold blanks provided in this application adopts the following technical solution: A high-precision intelligent machining method for mold blanks includes: By applying ultrasonic guided waves to the mold blank, the target stress distribution field inside the mold blank is constructed based on the feedback signal returned by the ultrasonic guided waves. Using the preset principal stress and target deformation in the preset processing process as optimization targets, the processing path is planned for the target stress distribution field, generating stress processing data including target cutting parameters and target stress release path; The control processing equipment performs processing operations on the mold blank based on stress processing data.

[0006] Furthermore, the steps for constructing the target stress distribution field inside the mold blank based on the feedback signal returned by the ultrasonic guided wave include: Based on the preset guided wave modes and multiple propagation direction parameters, the feedback signal is subjected to adaptive beamforming synthesis processing to obtain multiple directional wave field signals; After performing time-frequency synchronization and alignment processing on each directional wavefield signal to obtain multi-angle wavefield data, the multi-angle wavefield data is restructured according to the transmission and reception relationship of ultrasonic guided waves to obtain full matrix feedback data. Dispersion analysis and mode separation processing are performed on the full matrix feedback data to obtain a multimodal feedback feature set; The multimodal feedback feature set is input into the deep convolutional inversion network. The nonlinear mapping relationship between features and stress field established inside the deep convolutional inversion network is used to process the multimodal feedback feature set and output the initial stress distribution field. The initial stress distribution field is iterated and physically constrained based on the material physical information of the mold blank to obtain the target stress distribution field.

[0007] Furthermore, the steps for performing dispersion analysis and mode separation processing on the full matrix feedback data to obtain the multimodal feedback feature set include: After performing a two-dimensional Fourier transform on the full matrix feedback data to obtain the wavenumber-frequency domain energy spectrum distribution, parameter matching is performed based on the measured energy spectrum peak values ​​extracted from the wavenumber-frequency domain energy spectrum distribution to obtain the dispersion curves of each feedback signal. Based on the dispersion curves, filters corresponding to each feedback signal are selected to filter the full matrix feedback data, resulting in multiple single-mode time-domain signals. Based on the physical relationship between the propagation speed and energy attenuation coefficient of each single-mode time-domain signal and stress, feature extraction is performed on each single-mode time-domain signal to obtain a multi-mode feedback feature set.

[0008] Furthermore, the steps for processing the multimodal feedback feature set and outputting the initial stress distribution field by leveraging the nonlinear mapping relationship between features and stress field established within the deep convolutional inversion network include: By using the encoder within the deep convolutional inversion network, multiple layers of convolution and pooling operations are performed on the multimodal feedback feature set to obtain multiple deep abstract feature tensors; Cross-modal feature interaction and spatial weight focusing are performed on each deep abstract feature tensor to obtain multiple context-aware features; Multiple context-aware features are input into the decoder within the deep convolutional inversion network for decoding to obtain the initial stress distribution field.

[0009] Furthermore, the steps of using the preset principal stress and target deformation in the preset processing process as optimization objectives, planning the processing path for the target stress distribution field, and generating stress processing data including target cutting parameters and target stress release paths include: Based on the target stress distribution field, the processing process is simulated to obtain stress evolution-predicted deformation pairs under different candidate processing paths; Based on the optimization objective and the stress evolution-predicted deformation pairs, the corresponding candidate machining paths and their cutting parameters are iteratively searched to obtain the target machining path and its corresponding target cutting parameters. Identify the target stress concentration area in the target machining path, insert preset machining data into the target stress concentration area, and generate the target stress release path.

[0010] Furthermore, based on the optimization objective and each stress evolution-predicted deformation pair, the steps of iteratively searching for the corresponding candidate machining paths and their cutting parameters to obtain the target machining path and its corresponding target cutting parameters include: After constructing a multi-objective function based on the optimization objective, the multi-objective function is encoded according to each candidate machining path and the cutting parameters corresponding to each candidate machining path to generate a set of candidate solutions; The objective function values ​​of each candidate solution in the candidate scheme set are obtained by calculating the encoded multi-objective function through each stress evolution-predicted deformation pair. The candidate solution set is iterated based on the objective function values ​​to obtain the candidate path parameter set; Based on a preset processing efficiency threshold, the target processing path and the target cutting parameters corresponding to the target processing path are determined from the candidate path parameter set.

[0011] Furthermore, the steps of identifying target stress concentration areas in the target machining path and inserting preset machining data into these areas to generate a target stress relief path include: Geometric process feature analysis is performed on the target machining path to obtain the path features of the target machining path; Based on path characteristics, after identifying the target stress concentration area from the target processing path, the corresponding preset processing data is obtained according to the stress relief strategy that matches the target stress concentration area. The original toolpath segment in the target machining path is replaced with preset machining data to generate the target stress relief path.

[0012] This application also provides a high-precision intelligent machining system for mold blanks, including: The information construction module is used to construct the target stress distribution field inside the mold blank based on the feedback signal returned by the ultrasonic guided wave after applying ultrasonic guided wave to the mold blank. The information generation module is used to plan the processing path of the target stress distribution field with the preset principal stress and target deformation in the preset processing process as the optimization target, and generate stress processing data including target cutting parameters and target stress release path; The machining control module is used to control the machining equipment to perform machining operations on the mold blank based on stress machining data.

[0013] Beneficial effects achieved: This application provides a high-precision intelligent machining method for mold blanks, comprising: applying ultrasonic guided waves to the mold blank and constructing a target stress distribution field inside the mold blank based on the feedback signal returned by the ultrasonic guided waves; using preset principal stresses and target deformation amounts in the preset machining process as optimization targets, performing machining path planning on the target stress distribution field to generate stress machining data including target cutting parameters and target stress release paths; and controlling the machining equipment to perform machining operations on the mold blank based on the stress machining data.

[0014] In this application, by utilizing ultrasonic guided waves to non-destructively construct the target stress distribution field inside the mold blank, the magnitude and location distribution of latent residual stress are determined before processing, thereby transforming uncontrollable internal stress into quantifiable and locatable clear processing constraints. Next, in the processing path planning stage, with the optimization objective of suppressing principal stress and final deformation during processing, the target stress distribution field is actively planned. The generated stress processing data not only includes conventional target cutting parameters but also a target stress release path customized based on the stress field. This allows the tool to actively guide and release stress according to a preset strategy while processing the material. Finally, the processing equipment strictly follows the stress processing data, ensuring that the material removal process and stress control process are carried out simultaneously. While the workpiece is being formed, its internal stress has already been pre-homogenized and reduced, thus preventing uncontrollable deformation caused by the spontaneous release of residual stress after processing, fundamentally guaranteeing the final dimensional accuracy and long-term stability of the workpiece. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the steps of a high-precision intelligent machining method for mold blanks according to this application; Figure 2 This is a schematic diagram of a high-precision mold base intelligent machining system according to this application.

[0016] Explanation of reference numerals in the attached figures: 10. Information Construction Module; 20. Information Generation Module; 30. Processing Control Module. Detailed Implementation

[0017] The following combination Figure 1 and Figure 2 This application will be described in further detail.

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0019] It should be noted that if the embodiments of the present invention involve directional indications (such as up, down, left, right, front, back, etc.), the directional indications are only used to explain the relative positional relationship and movement of the components in a specific posture. If the specific posture changes, the directional indications will also change accordingly.

[0020] Furthermore, if the embodiments of this invention involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the use of "and / or" or "and / or" throughout the text includes three parallel solutions. For example, "A and / or B" includes solution A, solution B, or a solution where both A and B are satisfied simultaneously. Furthermore, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.

[0021] This application discloses a high-precision intelligent machining method for mold blanks.

[0022] Please refer to Figure 1 The high-precision intelligent machining method for mold blanks proposed in this embodiment includes steps S10 to S30: Step S10: After applying ultrasonic guided waves to the mold blank, the target stress distribution field inside the mold blank is constructed based on the feedback signal returned by the ultrasonic guided waves.

[0023] This step involves applying ultrasonic guided waves to the mold blank and constructing the target stress distribution field inside the mold blank based on the returned feedback signal. This transforms the residual stress inside the mold blank, which is difficult to predict and quantify in traditional machining, from a latent state into a visualized data map. This provides initial input and decision-making basis for the entire intelligent machining process, avoiding the passive mode of relying on experience or post-event detection. It enables subsequent machining path planning to be based on stress distribution data rather than guesswork, laying a reliable data foundation for proactively intervening in and controlling machining deformation caused by internal stress from the source.

[0024] In this process, an ultrasonic sensor is tightly attached to a specific detection point on the surface of the mold using a coupling agent (such as grease or water). Then, an ultrasonic excitation device sends an electrical pulse signal of a specific frequency and waveform to the ultrasonic sensor. The ultrasonic sensor converts this electrical pulse signal into mechanical vibration, thereby exciting and coupling a specific mode of ultrasonic guided wave inside the mold. As this ultrasonic guided wave propagates inside the mold, its propagation speed, attenuation coefficient, frequency spectrum, and other characteristics change systematically due to the influence of the residual stress field within the material. A receiving ultrasonic sensor, in receiving mode, captures these stress-modulated ultrasonic guided waves and converts them into a feedback signal in the form of an electrical signal.

[0025] After completing step S10, i.e. successfully constructing the target stress distribution field, the ultrasonic sensor will be removed from the mold.

[0026] Step S20: Using the preset principal stress and target deformation in the preset processing process as optimization targets, perform processing path planning on the target stress distribution field to generate stress processing data including target cutting parameters and target stress release path.

[0027] This step uses preset principal stress and target deformation as optimization objectives to perform intelligent machining path planning on the constructed target stress distribution field. This transforms the abstract stress distribution data into an executable sequence of machining instructions, thereby generating stress machining data. This stress machining data not only includes target cutting parameters for conventional control of size and shape, but also target stress release paths designed based on the target stress distribution field. This transforms the machining strategy from the traditional "geometry-driven" to "stress-geometry co-driven," ensuring that while the tool removes material according to the stress machining data, it accurately executes corresponding machining movements and cutting operations in high-stress concentration areas. This guides and homogenizes internal stress in real time during the mold blank machining process, proactively suppressing workpiece deformation caused by stress redistribution from the process source. This provides a crucial process guarantee for achieving high-precision and high-stability final machining results.

[0028] Step S30: Control the processing equipment to perform processing operations on the mold blank based on the stress processing data.

[0029] First, based on the generated stress processing data containing the target cutting parameters and the target stress release path, the post-processor converts it into a standard G-code program that the processing equipment can recognize. This standard G-code program not only defines the motion trajectory of the tool in three-dimensional space, but also embeds specific process instructions corresponding to different path segments.

[0030] During machining, the machining equipment interprets and executes the standard G-code program. The motion controller on the machining equipment, in coordination with the servo drives, controls the tool to move along the predetermined target stress release path based on the motion trajectory. At the same time, the spindle control system on the machining equipment adjusts the speed and power in real time according to the target cutting parameters. It should be noted that when the standard G-code program executes to the target stress release path segment for a high stress concentration area, the spindle control system will automatically trigger and switch to the preset local finishing cycle mode according to the embedded process instructions. In this mode, the spindle control system will adjust the feed rate, cutting depth, and may even activate additional ultrasonic vibration or micro-lubrication devices to perform special machining operations aimed at equalizing stress, thereby ensuring that the intention of stress management is achieved. This transforms the strategy planned based on the target stress distribution field in the early stage into machining actions in the physical world, so that the material removal process and the active guidance and release process of internal stress are precisely synchronized in time and space. Thus, the root cause of internal stress that leads to subsequent deformation is eliminated in advance while the mold is being formed, ultimately ensuring that the mold blank obtains high dimensional accuracy and long-term stability.

[0031] In one feasible implementation, step S10 may specifically include steps S11 to S15: Step S11: Based on the preset guided wave mode and multiple propagation direction parameters, the feedback signal is subjected to adaptive beamforming synthesis processing to obtain multiple directional wave field signals.

[0032] It should be noted that the preset guided wave mode refers to the ultrasonic guided wave mode selected in advance based on the material properties, geometric thickness, and stress detection sensitivity requirements of the mold blank. For example, in plate-shaped mold blanks, stress-sensitive Lamb wave symmetric or antisymmetric modes are often selected. The multiple propagation direction parameters are a set of discretized wave propagation direction angles determined by theoretical wave propagation simulation or experimental calibration based on the detection area division of the mold blank and the topology of the ultrasonic sensor array. These parameters are used to cover all spatial orientations that need to be detected inside the mold blank.

[0033] First, based on the material type, geometric thickness, and selected preset waveguide mode of the mold blank, and combined with the dispersion curve of the preset waveguide mode in the mold blank, a frequency point that makes the ultrasonic guided wave sensitive to changes in internal stress and has a high signal-to-noise ratio is selected as the center excitation frequency. This central excitation frequency Typically, within the operating bandwidth of the ultrasonic sensor, the phase velocity is then obtained directly from or interpolated from the dispersion curve of the preset guided wave mode. (The dispersion curve is obtained in advance through experimental measurements on standard samples and describes the phase velocity.) (Correspondence between frequency and other parameters). In obtaining the central excitation frequency... and the corresponding phase velocity Then, the magnitude of the wave number k can be determined using the fundamental formula of physics, Formula 1: The wave number at the central excitation frequency was calculated. .

[0034] Given the geometric coordinates of the ultrasonic sensor on the mold surface, for the nth sensor in an ultrasonic sensor array mounted on a mold, the element value corresponding to its guide vector is obtained using the formula... Calculate, where, It is the nth guiding vector; The imaginary unit; It is a wavenumber vector, derived from the wavenumber. The propagation direction vector corresponding to the propagation direction parameter Decision, that is ; The position vector of the nth sensor with the array center as the origin; symbol " "" represents the dot product of vectors. This yields a guide vector for the propagation direction parameter, which characterizes the theoretical phase difference when the wavefront arrives at each sensor. The wavefront refers to the equiphase surface of the ultrasonic guided wave that propagates from the mold interior in the corresponding propagation direction and arrives at the sensor array.

[0035] As described above, after calculating the steering vector corresponding to each propagation direction parameter, the covariance matrix of the feedback signal received by the ultrasonic sensor array at the current moment (or within a short time window) is estimated: The covariance matrix represents time The sample covariance matrix is ​​formed by the snapshot vector of the feedback signals from the ultrasonic sensors. Through formula In practical calculations, the mathematical expectation is usually approximated by averaging L consecutive or overlapping snapshot vectors. Next, for the i-th propagation direction, the optimal complex weight coefficients are solved. The corresponding optimization problem is to ensure that the signal from the i-th propagation direction is distortion-free (i.e., satisfying the constraint condition that the conjugate transpose of the weight vector wn and the steering vector...). Given that the inner product is equal to 1, the output power (i.e., variance) Its analytical solution (optimal weight vector) is given by the formula. Provided.

[0036] Finally, for each preset propagation direction, a corresponding optimal weight vector is calculated. The weighting coefficients of this group are summed with the feedback signal using a complex weighted sum (i.e., ... This generates a directional wavefield signal focused on the corresponding direction of the innovation, thereby separating and reconstructing the directional wavefield signal that propagates along a series of specific propagation directions and has a significantly improved signal-to-noise ratio from the feedback signal containing multipath aliasing and noise. This lays a high-fidelity data foundation for subsequent multi-angle wavefield data fusion and accurate stress inversion.

[0037] in, This represents the snapshot vector at time t, composed of the feedback signals from all the ultrasonic sensors. This represents the feedback signal received by the Nth ultrasonic sensor at time t; Representing vectors The conjugate transpose of is obtained by first taking the conjugate complex number and then transposing it; L represents the number of snapshots used to estimate the covariance matrix; This represents the l-th time sampling point used for estimation; This represents the optimal complex weight vector corresponding to the i-th propagation direction, used to weight the sensor signal; This represents the steering vector corresponding to the i-th propagation direction, whose elements encode the theoretical phase difference of the signal arriving at each sensor from that direction; Represents the sample covariance matrix The inverse matrix; This indicates that at time t, after beamforming processing, the output is a directional wave field signal corresponding to the i-th propagation direction; Represents the optimal weight vector The conjugate transpose of; This indicates that the weight vector conjugate transpose With the snap vector of the feedback signal Perform complex inner product operations (i.e., weighted summation).

[0038] Step S12: Perform time-frequency synchronization alignment processing on each directional wave field signal to obtain multi-angle wave field data. Then, reorganize the multi-angle wave field data according to the transmission and reception relationship of the ultrasonic guided wave to obtain full matrix feedback data.

[0039] ① Because each directional wave field signal has a systematic deviation in its time start point and initial phase, it is necessary to use a unified reference time base and phase center for calibration. Specifically, this can be done by selecting a reference signal, such as a directional wave field signal in the normal direction or a theoretical simulation signal, as the time-frequency reference base, and then using the directional wave field signal... With reference signal Discretize using the same sampling frequency to obtain the corresponding discrete-time series. and ,in The integer sample index is used, and then, within the preset integer delay search range [-a, a], for each possible integer delay amount b, the discrete complex cross-correlation function is calculated. The specific calculation formula is as follows: Therefore, the discrete complex cross-correlation function is obtained. It is itself a complex sequence, and each value of it contains the amplitude correlation and phase difference information between the directional wave field signal and the reference signal at a specific delay b. After completing the calculation by traversing all b values, a complete discrete complex cross-correlation function sequence is obtained. The peak position of this sequence initially indicates the integer sample delay value between the corresponding directional wave field signal and the reference signal.

[0040] After obtaining the discrete complex cross-correlation function sequence and initially determining the integer sample delay corresponding to the peak position of the discrete complex cross-correlation function sequence, since the relative delay between the actual directional wave field signal and the reference signal may not be an integer multiple of the sampling interval, it is necessary to perform interpolation processing on the data points near the peak position to obtain a more accurate fractional delay value. For example, this can be achieved through parabolic interpolation: selecting the absolute amplitude value at the peak index bp and the two cross-correlation function values ​​adjacent to it in the discrete complex cross-correlation function sequence, we obtain... , and By fitting a parabola to these three points, the fractional delay estimate can be obtained by acquiring the delay coordinates corresponding to the vertex of the parabola. Its calculation formula is Meanwhile, the cross-correlation function value read at the peak position is itself a complex number, and the cross-correlation function value can be specifically expressed as: (Among them, the real part) and the virtual part The modulus and direction together determine the argument of the complex number. The relative phase difference between the reference signal and the corresponding directional wave field signal is given directly.

[0041] Directional wave field signal is resampled Convert to new signal A fractional delay filter based on a finite-length unit impulse response (FIR) structure is employed. The coefficients of this fractional delay filter are dynamically designed according to the fractional delay value. Specific applications include sinc interpolation filters designed using the window function method or Farrow structure filters using Lagrange interpolation polynomial coefficients. The time-shift compensation process for directional wave field signals first involves... An intermediate signal is obtained by cyclically shifting the integer sample portion. The intermediate signal is then convolved with a fractional delay filter. The group delay of the fractional delay filter introduces the required fractional sample offset, thereby synthesizing a directional wavefield signal that has undergone a full delay b. This completes the time alignment of the directional wave field signal.

[0042] Then, based on the estimated complex value of the cross-correlation function relative phase difference For time-aligned directional wavefield signals Phase compensation is performed by multiplying the time-aligned directional wavefield signal by the corresponding complex phase rotation factor. The complex phase rotation factor is derived from the estimated relative phase difference. Direct construction, its specific form is as follows ), to obtain the phase-compensated directional wave field signal .

[0043] Ultimately, all directional wavefield signals that have undergone the above two compensation steps are consistent with the selected reference signal in both time and initial phase, thus completing time-frequency synchronization alignment. By coherently integrating the directional wavefield signals that have completed time-frequency synchronization alignment, an enhanced wavefield signal corresponding to a specific angle or focal point can be obtained. Then, by changing the scanning angle and repeating this process, complete multi-angle wavefield data can be obtained.

[0044] ② It should be noted that the transmit-receive pair relationship refers to the fact that in the actual ultrasonic guided wave array detection, each ultrasonic sensor can act as both a transmitting ultrasonic sensor and a receiving ultrasonic sensor in the "one transmit, one receive" mode, thus forming N transmitting array elements × N receiving array elements as independent transmit-receive channel pairs in the array of N array elements.

[0045] Based on the transmit element index and receive element index corresponding to each enhanced wavefield signal source in the multi-angle wavefield data, the data are rearranged and filled into a three-dimensional data structure. The three dimensions of this three-dimensional data structure correspond to the transmit element number, the receive element number, and the time sampling point, respectively. In this way, a full matrix feedback data containing the time series signal and wave propagation path information between all transmit elements and all receive elements is constructed.

[0046] Step S13: Perform dispersion analysis and mode separation processing on the full matrix feedback data to obtain a multimodal feedback feature set.

[0047] In this step, dispersion analysis and mode separation processing are performed on the full matrix feedback data to analyze and decouple the inherent multimode and dispersion characteristics of ultrasonic guided waves propagating in complex structures. This decomposes the mixed and overlapping complex wavefield signals into single-mode waves, overcoming the interference caused by multimode coupling and waveform distortion on defect identification. Based on the inherent dispersion curves of different guided wave modes, it can separate symmetric and antisymmetric modes from the full matrix feedback data, such as Lamb waves, to obtain a multimode feedback feature set. This makes it possible to extract features from the wavefield data that are more sensitive to minor damage and have stronger anti-interference capabilities, laying a key data foundation for achieving high-precision structural health diagnosis.

[0048] Furthermore, step S13 may also include steps S131 to S133: Step S131: Perform a two-dimensional Fourier transform on the full matrix feedback data to obtain the wavenumber-frequency domain energy spectrum distribution. Then, perform parameter matching based on the measured energy spectrum peak values ​​extracted from the wavenumber-frequency domain energy spectrum distribution to obtain the dispersion curves of each feedback signal.

[0049] First, the received signals corresponding to fixed transmitting elements or the average of all transmitting elements are extracted from the full matrix feedback data to form a two-dimensional spatiotemporal signal matrix. Each row of this matrix corresponds to a fixed spatial location of the ultrasonic sensor (i.e., a receiving channel), and each column corresponds to a unified time sampling point. Then, a one-dimensional fast Fourier transform is performed on each row of this matrix, i.e., the discrete time series corresponding to each spatial channel, to independently transform the time-domain signal of each spatial channel to the frequency domain, thus obtaining the complex frequency spectrum corresponding to each spatial channel. After this step, the original two-dimensional spatiotemporal signal matrix is ​​transformed into a new two-dimensional complex matrix. Each row of this new matrix still corresponds to the original spatial channel index, but the meaning of each column has changed from a time sampling point index to a discrete frequency index; that is, each column represents a specific frequency component, and each element value in the matrix represents the corresponding spatial frequency component. The complex spectral values ​​of the channel and frequency components are obtained. Then, for each column of the two-dimensional complex matrix, it is regarded as a discrete sequence with the spatial channel index as the independent variable. A one-dimensional fast Fourier transform is performed on this discrete sequence to transform the spatial channel distribution information of each frequency component from the spatial domain to the wavenumber domain. After this transformation is performed on each column, the original two-dimensional complex matrix is ​​further mapped into a new two-dimensional complex matrix. The row dimension of this new matrix is ​​transformed into discrete wavenumber index, while the column dimension remains discrete frequency index. Each element value in the matrix is ​​the complex spectral value of the corresponding wavenumber component and frequency component. This matrix is ​​the wavenumber-frequency domain complex matrix. After calculating the square of the amplitude of the wavenumber-frequency domain complex matrix to obtain the wavenumber-frequency domain energy spectrum distribution, local maximum detection is used to identify ridges or discrete peak points with significantly higher energy than the background in the wavenumber-frequency domain energy spectrum distribution. The coordinates of these points correspond to the measured wavenumber-frequency data pairs.

[0050] Based on the material configuration (such as plate, tube, etc.) and boundary conditions of the mold blank being tested, the corresponding governing differential equations and characteristic equations are established. The governing differential equations are as follows: The characteristic equations include: in, and Lamé constants are two fundamental constants characterizing the elastic properties of materials; Let be the displacement vector of a medium particle, describing its motion during wave propagation; t represents the density of the material; t represents time. This is half the thickness of the mold blank; The wavenumber is the wave number propagating along the mold blank surface. ; For the transverse wavenumber components associated with the P-wave, ; The transverse wavenumber component associated with the shear wave is defined as follows: ; For the longitudinal wave velocity of the material, ; The transverse wave velocity of the material, ; The angular frequency of the wave. , For frequency.

[0051] The governing differential equation and characteristic equation describe the intrinsic relationship between wavenumber, frequency, and material parameters. The material parameters in this equation are explicitly represented as unknown variables to be solved, thus obtaining a parameterized dispersion relation function, which defines the mathematical form of the theoretical dispersion curve. Subsequently, the material parameters in this parameterized dispersion curve model are adjusted using least-squares fitting. The specific process is as follows: First, the wavenumber and frequency values ​​of all measured peak points in the wavenumber-frequency domain energy spectrum distribution are organized into observed wavenumber vectors and observed frequency vectors, respectively, forming an observation dataset. Based on the dispersion relation function containing the material parameters to be solved, the theoretical wavenumber value corresponding to the dispersion relation function at the observed frequency vector is calculated, forming a predicted wavenumber vector. Then, the difference between each corresponding point in the observed wavenumber vector and the predicted wavenumber vector is calculated to obtain a residual vector. The sum of squared residuals is used as the objective function to measure the degree of agreement between the theoretical curve and the measured data, initializing the guessed material parameter values, and then using Levenbe... Iterative optimization algorithms such as the rg-Marquardt algorithm calculate the gradient of the objective function with respect to inverted material parameters such as elastic modulus and density, and approximate it with the Hessian matrix in each iteration. This determines the parameter update direction and step size, thereby generating a new set of material parameter estimates. After each parameter update, the theoretical dispersion curve and the corresponding objective function value are recalculated. Through continuous iteration, the objective function value is continuously reduced until the change in the objective function value between two adjacent iterations is less than a preset threshold, indicating that the optimization process has converged. When the matching error reaches its minimum, that is, when the objective function no longer decreases significantly, the theoretical dispersion curve corresponding to the final material parameter values ​​output by the optimization algorithm is the inversion result that achieves the best match with the measured energy spectrum peak point, which is the final dispersion curve of each feedback signal.

[0052] Step S132: Based on each dispersion curve, select the filter corresponding to each feedback signal to filter the full matrix feedback data to obtain multiple single-mode time-domain signals.

[0053] Each dispersion curve obtained from the inversion (i.e., the wavenumber-frequency relationship curve of a specific guided wave mode) is transformed into a filter template in the wavenumber-frequency domain. This filter template defines a passband region with a certain bandwidth centered on the theoretical dispersion curve in the wavenumber-frequency plane. For each target mode that needs to be separated, its corresponding filter template is multiplied by a complex matrix in the wavenumber-frequency domain. This operation is equivalent to a two-dimensional bandpass filter operation in the frequency domain, which can retain the energy of the wavenumber-frequency components near the dispersion curve of the target mode, while attenuating the energy of other modal components and noise. Then, a two-dimensional inverse discrete Fourier transform is performed on the wavenumber-frequency domain complex matrix after processing by the specific filter template for this mode. That is, first, a one-dimensional inverse fast Fourier transform is performed on each column of the filtered wavenumber-frequency domain complex matrix to convert it from the wavenumber domain back to the spatial channel domain. Then, a one-dimensional inverse fast Fourier transform is performed on each row of the resulting wavenumber-frequency domain complex matrix to convert it from the frequency domain back to the time domain. The final output is multiple single-mode time-domain signals that are consistent with the spatial channel arrangement of the original full matrix feedback data, but only contain the wave components of the target mode. This provides a physically meaningful signal basis for subsequent defect localization and quantitative assessment based on single-mode wave velocity, attenuation, or waveform characteristics.

[0054] Step S133: Based on the physical relationship between the propagation speed and energy attenuation coefficient of each single-mode time-domain signal and the stress, feature extraction is performed on each single-mode time-domain signal to obtain a multi-mode feedback feature set.

[0055] A single-mode time-domain signal contains one or more modal wave packets on the time axis.

[0056] For each single-mode time-domain signal, based on the spatial geometry of the ultrasonic sensor array and the wave propagation path length, the arrival time difference of the corresponding mode wave packet between the transmitting and receiving ultrasonic sensors is identified by cross-correlation or threshold detection methods. The absolute group velocity or phase velocity of the mode wave packet propagating in the model is calculated using the known propagation path length as the propagation velocity feature. At the same time, by comparing the attenuation of the mode wave packet amplitude on multiple receiving channels along the wave propagation path, the energy attenuation coefficient feature of the corresponding mode wave packet is extracted. The propagation velocity feature and energy attenuation coefficient feature extracted for each single-mode signal are combined to form a feature vector of the corresponding mode.

[0057] Finally, the feature vectors corresponding to all the separated time-domain signals of different modes are aggregated according to mode type to form a multimodal feedback feature set containing velocity and decay information.

[0058] Step S14: Input the multimodal feedback feature set into the deep convolutional inversion network. Process the multimodal feedback feature set through the nonlinear mapping relationship between features and stress field established inside the deep convolutional inversion network, and output the initial stress distribution field.

[0059] By inputting the multimodal feedback feature set into a deep convolutional inversion network, the limitations of traditional inversion methods in handling complex, high-dimensional, and nonlinear mapping relationships are overcome. This enables high-precision and high-efficiency quantitative reconstruction of the internal stress field of the mold blank. By utilizing the feature learning and spatial relationship modeling capabilities of the deep convolutional inversion network, an end-to-end nonlinear mapping function between the multimodal feedback feature set and the spatial continuous stress distribution is automatically learned and established. This allows for the direct inference and output of an initial stress distribution field covering the entire detection area from the multimodal feedback feature set without relying on manually preset simplified physical models or empirical formulas. This provides a comprehensive and visualized quantitative basis for subsequent stress state assessment and process optimization.

[0060] Furthermore, step S14 may also include steps S141 to S143: Step S141: The encoder in the deep convolutional inversion network performs multi-layer convolution and pooling operations on the multimodal feedback feature set to obtain multiple deep abstract feature tensors.

[0061] The encoder reconstructs the input multimodal feedback feature set in both spatial and feature channel dimensions, forming a two-dimensional or three-dimensional input tensor. The spatial dimension corresponds to the detection region grid of the modality, while the feature channel dimension corresponds to different modalities and their physical characteristics, such as the propagation speed and attenuation coefficient of different modalities. Next, after the multimodal feedback feature set is reconstructed into the input tensor, the encoder's convolutional layers use their learnable convolutional kernels to perform sliding computation and feature extraction on the input tensor. Each convolutional computation generates a response value representing a specific pattern (such as a specific combination of features or spatial relationships) in a local region of the input tensor. All these response values, arranged according to their spatial positions, form a two-dimensional or three-dimensional grid-like data matrix, and the output is a feature map. A convolutional layer typically contains multiple different convolutional kernels, thus generating multiple feature maps simultaneously. Each feature map is specifically responsible for extracting and representing a particular type of feature from the input data.

[0062] Next, a nonlinear activation function (such as ReLU) is applied to the generated feature map to introduce nonlinear transformation capability. Then, max pooling is performed on the activated feature map. The feature map is downsampled by taking the maximum value in the local receptive 2x2 window, thereby reducing the spatial size of the feature map to reduce computational complexity and expand the window of subsequent layers. At the same time, the most significant feature response is preserved and preliminary spatial invariance is provided.

[0063] After the stacking process of "convolution-activation-pooling" as described above, the spatial size of the feature map decreases layer by layer while the number of feature channels increases layer by layer. The representation of the features also gradually transforms from multimodal feedback features into deep abstract feature tensors containing complex spatial context and high-level abstract patterns. By repeating such layer stacking multiple times, the encoder finally outputs multiple deep abstract feature tensors.

[0064] Step S142 involves performing cross-modal feature interaction and spatial weight focusing on each deep abstract feature tensor to obtain multiple context-aware features.

[0065] Cross-modal feature interaction is achieved by concatenating and fusing multiple deep abstract feature tensors output by the encoder along the feature channel dimension. The fusion operation uses one-dimensional convolution or fully connected layers to perform global interaction and reweighting on the concatenated feature channels, so that deep abstract feature tensors from different modalities can complement and calibrate each other, thereby generating a fused feature tensor containing multi-modal information.

[0066] Next, spatial weights are focused onto the generated fusion feature tensor. This tensor generates a spatial attention weight map through a parallel attention branch. Each pixel value in the spatial attention weight map reflects the importance of the corresponding spatial location feature to the stress field reconstruction task. This is usually implemented through a lightweight sub-network. This sub-network first performs global average pooling on the fusion feature tensor to capture the global context, and then learns the importance weights of each spatial location through a fully connected layer and a non-linear activation function (such as ReLU, i.e., the linear rectified function). After obtaining the spatial attention weight map, the spatial attention weight map is multiplied element-wise with the fusion feature tensor to suppress the feature response of the background or irrelevant regions, while strengthening the feature representation of key regions. The fusion feature tensor after the above two steps of cross-modal interaction and spatial weight focusing is the context-aware feature.

[0067] The nonlinear activation function designed in this step is a mathematical function applied to the output of neurons in a deep convolutional inversion network. Its core characteristic is that there is no simple linear proportional relationship between its output and input. A simple threshold rule is used to introduce nonlinear transformation capability into the network. For example, for ReLU, it sets all negative input values ​​to zero, while positive input values ​​remain unchanged.

[0068] Step S143: Input multiple context-aware features into the decoder in the deep convolutional inversion network to perform decoding operations and obtain the initial stress distribution field.

[0069] The decoder receives context-aware features as input. In each decoding layer of the decoder, the input context-aware features are first upsampled (usually by transpose convolution or pixel shuffling) to gradually recover the spatial size of the feature map. The context-aware feature representation is then mapped onto a high-resolution spatial grid to obtain the upsampled feature tensor.

[0070] Next, by using skip connections, the deep abstract feature tensors and upsampled feature tensors in the corresponding layers of the encoder path are concatenated and fused in the channel dimension. This compensates for the shallow positional information and texture details retained in the encoding stage in the decoding process. After obtaining the fused feature tensor, convolution and nonlinear activation operations are performed on the fused feature tensor to further optimize and integrate the features, and gradually transform the context-aware features into feature representations directly related to stress physical quantities.

[0071] By repeating the above decoding layer operation of "upsampling-feature fusion-convolution optimization", the spatial size of the feature map is gradually restored to the same resolution as the target stress distribution field, while its feature channel number is gradually reduced. In the last decoding layer, a 1x1 convolution kernel is used to map the multi-channel feature map into a single-channel output, which is a two-dimensional initial stress distribution field in which the value of each pixel represents the stress magnitude at that location.

[0072] Step S15: Iterate and physically constrain the initial stress distribution field based on the material physical information of the mold blank to obtain the target stress distribution field.

[0073] It should be noted that material physical information refers to the set of inherent physical property parameters of the material constituting the mold blank, including elastic modulus, Poisson's ratio, density, and yield strength.

[0074] By identifying the elastic modulus and Poisson's ratio in the material's physical information, and using these parameters based on the generalized Hooke's law, a constitutive equation is constructed to describe the linear relationship between the stress tensor and strain tensor at any point within the mold. In the three-dimensional case, this constitutive equation can be specifically expressed as a matrix form containing the elastic modulus and Poisson's ratio, representing the stress components as a linear combination of the strain components. Simultaneously, based on the principles of continuum mechanics, an equilibrium differential equation that the stress must satisfy is introduced as another fundamental physical constraint. This equation requires that the divergence of all stress components on any tiny volume element within the mold be balanced with the volume forces. The resulting constitutive equation is combined with the equilibrium differential equation, and considering the geometric boundary conditions of the mold and the known external force boundary conditions, together forming a closed system of partial differential equations. This system of equations systematically expresses the complete mechanical relationship that must be followed between the stress field, strain field, and material physical parameters, i.e., the physical constraint.

[0075] Substituting the initial stress distribution field into the aforementioned physical constraints, a forward mechanical simulation is performed using the finite element method. The strain or displacement response of the mold blank under a given stress distribution field is calculated, and the simulation results are compared with the physical constraints. Based on the residuals, the initial stress distribution field is iteratively corrected using the gradient descent method until it satisfies all physical constraints and the residuals converge to a set threshold. The final stress distribution field output is the target stress distribution field, ensuring that the target stress distribution field has strict physical consistency. By introducing material physics information as constraints, the initial stress distribution field obtained by data-driven inversion is corrected to a range that conforms to the actual mechanical principles, thereby significantly improving the reliability of stress prediction results and its engineering application value.

[0076] In one feasible implementation, step S20 may specifically include steps S21 to S23: Step S21: Simulate the processing process based on the target stress distribution field to obtain stress evolution-predicted deformation pairs under different candidate processing paths.

[0077] It should be noted that candidate machining paths refer to multiple alternative CNC tool motion trajectory schemes that are pre-set or generated based on workpiece geometric parameters and process knowledge.

[0078] The target stress distribution field is used as the initial condition and assigned to each node of the finite element mesh model of the blank to be processed to characterize its initial residual stress state. For each candidate machining path, the gradual material removal process is simulated in the finite element mesh model based on the workpiece geometry parameters and corresponding cutting parameters of the candidate machining path. This process requires coupling a constitutive model reflecting the dynamic mechanical behavior of the material, the cutting force and cutting thermal load generated by the tool-workpiece interaction, and considering the fixture constraints in actual machining. While simulating the material being removed layer by layer, the redistribution of the internal stress field of the blank due to load changes and boundary condition changes after each step of material removal needs to be calculated and recorded in real time; this is the stress evolution process.

[0079] After the simulation of the entire candidate machining path is completed, all fixture constraints are removed in the finite element mesh model. The free deformation state of the workpiece under the action of only internal residual stress is simulated, and its size and shape deviations are calculated to obtain the corresponding predicted deformation results.

[0080] Each candidate machining path will go through the complete simulation process described above and output a pair consisting of the stress evolution process and the predicted deformation result, namely the stress evolution-predicted deformation pair. This allows for the prediction of the specific impact of different machining paths on the internal stress state of the mold blank and the final machining deformation in the virtual environment before physical machining begins.

[0081] Step S22: Based on the optimization objective and each stress evolution-predicted deformation pair, iteratively search the corresponding candidate machining paths and the cutting parameters of the candidate machining paths to obtain the target machining path and the target cutting parameters corresponding to the target machining path.

[0082] In this step, based on the optimization objective and each stress evolution-predicted deformation pair, the corresponding candidate machining paths and cutting parameters are iteratively searched to automatically optimize from numerous candidate machining paths, thus solving the decision-making problem under multiple objectives (such as minimum deformation and maximum efficiency). By constructing a feedback optimization closed loop, the stress evolution-predicted deformation pair simulated in step S21 is used as the evaluation basis. Among multiple candidate machining paths and corresponding cutting parameters, the result that best balances the conflicting objectives is searched, realizing a fundamental shift from trial and error based on experience to intelligent decision-making based on data and models. The final output is a comprehensive optimal target machining path and its corresponding target cutting parameters, thus ensuring the theoretical optimality of the process scheme in terms of controlling deformation and improving efficiency before machining is executed.

[0083] Furthermore, step S22 may also include steps S221 to S224: Step S221: After constructing a multi-objective function based on the optimization objective, the multi-objective function is encoded according to each candidate machining path and the cutting parameters corresponding to each candidate machining path to generate a candidate scheme set.

[0084] It should be noted that the optimization objectives include minimizing the maximum principal stress caused by the machining process, minimizing the final deformation of the workpiece, and possibly minimizing the machining time.

[0085] Define a sub-objective function for each optimization objective. For example, the predicted deformation result obtained from the processing simulation can be directly used as the deformation sub-objective function. (in, To predict deformation, the peak principal stress during the stress evolution process is used as the stress sub-objective function. (in, The peak principal stress during stress evolution is represented by the theoretical machining time obtained by dividing the total length of the preset machining path by the programmed feed rate in the cutting parameters, which is used as the efficiency sub-objective function. (in, Indicates the total length of the preset processing path. (representing the programmed feed rate), thereby constructing a multi-objective function that includes the above three sub-objective functions. ,in, represents the solution vector composed of the workpiece geometric parameters and cutting physical parameters encoded by the preset machining path; T represents the transpose, used to indicate that the multi-objective function is a column vector with three sub-objective functions as components; the optimization objective of this multi-objective function is to find the solution that makes each sub-objective function in the multi-objective function achieve the comprehensive optimum.

[0086] Next, each candidate machining path and its corresponding cutting parameters are encoded and transformed into a solution vector that can be directly manipulated by the optimization algorithm. In the algorithm implementation, real number encoding is usually used. The geometric key point coordinate sequence of the preset machining path, the tool motion type identifier, and cutting parameters such as cutting speed and feed rate are concatenated into a one-dimensional vector in a predetermined order. This one-dimensional vector represents a candidate solution. After encoding each candidate machining path and its corresponding cutting parameters in this way, they are integrated into a candidate solution set.

[0087] It should be noted that the programmed feed rate mentioned above refers to the command value set in the CNC machining program for the speed of the tool relative to the workpiece during the cutting process.

[0088] Step S222: Calculate the coded multi-objective function by using each stress evolution-predicted deformation pair to obtain the objective function value of each candidate solution in the candidate scheme set.

[0089] For each candidate solution in the candidate solution set, it needs to be matched with the corresponding stress evolution-predicted deformation pair. After extracting the stress evolution process and predicted deformation result describing the whole process of the preset processing path from the stress evolution-predicted deformation pair corresponding to the candidate solution, the corresponding stress evolution process and predicted deformation result are input into each sub-objective function of the multi-objective function for calculation.

[0090] Specifically, the predicted deformation result is directly assigned to the deformation sub-objective function. Peak stress values ​​are extracted from stress evolution data and assigned to the stress sub-objective function. Simultaneously, the theoretical processing time is calculated based on the path length and programmed feed rate contained in the candidate solution encoding vector, and assigned to the efficiency sub-objective function. .

[0091] Through the above calculations, each candidate solution obtains an objective function value consisting of three specific scalar values. This transforms each candidate solution in the candidate solution set into a set of evaluation indicators that can be directly compared and processed.

[0092] Step S223: Iterate through the candidate scheme set based on the objective function values ​​to obtain the candidate path parameter set.

[0093] The set of candidate solutions and their corresponding objective function values ​​are input into a multi-objective optimization algorithm, such as a fast non-dominated sorting genetic algorithm with an elite retention strategy. This algorithm treats each candidate solution as an individual, forming a population, and its objective function value is used to calculate the fitness of the individual.

[0094] In the iterative operation, all individuals are first sorted into non-dominated categories by comparing the objective function values ​​of any two individuals. If all objective function values ​​of an individual are no worse than those of another individual and at least one objective function value is better, then that individual dominates the other individual. Through this pairwise comparison, all individuals that are not dominated by any other individual are found and assigned to the first non-dominated layer. Then, the first non-dominated layer is temporarily removed from the population, and the above process is repeated among the remaining individuals to determine the second non-dominated layer. This process continues until all individuals are assigned to a specific non-dominated layer.

[0095] After stratification, individuals within the same non-dominated stratum are ranked according to their crowding distance. The crowding distance is calculated for each individual by examining the differences between its two adjacent individuals within the same non-dominated stratum across each objective function dimension, and then summing these differences dimensionally to obtain the distance. This distance characterizes the spatial distribution density of an individual within its non-dominated stratum; a larger distance indicates a sparser spatial distribution density and a greater contribution to diversity. Finally, the selection operation follows a priority order of "non-dominated stratum order first, then crowding distance," meaning individuals with smaller non-dominated stratum numbers are prioritized, and within the same non-dominated stratum, individuals with larger distances are prioritized. This mechanism selects several individuals with better overall performance and distribution as the parent set for subsequent crossover and mutation operations.

[0096] Next, crossover and mutation operations are performed on the selected parent individuals. The crossover operation generates new offspring by exchanging parts of the encoding vectors of the two parent individuals (e.g., using simulated binary crossover). The mutation operation introduces diversity by randomly perturbing certain gene positions in the individual's encoding vector (e.g., using multinomial mutation). After selection, crossover, and mutation, a new set of offspring individuals is generated. The offspring individuals are then merged with the parent individuals to form a new population, and the objective function values ​​of all individuals in the new population are recalculated.

[0097] The above process is repeated until the preset number of iterations is reached. At this point, the algorithm converges, and the resulting population is the candidate path parameter set. By simulating the evolutionary mechanism of "survival of the fittest", a set of solutions that achieve the best balance among multiple conflicting objectives such as deformation, stress, and efficiency are explored, thereby providing a set of preset machining paths and their corresponding cutting parameters for final selection.

[0098] The calculation steps for the objective function values ​​of all individuals in the new population are identical in core principle and execution logic to the calculation process in step S222. Whether it's the initial population or newly generated offspring individuals during the iteration process, each candidate solution corresponds to a one-dimensional vector of preset processing paths and cutting parameters.

[0099] Step S224: Based on the preset processing efficiency threshold, determine the target processing path and the target cutting parameters corresponding to the target processing path from the candidate path parameter set.

[0100] It should be noted that the preset processing efficiency threshold is an upper limit value for the theoretical processing time that is pre-set by the process requirements, and is used for the first round of screening of the candidate path parameter set.

[0101] By traversing each candidate solution in the candidate path parameter set and parsing the corresponding theoretical processing time from each candidate solution, candidate solutions whose theoretical processing time does not exceed the preset processing efficiency threshold are selected to form a feasible solution subset. Then, the candidate solution with the smallest deformable sub-objective function value or the candidate solution with the smallest stress sub-objective function value is directly selected from the feasible solution subset. The selected candidate solution is then decoded, and the geometric key point coordinate sequence and tool motion type identifier extracted are used to form the target processing path. The corresponding cutting parameters are extracted and determined as the target cutting parameters. This enables the determination of a final executable process plan under the premise of ensuring that the processing efficiency meets the hard requirements, providing clear input for generating processing instructions that can directly drive the machine tool.

[0102] Step S23: Identify the target stress concentration area in the target machining path, insert preset machining data into the target stress concentration area, and generate the target stress release path.

[0103] In this step, by identifying the target stress concentration area in the target machining path and inserting preset machining data into the target stress concentration area, a target stress release path is generated. Further, the local high stress problem that may occur during the machining process is refined. By introducing preset machining data into the target stress concentration area, such as specific tool paths, cutting parameter adjustments, or process interventions such as pausing cooling, the residual stress on the target stress concentration area is actively released or redistributed. This generates a target stress release path optimized for stress release, thereby further reducing the risk of stress concentration in critical areas of the workpiece while ensuring machining efficiency and overall deformation control.

[0104] Furthermore, step S23 may also include steps S231 to S233: Step S231: Perform geometric process feature analysis on the target machining path to obtain the path features of the target machining path.

[0105] It should be noted that path features include path curvature, gradient of material removal rate change, and abrupt change points in the cutting direction.

[0106] The path curvature is obtained by numerically differentiating the coordinate sequence of tool movement trajectory points on the target machining path. Specifically, for any trajectory point in the tool movement trajectory point coordinate sequence, the coordinates of the preceding and following points are obtained. The first and second derivatives of the trajectory at that point are calculated using the central difference method. Then, based on the fact that the curvature value is equal to the ratio of the magnitude of the derivative of the tangent vector at that point to the cube of the magnitude of the tangent vector, the calculated first and second derivatives are substituted into the formula for calculation to obtain the curvature of the tool movement trajectory point coordinate sequence. This calculation process is repeated for all trajectory point coordinates on the target machining path to obtain the complete path curvature, which is used to quantify the severity of the path bending.

[0107] The gradient of material removal rate variation is obtained by analyzing the line segments between the coordinates of adjacent tool movement trajectory points on the target machining path. After calculating the path segment length based on the coordinates of adjacent tool movement trajectory points, the cutting depth and cutting width corresponding to the path segment length are combined. The cutting depth and cutting width are multiplied to obtain the cutting cross-sectional area. Then, the cutting cross-sectional area is multiplied by the path segment length to obtain the theoretical material removal volume of the path segment. Next, the material removal volume is divided by the time required to cut the segment (calculated by dividing the path segment length by the programmed feed rate) to obtain the material removal rate of the path segment. Finally, by calculating the difference in material removal rates between adjacent path segments and dividing this difference by the path distance between the center points of the two path segments, the gradient of material removal rate variation along the tool movement trajectory can be obtained.

[0108] The abrupt change point of the cutting force direction is identified by using discrete locations on the target machining path where the theoretical principal direction of the cutting force changes significantly due to the tool entering / exiting the workpiece, drastic changes in the shape of the cutting section, or sudden changes in the tool posture.

[0109] Step S232: Based on path features, after identifying the target stress concentration area from the target processing path, obtain the corresponding preset processing data according to the stress relief strategy that matches the target stress concentration area.

[0110] It should be noted that the preset stress value is a threshold value that is pre-set based on the mechanical properties of the mold blank, historical process data and allowable stress safety range. It is used to determine whether the potential stress risk indicated by the path characteristics exceeds the standard.

[0111] By analyzing three path characteristics—path curvature, material removal rate gradient, and abrupt change in cutting force direction—when the path curvature of a certain path segment is too high, or the material removal rate gradient increases sharply, or the point itself is an abrupt change in cutting force direction, the corresponding area is identified as having a high risk of generating residual stress during processing, and is marked as a target stress concentration area.

[0112] Based on a pre-built strategy mapping library within the system, this library associates different types of path features with corresponding stress mitigation strategies. For example, for a target stress concentration region identified due to abrupt changes in path curvature, preset machining data for "reducing the programmed feed rate and increasing the finishing cycle" is matched; for a target stress concentration region identified due to a sharp increase in the material removal rate gradient, preset machining data for "segmented improvement of cutting parameters" is matched. The system retrieves and calls the corresponding preset machining data from the strategy mapping library based on the dominant path feature type of the currently identified target stress concentration region.

[0113] Step S233: Replace the original toolpath segment in the target machining path with preset machining data to generate the target stress relief path.

[0114] Based on the start and end positions of each target stress concentration area, the corresponding tool motion trajectory point segment is located in the tool motion trajectory point coordinate sequence of the target machining path, and alternative process instructions are parsed from the preset machining data obtained in step S232. The alternative process instructions include a new set of local tool path point coordinates, modified cutting parameters, or inserted specific auxiliary operations. Then, in the data structure of the target machining path, the data on the tool motion trajectory point segment is replaced with the alternative process instructions, and the target machining path is converted into a target stress relief path. This achieves effective intervention in high stress risk areas (i.e. target stress concentration areas) without changing the overall machining scheme framework. Thus, a stress relief mechanism is actively embedded in the final executed target machining path, generating a target stress relief path, which provides a guarantee for directly improving the machining accuracy of the mold blank and improving the dimensional stability of the workpiece from the process level.

[0115] This application also provides a high-precision intelligent machining system for mold blanks, as described above. Figure 2 As shown, this high-precision intelligent machining system for mold blanks includes: The information construction module 10 is used to construct the target stress distribution field inside the mold blank based on the feedback signal returned by the ultrasonic guided wave after applying ultrasonic guided wave to the mold blank. The information generation module 20 is used to plan the processing path of the target stress distribution field with the preset principal stress and target deformation in the preset processing process as the optimization target, and generate stress processing data including target cutting parameters and target stress release path. The machining control module 30 is used to control the machining equipment to perform machining operations on the mold blank based on stress machining data.

[0116] Optionally, the information building module 10 is also used for: Based on the preset guided wave modes and multiple propagation direction parameters, the feedback signal is subjected to adaptive beamforming synthesis processing to obtain multiple directional wave field signals; After performing time-frequency synchronization and alignment processing on each directional wavefield signal to obtain multi-angle wavefield data, the multi-angle wavefield data is restructured according to the transmission and reception relationship of ultrasonic guided waves to obtain full matrix feedback data. Dispersion analysis and mode separation processing are performed on the full matrix feedback data to obtain a multimodal feedback feature set; The multimodal feedback feature set is input into the deep convolutional inversion network. The nonlinear mapping relationship between features and stress field established inside the deep convolutional inversion network is used to process the multimodal feedback feature set and output the initial stress distribution field. The initial stress distribution field is iterated and physically constrained based on the material physical information of the mold blank to obtain the target stress distribution field.

[0117] Optionally, the information building module 10 is also used for: After performing a two-dimensional Fourier transform on the full matrix feedback data to obtain the wavenumber-frequency domain energy spectrum distribution, parameter matching is performed based on the measured energy spectrum peak values ​​extracted from the wavenumber-frequency domain energy spectrum distribution to obtain the dispersion curves of each feedback signal. Based on the dispersion curves, filters corresponding to each feedback signal are selected to filter the full matrix feedback data, resulting in multiple single-mode time-domain signals. Based on the physical relationship between the propagation speed and energy attenuation coefficient of each single-mode time-domain signal and stress, feature extraction is performed on each single-mode time-domain signal to obtain a multi-mode feedback feature set.

[0118] Optionally, the information building module 10 is also used for: By using the encoder within the deep convolutional inversion network, multiple layers of convolution and pooling operations are performed on the multimodal feedback feature set to obtain multiple deep abstract feature tensors; Cross-modal feature interaction and spatial weight focusing are performed on each deep abstract feature tensor to obtain multiple context-aware features; Multiple context-aware features are input into the decoder within the deep convolutional inversion network for decoding to obtain the initial stress distribution field.

[0119] Optionally, the information generation module 20 is also used for: Based on the target stress distribution field, the processing process is simulated to obtain stress evolution-predicted deformation pairs under different candidate processing paths; Based on the optimization objective and the stress evolution-predicted deformation pairs, the corresponding candidate machining paths and their cutting parameters are iteratively searched to obtain the target machining path and its corresponding target cutting parameters. Identify the target stress concentration area in the target machining path, insert preset machining data into the target stress concentration area, and generate the target stress release path.

[0120] Optionally, the information generation module 20 is also used for: After constructing a multi-objective function based on the optimization objective, the multi-objective function is encoded according to each candidate machining path and the cutting parameters corresponding to each candidate machining path to generate a set of candidate solutions; The objective function values ​​of each candidate solution in the candidate scheme set are obtained by calculating the encoded multi-objective function through each stress evolution-predicted deformation pair. The candidate solution set is iterated based on the objective function values ​​to obtain the candidate path parameter set; Based on a preset processing efficiency threshold, the target processing path and the target cutting parameters corresponding to the target processing path are determined from the candidate path parameter set.

[0121] Optionally, the information generation module 20 is also used for: Geometric process feature analysis is performed on the target machining path to obtain the path features of the target machining path; Based on path characteristics, after identifying the target stress concentration area from the target processing path, the corresponding preset processing data is obtained according to the stress relief strategy that matches the target stress concentration area. The original toolpath segment in the target machining path is replaced with preset machining data to generate the target stress relief path.

[0122] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A high-precision intelligent machining method for mold blanks, characterized in that, include: By applying ultrasonic guided waves to the mold blank, a target stress distribution field inside the mold blank is constructed based on the feedback signal returned by the ultrasonic guided waves. Using the preset principal stress and target deformation in the preset processing process as optimization targets, the processing path is planned for the target stress distribution field to generate stress processing data including target cutting parameters and target stress release path; The control processing equipment performs processing operations on the mold blank based on the stress processing data.

2. The high-precision mold blank intelligent machining method according to claim 1, characterized in that, The step of constructing the target stress distribution field inside the mold based on the feedback signal returned by the ultrasonic guided wave includes: Based on the preset guided wave modes and multiple propagation direction parameters, the feedback signal is subjected to adaptive beamforming synthesis processing to obtain multiple directional wave field signals; After performing time-frequency synchronization alignment processing on each of the directional wavefield signals to obtain multi-angle wavefield data, the multi-angle wavefield data is structurally reorganized according to the transmit / receive pair relationship of the ultrasonic guided wave to obtain full matrix feedback data. The full matrix feedback data is subjected to dispersion analysis and mode separation processing to obtain a multimodal feedback feature set; The multimodal feedback feature set is input into a deep convolutional inversion network. The multimodal feedback feature set is processed through the nonlinear mapping relationship between features and stress field established inside the deep convolutional inversion network to output the initial stress distribution field. The initial stress distribution field is iterated and physically constrained based on the material physical information of the mold blank to obtain the target stress distribution field.

3. The high-precision mold blank intelligent machining method according to claim 2, characterized in that, The step of performing dispersion analysis and mode separation processing on the full matrix feedback data to obtain a multimodal feedback feature set includes: After performing a two-dimensional Fourier transform on the full matrix feedback data to obtain the wavenumber-frequency domain energy spectrum distribution, parameter matching is performed based on the measured energy spectrum peak values ​​extracted from the wavenumber-frequency domain energy spectrum distribution to obtain the dispersion curves of each feedback signal. Based on the dispersion curves, filters corresponding to the feedback signals are selected to filter the full matrix feedback data, resulting in multiple single-mode time-domain signals. Based on the physical relationship between the propagation speed and energy attenuation coefficient of each single-mode time-domain signal and stress, feature extraction is performed on each single-mode time-domain signal to obtain the multimodal feedback feature set.

4. The high-precision mold blank intelligent machining method according to claim 2, characterized in that, The step of processing the multimodal feedback feature set and outputting the initial stress distribution field by using the nonlinear mapping relationship between features and stress field established within the deep convolutional inversion network includes: The encoder within the deep convolutional inversion network performs multi-layer convolution and pooling operations on the multimodal feedback feature set to obtain multiple deep abstract feature tensors. Cross-modal feature interaction and spatial weight focusing are performed on each of the deep abstract feature tensors to obtain multiple context-aware features; The multiple context-aware features are input into the decoder within the deep convolutional inversion network for decoding to obtain the initial stress distribution field.

5. The high-precision intelligent machining method for mold blanks according to claim 1, characterized in that, The step of using the preset principal stress and target deformation in the preset processing process as optimization targets, planning the processing path for the target stress distribution field, and generating stress processing data including target cutting parameters and target stress release paths includes: Based on the target stress distribution field, the processing process is simulated to obtain stress evolution-predicted deformation pairs under different candidate processing paths; Based on the optimization objective and each stress evolution-predicted deformation pair, the corresponding candidate machining paths and the cutting parameters of the candidate machining paths are iteratively searched to obtain the target machining path and the target cutting parameters corresponding to the target machining path; Identify the target stress concentration area in the target processing path, insert preset processing data into the target stress concentration area, and generate the target stress release path.

6. The high-precision intelligent machining method for mold blanks according to claim 5, characterized in that, The step of iteratively searching for the corresponding candidate machining paths and their cutting parameters based on the optimization objective and each stress evolution-predicted deformation pair to obtain the target machining path and its corresponding target cutting parameters includes: After constructing a multi-objective function based on the optimization objective, the multi-objective function is encoded according to each candidate machining path and the cutting parameters corresponding to each candidate machining path to generate a candidate scheme set; The objective function values ​​of each candidate solution in the candidate scheme set are obtained by calculating the encoded multi-objective function through each stress evolution-predicted deformation pair. Based on the objective function values, the candidate scheme set is iterated to obtain the candidate path parameter set; Based on a preset processing efficiency threshold, the target processing path and the target cutting parameters corresponding to the target processing path are determined from the candidate path parameter set.

7. The high-precision intelligent machining method for mold blanks according to claim 5, characterized in that, The step of identifying the target stress concentration region in the target processing path, inserting preset processing data into the target stress concentration region, and generating the target stress release path includes: Geometric process feature analysis is performed on the target processing path to obtain the path features of the target processing path; Based on the path characteristics, after identifying the target stress concentration area from the target processing path, the corresponding preset processing data is obtained according to the stress relief strategy that matches the target stress concentration area. The original toolpath segment in the target machining path is replaced with the preset machining data to generate the target stress relief path.

8. A high-precision intelligent machining system for mold blanks, characterized in that, A high-precision intelligent machining method for mold blanks, as described in any one of claims 1 to 7, comprises: An information construction module is used to construct a target stress distribution field inside the mold blank based on the feedback signal returned by the ultrasonic guided wave after applying an ultrasonic guided wave to the mold blank. The information generation module is used to plan the processing path for the target stress distribution field with the preset principal stress and target deformation in the preset processing process as optimization targets, and generate stress processing data including target cutting parameters and target stress release path; The machining control module is used to control the machining equipment to perform machining operations on the mold blank according to the stress machining data.

Images