Roller sawing machine dual-shaft decoupling linkage cutting method based on digital twinning

Abstract

The application provides a roll sawing machine double-shaft decoupling linkage cutting method based on digital twinning, relates to the numerical control technical field, and comprises the following steps: constructing a roll digital twinning model through three-dimensional laser scanning and ultrasonic detection, generating a cutting force prediction curve and a vibration risk diagram, and planning a double-shaft decoupling linkage track according to the cutting force prediction curve and the vibration risk diagram. In the cutting process, sensor signals are collected in real time to estimate the cutting state, a feedback control and a model prediction correction double-layer architecture is adopted to dynamically adjust double-shaft motion parameters, and a cutting force prediction model is optimized according to task performance, so that the cutting precision and efficiency are improved.

Description

Technical Field

[0001] This invention relates to numerical control technology, and more particularly to a dual-axis decoupled linkage cutting method for roll sawing machines based on digital twins. Background Technology

[0002] Existing roll sawing technology relies on fixed cutting parameters, which cannot adapt to complex working conditions such as uneven hardness and geometric eccentricity of roll materials. During the cutting process, sudden changes in material hardness often cause drastic fluctuations in cutting force, leading to increased saw blade vibration, decreased cutting accuracy, and equipment damage. These problems are particularly pronounced when cutting large rolls, due to their large mass and high inertia, and can even lead to saw blade breakage and processing accidents.

[0003] Existing control methods mainly employ single-axis feed control, requiring operators to adjust the feed speed based on experience, which cannot accurately respond to dynamic changes during the cutting process. Although some equipment has introduced load adaptive control, it lacks comprehensive consideration of material properties and geometric characteristics, thus failing to achieve precise predictive control. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a digital twin-based dual-axis decoupled linkage cutting method for roll sawing machines, which can solve the problems in existing technologies.

[0005] Acquire three-dimensional laser scanning data and ultrasonic material testing data of the roll to be cut, construct a digital twin model of the roll including material hardness distribution and geometric eccentricity, and generate a full-cycle cutting force prediction curve and vibration risk distribution map; Based on the full-cycle cutting force prediction curve and vibration risk distribution map, a dual-axis decoupled linkage trajectory command is generated. The transverse axis performs a tool deflection action in the low-frequency band, and the tool deflection displacement is determined by the predicted cutting force amplitude. The longitudinal axis generates micro-oscillations in the high-frequency band, and the oscillation frequency is adaptively adjusted according to the vibration risk index. During the cutting process, signals from multiple sensor sources are collected in real time to estimate cutting state parameters, which include the actual cutting force and the actual material hardness. The feedback control layer generates biaxial compensation commands based on the deviation between the actual cutting force and the predicted cutting force. When the actual vibration amplitude exceeds the target value, the longitudinal axis oscillation parameters are adjusted. The model prediction correction layer updates the actual material hardness in reverse to the uncut area in the digital twin model, dynamically correcting the biaxial feedforward commands for the uncut area. Based on the performance indicators after the cutting task is completed, optimize and store the cutting force prediction parameters.

[0006] In one optional implementation, the steps of acquiring three-dimensional laser scanning data and ultrasonic material testing data of the roll to be cut, constructing a digital twin model of the roll including material hardness distribution and geometric eccentricity, and generating a full-cycle cutting force prediction curve and vibration risk distribution map include: Collect roll point cloud data, extract the center coordinates of each measurement section, calculate the geometric eccentricity based on the position offset of the center coordinates, and calculate the cylindricity error based on the radial fluctuation of the radius of each section. By scanning the roll circumference at multiple angles using an ultrasonic probe, and based on the sound velocity attenuation characteristics and echo amplitude distribution of the ultrasonic echo signal, a three-dimensional distribution of material hardness in the roll circumference direction is established. The geometric eccentricity, the cylindricity error, and the three-dimensional distribution of material hardness are spatially registered to construct a digital twin model of the roll that integrates material properties and geometric properties. Based on the aforementioned digital twin model of the roll, the instantaneous cutting force during the cutting process is calculated according to the cutting process parameters and the roll spatial characteristic parameters. The cutting process parameters include the saw blade cutting depth and the roll rotation angle, and the roll spatial characteristic parameters include the material hardness and geometric eccentricity corresponding to the angle. A full-cycle cutting force prediction curve is generated based on the temporal variation of the instantaneous cutting force within the complete rotation cycle of the roll. The vibration excitation intensity is calculated based on the rate of change of the cutting force amplitude in the full-cycle cutting force prediction curve and the spatial distribution of the hardness gradient in the three-dimensional distribution of material hardness, and a vibration risk distribution map is generated.

[0007] In one optional implementation, a dual-axis decoupled linkage trajectory command is generated based on the full-cycle cutting force prediction curve and vibration risk distribution map. The transverse axis performs a tool deflection action in the low-frequency band, with the deflection displacement determined by the predicted cutting force amplitude. The longitudinal axis generates micro-oscillations in the high-frequency band, with the oscillation frequency adaptively adjusted according to the vibration risk index. This step includes: The full-cycle cutting force prediction curve is decomposed in the frequency domain to extract the low-frequency component and the high-frequency component of the cutting force. The low-frequency component of the cutting force represents the macroscopic change trend of the cutting force within the roll rotation cycle, and the high-frequency component of the cutting force represents the transient fluctuation of the cutting force caused by abrupt changes in material hardness and geometric eccentricity. A transverse shaft deflection curve is generated based on the amplitude of the low-frequency component of the cutting force. The transverse shaft deflection curve describes the displacement change law of the transverse shaft during the complete rotation cycle of the roll. Based on the vibration risk index at each circumferential angle position in the vibration risk distribution diagram, the target oscillation frequency of the longitudinal axis at the corresponding angle position is calculated. When the vibration risk index is higher than the vibration risk threshold, the target oscillation frequency is increased; when the vibration risk index is lower than the vibration risk threshold, the target oscillation frequency is decreased. The oscillation amplitude of the longitudinal axis is determined based on the frequency characteristics of the high-frequency component of the cutting force, and the oscillation amplitude is related to the energy distribution of the high-frequency component of the cutting force. The transverse axis tool deflection displacement curve and the longitudinal axis oscillation parameters are combined to calculate the composite trajectory. The longitudinal axis oscillation parameters include the target oscillation frequency and oscillation amplitude to obtain the composite displacement trajectory of the two axes at each moment. The synthesized displacement trajectory is smoothed to generate a dual-axis decoupled linkage trajectory command.

[0008] In one optional implementation, the step of calculating the composite trajectory of the transverse axis tool deflection displacement curve and the longitudinal axis oscillation parameters to obtain the composite displacement trajectory of the two axes at each moment includes: A dual-axis spatial coordinate system is established, and the transverse axis tool displacement curve is mapped to the transverse axis displacement sequence in the time domain. The target oscillation frequency and the oscillation amplitude are converted into the longitudinal axis oscillation displacement sequence in the time domain. Based on the correspondence between the roll rotation angle and time, the transverse axis displacement value and the longitudinal axis oscillation displacement value at each moment are determined. Using the transverse axis displacement value as the transverse component and the longitudinal axis oscillation displacement value as the longitudinal component, the vector sum of the two-axis displacements at each moment is calculated to obtain a composite displacement vector sequence. The composite displacement vector sequence is then subjected to kinematic constraint checks, including a maximum velocity limit and a maximum acceleration limit for both axes. When there is a displacement point in the composite displacement vector sequence that exceeds the kinematic constraints, the transverse and longitudinal components of that displacement point are scaled proportionally. The composite displacement vector sequence that satisfies the kinematic constraints is then converted into a composite displacement trajectory of both axes at each moment.

[0009] In one optional implementation, during the cutting process, the step of acquiring multi-source sensor signals in real time and estimating cutting state parameters, including the actual cutting force and the actual material hardness, includes: During the cutting process, cutting force signals are collected by force sensors, acoustic emission signals from the cutting area are collected by acoustic emission sensors, and vibration signals are collected by vibration sensors to obtain multi-source sensor signals. Feature extraction is performed on the cutting force signal, the acoustic emission signal, and the vibration signal to obtain cutting force features, acoustic emission features, and vibration features. The actual cutting force and the actual material hardness are estimated by multi-source information fusion.

[0010] In one optional implementation, a feedback control layer generates biaxial compensation commands based on the deviation between the actual cutting force and the predicted cutting force, and adjusts the longitudinal axis oscillation parameters when the actual vibration amplitude exceeds the target value; the model prediction correction layer updates the actual material hardness in reverse to the uncut area in the digital twin model, and the step of dynamically correcting the biaxial feedforward commands for the uncut area includes: In the feedback control layer, the deviation between the actual cutting force and the predicted cutting force is calculated and converted into transverse axis compensation commands and longitudinal axis compensation commands. The transverse axis compensation commands are used to correct the feed rate of the transverse axis, and the longitudinal axis compensation commands are used to correct the oscillating displacement of the longitudinal axis. When the actual vibration amplitude collected by the vibration sensor exceeds the vibration target value, the longitudinal axis oscillation amplitude is reduced. The longitudinal axis oscillation parameters include oscillation frequency and oscillation amplitude. In the model prediction correction layer, the actual material hardness is mapped to the spatial coordinate system of the digital twin model to determine the location of the cut area corresponding to the actual material hardness; Based on the correspondence between the location of the cut area and the actual material hardness, the material hardness distribution of the uncut area is calculated along the cutting path direction, and the material hardness distribution of the uncut area is corrected by the actual material hardness of the cut area. Based on the corrected material hardness distribution of the uncut region, the dual-axis feedforward command for the uncut region is recalculated. The dual-axis feedforward command includes the transverse axis feed rate and longitudinal axis oscillation parameters of the uncut region.

[0011] In one optional implementation, the step of optimizing and storing the cutting force prediction parameters based on the performance indicators after the cutting task is executed includes: Obtain performance metrics after the cutting task is executed, including cutting accuracy and cutting efficiency; Extract the actual cutting force during the cutting task execution process and the predicted cutting force at the corresponding position in the full-cycle cutting force prediction curve, and calculate the prediction error between the actual cutting force and the predicted cutting force; A parameter optimization objective function is constructed, which takes the performance index and the prediction error as inputs and outputs the adjustment amount of the cutting force prediction parameters. The cutting force prediction parameters are updated according to the adjustment amount, and the updated cutting force prediction parameters are associated with the working condition information of the cutting task and stored in the parameter library. The working condition information includes material type and geometric eccentricity characteristics.

[0012] This invention achieves accurate prediction of the cutting process by constructing a digital twin model of the roll that integrates material hardness distribution and geometric eccentricity, and generates dual-axis decoupled linkage trajectory commands based on the prediction results. The transverse axis performs low-frequency tool deflection to cope with changes in macroscopic cutting force, while the longitudinal axis generates high-frequency micro-oscillations to suppress cutting chatter, significantly improving the stability of the roll cutting process. Compared with traditional single-axis control, this method reduces the amplitude of cutting vibration and improves cutting accuracy.

[0013] This invention also achieves accurate estimation of cutting conditions and dynamic optimization of control strategies through multi-source sensor information fusion and a two-layer control architecture. The feedback control layer adjusts the dual-axis control parameters in real time, while the model prediction correction layer updates the digital twin model with the actual material hardness, dynamically correcting the feedforward commands for the uncut area. It can optimize and store the cutting force prediction parameters based on the cutting task execution results, forming a continuously optimized process knowledge base, effectively improving the intelligence level and processing efficiency of roll cutting. Attached Figure Description

[0014] Figure 1 This is a schematic flowchart of a dual-axis decoupled linkage cutting method for a roll saw based on digital twins, according to an embodiment of the present invention. Detailed Implementation

[0015] The technical solutions of the present invention will be described below with reference to the accompanying drawings. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0016] Figure 1 This is a schematic flowchart of the dual-axis decoupled linkage cutting method for roll sawing machines based on digital twins according to the present invention. Figure 1 As shown, the method includes: Acquire three-dimensional laser scanning data and ultrasonic material testing data of the roll to be cut, construct a digital twin model of the roll including material hardness distribution and geometric eccentricity, and generate a full-cycle cutting force prediction curve and vibration risk distribution map; Based on the full-cycle cutting force prediction curve and vibration risk distribution map, a dual-axis decoupled linkage trajectory command is generated. The transverse axis performs a tool deflection action in the low-frequency band, and the tool deflection displacement is determined by the predicted cutting force amplitude. The longitudinal axis generates micro-oscillations in the high-frequency band, and the oscillation frequency is adaptively adjusted according to the vibration risk index. During the cutting process, signals from multiple sensor sources are collected in real time to estimate cutting state parameters, which include the actual cutting force and the actual material hardness. The feedback control layer generates biaxial compensation commands based on the deviation between the actual cutting force and the predicted cutting force. When the actual vibration amplitude exceeds the target value, the longitudinal axis oscillation parameters are adjusted. The model prediction correction layer updates the actual material hardness in reverse to the uncut area in the digital twin model, dynamically correcting the biaxial feedforward commands for the uncut area. Based on the performance indicators after the cutting task is completed, optimize and store the cutting force prediction parameters.

[0017] In one optional implementation, the steps of acquiring three-dimensional laser scanning data and ultrasonic material testing data of the roll to be cut, constructing a digital twin model of the roll including material hardness distribution and geometric eccentricity, and generating a full-cycle cutting force prediction curve and vibration risk distribution map include: Collect roll point cloud data, extract the center coordinates of each measurement section, calculate the geometric eccentricity based on the position offset of the center coordinates, and calculate the cylindricity error based on the radial fluctuation of the radius of each section. By scanning the roll circumference at multiple angles using an ultrasonic probe, and based on the sound velocity attenuation characteristics and echo amplitude distribution of the ultrasonic echo signal, a three-dimensional distribution of material hardness in the roll circumference direction is established. The geometric eccentricity, the cylindricity error, and the three-dimensional distribution of material hardness are spatially registered to construct a digital twin model of the roll that integrates material properties and geometric properties. Based on the aforementioned digital twin model of the roll, the instantaneous cutting force during the cutting process is calculated according to the cutting process parameters and the roll spatial characteristic parameters. The cutting process parameters include the saw blade cutting depth and the roll rotation angle, and the roll spatial characteristic parameters include the material hardness and geometric eccentricity corresponding to the angle. A full-cycle cutting force prediction curve is generated based on the temporal variation of the instantaneous cutting force within the complete rotation cycle of the roll. The vibration excitation intensity is calculated based on the rate of change of the cutting force amplitude in the full-cycle cutting force prediction curve and the spatial distribution of the hardness gradient in the three-dimensional distribution of material hardness, and a vibration risk distribution map is generated.

[0018] For example, a laser scanning device is used to acquire three-dimensional point cloud data around the axial direction of the roll. The scanning resolution is set to at least 10 sampling points per millimeter, the scanning angle covers the entire 360-degree circumference, and the axial sampling interval is adjustable from 5 mm to 20 mm. The point cloud data is expressed in a three-dimensional coordinate system, with each data point containing X-axis, Y-axis, and Z-axis coordinates, achieving a coordinate accuracy of 0.01 mm. After acquisition, the point cloud data is grouped according to axial position to form several measurement sections, each corresponding to a circumferential point cloud set at a certain axial position of the roll.

[0019] For each measurement section, the least squares fitting method is used to extract the center coordinates. Specifically, the coordinates of all point clouds within the section are substituted into the circle fitting algorithm, and iterative calculations are performed to minimize the sum of the squares of the distances from each point to the fitted circle center, with a convergence accuracy set to 0.001 mm. The fitted circle center coordinates are recorded as the geometric center position of that section. The line connecting the center coordinates of the two ends of the roll section is selected as the theoretical axis reference. The perpendicular distance from the center of each measurement section to this reference axis is calculated; this distance represents the geometric eccentricity at the corresponding position. When the eccentricity exceeds 0.5 mm, it is marked as a high eccentricity region. Simultaneously, the distance from all points within each section to the fitted circle center is calculated, and the difference between the maximum and minimum distances is extracted as the radial fluctuation of the section's radius. The total cylindricity error is obtained by accumulating the radial fluctuations of all sections along the axial direction.

[0020] Ultrasonic material testing employs a contact probe with a frequency ranging from 5 MHz to 10 MHz. The probe moves at equal angular intervals along the circumference of the roll, with intervals ranging from 10 to 30 degrees. Ultrasonic pulses are emitted and echo signals are received at each angular position. The echo signals are acquired via an analog-to-digital converter with a sampling rate of 100 MHz, and the acquisition time window covers the complete round-trip time of the ultrasonic wave within the roll material. The sound velocity attenuation characteristics are extracted based on the echo signal's flight time and amplitude. The sound velocity attenuation rate corresponds to the material's internal density and hardness; a higher attenuation rate indicates greater material hardness. The echo amplitude distribution reflects the reflection intensity at the material interface; higher interface clarity indicates better hardness uniformity. Spatially interpolating the sound velocity attenuation rate and echo amplitude at each angular position, a three-dimensional mesh is constructed along the roll's axial and circumferential directions. Each mesh node stores the estimated material hardness value at its corresponding location. The hardness unit is Rockwell hardness (HRC), ranging from 20 HRC to 65 HRC.

[0021] During spatial registration, geometric eccentricity data and the three-dimensional distribution of material hardness are uniformly mapped to the roll cylindrical coordinate system. The origin of the coordinate system is set at the axial center of the roll, the radial coordinate corresponds to the roll radius, the axial coordinate corresponds to the roll length direction, and the angular coordinate corresponds to the circumferential position. Geometric eccentricity data is labeled on the corresponding axial coordinate according to the measurement section position, and cylindricity error is superimposed on the radial coordinate offset. The grid nodes of the three-dimensional distribution of material hardness are aligned with the grid of the cylindrical coordinate system, and the interpolation error is controlled within 5%. The registered digital twin model of the roll contains three attribute fields for each spatial location: geometric eccentricity, cylindricity error, and material hardness. The data storage adopts a hierarchical index structure with an axial resolution of 10 mm and an angular resolution of 5 degrees, supporting fast querying of attribute values ​​at any spatial location.

[0022] When calculating the instantaneous cutting force during the cutting process based on the digital twin model of the roll, the input parameters include the saw blade penetration depth and the current rotation angle of the roll. The saw blade penetration depth is defined as the radial distance from the saw blade edge into the roll surface, ranging from 0.5 mm to 5 mm. The roll rotation angle is based on the initial position as zero degrees, with one rotation corresponding to 360 degrees. The material hardness and geometric eccentricity at the corresponding angular position are retrieved from the digital twin model based on the current rotation angle, using a nearest neighbor interpolation strategy with a query latency of less than 1 millisecond. The instantaneous cutting force consists of three components: tangential force, radial force, and axial force. Each component is correlated with the penetration depth, material hardness, and geometric eccentricity through an empirical model. The tangential force calculation uses a reference tangential force as the starting point. The reference tangential force corresponds to a standard material hardness of 40 HRC and a unit penetration depth of 1 mm, with a typical value of 280 Newtons. The actual tangential force is linearly scaled by the penetration depth based on the reference tangential force and multiplied by the material hardness coefficient. The material hardness coefficient is based on 40 HRC as a baseline value of 1.0. For every 10 HRC increase in hardness, the coefficient increases by 0.15 to 0.25. For example, when the material hardness is 50 HRC, the coefficient ranges from 1.15 to 1.25. When the penetration depth is 3 mm, the tangential force is calculated as 280 N × 3 × hardness coefficient, resulting in a range of 966 N to 1050 N. The radial force consists of two superimposed parts. The first part is the steady-state radial force, which is 30% to 40% of the tangential force. The second part is the eccentric disturbance force, the amplitude of which is proportional to the geometric eccentricity. For every 0.1 mm increase in eccentricity, the disturbance force increases by 8% to 12% of the steady-state radial force. When the eccentricity is 0.5 mm, the disturbance force reaches 40% to 60% of the steady-state radial force. Axial force is mainly caused by the saw blade installation tilt angle and the unevenness of the roll end face. For every 0.1 degree increase in tilt angle, the axial force increases by 2% to 3% of the tangential force. For every 0.05 mm increase in end face unevenness, the axial force increases by 1% to 2% of the tangential force. The total amplitude of axial force is generally 10% to 20% of the tangential force. In the aforementioned case of 1000 Newtons of tangential force, the axial force ranges from 100 Newtons to 200 Newtons.

[0023] During a complete rotation cycle, the instantaneous cutting force is calculated for every 1 degree of rotation of the roll, generating 360 cutting force sampling points. These 360 ​​sampling points are arranged in chronological order, with the horizontal axis representing the roll rotation angle and the vertical axis representing the instantaneous cutting force amplitude, forming a full-cycle cutting force prediction curve. The curve data is stored using floating-point numbers with an accuracy maintained to 0.01 Newtons. The rate of change of the cutting force amplitude in the curve is calculated by the difference between adjacent sampling points, with the difference interval adjustable from 1 to 5 degrees, and the unit of the rate of change is Newtons per degree. The hardness gradient in the three-dimensional distribution of material hardness is calculated using the spatial derivative, with the derivative direction along the circumferential tangent, and the gradient unit is HRC per millimeter. The vibration excitation intensity is defined as a weighted combination of the rate of change of cutting force and the hardness gradient. The weighting coefficients are calibrated according to the saw blade stiffness and damping characteristics, with the weight of the rate of change of cutting force ranging from 0.6 to 0.8 and the weight of the hardness gradient ranging from 0.2 to 0.4. When the vibration excitation intensity exceeds a threshold, the location is marked as a high-risk area. The typical threshold value is 50 to 80 Newtons per degree. The vibration risk distribution map uses the roll circumference angle as the horizontal axis and the vibration excitation intensity as the vertical axis, and uses color coding to distinguish the risk level: green indicates low risk, yellow indicates medium risk, and red indicates high risk.

[0024] This method constructs a digital twin model of the roll using three-dimensional laser scanning and ultrasonic detection, accurately capturing the material hardness distribution and geometric eccentricity characteristics, and providing high-precision spatial data support for cutting force prediction.

[0025] In one optional implementation, a dual-axis decoupled linkage trajectory command is generated based on the full-cycle cutting force prediction curve and vibration risk distribution map. The transverse axis performs a tool deflection action in the low-frequency band, with the deflection displacement determined by the predicted cutting force amplitude. The longitudinal axis generates micro-oscillations in the high-frequency band, with the oscillation frequency adaptively adjusted according to the vibration risk index. This step includes: The full-cycle cutting force prediction curve is decomposed in the frequency domain to extract the low-frequency component and the high-frequency component of the cutting force. The low-frequency component of the cutting force represents the macroscopic change trend of the cutting force within the roll rotation cycle, and the high-frequency component of the cutting force represents the transient fluctuation of the cutting force caused by abrupt changes in material hardness and geometric eccentricity. A transverse shaft deflection curve is generated based on the amplitude of the low-frequency component of the cutting force. The transverse shaft deflection curve describes the displacement change law of the transverse shaft during the complete rotation cycle of the roll. Based on the vibration risk index at each circumferential angle position in the vibration risk distribution diagram, the target oscillation frequency of the longitudinal axis at the corresponding angle position is calculated. When the vibration risk index is higher than the vibration risk threshold, the target oscillation frequency is increased; when the vibration risk index is lower than the vibration risk threshold, the target oscillation frequency is decreased. The oscillation amplitude of the longitudinal axis is determined based on the frequency characteristics of the high-frequency component of the cutting force, and the oscillation amplitude is related to the energy distribution of the high-frequency component of the cutting force. The transverse axis tool deflection displacement curve and the longitudinal axis oscillation parameters are combined to calculate the composite trajectory. The longitudinal axis oscillation parameters include the target oscillation frequency and oscillation amplitude to obtain the composite displacement trajectory of the two axes at each moment. The synthesized displacement trajectory is smoothed to generate a dual-axis decoupled linkage trajectory command, which includes a lateral axis displacement command sequence and a longitudinal axis displacement command sequence.

[0026] For example, the full-cycle cutting force prediction curve is processed by a frequency domain decomposition module, which uses a fast Fourier transform algorithm to convert the time-domain cutting force signal to the frequency domain. The input data consists of cutting force sampling points corresponding to 360 angular positions within the complete rotation cycle of the roll, with a sampling interval of 1 degree, and the data format is a floating-point array. The frequency domain decomposition sets a cutoff frequency as the boundary between low and high frequencies. A typical cutoff frequency is 3 to 5 times the fundamental frequency of the roll rotation. The fundamental frequency of the roll rotation is determined by the rotation speed; at a rotation speed of 10 revolutions per minute, the fundamental frequency is 0.167 Hz, and the cutoff frequency is correspondingly set to 0.5 Hz to 0.835 Hz. Frequency components below the cutoff frequency are classified as low-frequency components of the cutting force, and frequency components above the cutoff frequency are classified as high-frequency components. The low-frequency components of the cutting force are extracted using a low-pass filter, with the filter order set to 4th to 8th order and the transition band width being 10% to 20% of the cutoff frequency. The filtered low-frequency component curve retains the macroscopic variation trend of the cutting force within the roll rotation cycle, with a smooth waveform and obvious periodicity. The high-frequency component of the cutting force is extracted by a high-pass filter with the filter parameters configured symmetrically with the low-pass filter. The extracted high-frequency component curve exhibits rapid fluctuation characteristics, reflecting the transient impact of the cutting force caused by the sudden change in material hardness and the geometric eccentricity.

[0027] The transverse axis tool retraction curve generation module receives the low-frequency component of the cutting force as input and establishes a mapping relationship between the cutting force amplitude and the tool retraction amount. This mapping relationship uses a linear proportional model, with the proportionality coefficient called the tool retraction compliance coefficient, measured in millimeters per kilonewton (m / kN), typically ranging from 0.05 mm / kN to 0.15 mm / kN. The tool retraction compliance coefficient is calibrated based on a combination of saw blade stiffness and bed structure stiffness; higher stiffness results in a smaller coefficient. Multiplying the cutting force amplitude at each angular position on the low-frequency component curve of the cutting force by the tool retraction compliance coefficient yields the transverse axis tool retraction amount at that corresponding angular position. For example, if the low-frequency component of the cutting force at a certain angular position is 950 N, and the tool retraction compliance coefficient is 0.1 mm / kN, then the transverse axis tool retraction amount at that position is 0.095 mm. The deflection displacements at 360 angular positions are arranged in angular order to form a transverse axis deflection displacement curve. The horizontal axis of the curve represents the roll rotation angle, and the vertical axis represents the transverse axis displacement. The curve describes the displacement variation of the transverse axis within a complete roll rotation cycle. The displacement curve data storage accuracy is maintained to 0.001 mm, and the update cycle is synchronized with the roll rotation angle sampling cycle, with a 1-degree angular interval.

[0028] The longitudinal axis oscillation parameter calculation module reads the vibration risk index at each circumferential angle position from the vibration risk distribution map. The vibration risk index is a dimensionless value, ranging from 0 to 100, with higher values ​​indicating higher vibration risk. The module sets a vibration risk threshold, typically between 50 and 70, which can be adjusted according to machining accuracy requirements and equipment vibration characteristics. When the vibration risk index at a certain angle position is higher than the vibration risk threshold, the target oscillation frequency of the longitudinal axis at that position is set to a high-frequency level, ranging from 80 Hz to 120 Hz. When the vibration risk index is lower than the vibration risk threshold, the target oscillation frequency is set to a low-frequency level, ranging from 30 Hz to 60 Hz. The oscillation frequency adjustment adopts a piecewise constant value strategy, keeping the oscillation frequency constant within adjacent 5-degree to 10-degree angle intervals to avoid frequency abrupt changes causing oscillations in the control system. The oscillation amplitude is determined based on the energy distribution of the high-frequency component of the cutting force. The energy distribution is calculated by square integration of the high-frequency component curve, with an adjustable integration window width of 10 to 30 degrees. A larger energy distribution value indicates stronger high-frequency disturbances in that region, and the longitudinal axis oscillation amplitude is correspondingly increased to counteract the disturbances. The mapping relationship between the oscillation amplitude and the energy distribution value is linearly proportional, and the proportionality coefficient is called the oscillation gain coefficient, which ranges from 0.02 mm to 0.08 mm. The energy distribution value is in Newton-square degrees. If the energy distribution value in a certain angular interval is 500 Newton-square degrees and the oscillation gain coefficient is 0.05 mm, then the longitudinal axis oscillation amplitude in that interval is 0.025 mm. The upper limit of the oscillation amplitude is set to 0.1 mm. When the amplitude exceeds the upper limit, it is forcibly truncated to the upper limit value to prevent the oscillation amplitude from being too large and affecting the machining accuracy.

[0029] The synthetic trajectory calculation module establishes a dual-axis spatial coordinate system, with the horizontal axis defined as the X-axis and the vertical axis as the Y-axis, with the origin located at the initial feed position of the saw blade. The horizontal axis deflection displacement curve is mapped to the X-axis displacement sequence in the time domain. Time and roll rotation angle are correlated through rotational speed; at a rotational speed of 10 revolutions per minute, 1 degree angle corresponds to 0.0167 seconds. The vertical axis oscillation parameters are converted into an oscillating displacement sequence in the Y-axis time domain. The oscillation displacement adopts a sine wave form, with the oscillation frequency and amplitude serving as the frequency and amplitude parameters of the sine function. The dual-axis displacements at each moment are calculated through vector superposition. The X-axis displacement value comes from the sampled value of the horizontal axis deflection displacement curve at that moment, and the Y-axis displacement value comes from the calculated value of the vertical axis oscillation function at that moment. The two constitute the dual-axis displacement vector at that moment. The displacement vector sequence covers the complete rotation cycle of the roll, and the sequence length is equal to the number of angle sampling points, 360. The vector sequence undergoes kinematic constraint verification, including dual-axis maximum speed limits and dual-axis maximum acceleration limits. The maximum velocity limit is set at 5 mm / s to 20 mm / s for the X-axis and 10 mm / s to 40 mm / s for the Y-axis. The maximum acceleration limit is set at 500 mm / s² to 2000 mm / s² for the X-axis and 1000 mm / s² to 5000 mm / s² for the Y-axis. During verification, the velocity vector is obtained by calculating the difference between displacement vectors at adjacent time points, and the acceleration vector is obtained by differentiating the velocity vectors. The axial components of the velocity and acceleration vectors are compared with the limit values. If a component of a certain axis exceeds the limit value at a certain time point, the X-axis and Y-axis displacements at that time point are scaled proportionally by the limit value divided by the excess value, ensuring that the scaled velocity or acceleration is exactly equal to the limit value. The scaling operation maintains the proportional relationship between the X-axis and Y-axis displacements to avoid distortion of the trajectory direction. The displacement vector sequence that has passed the constraint verification is the composite displacement trajectory of the two axes at each time point.

[0030] The smoothing module filters the synthesized displacement trajectory using either a moving average filter or a Gaussian filter, with a filtering window width of 3 to 7 sampling points. The filtering aims to eliminate high-frequency noise and abrupt changes in the trajectory, improving its continuity and executability. The filtered trajectory data is decomposed into a lateral axis displacement command sequence and a longitudinal axis displacement command sequence. Both sequences are synchronized in time, with a sampling period of 0.0167 seconds. The command sequence data format is a key-value pair of timestamp and displacement value, with displacement value accuracy maintained to 0.001 mm and timestamp accuracy maintained to 0.001 seconds. The command sequence is sent to the dual-axis drive system via the motion controller interface. The interface protocol uses pulse direction signals or analog voltage signals, with pulse frequencies ranging from 1 kHz to 100 kHz and analog voltages ranging from -10 V to +10 V.

[0031] This method achieves precise control of the cutting process. The transverse axis performs the tool deflection action based on the low-frequency cutting force component, while the longitudinal axis adaptively adjusts the oscillation frequency according to the vibration risk index. This decoupled linkage method can simultaneously cope with macroscopic cutting force changes and microscopic vibration risks, significantly improving the stability and disturbance resistance during the roll cutting process.

[0032] In one optional implementation, the step of calculating the composite trajectory of the transverse axis tool deflection displacement curve and the longitudinal axis oscillation parameters to obtain the composite displacement trajectory of the two axes at each moment includes: A dual-axis spatial coordinate system is established, and the transverse axis tool displacement curve is mapped to the transverse axis displacement sequence in the time domain. The target oscillation frequency and the oscillation amplitude are converted into the longitudinal axis oscillation displacement sequence in the time domain. Based on the correspondence between the roll rotation angle and time, the transverse axis displacement value and the longitudinal axis oscillation displacement value at each moment are determined. Using the transverse axis displacement value as the transverse component and the longitudinal axis oscillation displacement value as the longitudinal component, the vector sum of the two-axis displacements at each moment is calculated to obtain a composite displacement vector sequence. The composite displacement vector sequence is then subjected to kinematic constraint checks, including a maximum velocity limit and a maximum acceleration limit for both axes. When there is a displacement point in the composite displacement vector sequence that exceeds the kinematic constraints, the transverse and longitudinal components of that displacement point are scaled proportionally. The composite displacement vector sequence that satisfies the kinematic constraints is then converted into a composite displacement trajectory of both axes at each moment.

[0033] For example, the dual-axis spatial coordinate system is established using a Cartesian coordinate system model. The horizontal axis is defined as the X-axis direction, and the vertical axis is defined as the Y-axis direction. The origin of the coordinate system is set at the initial feed position of the saw blade. The positive direction of the X-axis points the direction in which the saw blade moves away from the roll, and the positive direction of the Y-axis points in the direction in which the saw blade moves axially. The coordinate system units are uniformly set to millimeters, the time unit to seconds, and the angle unit to degrees. The data structure of the horizontal axis deflection displacement curve is a mapping table of angles and displacements. The mapping table contains 360 data points, each corresponding to an integer angle position within the roll rotation cycle, with the angle range from 0 degrees to 359 degrees. The displacement value is a floating-point number, with precision maintained to three decimal places. The time domain mapping module receives the horizontal axis deflection displacement curve as input and establishes a correspondence between angle and time based on the roll rotation speed. The rotation speed unit is revolutions per minute, with a typical range of 8 to 15 revolutions per minute. The correspondence between angle and time is calculated using rotational speed. At a rotational speed of 10 revolutions per minute, the roll takes 6 seconds to complete a 360-degree rotation, with 1 degree angle corresponding to a time interval of 0.0167 seconds. The time-domain mapping module iterates through the 360 ​​angular positions of the transverse axis deflection curve, multiplies each angular position by the unit angular time interval to obtain the corresponding time, and assigns the displacement value of that angular position to that time, forming the transverse axis displacement sequence in the time domain. The transverse axis displacement sequence data structure is a key-value pair array of time and displacement, with an array length of 360. Time values ​​start from 0 seconds and increment at 0.0167-second intervals. Displacement values ​​directly inherit the displacement data of the corresponding angle in the transverse axis deflection curve.

[0034] The longitudinal axis oscillation parameters include the target oscillation frequency and oscillation amplitude. The target oscillation frequency is an array of oscillation frequencies at each circumferential angular position, with an array length of 360 and a unit of Hertz, ranging from 30 Hertz to 120 Hertz. The oscillation amplitude is an array of oscillation amplitudes at each circumferential angular position, with an array length of 360 and a unit of millimeters, ranging from 0.01 mm to 0.1 mm. The longitudinal axis oscillation displacement sequence is generated using a sine wave model. The frequency and amplitude parameters of the sine wave are dynamically adjusted at different times according to the target oscillation frequency and amplitude at the corresponding angular position. The time base of the oscillation displacement sequence is synchronized with the transverse axis displacement sequence, with 360 time points and a time interval of 0.0167 seconds. The longitudinal axis oscillation displacement value at each moment is calculated as follows: the corresponding angular position at that moment is obtained by querying the correspondence between time and angle. After obtaining the angular position, the frequency and amplitude values ​​of that angular position are extracted from the target oscillation frequency array and the oscillation amplitude array. The oscillation displacement value is calculated according to y = A × sin(2πft), where y is the oscillation displacement value, A is the amplitude value, f is the frequency value, and t is the time accumulation value. The time accumulation value starts from the beginning of the angular interval where the moment is located, and the accumulation step size is the time interval. To avoid abrupt displacement changes when the frequency switches between adjacent angular intervals, the phase of the sine function remains continuous when the angular interval switches. Continuity is achieved by recording the phase value at the end of the previous interval and using it as the starting phase of the new interval. The longitudinal axis oscillation displacement sequence data structure is a key-value pair array of time and oscillation displacement, with an array length of 360. The time values ​​are completely consistent with the time values ​​of the transverse axis displacement sequence, and the oscillation displacement value is the result of sine wave calculation, with the positive and negative range determined by the oscillation amplitude.

[0035] The determination of the transverse axis displacement value and the longitudinal axis oscillation displacement value at each moment is achieved through time indexing. The time index traverses the transverse axis displacement sequence and the longitudinal axis oscillation displacement sequence, keeping the time values ​​of the two sequences synchronized. The time index starts from 0 seconds and increases in increments of 0.0167 seconds until it covers the complete rotation cycle of the roll. The transverse axis displacement value at a certain moment is read from the transverse axis displacement sequence, and the longitudinal axis oscillation displacement value is read from the longitudinal axis oscillation displacement sequence. The vector sum of the dual-axis displacements is calculated using the vector addition rule. The transverse axis displacement value is used as the transverse component of the vector, and the longitudinal axis oscillation displacement value is used as the longitudinal component of the vector. The composite displacement vector at each moment is composed of the transverse component and the longitudinal component. The composite displacement vector sequence data structure is a key-value pair array of time and displacement vectors, with an array length of 360. The displacement vector contains a transverse component field and a longitudinal component field. The transverse component field stores the transverse axis displacement value at that moment, and the longitudinal component field stores the longitudinal axis oscillation displacement value at that moment. Both component fields are floating-point numbers, with precision maintained to three decimal places.

[0036] The kinematic constraint verification module receives the synthesized displacement vector sequence as input. The verification includes dual-axis maximum velocity limits and dual-axis maximum acceleration limits. The dual-axis maximum velocity limits are divided into lateral axis maximum velocity limits and longitudinal axis maximum velocity limits. Typical values ​​for the lateral axis maximum velocity limit are 5 mm / s to 20 mm / s, and for the longitudinal axis maximum velocity limit are 10 mm / s to 40 mm / s. The dual-axis maximum acceleration limits are also divided into lateral axis maximum acceleration limits and longitudinal axis maximum acceleration limits. Typical values ​​for the lateral axis maximum acceleration limit are 500 mm / s² to 2000 mm / s², and for the longitudinal axis maximum acceleration limit are 1000 mm / s² to 5000 mm / s². Velocity calculation uses the displacement difference method; the lateral axis velocity value at a given moment is calculated according to v... x =(x i -x i-1 ) / Δt is calculated, where v x The lateral axis velocity value, x i Let x be the horizontal component at that moment. i-1 The horizontal component represents the previous moment, Δt represents the time interval, and the longitudinal axis velocity value is determined according to v. y =(y i -y i-1 ) / Δt is calculated, where v y The longitudinal axis velocity value, y i For the longitudinal component at that moment, y i-1 This represents the longitudinal component from the previous moment. Acceleration is calculated using the velocity difference method; the transverse axis acceleration value at a given moment is determined according to a... x =(v_{x,i}-v_{x,i-1}) / Δt is calculated, where a x Let v_{x,i} be the lateral axis acceleration value, v_{x,i} be the lateral axis velocity value at this moment, and v_{x,i-1} be the lateral axis velocity value at the previous moment. The longitudinal axis acceleration value is calculated according to a. y =(v_{y,i}-v_{y,i-1}) / Δt is calculated, where a y Let v_{y,i} be the longitudinal axis acceleration value, v_{y,i-1} be the longitudinal axis velocity value at that moment, and v_{y,i-1} be the longitudinal axis velocity value at the previous moment. The verification module iterates through all moments of the synthesized displacement vector sequence, calculating the lateral axis velocity value, longitudinal axis velocity value, lateral axis acceleration value, and longitudinal axis acceleration value at each moment. The calculation results are compared with the corresponding maximum velocity limit and maximum acceleration limit. The absolute values ​​of the velocity and acceleration values ​​are used in the comparison. If the absolute value of the lateral axis velocity exceeds the maximum lateral axis velocity limit, or the absolute value of the longitudinal axis velocity exceeds the maximum longitudinal axis velocity limit, or the absolute value of the lateral axis acceleration exceeds the maximum lateral axis acceleration limit, or the absolute value of the longitudinal axis acceleration exceeds the maximum longitudinal axis acceleration limit, then that moment is determined to be a displacement point exceeding the kinematic constraints.

[0037] Displacement points exceeding kinematic constraints require scaling. The goal of scaling is to adjust the exceeding velocity or acceleration values ​​to exactly equal the corresponding limit values, while maintaining the proportional relationship between the lateral and longitudinal components. The scaling ratio is calculated as follows: if the velocity exceeds the limit, the scaling ratio is k = v_max / v_actual, where k is the scaling ratio, v_max is the maximum velocity limit, and v_actual is the actual velocity value. If the acceleration exceeds the limit, the scaling ratio is k = a_max / a_actual, where a_max is the maximum acceleration limit, and a_actual is the actual acceleration value. When both the lateral and longitudinal axes exceed the limits, the smaller of the two scaling ratios is used. The scaling operation is performed simultaneously on the lateral and longitudinal components of the displacement point. The scaled lateral component is calculated as x' = x × k, and the scaled longitudinal component is calculated as y' = y × k, where x' is the scaled lateral component, x is the original lateral component, and y' is the scaled longitudinal component, y is the original longitudinal component. After scaling is completed, the verification module recalculates the velocity and acceleration values ​​at the displacement point to confirm that neither the velocity nor acceleration values ​​exceed the limits. Scaling causes continuous changes in displacement values ​​between adjacent time points. The verification module employs an iterative optimization strategy, with 3 to 5 iterations. Each iteration re-verifies and scales all time points until the velocity and acceleration values ​​at all time points satisfy the kinematic constraints.

[0038] The synthetic displacement vector sequence satisfying kinematic constraints is converted into a synthetic displacement trajectory along two axes at various time points. The conversion process involves data structure reorganization. The synthetic displacement trajectory data structure is a triplet array of time, lateral axis displacement, and longitudinal axis displacement, with an array length of 360. The time field inherits the time value of the synthetic displacement vector sequence, the lateral axis displacement field inherits the lateral component field, and the longitudinal axis displacement field inherits the longitudinal component field. The synthetic displacement trajectory data precision is maintained to four decimal places for time and three decimal places for displacement. The data storage format is a comma-separated value file or a binary array, with a file size of approximately 30 to 50 kilobytes. The synthetic displacement trajectory is output to a motion controller interface. The interface supports pulse direction signals or analog voltage signals, with a pulse frequency range of 1 kHz to 100 kHz and an analog voltage range of -10 V to +10 V. The interface communication cycle is 1 to 10 milliseconds, and data transmission uses direct memory access, with a transmission latency of less than 0.5 milliseconds.

[0039] This invention establishes a dual-axis spatial coordinate system and performs vector synthesis calculations to achieve precise decoupling and coordinated control of the transverse axis tool deflection action and the longitudinal axis oscillation action. Kinematic constraint verification and proportional scaling strategies ensure that the synthesized displacement trajectory meets the speed and acceleration limits of the mechanical system, avoiding overshoot or oscillation during trajectory execution and improving the stability and accuracy of cutting.

[0040] In one optional implementation, during the cutting process, the step of acquiring multi-source sensor signals in real time and estimating cutting state parameters, including the actual cutting force and the actual material hardness, includes: During the cutting process, cutting force signals are collected by force sensors, acoustic emission signals from the cutting area are collected by acoustic emission sensors, and vibration signals are collected by vibration sensors to obtain multi-source sensor signals. Feature extraction is performed on the cutting force signal, the acoustic emission signal, and the vibration signal to obtain cutting force features, acoustic emission features, and vibration features. The actual cutting force and the actual material hardness are estimated by multi-source information fusion.

[0041] For example, a force sensor is installed at the connection between the saw blade feed mechanism and the machine bed. It is a piezoelectric force sensor or a strain gauge force sensor, with a measurement range of 0 Newtons to 5000 Newtons, a sensitivity of 2 mV to 5 mV per Newton, and a natural frequency higher than 1 kHz to capture dynamic changes in cutting force. The force sensor outputs an analog voltage signal, which is transmitted to a signal conditioning module via a shielded cable. The conditioning module includes an amplification circuit and a filtering circuit. The amplification factor is adjustable from 10x to 100x, and the filtering circuit uses a low-pass filter with a cutoff frequency set to 500 Hz to 800 Hz to suppress high-frequency noise and electromagnetic interference. The conditioned signal enters an analog-to-digital converter with a 16-bit resolution and a sampling frequency of 2 kHz to 5 kHz. The sampling frequency is determined based on the roll rotation speed and the required time resolution. A typical sampling frequency is 3 kHz at a rotation speed of 10 revolutions per minute, corresponding to 1800 data points collected per revolution. The cutting force signal data structure is a key-value pair sequence of timestamp and cutting force amplitude. The timestamp accuracy is 0.001 seconds, the cutting force amplitude unit is Newton, and the data storage accuracy is one decimal place.

[0042] The acoustic emission sensor is mounted on the side of the saw blade holder or on the roll support. It employs a piezoelectric ceramic acoustic emission sensor with a resonant frequency of 150 kHz to 300 kHz and a sensitivity of 60 to 80 dB / dB. The acoustic emission signal reflects microscopic dynamic processes in the cutting area, such as microcrack propagation, grain fracture, and frictional contact. The signal frequency range is 50 kHz to 500 kHz. The acoustic emission sensor outputs a high-frequency voltage signal, which is amplified by a preamplifier with a gain of 40 to 60 dB. The preamplifier bandwidth matches the sensor's resonant frequency range. After amplification, the signal enters the acoustic emission signal processing unit. The processing unit performs envelope detection to extract the acoustic emission amplitude envelope. Envelope detection uses a combination of full-wave rectification and low-pass filtering, with the low-pass filter cutoff frequency ranging from 10 kHz to 20 kHz. The envelope signal is sampled by an analog-to-digital converter with a sampling frequency of 50 kHz to 100 kHz and a converter resolution of 14 to 16 bits. The acoustic emission signal data structure is a key-value pair sequence of timestamps and acoustic emission amplitudes. The acoustic emission amplitude is in decibels or millivolts, and the data storage precision is one decimal place.

[0043] The vibration sensor is installed near the saw blade spindle bearing housing or bed guide rail. It is an acceleration-type vibration sensor with a measurement range of -50 m / s² to +50 m / s², a frequency response range of 1 Hz to 5 kHz, and a sensitivity of 10 mV / s² to 50 mV / s². The vibration sensor outputs an analog voltage signal, which is amplified by a charge amplifier or voltage amplifier at a magnification of 10 to 50 times. The amplified signal then enters a bandpass filter with a lower cutoff frequency of 10 Hz and an upper cutoff frequency of 2 kHz, filtering out low-frequency drift and ultra-high-frequency noise. The filtered signal is then sampled by an analog-to-digital converter at a sampling frequency of 5 kHz to 10 kHz with a 16-bit resolution. The vibration signal data structure is a key-value sequence of timestamps and vibration acceleration amplitudes, with the vibration acceleration amplitude unit being m / s². The data storage precision is two decimal places. The time reference of the multi-source sensor signals is unified by a hardware synchronization trigger. The trigger receives the pulse signal from the roll rotary encoder as a synchronization reference to ensure that the timestamps of the cutting force signal, acoustic emission signal and vibration signal are aligned, with a timestamp alignment error of less than 1 millisecond.

[0044] The cutting force feature extraction module receives the cutting force signal as input. Feature extraction is performed within a time window, with the window length set to 0.1 to 0.5 seconds. The window sliding step size is 50% to 75% of the window length, ensuring a balance between the temporal continuity and computational efficiency of feature calculation. Cutting force features include the mean cutting force, peak cutting force, standard deviation of cutting force, and peak-to-peak cutting force. The mean cutting force is calculated as the arithmetic mean of the cutting force amplitudes at all sampling points within the window. The peak cutting force is the maximum value of the cutting force amplitude within the window. The standard deviation of the cutting force is calculated by taking the square root of the average of the squared deviations of the cutting force amplitudes from the mean within the window. The peak-to-peak cutting force is the difference between the maximum and minimum values ​​of the cutting force amplitude within the window. The cutting force feature data structure is a tuple containing the start time of the time window, the mean cutting force, the peak cutting force, the standard deviation of the cutting force, and the peak-to-peak cutting force. The feature value precision is maintained to one decimal place.

[0045] The acoustic emission feature extraction module receives the acoustic emission signal as input. The feature extraction time window length is consistent with the cutting force feature extraction, and the window sliding step size is synchronized. Acoustic emission features include acoustic emission energy, acoustic emission count, root mean square (RMS) acoustic emission amplitude, and acoustic emission peak frequency. Acoustic emission energy is calculated as the sum of the squares of the acoustic emission amplitudes within the window. The acoustic emission count is the number of times the acoustic emission amplitude exceeds a set threshold within the window; the threshold is typically 10% to 20% of the dynamic range of the acoustic emission signal. The RMS acoustic emission amplitude is calculated by taking the square root of the average of the squares of the acoustic emission amplitudes within the window. The acoustic emission peak frequency is obtained by performing a Fast Fourier Transform on the acoustic emission signal within the window to obtain the spectrum; the frequency component with the largest amplitude in the spectrum is the peak frequency. The acoustic emission feature data structure is a tuple containing the start time of the time window, acoustic emission energy, acoustic emission count, RMS acoustic emission amplitude, and acoustic emission peak frequency. The feature value precision is maintained to zero decimal places for acoustic emission energy, integer for acoustic emission count, one decimal place for RMS acoustic emission amplitude, and zero decimal places for acoustic emission peak frequency.

[0046] The vibration feature extraction module receives vibration signals as input, and its feature extraction time window length and sliding step size are synchronized with the aforementioned module. Vibration features include the root mean square (RMS) of vibration amplitude, peak value, vibration frequency band energy distribution, and peak frequency. The RMS of vibration amplitude is calculated by taking the square root of the average of the squares of the vibration acceleration amplitudes within the window. The peak value is the maximum value of the vibration acceleration amplitude within the window. The vibration frequency band energy distribution is obtained by performing a Fast Fourier Transform on the vibration signal within the window. The spectrum is divided into a low-frequency band (0 Hz to 100 Hz), a mid-frequency band (100 Hz to 500 Hz), and a high-frequency band (500 Hz to 2000 Hz). The energy of each frequency band is calculated by summing the squares of the amplitudes of the frequency components within that band. The peak frequency is the frequency component with the largest amplitude in the spectrum. The vibration characteristic data structure is a multivariate set of the start time of the time window, root mean square vibration amplitude, peak vibration, low-frequency energy, mid-frequency energy, high-frequency energy and peak vibration frequency. The accuracy of the characteristic values ​​is maintained to two decimal places for the root mean square vibration amplitude, two decimal places for the peak vibration, zero decimal places for the frequency band energy and zero decimal places for the peak vibration frequency.

[0047] The multi-source information fusion module receives cutting force characteristics, acoustic emission characteristics, and vibration characteristics as inputs. The fusion method employs a weighted fusion model. This model establishes a mapping relationship between cutting force characteristics, acoustic emission characteristics, and vibration characteristics to actual cutting force and actual material hardness. The weighting coefficients for this mapping relationship are determined through offline calibration experiments. The estimated value of the actual cutting force is calculated as: (mean cutting force × 0.7) + (acoustic emission energy × 0.00015) + (root mean square vibration amplitude × 15) + 50 Newtons. The estimated value of the actual material hardness is calculated as: (peak cutting force × 0.04) + (peak acoustic emission frequency × 0.0001) + (mid-frequency vibration energy × 0.005) + 10 HRC. The fusion result data structure consists of a triplet containing the start time of the time window, the estimated value of the actual cutting force, and the estimated value of the actual material hardness. The accuracy of the estimates is maintained to one decimal place for the actual cutting force and zero decimal places for the actual material hardness.

[0048] This method provides a reliable basis for adaptive control and quality monitoring of the cutting process through multi-source sensor signal acquisition and feature extraction.

[0049] In one optional implementation, a feedback control layer generates biaxial compensation commands based on the deviation between the actual cutting force and the predicted cutting force, and adjusts the longitudinal axis oscillation parameters when the actual vibration amplitude exceeds the target value; the model prediction correction layer updates the actual material hardness in reverse to the uncut area in the digital twin model, and the step of dynamically correcting the biaxial feedforward commands for the uncut area includes: In the feedback control layer, the deviation between the actual cutting force and the predicted cutting force is calculated, and a transverse axis compensation command and a longitudinal axis compensation command are generated based on the deviation. The transverse axis compensation command is used to correct the feed rate of the transverse axis, and the longitudinal axis compensation command is used to correct the oscillating displacement of the longitudinal axis. When the actual vibration amplitude collected by the vibration sensor exceeds the vibration target value, the longitudinal axis oscillation amplitude is reduced. The longitudinal axis oscillation parameters include oscillation frequency and oscillation amplitude. In the model prediction correction layer, the actual material hardness is mapped to the spatial coordinate system of the digital twin model to determine the location of the cut area corresponding to the actual material hardness; Based on the correspondence between the location of the cut area and the actual material hardness, the material hardness distribution of the uncut area is calculated along the cutting path direction, and the material hardness distribution of the uncut area is corrected by the actual material hardness of the cut area. Based on the corrected material hardness distribution of the uncut region, the dual-axis feedforward command for the uncut region is recalculated. The dual-axis feedforward command includes the transverse axis feed rate and longitudinal axis oscillation parameters of the uncut region.

[0050] For example, the feedback control layer receives the actual cutting force estimate and the predicted cutting force curve as input. The actual cutting force estimate comes from the output of the multi-source information fusion module, and the predicted cutting force curve comes from the full-cycle cutting force prediction result of the digital twin model. The deviation calculation module performs time synchronization matching between the actual and predicted cutting forces. Matching is achieved through the roll rotation angle. The actual cutting force estimate carries the start time information of the time window. The start time of the time window is converted into the corresponding roll rotation angle through the roll rotation speed. The predicted cutting force value at the same angle position is extracted from the predicted cutting force curve. The deviation value is calculated by subtracting the predicted cutting force from the actual cutting force. A positive deviation value indicates that the actual cutting force is higher than the predicted value, and a negative deviation value indicates that the actual cutting force is lower than the predicted value. The deviation value data structure is a key-value pair of roll rotation angle and deviation amplitude. The deviation amplitude is in Newtons, and the accuracy is maintained to one decimal place.

[0051] The transverse axis compensation command generation module receives the deviation value as input. The compensation command uses a proportional control algorithm with a proportional control gain coefficient set to 0.001 to 0.005, in millimeters per second per Newton. The transverse axis compensation command is calculated as deviation value × proportional control gain coefficient. The result is the correction amount for the transverse axis feed speed. A positive correction amount indicates that the feed speed needs to be reduced to decrease the cutting force, while a negative correction amount indicates that the feed speed needs to be increased to increase the cutting force. The transverse axis feed speed correction amount is superimposed on the current transverse axis feed speed command. The superimposed feed speed command must not exceed the maximum feed speed limit of the transverse axis, which is typically 30 mm to 80 mm per second. The transverse axis compensation command data structure is a key-value pair between the roll rotation angle and the feed speed correction amount, in millimeters per second, with accuracy maintained to two decimal places. The transverse axis compensation command is sent to the transverse axis driver through the motion controller interface. The compensation command update cycle is 0.05 seconds to 0.2 seconds, synchronized with the output cycle of the actual cutting force estimate.

[0052] The longitudinal axis compensation command generation module receives the deviation value as input, and the compensation command fine-tunes the longitudinal axis oscillation displacement. The longitudinal axis oscillation displacement correction is calculated by multiplying the deviation value by the oscillation displacement correction gain coefficient, with a typical value of 0.00001 to 0.00005, in millimeters per Newton. The correction is added to the longitudinal axis oscillation amplitude, and the resulting oscillation amplitude must not exceed the upper limit of the longitudinal axis oscillation amplitude by 0.1 millimeters. The longitudinal axis compensation command data structure is a key-value pair between the roll rotation angle and the oscillation displacement correction, with the oscillation displacement correction in millimeters and accuracy maintained to three decimal places. The longitudinal axis compensation command is sent to the longitudinal axis driver through the motion controller interface, and the compensation command update cycle is consistent with that of the transverse axis compensation command.

[0053] The vibration amplitude monitoring module receives the actual vibration amplitude collected by the vibration sensor. The actual vibration amplitude is extracted from the vibration peak value or the root mean square value of the vibration amplitude of the vibration signal, with an extraction period of 0.1 to 0.5 seconds. The vibration target value is determined by the machining accuracy requirements and the vibration characteristics of the equipment, with a typical range of 5 m / s² to 15 m / s². The monitoring module compares the actual vibration amplitude with the vibration target value. When the actual vibration amplitude exceeds the vibration target value, the longitudinal axis oscillation parameter adjustment mechanism is triggered. The adjustment mechanism reduces the longitudinal axis oscillation amplitude. The reduction is calculated by subtracting the vibration target value from the actual vibration amplitude and then multiplying it by the oscillation amplitude adjustment coefficient. The typical value of the oscillation amplitude adjustment coefficient is 0.001 to 0.003, in millimeters per second squared. The reduced longitudinal axis oscillation amplitude must not be lower than the lower limit of 0.01 mm to avoid the oscillation amplitude being too small and losing its vibration suppression effect. The longitudinal axis oscillation amplitude adjustment command is sent to the longitudinal axis driver via the motion controller interface. The adjustment command has a higher priority than the longitudinal axis compensation command, and the adjustment command takes effect in 0.02 to 0.1 seconds. The oscillation frequency remains unchanged during the vibration amplitude adjustment process, and the oscillation frequency continues to be executed according to the target oscillation frequency in the vibration risk distribution map.

[0054] The model prediction correction layer receives the actual material hardness estimate as input, which comes from the output of the multi-source information fusion module and carries the start time information of the time window. The spatial coordinate mapping module maps the actual material hardness to the spatial coordinate system of the digital twin model. The spatial coordinate system of the digital twin model adopts a two-dimensional polar coordinate system of the roll circumferential angle and axial position. The circumferential angle ranges from 0 degrees to 360 degrees, and the axial position ranges from the axial length of the roll from the left end face to the right end face. The start time of the time window is converted into the roll circumferential angle by the roll rotation speed, and the axial position is determined by the current axial feed position of the saw blade, which is obtained from the position feedback signal of the longitudinal axis driver. The cut area position corresponding to the actual material hardness is the coordinate point of the circumferential angle and axial position. The data structure of the cut area position is a triplet of circumferential angle, axial position, and actual material hardness. The unit of actual material hardness is HRC, and the accuracy is maintained to zero decimal places.

[0055] The module for estimating the material hardness distribution in the uncut area is based on the correspondence between the location of the cut area and the actual material hardness. The estimation strategy employs a spatial interpolation method along the cutting path. The cutting path direction is the tangential direction of the roll circumference. Based on the circumferential angle of the cut area, the module extrapolates the circumferential angle position of the uncut area along the tangential direction. The extrapolated angle ranges from the circumferential angle of the cut area plus 1 degree to plus 180 degrees. The material hardness at each circumferential angle position in the uncut area is corrected using the actual material hardness of the cut area. The correction method is a weighted average, with the weight decreasing with the circumferential angle distance; the closer the distance, the greater the weight. The formula for calculating the corrected material hardness at a certain circumferential angle position in the uncut area is: actual material hardness of the cut area × weight coefficient + predicted material hardness from the digital twin model × (1 - weight coefficient). The weight coefficient is calculated as 1 / (1 + circumferential angle distance / 10 degrees), where the circumferential angle distance is the absolute value of the circumferential angle of the uncut area minus the circumferential angle of the cut area. The data structure for the material hardness distribution in the uncut area is a triplet array of circumferential angle, axial position, and corrected material hardness. The array length is equal to the number of circumferential angle positions in the uncut area, and the material hardness accuracy is maintained to zero decimal places.

[0056] The dual-axis feedforward command recalculation module receives the corrected material hardness distribution of the uncut area as input. The recalculation process reuses the cutting force prediction and dual-axis trajectory planning processes from the digital twin model. The corrected material hardness distribution replaces the original material hardness data of the uncut area in the digital twin model, re-triggers the cutting force prediction module to calculate the predicted cutting force values ​​at each circumferential angle position of the uncut area. The predicted cutting force values ​​are calculated using an empirical model based on the corrected material hardness and depth of cut. The regenerated cutting force prediction curve is input to the dual-axis trajectory planning module, which recalculates the transverse axis feed rate and longitudinal axis oscillation parameters of the uncut area. The transverse axis feed rate is adjusted based on the corrected cutting force prediction value; the higher the predicted cutting force value, the lower the feed rate. The typical adjustment ratio is 0.01 mm / s to 0.03 mm / s. The longitudinal axis oscillation parameters are adjusted based on the corrected vibration risk distribution map, which is recalculated using the corrected material hardness distribution and geometric eccentricity. The calculation methods for oscillation frequency and amplitude remain consistent with the original process. The recalculated dual-axis feedforward command data structure is a multi-element array of circumferential angle, transverse axis feed rate, longitudinal axis oscillation frequency, and longitudinal axis oscillation amplitude. The array covers all circumferential angle positions in the uncut area, and the command accuracy is maintained to two decimal places for the feed rate, zero decimal places for the oscillation frequency, and three decimal places for the oscillation amplitude.

[0057] This method enables dynamic correction of dual-axis feedforward commands, forming a closed-loop control architecture that combines feedforward and feedback. This enhances the cutting process's adaptability to material hardness fluctuations and working condition changes, ensuring cutting accuracy and stability.

[0058] In one optional implementation, the step of optimizing and storing the cutting force prediction parameters based on the performance indicators after the cutting task is executed includes: Obtain performance metrics after the cutting task is executed, including cutting accuracy and cutting efficiency; Extract the actual cutting force during the cutting task execution process and the predicted cutting force at the corresponding position in the full-cycle cutting force prediction curve, and calculate the prediction error between the actual cutting force and the predicted cutting force; A parameter optimization objective function is constructed, which takes the performance index and the prediction error as inputs and outputs the adjustment amount of the cutting force prediction parameters. The cutting force prediction parameters are updated according to the adjustment amount, and the updated cutting force prediction parameters are associated with the working condition information of the cutting task and stored in the parameter library. The working condition information includes material type and geometric eccentricity characteristics.

[0059] For example, the performance index acquisition module is activated after the cutting task is completed. The completion condition for the cutting task is that the saw blade has completed the cutting of the entire circumference of the roll and retracted to the initial position. Performance indices include cutting accuracy and cutting efficiency. Cutting accuracy is obtained by measuring the flatness and roundness of the roll end face after cutting. Flatness is measured using a coordinate measuring machine or a laser profilometer, with at least 36 measurement points evenly distributed on the circumference. Flatness is calculated as the maximum deviation between all measurement points and the fitted plane, in millimeters, with accuracy maintained to three decimal places. Roundness is also measured using a coordinate measuring machine, with at least 72 measurement points. Roundness is calculated as the maximum deviation between all measurement points and the fitted circle, in millimeters, with accuracy maintained to three decimal places. The cutting accuracy index is calculated as a weighted sum of flatness and roundness, with weighting coefficients of 0.6 and 0.4 respectively. The unit of the cutting accuracy index is millimeters; a smaller value indicates higher accuracy. Cutting efficiency is calculated as the ratio of the total cutting time to the roll cutting length. The total cutting time is the time difference from task start to task completion, expressed in seconds. The roll cutting length is the circumference of the roll, expressed in millimeters. The cutting efficiency is expressed as seconds per millimeter; a smaller value indicates higher efficiency. The performance index data structure is a triple consisting of a cutting task identifier, a cutting accuracy index, and a cutting efficiency index. The cutting task identifier is a unique string used to associate the cutting task's operating information with its execution data.

[0060] The actual cutting force extraction module extracts a sequence of estimated actual cutting forces from the output of the multi-source information fusion module during the cutting task execution process. The data structure of this sequence is an array of key-value pairs between the start time of the time window and the estimated actual cutting force. The array length equals the number of time windows during the cutting task execution process, typically ranging from 30 to 180. The predicted cutting force extraction module extracts the predicted cutting force value corresponding to the actual cutting force from the full-cycle cutting force prediction curve. The corresponding position is determined by converting the start time of the time window into the roll rotation angle. The conversion method is to multiply the start time of the time window by the roll rotation speed and then convert it into an angle. The roll rotation speed is measured in degrees per second. The full-cycle cutting force prediction curve data structure is an array of key-value pairs between the roll rotation angle and the predicted cutting force. The array length is 360, corresponding to sampling points along the 360-degree circumference of the roll. The predicted cutting force value is directly read from the prediction curve using an angle index. If the angle value is not an integer, linear interpolation is used. The interpolation method is a weighted average of the predicted cutting forces at two adjacent integer angle positions, with weights distributed inversely proportional to the distance between the angle value and the two integer angles.

[0061] The prediction error calculation module receives the actual cutting force estimation sequence and the predicted cutting force value sequence as input. The two sequences have the same length and their elements correspond one-to-one. The prediction error is calculated by subtracting the predicted cutting force value from the actual cutting force estimation value. The result is a prediction error sequence. The prediction error sequence data structure is a key-value pair array of roll rotation angles and prediction error values. The unit of the prediction error value is Newtons, and the precision is maintained to one decimal place. The statistical indicators of the prediction error include the prediction error mean, prediction error standard deviation, and prediction error maximum absolute value. The prediction error mean is calculated as the arithmetic mean of all elements in the prediction error sequence. The prediction error standard deviation is calculated by taking the square root of the sum of the squares of the deviations of each element from the mean. The maximum absolute value of the prediction error is the element with the largest absolute value in the prediction error sequence. The prediction error statistical indicator data structure is a triplet of the prediction error mean, prediction error standard deviation, and prediction error maximum absolute value, all in Newtons, with precision maintained to one decimal place.

[0062] The parameter optimization objective function construction module receives performance indicators and prediction error statistics as inputs. The objective function employs a weighted multi-objective optimization model. The optimization objectives are to minimize the cutting accuracy index, maximize the reciprocal of the cutting efficiency index, minimize the absolute value of the mean prediction error, and minimize the standard deviation of the prediction error. These four optimization objectives are normalized to a dimensionless range of 0 to 1. The normalization method is (each index - historical minimum) / (historical maximum - historical minimum). The historical minimum and maximum values ​​are obtained statistically from historical cutting task data stored in the parameter library. The four normalized optimization objectives are then weighted and summed according to weight coefficients of 0.4, 0.3, 0.2, and 0.1, with a sum of 1.0. The weighted sum is the objective function value, ranging from 0 to 1, with smaller values ​​indicating better overall performance. The objective function uses a gradient descent algorithm or a genetic algorithm to solve for the adjustment of the cutting force prediction parameters. These parameters include the tangential force coefficient, material hardness coefficient, and geometric eccentricity influence coefficient. The adjustment is the increment or scaling ratio of each parameter. The gradient descent algorithm has a learning rate of 0.01 to 0.1, an iteration count of 10 to 50, and terminates when the change in the objective function value is less than 0.001 or the maximum number of iterations is reached. The genetic algorithm has a population size of 20 to 100, a crossover probability of 0.6 to 0.9, a mutation probability of 0.01 to 0.1, and a generation count of 20 to 100. The termination condition is also 0.001 or the maximum number of generations is reached. The adjustment data structure consists of triplets of tangential force coefficient adjustment, material hardness coefficient adjustment, and geometric eccentricity influence coefficient adjustment. The adjustment amounts are relative increments in percentage form, with precision maintained to one decimal place.

[0063] The cutting force prediction parameter update module receives the adjustment amount as input. The update method is to multiply the original parameter value by a percentage of the adjustment amount and then add the original parameter value. The typical range of the original tangential force coefficient is 200 N / mm to 400 N / mm, the typical range of the original material hardness coefficient is 0.8 to 1.5, and the typical range of the original geometric eccentricity influence coefficient is 50 N / mm to 150 N / mm. The updated cutting force prediction parameters must meet physical constraints: the tangential force coefficient must not be lower than 100 N / mm or higher than 500 N / mm, the material hardness coefficient must not be lower than 0.5 or higher than 2.0, and the geometric eccentricity influence coefficient must not be lower than 20 N / mm or higher than 200 N / mm. If the updated parameters exceed the constraint range, they are forcibly truncated to the constraint boundary. The updated cutting force prediction parameter data structure is a triplet of the tangential force coefficient, material hardness coefficient, and geometric eccentricity influence coefficient, with accuracy maintained to one decimal place.

[0064] The parameter library storage module receives updated cutting force prediction parameters and cutting task operating condition information as input. The operating condition information includes material type and geometric eccentricity characteristics. The material type is obtained from the roll material properties and is encoded as a string. Typical values ​​include high-speed steel, tool steel, and alloy steel. The material type is associated with the material hardness range and cutting performance benchmark values. Geometric eccentricity characteristics are extracted from the eccentricity analysis results of the roll geometric model. These characteristics include the mean eccentricity and the peak-to-peak eccentricity. The mean eccentricity is the arithmetic mean of the eccentricity at all angular positions around the roll circumference, and the peak-to-peak eccentricity is the difference between the maximum and minimum eccentricity values. Both are in millimeters, with precision maintained to two decimal places. The parameter library uses a relational database or a key-value database. The database table structure includes fields such as cutting task identifier, material type, mean eccentricity, peak-to-peak eccentricity, tangential force coefficient, material hardness coefficient, geometric eccentricity influence coefficient, cutting accuracy index, cutting efficiency index, mean prediction error, standard deviation of prediction error, and storage timestamp. The parameter library supports querying historical parameters by material type and geometric eccentricity feature range. The query interface takes the material type and geometric eccentricity feature range as input and outputs a set of historical parameter records that meet the conditions. The query response time is less than 100 milliseconds. The parameter library periodically cleans up historical records that have exceeded the retention period, which is typically 6 to 24 months. The cleanup strategy is to retain the top 10 records with the best performance indicators for each combination of material type and geometric eccentricity feature.

[0065] This method stores the optimized parameters associated with working conditions in a parameter library, enabling the self-learning and self-optimization of the cutting force prediction model, thereby improving prediction accuracy and control performance.

[0066] In a second aspect, a computer-readable storage medium is provided, having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

Claims

1. A dual-axis decoupled linkage cutting method for roll sawing machines based on digital twins, characterized in that, include: Acquire three-dimensional laser scanning data and ultrasonic material testing data of the roll to be cut, construct a digital twin model of the roll including material hardness distribution and geometric eccentricity, and generate a full-cycle cutting force prediction curve and vibration risk distribution map; Based on the full-cycle cutting force prediction curve and vibration risk distribution map, a dual-axis decoupled linkage trajectory command is generated. The transverse axis performs a tool deflection action in the low-frequency band, and the tool deflection displacement is determined by the predicted cutting force amplitude. The longitudinal axis generates micro-oscillations in the high-frequency band, and the oscillation frequency is adaptively adjusted according to the vibration risk index. During the cutting process, signals from multiple sensor sources are collected in real time to estimate cutting state parameters, which include the actual cutting force and the actual material hardness. The feedback control layer generates biaxial compensation commands based on the deviation between the actual cutting force and the predicted cutting force. When the actual vibration amplitude exceeds the target value, the longitudinal axis oscillation parameters are adjusted. The model prediction correction layer updates the actual material hardness in reverse to the uncut area in the digital twin model, dynamically correcting the biaxial feedforward commands for the uncut area. Based on the performance indicators after the cutting task is completed, optimize and store the cutting force prediction parameters.

2. The method according to claim 1, characterized in that, The steps for acquiring 3D laser scanning data and ultrasonic material testing data of the roll to be cut, constructing a digital twin model of the roll including material hardness distribution and geometric eccentricity, and generating a full-cycle cutting force prediction curve and vibration risk distribution map include: Collect roll point cloud data, extract the center coordinates of each measurement section, calculate the geometric eccentricity based on the position offset of the center coordinates, and calculate the cylindricity error based on the radial fluctuation of the radius of each section. By scanning the roll circumference at multiple angles using an ultrasonic probe, and based on the sound velocity attenuation characteristics and echo amplitude distribution of the ultrasonic echo signal, a three-dimensional distribution of material hardness in the roll circumference direction is established. The geometric eccentricity, the cylindricity error, and the three-dimensional distribution of material hardness are spatially registered to construct a digital twin model of the roll that integrates material properties and geometric properties. Based on the aforementioned digital twin model of the roll, the instantaneous cutting force during the cutting process is calculated according to the cutting process parameters and the roll spatial characteristic parameters. The cutting process parameters include the saw blade cutting depth and the roll rotation angle, and the roll spatial characteristic parameters include the material hardness and geometric eccentricity corresponding to the angle. A full-cycle cutting force prediction curve is generated based on the temporal variation of the instantaneous cutting force within the complete rotation cycle of the roll. The vibration excitation intensity is calculated based on the rate of change of the cutting force amplitude in the full-cycle cutting force prediction curve and the spatial distribution of the hardness gradient in the three-dimensional distribution of material hardness, and a vibration risk distribution map is generated.

3. The method according to claim 1, characterized in that, Based on the full-cycle cutting force prediction curve and vibration risk distribution map, a dual-axis decoupled linkage trajectory command is generated. The transverse axis performs a tool deflection action in the low-frequency band, with the deflection displacement determined by the predicted cutting force amplitude. The longitudinal axis generates micro-oscillations in the high-frequency band, with the oscillation frequency adaptively adjusted according to the vibration risk index. The steps include: The full-cycle cutting force prediction curve is decomposed in the frequency domain to extract the low-frequency component and the high-frequency component of the cutting force. The low-frequency component of the cutting force represents the macroscopic change trend of the cutting force within the roll rotation cycle, and the high-frequency component of the cutting force represents the transient fluctuation of the cutting force caused by abrupt changes in material hardness and geometric eccentricity. A transverse shaft deflection curve is generated based on the amplitude of the low-frequency component of the cutting force. The transverse shaft deflection curve describes the displacement change law of the transverse shaft during the complete rotation cycle of the roll. Based on the vibration risk index at each circumferential angle position in the vibration risk distribution diagram, the target oscillation frequency of the longitudinal axis at the corresponding angle position is calculated. When the vibration risk index is higher than the vibration risk threshold, the target oscillation frequency is increased; when the vibration risk index is lower than the vibration risk threshold, the target oscillation frequency is decreased. The oscillation amplitude of the longitudinal axis is determined based on the frequency characteristics of the high-frequency component of the cutting force, and the oscillation amplitude is related to the energy distribution of the high-frequency component of the cutting force. The transverse axis tool deflection displacement curve and the longitudinal axis oscillation parameters are combined to calculate the composite trajectory. The longitudinal axis oscillation parameters include the target oscillation frequency and oscillation amplitude to obtain the composite displacement trajectory of the two axes at each moment. The synthesized displacement trajectory is smoothed to generate a dual-axis decoupled linkage trajectory command.

4. The method according to claim 3, characterized in that, The steps for calculating the composite trajectory of the transverse axis tool deflection displacement curve and the longitudinal axis oscillation parameters to obtain the composite displacement trajectory of the two axes at each moment include: A dual-axis spatial coordinate system is established, and the transverse axis tool displacement curve is mapped to the transverse axis displacement sequence in the time domain. The target oscillation frequency and the oscillation amplitude are converted into the longitudinal axis oscillation displacement sequence in the time domain. Based on the correspondence between the roll rotation angle and time, the transverse axis displacement value and the longitudinal axis oscillation displacement value at each moment are determined. Using the transverse axis displacement value as the transverse component and the longitudinal axis oscillation displacement value as the longitudinal component, the vector sum of the two-axis displacements at each moment is calculated to obtain a composite displacement vector sequence. The composite displacement vector sequence is then subjected to kinematic constraint checks, including a maximum velocity limit and a maximum acceleration limit for both axes. When there is a displacement point in the composite displacement vector sequence that exceeds the kinematic constraints, the transverse and longitudinal components of that displacement point are scaled proportionally. The composite displacement vector sequence that satisfies the kinematic constraints is then converted into a composite displacement trajectory of both axes at each moment.

5. The method according to claim 1, characterized in that, During the cutting process, the steps of acquiring multi-source sensor signals in real time and estimating cutting state parameters, including the actual cutting force and the actual material hardness, include: During the cutting process, cutting force signals are collected by force sensors, acoustic emission signals from the cutting area are collected by acoustic emission sensors, and vibration signals are collected by vibration sensors to obtain multi-source sensor signals. Feature extraction is performed on the cutting force signal, the acoustic emission signal, and the vibration signal to obtain cutting force features, acoustic emission features, and vibration features. The actual cutting force and the actual material hardness are estimated by multi-source information fusion.

6. The method according to claim 1, characterized in that, The feedback control layer generates biaxial compensation commands based on the deviation between the actual cutting force and the predicted cutting force. When the actual vibration amplitude exceeds the target value, the longitudinal axis oscillation parameters are adjusted. The model prediction correction layer updates the actual material hardness in reverse to the uncut area in the digital twin model. The steps of dynamically correcting the biaxial feedforward commands for the uncut area include: In the feedback control layer, the deviation between the actual cutting force and the predicted cutting force is calculated, and a transverse axis compensation command and a longitudinal axis compensation command are generated. The transverse axis compensation command is used to correct the feed rate of the transverse axis, and the longitudinal axis compensation command is used to correct the oscillating displacement of the longitudinal axis. When the actual vibration amplitude collected by the vibration sensor exceeds the vibration target value, the longitudinal axis oscillation amplitude is reduced. The longitudinal axis oscillation parameters include oscillation frequency and oscillation amplitude. In the model prediction correction layer, the actual material hardness is mapped to the spatial coordinate system of the digital twin model to determine the location of the cut area corresponding to the actual material hardness; Based on the correspondence between the location of the cut area and the actual material hardness, the material hardness distribution of the uncut area is calculated along the cutting path direction, and the material hardness distribution of the uncut area is corrected by the actual material hardness of the cut area. Based on the corrected material hardness distribution of the uncut region, the dual-axis feedforward command for the uncut region is recalculated. The dual-axis feedforward command includes the transverse axis feed rate and longitudinal axis oscillation parameters of the uncut region.

7. The method according to claim 1, characterized in that, The steps for optimizing and storing cutting force prediction parameters based on performance metrics after the cutting task is completed include: Obtain performance metrics after the cutting task is executed, including cutting accuracy and cutting efficiency; Extract the actual cutting force during the cutting task execution process and the predicted cutting force at the corresponding position in the full-cycle cutting force prediction curve, and calculate the prediction error between the actual cutting force and the predicted cutting force; A parameter optimization objective function is constructed, which takes the performance index and the prediction error as inputs and outputs the adjustment amount of the cutting force prediction parameters. The cutting force prediction parameters are updated according to the adjustment amount, and the updated cutting force prediction parameters are associated with the working condition information of the cutting task and stored in the parameter library. The working condition information includes material type and geometric eccentricity characteristics.

8. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images