Honeycomb core six-axis ultrasonic cutting tool path dynamic compensation and trajectory optimization control method

Abstract

The invention provides a honeycomb core six-axis ultrasonic cutting tool path dynamic compensation and trajectory optimization control method, which relates to the technical field of intelligent control, and comprises the following steps: obtaining honeycomb core structure parameters and cutting contour lines, establishing an interaction model based on deep reinforcement learning, and generating stress-strain prediction data; according to the method, feature regions are divided, differential initial machining parameters are set, an initial tool path track is generated, a digital twin model is established, machining data dynamic optimization parameters are collected in real time, and a tool path is compensated, so that the cutting precision and efficiency of the honeycomb core structure are improved, and machining deformation and material damage are reduced.

Description

Technical Field

[0001] The present invention relates to intelligent control technology, and in particular to a control method for dynamic compensation and trajectory optimization of six-axis ultrasonic cutting tool paths for honeycomb cores. Background Art

[0002] Honeycomb core structure is a lightweight structural material widely used in aerospace, rail transportation, shipbuilding and other fields. The six-axis ultrasonic cutting system has gradually become the main equipment for honeycomb core structure processing due to its flexibility and adaptability. Honeycomb core structure presents many challenges during processing due to its special geometric characteristics and heterogeneity. Traditional ultrasonic cutting processing methods usually use fixed processing parameters and cannot be dynamically adjusted according to the anisotropic characteristics and complex contour shapes of the honeycomb core material. This static processing parameter setting method is difficult to ensure consistency in processing quality when processing honeycomb core structures with different curvatures and different density areas, and is prone to problems such as cutting deformation, uneven surface roughness and increased tool wear.

[0003] Existing technologies for ultrasonic cutting of honeycomb cores still have the following defects and deficiencies: The lack of a toolpath planning method that considers the structural characteristics of the honeycomb core makes it impossible to optimize parameters specifically for different contour feature locations, which leads to local stress concentration and cutting defects when cutting complex curved surfaces. The lack of a real-time monitoring and dynamic compensation mechanism prevents timely adjustment of processing parameters based on the actual processing status, making it difficult to strike a balance between processing accuracy and efficiency. The lack of a comprehensive optimization method that combines material properties, processing parameters, and processing conditions such as cutting force, vibration, and temperature makes it difficult to achieve intelligent control and optimization of the entire processing process, hindering further improvements in the processing quality and efficiency of honeycomb core structures. Summary of the Invention

[0004] The embodiment of the present invention provides a control method for dynamic compensation and trajectory optimization of a six-axis ultrasonic cutting tool path for a honeycomb core, which can solve the problems in the prior art.

[0005] A first aspect of an embodiment of the present invention provides a control method for dynamic compensation and trajectory optimization of a six-axis ultrasonic cutting tool path for a honeycomb core, comprising: Obtain material characteristic parameters and cutting contour lines of the honeycomb core structure; Based on deep reinforcement learning, a tool-material interaction model is established, the material characteristic parameters and cutting contour line are input into the interaction model, the curvature change and tangent angle of the cutting contour line are analyzed, and the stress and strain prediction data of the honeycomb core structure at different contour feature positions are generated in combination with the arrangement of the honeycomb units. The entire processing area is divided into multiple feature areas based on the stress and strain prediction data, and differentiated initial processing parameters are set for different areas. The initial processing parameters include tool feed speed, spindle speed, ultrasonic amplitude and frequency; generating an initial tool path trajectory based on the area division result and the initial processing parameters; Establishing a digital twin model of the honeycomb core structure, collecting cutting force, vibration spectrum, and temperature field distribution data of the current processing area in real time, updating the processing state of the digital twin model, comparing the processing state with the stress and strain prediction data, calculating the deviation between the actual processing state and the predicted state, dynamically optimizing the processing parameters of the current position based on the deviation, compensating the initial tool path trajectory, and generating a compensated tool path trajectory; Processing is performed according to the compensated tool path trajectory, and actual processing data is fed back to the digital twin model in real time to continuously optimize processing parameters until the entire processing process is completed.

[0006] In an optional embodiment, Based on deep reinforcement learning, an interaction model between a tool and a material is established, the material characteristic parameters and the cutting contour line are input into the interaction model, the curvature change and the tangent angle of the cutting contour line are analyzed, and the stress and strain prediction data of the honeycomb core structure at different contour feature positions are generated in combination with the arrangement of the honeycomb units. The entire processing area is divided into multiple feature areas according to the stress and strain prediction data, and differentiated initial processing parameters are set for different areas, wherein the initial processing parameters include tool feed speed, spindle speed, ultrasonic amplitude and frequency. The steps include: Calculating the curvature value and tangent angle of the cutting contour line to obtain the wall thickness distribution data of the honeycomb core structure and the arrangement data of the honeycomb units; Constructing a state space of the interaction model, wherein the state space includes a material state vector constructed based on the material characteristic parameters and a geometric characteristic vector constructed based on the curvature value and tangent angle, wall thickness distribution data, and arrangement data; Constructing an action space of the interaction model, wherein the action space includes a tool feed speed adjustment amount, a spindle speed adjustment amount, an ultrasonic amplitude adjustment amount, and a frequency adjustment amount; Inputting the material state vector and the geometric characteristic vector into the interactive model to predict stress and strain data; Calculating a stress gradient value based on the stress-strain data, and dividing the processing area into a plurality of characteristic areas according to the stress gradient value; for each characteristic area, calculating a tool feed speed based on the curvature value, calculating a spindle speed based on the arrangement data, calculating an ultrasonic amplitude based on the stress gradient value, and calculating the frequency based on the wall thickness distribution data and the arrangement data, to obtain differentiated initial processing parameters; A comprehensive reward value is calculated based on the cutting force, machining quality parameters, and machining time, and the comprehensive reward value is fed back to the interactive model for online learning to obtain optimized initial machining parameters.

[0007] In an optional embodiment, The steps of calculating the tool feed speed based on the curvature value, calculating the spindle speed based on the arrangement data, calculating the ultrasonic amplitude based on the stress gradient value, and calculating the frequency based on the wall thickness distribution data and the arrangement data include: Calculating a curvature change rate of each point on the cutting contour line, and determining a discrete feed speed for each point based on the curvature change rate; when the curvature change rate is less than a first preset threshold, the discrete feed speed takes a maximum feed speed value; when the curvature change rate is greater than the first preset threshold and less than a second preset threshold, the discrete feed speed decreases according to a quadratic function with the curvature change rate; when the curvature change rate is greater than the second preset threshold, the discrete feed speed takes a minimum feed speed value; when the curvature gradient between adjacent points exceeds the second preset threshold, performing a smooth transition process on the discrete feed speed to obtain a continuous feed speed curve; Calculating the matching relationship between the ultrasonic period and the spindle speed based on the arrangement data of the honeycomb units, and taking the frequency multiplication value of the reference spindle speed and the arrangement data as the spindle speed; A stress segment value is calculated based on the maximum and minimum values ​​of the stress gradient value, the stress gradient value is divided into three segments based on the stress segment value, and the ultrasonic amplitude is adjusted to a corresponding preset value according to the position of the segment where the stress gradient value is located; The square root of the sum of the squares of the changes in the wall thickness distribution data in the horizontal and vertical directions is calculated to obtain the spatial variation rate, the arrangement period is determined according to the arrangement data, the natural frequency of the processing system is used as the reference frequency, and the frequency value is obtained based on the reference frequency, the spatial variation rate and the arrangement period.

[0008] In an optional embodiment, The steps for calculating the comprehensive reward value based on cutting force, machining quality parameters and machining time include: The synthetic cutting force and cumulative stress gradient during the machining process are calculated. The ratio of the actual material removal volume to the theoretical material removal volume per unit time is calculated to obtain the time efficiency index. The weighted average of the local surface roughness and the local surface integrity score is obtained to obtain the machining quality. The comparison result between the synthetic cutting force and the force threshold, and the comparison result between the cumulative stress gradient and the stress threshold are used as the basic process constraint layer reward value, and when the corresponding thresholds are exceeded, the basic process constraint layer reward value is calculated based on the difference using an exponential function; Calculate the ratio of the product of feed speed and amplitude to its reference value, and the ratio of the product of frequency and amplitude to its reference value, and use the weighted exponential sum of the differences between the two ratios and the reference value as the parameter coupling layer reward value; The comparison result between the processing quality and the quality threshold, and the comparison result between the time efficiency index and the efficiency threshold are used as the quality efficiency layer reward value, and when the corresponding thresholds are exceeded, the quality efficiency layer reward value is calculated based on the difference using an exponential function; The basic process constraint layer reward value, the parameter coupling layer reward value, and the quality efficiency layer reward value are multiplied together to obtain a comprehensive reward value.

[0009] In an optional embodiment, Based on the area division result and the initial processing parameters, the step of generating an initial tool path trajectory includes: Obtaining stress distribution data, geometric feature data, and processing difficulty data for each region in the region division result, and constructing the stress distribution data, geometric feature data, and processing difficulty data into a regional feature vector; analyzing the stress gradient, stress distribution, and curvature change in the regional feature vector to obtain a regional complexity evaluation result; A trajectory point distribution strategy is determined based on the regional complexity evaluation result, trajectory point arrangements are increased in the region where the stress gradient exceeds a preset gradient threshold, the tool posture is adjusted in the region where the curvature change exceeds a preset curvature change threshold, and the feed speed is optimized in the region where the stress distribution exceeds a preset stress distribution threshold; the trajectory point distribution strategy is combined with the initial machining parameters to form a trajectory point sequence that satisfies the six-axis motion constraints; the trajectory point sequence is converted into a continuous trajectory segment, and a transition trajectory is generated at the trajectory mutation position; the continuous trajectory segment and the transition trajectory are combined to form a complete initial tool path trajectory.

[0010] In an optional embodiment, The steps of establishing a digital twin model of a honeycomb core structure, collecting cutting force, vibration spectrum, and temperature field distribution data of a current processing area in real time, updating the processing state of the digital twin model, comparing the processing state with the stress and strain prediction data, calculating the deviation between the actual processing state and the predicted state, dynamically optimizing the processing parameters of the current position according to the deviation, and compensating the initial tool path trajectory to generate a compensated tool path trajectory include: Generate the initial state of the digital twin model based on the geometric parameters of the honeycomb core structure and the material characteristic parameters; calculate the square root of the sum of the squares of the cutting force in three directions to obtain the synthetic cutting force, and calculate the inverse tangent of the cutting force in the Y direction and the X direction to obtain the cutting force direction angle; perform Fourier transform on the vibration spectrum to extract the main frequency band and bandwidth; reconstruct the temperature field distribution data based on the spatial basis function and the time response function; construct the synthetic cutting force, cutting force direction angle, main frequency band, bandwidth and reconstructed temperature field data into a real-time state vector; input the real-time state vector into a Kalman filter to obtain the processing state of the digital twin model; Calculating the Euclidean distance between the processing state and the stress-strain prediction data to obtain a deviation; An adaptive adjustment coefficient is calculated based on the deviation, a parameter adjustment amount is obtained based on the adaptive adjustment coefficient and the deviation, and a normal compensation amount and a tangential compensation amount are calculated based on the parameter adjustment amount; the current machining position on the initial tool path trajectory is offset by the normal compensation amount along the normal unit vector and by the tangential compensation amount along the tangential unit vector to generate a compensated tool path trajectory.

[0011] In an optional embodiment, The steps of obtaining a parameter adjustment amount based on the adaptive adjustment coefficient and the deviation, and calculating a normal compensation amount and a tangential compensation amount based on the parameter adjustment amount include: Calculating the proportional term, integral term, and differential term based on the adaptive adjustment coefficient, and weighting the proportional term, integral term, and differential term in each direction to obtain a feed speed adjustment amount, a spindle speed adjustment amount, and a cutting depth adjustment amount; Converting the feed speed adjustment, spindle speed adjustment and cutting depth adjustment into normal compensation and tangential compensation; Collect machining surface profile data, cutting force fluctuation data and energy consumption data, take the deviation of the machining surface profile data from the target profile as the accuracy evaluation index, take the deviation of the cutting force fluctuation data from the set threshold as the stability evaluation index, and take the ratio of the energy consumption data to the benchmark energy consumption as the efficiency evaluation index; construct a compensation optimization target based on the accuracy evaluation index, stability evaluation index and efficiency evaluation index; construct a meta-learning network, the meta-learning network includes a task adaptation module for optimizing the compensation strategy and a strategy optimization module for online adjustment of the compensation amount; input the compensation optimization target into the task adaptation module to obtain the optimization parameters of the current machining condition, and adjust the normal compensation amount and the tangential compensation amount through the strategy optimization module based on the optimization parameters; adopt an iterative optimization method to update the parameters of the meta-learning network, and output the updated meta-learning network as the final normal compensation amount and the tangential compensation amount.

[0012] According to a second aspect of an embodiment of the present invention, an electronic device is provided, including: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0013] According to a third aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0014] The present invention establishes an interaction model between the tool and the material through deep reinforcement learning, which can accurately predict the stress and strain distribution at different contour feature positions, realize precise processing control of complex honeycomb core structures, and effectively avoid the precision loss caused by tool offset and material deformation in traditional processing methods.

[0015] The present invention uses digital twin technology to monitor and adjust processing parameters in real time, constructing a closed-loop feedback control system that can dynamically compensate the tool path trajectory according to the actual cutting state during the processing process, significantly improving the processing accuracy and surface quality of the honeycomb core structure, especially in areas with large curvature changes, and reducing the occurrence of processing defects.

[0016] The present invention sets differentiated initial processing parameters for different feature areas and performs real-time optimization during the processing process. This method greatly improves the processing efficiency of the honeycomb core structure, reduces tool wear, extends tool life, and reduces scrap rate, providing reliable technical support for the manufacture of high-precision honeycomb core structure components in aerospace and other fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 Schematic diagram of the flow of a control method for dynamic compensation and trajectory optimization of a six-axis ultrasonic cutting tool path for a honeycomb core according to an embodiment of the present invention; Figure 2 Schematic diagram of the relationship between ultrasonic cutting of honeycomb core and ultrasonic amplitude and stress gradient; Figure 3 This is the architecture diagram of the dynamic compensation and trajectory optimization system for the digital twin model of the honeycomb core structure; Figure 4 Schematic diagram comparing the compensation optimization effects of the meta-learning network for six-axis ultrasonic cutting of honeycomb cores. DETAILED DESCRIPTION

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0019] The following specific embodiments are used to describe the technical solution of the present invention in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0020] Figure 1 Schematic diagram of the control method for dynamic compensation and trajectory optimization of six-axis ultrasonic cutting tool path for honeycomb core according to an embodiment of the present invention. Figure 1 As shown, the method includes: Obtain material characteristic parameters and cutting contour lines of the honeycomb core structure; Based on deep reinforcement learning, a tool-material interaction model is established, the material characteristic parameters and cutting contour line are input into the interaction model, the curvature change and tangent angle of the cutting contour line are analyzed, and the stress and strain prediction data of the honeycomb core structure at different contour feature positions are generated in combination with the arrangement of the honeycomb units. The entire processing area is divided into multiple feature areas based on the stress and strain prediction data, and differentiated initial processing parameters are set for different areas. The initial processing parameters include tool feed speed, spindle speed, ultrasonic amplitude and frequency; generating an initial tool path trajectory based on the area division result and the initial processing parameters; Establishing a digital twin model of the honeycomb core structure, collecting cutting force, vibration spectrum, and temperature field distribution data of the current processing area in real time, updating the processing state of the digital twin model, comparing the processing state with the stress and strain prediction data, calculating the deviation between the actual processing state and the predicted state, dynamically optimizing the processing parameters of the current position based on the deviation, compensating the initial tool path trajectory, and generating a compensated tool path trajectory; Processing is performed according to the compensated tool path trajectory, and actual processing data is fed back to the digital twin model in real time to continuously optimize processing parameters until the entire processing process is completed.

[0021] In an optional embodiment, a tool-material interaction model is established based on deep reinforcement learning, the material characteristic parameters and the cutting contour are input into the interaction model, the curvature change and the tangent angle of the cutting contour are analyzed, and stress and strain prediction data of the honeycomb core structure at different contour feature positions are generated in combination with the arrangement of the honeycomb units. The entire processing area is divided into multiple feature areas based on the stress and strain prediction data, and differentiated initial processing parameters are set for different areas, wherein the initial processing parameters include tool feed speed, spindle speed, ultrasonic amplitude, and frequency. The steps include: Calculating the curvature value and tangent angle of the cutting contour line to obtain the wall thickness distribution data of the honeycomb core structure and the arrangement data of the honeycomb units; Constructing a state space of the interaction model, wherein the state space includes a material state vector constructed based on the material characteristic parameters and a geometric characteristic vector constructed based on the curvature value and tangent angle, wall thickness distribution data, and arrangement data; Constructing an action space of the interaction model, wherein the action space includes a tool feed speed adjustment amount, a spindle speed adjustment amount, an ultrasonic amplitude adjustment amount, and a frequency adjustment amount; Inputting the material state vector and the geometric characteristic vector into the interactive model to predict stress and strain data; Calculating a stress gradient value based on the stress-strain data, and dividing the processing area into a plurality of characteristic areas according to the stress gradient value; for each characteristic area, calculating a tool feed speed based on the curvature value, calculating a spindle speed based on the arrangement data, calculating an ultrasonic amplitude based on the stress gradient value, and calculating the frequency based on the wall thickness distribution data and the arrangement data, to obtain differentiated initial processing parameters; A comprehensive reward value is calculated based on the cutting force, machining quality parameters, and machining time, and the comprehensive reward value is fed back to the interactive model for online learning to obtain optimized initial machining parameters.

[0022] For example, the characteristic parameters of the honeycomb core material are collected, including elastic modulus 150GPa, Poisson's ratio 0.32, yield strength 260MPa, and hardness HRC45. The cutting contour line data of the honeycomb core structure to be processed is obtained. The contour line is obtained by digital scanning and contains 3000 discrete point coordinates. These discrete point coordinates are processed to calculate the curvature value of the contour line. The maximum curvature reaches 0.85 / mm at the concave corner and the minimum curvature is 0.05 / mm at the convex corner. The tangent angle is calculated at the same time, and the angle range is 0° to 175°. The wall thickness distribution data of the honeycomb core structure is obtained. The wall thickness of the central area is 0.8mm and the wall thickness of the edge area is 1.2mm. The honeycomb unit arrangement data is obtained. The central area is a regular hexagonal arrangement and the edge area is a deformed hexagonal arrangement. The unit size ranges from 4mm to 8mm.

[0023] When constructing the state space of the interaction model, the material state vector includes four dimensions: elastic modulus, Poisson's ratio, yield strength, and hardness; the geometric eigenvector includes four dimensions: curvature, tangent angle, wall thickness, and cell density. The state space is represented by an 8-dimensional vector, and the state of each sample point is expressed as [150, 0.32, 260, 45, κ, θ, t, d], where κ is the curvature value at the current point, θ is the tangent angle, t is the wall thickness, and d is the cell density. For the action space of the interaction model, the feed rate adjustment range is defined as -20 mm / min to +20 mm / min, the spindle speed adjustment range is -500 rpm to +500 rpm, the ultrasonic amplitude adjustment range is -5 μm to +5 μm, and the frequency adjustment range is -2 kHz to +2 kHz. The action space is represented by a 4-dimensional vector.

[0024] The interaction model is implemented using the Deep Deterministic Policy Gradient (DDPG) algorithm. Both the policy network and the value network are four-layer fully connected neural networks with 128 and 64 hidden nodes, respectively. The policy network takes a state vector as input and outputs an action vector; the value network takes a state vector and an action vector as input and outputs a Q value. The network is trained using the Adam optimizer with a learning rate of 0.001 and a discount factor of 0.95.

[0025] After inputting the material state vector and geometric eigenvector into the interactive model, stress and strain prediction data were obtained. The stress distribution range was 50 MPa to 230 MPa, and the strain distribution range was 0.05% to 0.8%. The stress gradient value was calculated based on the stress distribution, ranging from 0.5 MPa / mm to 8 MPa / mm. The processing area was divided into three characteristic regions based on the stress gradient threshold: low stress region (gradient <2 MPa / mm), medium stress region (gradient 2-5 MPa / mm), and high stress region (gradient >5 MPa / mm).

[0026] For each characteristic area, the tool feed rate, spindle speed, ultrasonic amplitude and frequency are calculated separately to obtain the differentiated initial processing parameters for each characteristic area. Low stress area: feed rate 90-100mm / min, spindle speed 6000-6500rpm, ultrasonic amplitude 13-15μm, frequency 20-21kHz. Medium stress area: feed rate 70-90mm / min, spindle speed 6500-7000rpm, ultrasonic amplitude 10-13μm, frequency 19-20kHz. High stress area: feed rate 50-70mm / min, spindle speed 7000-8000rpm, ultrasonic amplitude 5-10μm, frequency 18-19kHz.

[0027] During the machining process, cutting force, surface roughness, and machining time are monitored in real time to generate a comprehensive reward value. This calculated comprehensive reward value is fed back into the interactive model for online learning. The parameters of the policy network and value network are updated through an experience replay mechanism. A batch update is performed every 100 sample points, with a batch size of 64.

[0028] The present invention achieves accurate prediction and control of the honeycomb core structure processing process by constructing a material-tool interaction model; the design of the state space and action space enables the system to adaptively adjust the processing parameters according to the material properties and geometric features of different areas, avoiding processing defects caused by improper parameter settings in traditional methods; the comprehensive reward mechanism supports the continuous learning and optimization of the model, improving the system's adaptability to complex processing environments.

[0029] In an optional embodiment, the steps of calculating the tool feed speed based on the curvature value, calculating the spindle speed based on the arrangement data, calculating the ultrasonic amplitude based on the stress gradient value, and calculating the frequency based on the wall thickness distribution data and the arrangement data include: Calculating a curvature change rate of each point on the cutting contour line, and determining a discrete feed speed for each point based on the curvature change rate; when the curvature change rate is less than a first preset threshold, the discrete feed speed takes a maximum feed speed value; when the curvature change rate is greater than the first preset threshold and less than a second preset threshold, the discrete feed speed decreases according to a quadratic function with the curvature change rate; when the curvature change rate is greater than the second preset threshold, the discrete feed speed takes a minimum feed speed value; when the curvature gradient between adjacent points exceeds the second preset threshold, performing a smooth transition process on the discrete feed speed to obtain a continuous feed speed curve; Calculating the matching relationship between the ultrasonic period and the spindle speed based on the arrangement data of the honeycomb units, and taking the frequency multiplication value of the reference spindle speed and the arrangement data as the spindle speed; A stress segment value is calculated based on the maximum and minimum values ​​of the stress gradient value, the stress gradient value is divided into three segments based on the stress segment value, and the ultrasonic amplitude is adjusted to a corresponding preset value according to the position of the segment where the stress gradient value is located; The square root of the sum of the squares of the changes in the wall thickness distribution data in the horizontal and vertical directions is calculated to obtain the spatial variation rate, the arrangement period is determined according to the arrangement data, the natural frequency of the processing system is used as the reference frequency, and the frequency value is obtained based on the reference frequency, the spatial variation rate and the arrangement period.

[0030] For example, for the calculation of the feed speed, the system first obtains the curvature value of each point on the cutting contour line, and obtains the curvature change rate by dividing the curvature difference between adjacent points by the distance between the points. The system sets the discrete feed speed according to the curvature change rate. In practical applications, the first preset threshold can be set to 0.15, and the second preset threshold can be set to 0.45. When the curvature change rate is less than 0.15, the discrete feed speed takes the maximum value of 120mm / min; when the curvature change rate is between 0.15 and 0.45, the discrete feed speed decreases according to the quadratic function law, and the specific calculation is: discrete feed speed = 120-[(curvature change rate-0.15) / (0.45-0.15)]²×(120-40), where 120 represents the maximum feed speed and 40 represents the minimum feed speed; when the curvature change rate is greater than 0.45, the discrete feed speed takes the minimum value of 40mm / min. For adjacent points with a curvature gradient exceeding 0.45, the system uses a smooth transition process to avoid sudden changes in feed rate. The smooth transition process uses the Bezier curve interpolation method to generate a third-order Bezier curve between adjacent discrete points to ensure continuous and smooth speed changes. For example, for points P1 (20, 30) and P2 (25, 40) on the contour line, if their curvature change rates are 0.2 and 0.5 respectively, the discrete feed rate of point P1 is 105 mm / min and the discrete feed rate of point P2 is 40 mm / min. The system will generate a continuous feed rate curve between these two points to achieve a smooth transition from 105 mm / min to 40 mm / min.

[0031] The spindle speed is calculated based on the honeycomb cell layout data. The system first determines the honeycomb cell layout period, such as one honeycomb cell every 20 mm. The ultrasonic machining period must match the spindle speed so that each honeycomb cell receives uniform ultrasonic energy. The system selects a baseline spindle speed of 1000 rpm and then adjusts the spindle speed based on the frequency multiplication factor of the layout data. The frequency multiplication factor is calculated by dividing the layout period by the baseline period (e.g., 5 mm), and the resulting quotient is used as the frequency multiplication factor. For example, when the layout period is 20 mm, the frequency multiplication factor is 4, and the spindle speed is set to 1000 × 4 = 4000 rpm. When the layout period is 15 mm, the frequency multiplication factor is 3, and the spindle speed is set to 1000 × 3 = 3000 rpm. This ensures uniform distribution of ultrasonic energy across the honeycomb cells.

[0032] The calculation of the ultrasonic amplitude is based on the stress gradient value. The system first obtains the maximum and minimum stress gradient values ​​of the entire processing area, such as the maximum value is 85MPa / mm and the minimum value is 25MPa / mm. The stress segment value is obtained by dividing the difference between the maximum and minimum values ​​by 3, that is, (85-25) / 3=20MPa / mm. The system divides the stress gradient value into three segments: 25-45MPa / mm, 45-65MPa / mm and 65-85MPa / mm. When the stress gradient value is in the range of 25-45MPa / mm, the ultrasonic amplitude is set to the low range of 15μm; when the stress gradient value is in the range of 45-65MPa / mm, the ultrasonic amplitude is set to the mid-range of 25μm; when the stress gradient value is in the range of 65-85MPa / mm, the ultrasonic amplitude is set to the high range of 35μm. For example, for an area with a stress gradient value of 55 MPa / mm, the system sets the ultrasonic amplitude to a mid-range of 25 μm; for an area with a stress gradient value of 70 MPa / mm, the system sets the ultrasonic amplitude to a high-range of 35 μm.

[0033] The frequency calculation is based on the wall thickness distribution data and arrangement data. The system first calculates the changes in the wall thickness distribution data in the horizontal and vertical directions, recorded as Δx and Δy respectively. The spatial variation rate is calculated as the square root of (Δx²+Δy²). For example, if the horizontal wall thickness change at a point is 0.3mm and the vertical wall thickness change is 0.4mm, then the spatial variation rate of the point is the square root of (0.3²+0.4²), that is, 0.5mm. The system determines the arrangement period based on the arrangement data, such as 20mm. The natural frequency of the processing system, 22kHz, is used as the reference frequency, and the frequency value is calculated as: reference frequency × (1+spatial variation rate / arrangement period). For the above example, the frequency value is 22×(1+0.5 / 20)=22.55kHz. This calculation method can dynamically adjust the frequency according to the wall thickness change and arrangement characteristics to ensure processing accuracy and efficiency.

[0034] The selection of processing parameters for traditional honeycomb core ultrasonic cutting mainly relies on empirical settings or simple linear relationship adjustments. These methods fail to fully consider the synergistic relationship between the complex geometric characteristics and material properties of the honeycomb core structure, resulting in problems such as overcutting, undercutting, and edge tearing when cutting complex contours. The present invention achieves a precise matching of processing parameters with material properties and geometric features, especially the adaptive adjustment of feed speed for areas with curvature changes, which effectively avoids overcutting or undercutting during the processing process. The matching of spindle speed and honeycomb arrangement improves cutting stability, the coordination of ultrasonic amplitude and stress gradient reduces material deformation, and frequency optimization improves processing efficiency, thereby comprehensively improving cutting quality and efficiency.

[0035] Figure 2This is a schematic diagram of the relationship between ultrasonic cutting of honeycomb cores - ultrasonic amplitude and stress gradient. In the low stress area (25-45MPa / mm), the material is subjected to less force, and the system uses a low amplitude setting of 15μm to ensure processing stability while avoiding energy waste. When the stress gradient value enters the medium stress area (45-65MPa / mm), as shown at 55MPa / mm in the figure, the system automatically adjusts the amplitude to a mid-range 25μm to balance processing efficiency and surface quality. In the high stress area (65-85MPa / mm), as shown at the 70MPa / mm working point, the system increases the amplitude to a high-end of 35μm to provide sufficient energy to cope with high stress concentration areas in the material.

[0036] The segment boundary values ​​are determined using a scientific method. By calculating the difference between the maximum stress gradient (85MPa / mm) and the minimum stress gradient (25MPa / mm) and dividing it by 3, the stress segment value of 20MPa / mm is obtained, achieving uniform zoning. This adaptive amplitude adjustment strategy based on stress gradient overcomes the limitations of a single amplitude setting in traditional ultrasonic cutting, significantly reducing material deformation. It is particularly suitable for processing materials with complex geometric features and uneven stress distribution, such as honeycomb core structures, effectively improving cutting accuracy and surface quality.

[0037] In an optional embodiment, the step of calculating the comprehensive reward value based on the cutting force, the machining quality parameter and the machining time includes: The synthetic cutting force and cumulative stress gradient during the machining process are calculated. The ratio of the actual material removal volume to the theoretical material removal volume per unit time is calculated to obtain the time efficiency index. The weighted average of the local surface roughness and the local surface integrity score is obtained to obtain the machining quality. The comparison result between the synthetic cutting force and the force threshold, and the comparison result between the cumulative stress gradient and the stress threshold are used as the basic process constraint layer reward value, and when the corresponding thresholds are exceeded, the basic process constraint layer reward value is calculated based on the difference using an exponential function; Calculate the ratio of the product of feed speed and amplitude to its reference value, and the ratio of the product of frequency and amplitude to its reference value, and use the weighted exponential sum of the differences between the two ratios and the reference value as the parameter coupling layer reward value; The comparison result between the processing quality and the quality threshold, and the comparison result between the time efficiency index and the efficiency threshold are used as the quality efficiency layer reward value, and when the corresponding thresholds are exceeded, the quality efficiency layer reward value is calculated based on the difference using an exponential function; The basic process constraint layer reward value, the parameter coupling layer reward value, and the quality efficiency layer reward value are multiplied together to obtain a comprehensive reward value.

[0038] For example, the synthetic cutting force calculation of the machining process adopts the three-axis force measurement method, and the cutting force components in the X, Y, and Z directions are vector-synthesized to obtain the total resultant force value. For example, assuming that the measured X-direction force is 25N, the Y-direction force is 18N, and the Z-direction force is 32N, the synthetic cutting force is 44.3N. The cumulative stress gradient is obtained by collecting the stress distribution on the workpiece surface in real time during the machining process, calculating the stress difference between two adjacent points at intervals of 10 microseconds, and accumulating the stress difference during the entire machining process. For example, in an ultrasonic cutting process, the measured cumulative stress gradient value is 1850MPa / mm.

[0039] The calculation of the time efficiency index involves comparing actual material removal with theoretical values. The actual material removal volume is calculated by dividing the difference in workpiece weight before and after machining by the material density. The theoretical material removal volume is calculated based on the cutting parameters and path defined in the machining program. For example, a honeycomb core workpiece weighs 85g before machining and 72g after machining, with a material density of 0.12g / cm³. The calculated actual material removal volume is 108.3cm³, while the theoretical removal volume is 115cm³, resulting in a time efficiency index of 0.942.

[0040] Machining quality assessment encompasses two aspects: surface roughness and surface integrity. Surface roughness is measured at multiple points along the cut edge of the workpiece using a laser confocal microscope and the average is taken. Surface integrity is scored based on the ratio of defect-free edge area to the total edge length. Surface roughness is weighted 60% and surface integrity 40%. For example, if the measured average surface roughness Ra is 0.8μm (a full score of 1.0 translates to a linear score of 0.85) and the surface integrity score is 0.92, the machining quality score is 0.85 × 0.6 + 0.92 × 0.4 = 0.878.

[0041] The calculation of the reward value of the basic process constraint layer involves the comparison of the synthetic cutting force and the cumulative stress gradient with their corresponding thresholds. The cutting force threshold is set to 50N, and the stress gradient threshold is set to 2000MPa / mm. When the actual value is lower than the threshold, the reward value of this layer is 1.0; when it exceeds the threshold, the penalty coefficient is calculated using an exponential decay function. Taking the above case as an example, the synthetic cutting force of 44.3N is less than the threshold of 50N, and the cumulative stress gradient of 1850MPa / mm is less than the threshold of 2000MPa / mm, so the reward value of the basic process constraint layer is 1.0. If the cutting force exceeds the threshold, such as the measured value of 55N, which exceeds the threshold by 5N, then the exponential function e is used to calculate the penalty coefficient. (-0.05×5) =0.779 to calculate the score, and the final reward value for this layer is 0.779.

[0042] The calculation of the parameter coupling layer reward value considers two indicators: the product of feed speed and amplitude and the product of frequency and amplitude. Assuming the feed speed is 80mm / min and the amplitude is 15μm, the product of the reference value is 1200μm·mm / min, and the ratio is 1.0; the frequency is 20kHz, the amplitude is 15μm, and the product of the reference value is 300kHz·μm, then the ratio is 1.0. The reference value is set to 0.9, and the two differences are 0.1 and 0.1 respectively. Assuming the weights are 0.6 and 0.4 respectively, the weighted exponential function e is used. (-0.5×(0.6×0.1+0.4×0.1)) =0.951 to calculate the reward value of the parameter coupling layer.

[0043] The quality efficiency layer reward value is based on the comparison between the processing quality and the quality threshold, and the time efficiency index and the efficiency threshold. Assume that the quality threshold is 0.85 and the efficiency threshold is 0.9. In the above case, the processing quality is 0.878, which is higher than the quality threshold; the time efficiency index is 0.942, which is higher than the efficiency threshold. Therefore, the quality efficiency layer reward value is 1.0. If the processing quality is lower than the threshold, for example, 0.82, the difference is 0.03, then according to the exponential function e (-2×0.03) =0.942 to calculate the score; if the time efficiency is also lower than the threshold, such as 0.88, the difference is 0.02, then according to the exponential function e (-3×0.02) =0.941 to calculate the score, and multiply the two to get the quality efficiency layer reward value of 0.886.

[0044] The comprehensive reward value is calculated by multiplying the three reward values. Using the data from the above example, the reward value for the basic process constraint layer is 1.0, the reward value for the parameter coupling layer is 0.951, and the reward value for the quality efficiency layer is 1.0. The final comprehensive reward value is 1.0 × 0.951 × 1.0 = 0.951. This comprehensive reward value can be used as a basis for optimizing processing parameters in the ultrasonic cutting system. A higher value indicates a better parameter combination.

[0045] In practical applications, the system can adapt to different process requirements by adjusting various threshold and weight parameters. The system calculates reward values ​​in real time during processing and uses iterative optimization to find the optimal parameter combination, thereby achieving a balance between quality and efficiency in ultrasonic cutting of honeycomb cores. This comprehensive evaluation method can effectively guide the parameter optimization and decision-making process of the ultrasonic cutting system.

[0046] The design of reward functions in traditional honeycomb core ultrasonic cutting is mainly based on single or simple combinations of evaluation indicators. These methods ignore the coupling effects between parameters and multi-level process constraints, resulting in poor optimization results in practical applications. The three-tier reward evaluation system constructed by the present invention comprehensively considers process constraints, parameter coupling, and quality efficiency, forming a comprehensive evaluation mechanism that avoids the local optimality problem caused by single-indicator optimization. This mechanism ensures that the processing process not only ensures quality but also takes into account efficiency, and maintains cutting forces and stresses within a safe range, achieving global optimization of processing parameters and improving overall processing performance.

[0047] In an optional embodiment, the step of generating an initial tool path trajectory based on the region division result and the initial processing parameters includes: Obtaining stress distribution data, geometric feature data, and processing difficulty data for each region in the region division result, and constructing the stress distribution data, geometric feature data, and processing difficulty data into a regional feature vector; analyzing the stress gradient, stress distribution, and curvature change in the regional feature vector to obtain a regional complexity evaluation result; A trajectory point distribution strategy is determined based on the regional complexity evaluation result, trajectory point arrangements are increased in the region where the stress gradient exceeds a preset gradient threshold, the tool posture is adjusted in the region where the curvature change exceeds a preset curvature change threshold, and the feed speed is optimized in the region where the stress distribution exceeds a preset stress distribution threshold; the trajectory point distribution strategy is combined with the initial machining parameters to form a trajectory point sequence that satisfies the six-axis motion constraints; the trajectory point sequence is converted into a continuous trajectory segment, and a transition trajectory is generated at the trajectory mutation position; the continuous trajectory segment and the transition trajectory are combined to form a complete initial tool path trajectory.

[0048] For example, the stress distribution data, geometric feature data and processing difficulty data of each area in the area division result are obtained, and these data are constructed into a regional feature vector. Specifically, the stress distribution data includes the principal stress value, stress direction and stress concentration coefficient of each point; the geometric feature data includes the surface curvature, normal vector and connection characteristics of adjacent areas; the processing difficulty data includes the honeycomb core wall thickness distribution, unit arrangement density and honeycomb core contact stiffness estimation value. For a typical honeycomb core workpiece area, the data that can be collected include: the maximum stress in the area is 25MPa, the minimum is 5MPa, the average stress gradient is 1.2MPa / mm; the maximum surface curvature is 0.35mm -1 , the minimum curvature is 0.05mm -1 The honeycomb core wall thickness is 0.6mm, and the cell density is 8 per 100mm².

[0049] After the regional feature vector is constructed, the system will analyze the stress gradient, stress distribution and curvature change in the vector to obtain the regional complexity evaluation results. Stress gradient analysis is achieved by calculating the stress change rate between adjacent sampling points. When the stress change rate between two points exceeds 1.5MPa / mm, the area is marked as a high stress gradient area; stress distribution analysis is evaluated by statistically analyzing the degree to which stress deviates from the average value in the area. When the stress value of any point in the area deviates from the average value by more than 30%, the area is marked as an area with uneven stress distribution; curvature change analysis is performed by calculating the change rate of the surface curvature of the area. When the curvature change rate exceeds 0.1mm -1 / mm, the area is marked as a high curvature change area. By combining these indicators, the system divides the regional complexity into five levels: very low, low, medium, high, and very high.

[0050] Based on the regional complexity evaluation results, the system determines the trajectory point distribution strategy. In areas where the stress gradient exceeds a preset gradient threshold (e.g., 1.8MPa / mm), the system increases the trajectory point layout density from the standard 2 points / mm to 5 points / mm to ensure accurate machining of high stress gradient areas. In areas where the curvature change exceeds a preset curvature change threshold (e.g., 0.12mm -1 / mm), the system adjusts the tool's posture to maintain the optimal contact angle between the ultrasonic cutting tool and the honeycomb core surface (typically between 75° and 85°) to optimize cutting conditions and reduce honeycomb edge damage. In areas where the stress distribution exceeds a preset stress distribution threshold (for example, the ratio of maximum to minimum stress in a region is greater than 3), the system optimizes the feed speed from the standard 100mm / min to 60mm / min to ensure cutting quality.

[0051] By combining the trajectory point distribution strategy with the initial machining parameters, the system generates a trajectory point sequence that satisfies the six-axis motion constraints. This process takes into account the dynamic characteristics of the ultrasonic cutting machine, such as the maximum speed of each axis (10m / min for the X, Y, and Z axes and 20rpm for the A, B, and C axes), maximum acceleration (1m / s² for the X, Y, and Z axes and 2rad / s² for the A, B, and C axes), and jitter limits (10m / s³ for the X, Y, and Z axes and 15rad / s³ for the A, B, and C axes). Using an interpolation algorithm, the system generates a trajectory point sequence that satisfies these constraints, with a typical trajectory point spacing of 0.2mm in complex areas and 0.8mm in simple areas.

[0052] After the trajectory point sequence is generated, the system converts it into a continuous trajectory segment. This process uses the cubic B-spline fitting method to connect the discrete trajectory points into a smooth curve. The fitting accuracy is controlled within ±0.01mm to ensure cutting accuracy. At the location of trajectory mutation, such as the change of honeycomb core structure or the unit arrangement conversion area, the system will generate a transition trajectory. The transition trajectory usually adopts arc transition or spiral transition, and the transition radius is set between 0.5mm and 2mm according to the characteristics of the honeycomb core. For example, when the tangential angle between adjacent trajectory segments exceeds 30°, the system will automatically insert an arc transition trajectory with a radius of 1mm to make the ultrasonic cutting tool movement smooth transition.

[0053] Finally, the system combines the continuous trajectory segments and transition trajectories to form a complete initial toolpath trajectory. The complete trajectory is stored in a standard G-code format, including spatial position coordinates (X, Y, Z) and tool attitude angles (A, B, C), along with process parameters such as feed rate (F), spindle speed (S), ultrasonic amplitude and frequency. For a honeycomb core workpiece of medium complexity, the generated G-code file size is typically between 3MB and 8MB and contains thousands of control points. The initial toolpath trajectory generated by this method can fully adapt to the cutting requirements of different areas of the honeycomb core workpiece, providing a good foundation for subsequent toolpath optimization.

[0054] In actual application, a 300mm x 200mm honeycomb core panel was divided into 12 processing zones. System analysis revealed four areas with high stress gradients, three with high curvature variations, and two with uneven stress distribution. By applying this method, the system successfully generated a toolpath totaling 25 meters in length, comprising 12,500 points, 85 continuous paths, and 42 transition paths. This fully captures the honeycomb core cutting characteristics of each zone, ensuring both quality and efficiency.

[0055] This invention achieves a precise match between machining trajectories and material properties. It employs differentiated trajectory strategies for areas of varying complexity, increasing trajectory point density in high-stress areas and optimizing tool posture in high-curvature areas. This significantly improves machining accuracy for complex contours while simultaneously reducing vibration and impact caused by sudden trajectory changes through transition trajectory processing.

[0056] In an optional embodiment, a digital twin model of a honeycomb core structure is established, cutting force, vibration spectrum, and temperature field distribution data of a current processing area are collected in real time, the processing state of the digital twin model is updated, the processing state is compared with the stress and strain prediction data, the deviation between the actual processing state and the predicted state is calculated, the processing parameters of the current position are dynamically optimized based on the deviation, and the initial tool path trajectory is compensated. The steps of generating a compensated tool path trajectory include: Generate the initial state of the digital twin model based on the geometric parameters of the honeycomb core structure and the material characteristic parameters; calculate the square root of the sum of the squares of the cutting force in three directions to obtain the synthetic cutting force, and calculate the inverse tangent of the cutting force in the Y direction and the X direction to obtain the cutting force direction angle; perform Fourier transform on the vibration spectrum to extract the main frequency band and bandwidth; reconstruct the temperature field distribution data based on the spatial basis function and the time response function; construct the synthetic cutting force, cutting force direction angle, main frequency band, bandwidth and reconstructed temperature field data into a real-time state vector; input the real-time state vector into a Kalman filter to obtain the processing state of the digital twin model; Calculating the Euclidean distance between the processing state and the stress-strain prediction data to obtain a deviation; An adaptive adjustment coefficient is calculated based on the deviation, a parameter adjustment amount is obtained based on the adaptive adjustment coefficient and the deviation, and a normal compensation amount and a tangential compensation amount are calculated based on the parameter adjustment amount; the current machining position on the initial tool path trajectory is offset by the normal compensation amount along the normal unit vector and by the tangential compensation amount along the tangential unit vector to generate a compensated tool path trajectory.

[0057] Figure 3 This is the architecture diagram of the dynamic compensation and trajectory optimization system for the digital twin model of the honeycomb core structure. For example, to establish the initial state of the digital twin model of the honeycomb core structure, the system first obtains the geometric parameters of the honeycomb core structure, including the honeycomb unit size, wall thickness, and overall size; the material property parameters include elastic modulus, Poisson's ratio, density of 0.12g / cm³, yield strength of 65MPa, and thermal conductivity of 0.8W / (m·K). The system uses finite element analysis software to establish a grid model of the honeycomb core structure, setting the grid unit size to 0.2mm and the total number of grids to approximately 1 million. The initial boundary conditions are set to have the fixed end constrained to one edge of the honeycomb structure, and the free edges to the other edges.

[0058] When collecting machining data in real time, a three-axis force sensor is used to collect cutting forces, with a sampling frequency set to 10kHz. For the collected cutting force data in the X, Y, and Z directions, the system calculates the square root of the sum of the squares of the components in the three directions to obtain the composite cutting force. For example, when the cutting force in the X direction is 15N, the cutting force in the Y direction is 12N, and the cutting force in the Z direction is 8N, the calculated composite cutting force is 20.88N. Simultaneously, the inverse tangent of the cutting forces in the Y and X directions is calculated to obtain the cutting force direction angle. Under these data conditions, the cutting force direction angle is 38.7 degrees.

[0059] For vibration signals, the system uses a triaxial accelerometer to collect vibration data at a sampling frequency of 20kHz. The collected time-domain vibration signals are converted to frequency-domain signals using a fast Fourier transform. The system then identifies the dominant frequency band and bandwidth within the spectrum. In actual cases, the dominant frequency band is concentrated in the 18kHz-22kHz range, with a dominant frequency of 20kHz and a bandwidth of 4kHz, corresponding to the operating frequency of ultrasonic cutting tools.

[0060] The temperature field data is collected by an infrared thermal imager. However, due to the large amount of raw data and the noise, the system uses spatial basis functions and time response functions to reconstruct the data. In the specific implementation, the collected temperature field data is first subjected to singular value decomposition to extract 10 main spatial basis functions {φ1(x,y),φ2(x,y),...,φ 10 (x, y)}, these basis functions reflect the main spatial patterns of the temperature distribution in the cutting area. At the same time, by performing Fourier analysis on the temperature time series, four main time response functions {ψ1(t), ψ2(t), ψ3(t), ψ4(t)} are extracted, corresponding to constant response, linear change, exponential decay and periodic fluctuation respectively. The temperature field reconstruction expression is T(x, y, t) = Σ ij a ij φ i (x,y)ψ j (t), where a ij is the weight coefficient of each combination. For example, the temperature distribution in the cutting area reaches a maximum of 65°C at the contact point, and then decreases gradually outwards, with the temperature dropping to 45°C at 5mm from the contact point and to 35°C at 10mm from the contact point.

[0061] The system constructs a real-time state vector based on the synthetic cutting force of 20.88N, the cutting force direction angle of 38.7 degrees, the vibration dominant frequency of 20kHz, the bandwidth of 4kHz, and the reconstructed temperature field data. This state vector is then fed into a Kalman filter for processing. The process noise covariance matrix of the filter is set to a diagonal matrix with a diagonal element value of 0.05, and the observation noise covariance matrix is ​​set to a diagonal matrix with a diagonal element value of 0.02. After Kalman filtering, the updated digital twin model processing state is obtained, including the estimated stress distribution, strain distribution, and displacement field.

[0062] The system compares the updated machining state with the predicted stress and strain data. The predicted data shows a maximum equivalent stress of 35 MPa and a maximum strain of 0.004 at the current machining position. The deviation is calculated by calculating the Euclidean distance, which in the actual case is 8 MPa.

[0063] Based on the deviation, the system calculates an adaptive adjustment coefficient. When the deviation is within the 0-5 MPa range, the adjustment coefficient is 0.8; when the deviation is within the 5-10 MPa range, the adjustment coefficient is 1.2; and when the deviation is greater than 10 MPa, the adjustment coefficient is 1.5. In this example, the deviation is 8 MPa, and the adjustment coefficient is 1.2.

[0064] The system calculates the parameter adjustment based on the adaptive adjustment coefficient and the deviation. The parameter adjustment is equal to the deviation multiplied by the adjustment coefficient. In this example, it is 8 MPa × 1.2 = 9.6 MPa. Based on the parameter adjustment, the system calculates the normal compensation and tangential compensation. The normal compensation is equal to the parameter adjustment multiplied by 0.7, and the tangential compensation is equal to the parameter adjustment multiplied by 0.3. In this example, the normal compensation is 9.6 MPa × 0.7 = 6.72 MPa, which corresponds to a displacement of 0.15 mm, and the tangential compensation is 9.6 MPa × 0.3 = 2.88 MPa, which corresponds to a displacement of 0.06 mm.

[0065] The system obtains the current machining position coordinates (25.5mm, 40.2mm, 5.0mm), the normal unit vector (0.866, 0.5, 0), and the tangential unit vector (-0.5, 0.866, 0) on the initial toolpath. The current position is offset by 0.15mm along the normal unit vector and 0.06mm along the tangential unit vector, resulting in the compensated position coordinates of (25.63mm, 40.33mm, 5.0mm). The system sequentially processes all points on the initial toolpath to generate a complete compensated toolpath.

[0066] The present invention uses a digital twin model to reflect the processing status in real time, realizing the synchronous interaction between the physical world and the digital space; multi-source sensor data fusion improves the accuracy of state estimation, and deviation analysis enables the system to timely detect and correct processing deviations; the normal and tangential bidirectional compensation strategy ensures trajectory accuracy, improves the processing surface quality, and effectively copes with the nonlinear deformation of honeycomb core materials during the processing process.

[0067] In an optional embodiment, the step of obtaining a parameter adjustment amount based on the adaptive adjustment coefficient and the deviation, and calculating a normal compensation amount and a tangential compensation amount based on the parameter adjustment amount includes: Calculating the proportional term, integral term, and differential term based on the adaptive adjustment coefficient, and weighting the proportional term, integral term, and differential term in each direction to obtain a feed speed adjustment amount, a spindle speed adjustment amount, and a cutting depth adjustment amount; Converting the feed speed adjustment, spindle speed adjustment and cutting depth adjustment into normal compensation and tangential compensation; Collect machining surface profile data, cutting force fluctuation data and energy consumption data, take the deviation of the machining surface profile data from the target profile as the accuracy evaluation index, take the deviation of the cutting force fluctuation data from the set threshold as the stability evaluation index, and take the ratio of the energy consumption data to the benchmark energy consumption as the efficiency evaluation index; construct a compensation optimization target based on the accuracy evaluation index, stability evaluation index and efficiency evaluation index; construct a meta-learning network, the meta-learning network includes a task adaptation module for optimizing the compensation strategy and a strategy optimization module for online adjustment of the compensation amount; input the compensation optimization target into the task adaptation module to obtain the optimization parameters of the current machining condition, and adjust the normal compensation amount and the tangential compensation amount through the strategy optimization module based on the optimization parameters; adopt an iterative optimization method to update the parameters of the meta-learning network, and output the updated meta-learning network as the final normal compensation amount and the tangential compensation amount.

[0068] For example, key adjustment parameters are derived by calculating the relationship between the adaptive adjustment coefficient and the proportional, integral, and differential terms. Specifically, the adaptive adjustment coefficient can be expressed as a function of the rate of change of the deviation monitored in real time during the machining process. For example, when the rate of change of the deviation exceeds a preset threshold, the adaptive adjustment coefficient will increase according to a predetermined rule, and vice versa. For example, when the detected surface profile deviation rate of change exceeds 5% / second, the adaptive adjustment coefficient is adjusted from the default value of 1.0 to 1.2 to more quickly respond to machining anomalies.

[0069] After obtaining the adaptive adjustment coefficient, the system calculates the proportional, integral, and differential terms in each direction. The proportional term is directly proportional to the current deviation, the integral term accumulates historical deviations, and the differential term reflects the changing trend of the deviation. For the X-direction proportional term, if the current profile deviation is 0.1mm and the adaptive adjustment coefficient is 1.2, the X-direction proportional term is 0.12mm. For the adjustment terms in all three directions, the system performs a weighted summation using preset weights to determine the feed rate adjustment, spindle speed adjustment, and depth of cut adjustment. For example, the feed rate adjustment may be composed of a proportional term weighted by 0.5 in the X direction, an integral term weighted by 0.3 in the Y direction, and a differential term weighted by 0.2 in the Z direction. When these terms are 0.12 mm, 0.08 mm, and -0.05 mm, respectively, the calculated feed rate adjustment is (0.12 × 0.5) + (0.08 × 0.3) + (-0.05 × 0.2) = 0.06 + 0.024 - 0.01 = 0.074 mm / s, corresponding to 4.44 mm / min.

[0070] Feed rate adjustments, spindle speed adjustments, and depth of cut adjustments need to be converted to normal and tangential compensations. This conversion is based on the honeycomb core geometry and the planned cutting tool path. The normal direction is perpendicular to the honeycomb core wall, and the tangential direction is along the core edge. If the feed rate adjustment is 4.44 mm / min, the spindle speed adjustment is 200 rpm, and the depth of cut adjustment is 0.05 mm, the conversion matrix yields a normal compensation of 0.12 mm and a tangential compensation of 0.08 mm.

[0071] The accuracy of the compensation amount is crucial to the cutting quality, so the system has established a multi-dimensional evaluation mechanism. The system collects the cutting surface contour data and compares it with the target contour. For example, if the target contour is 0.05mm and the actual measured value is 0.08mm, the accuracy evaluation index is 0.03mm. At the same time, the system monitors the cutting force fluctuation data and compares it with the set threshold to obtain a stability evaluation index. In one case, the cutting force fluctuation was 25N, and the set threshold was 20N, so the stability evaluation index was 5N. The ratio of energy consumption data to the benchmark energy consumption constitutes an efficiency evaluation index. For example, if the actual energy consumption is 0.8 kWh and the benchmark energy consumption is 0.75 kWh, the efficiency evaluation index is 1.07.

[0072] Based on the three evaluation metrics above, the system constructs a comprehensive compensation optimization objective. This objective function can be expressed as a weighted sum of the three metrics, for example, assigning a weight of 0.5 to the accuracy evaluation metric, 0.3 to the stability evaluation metric, and 0.2 to the efficiency evaluation metric. In the above example, the compensation optimization objective value is (0.03 × 0.5) + (5 × 0.3) + (1.07 × 0.2) = 0.015 + 1.5 + 0.214 = 1.729.

[0073] To adapt to the characteristics of different cutting tasks, the system constructs a meta-learning network with a two-stage architecture design, including a task adaptation module and a policy optimization module, which can quickly learn and adjust the compensation strategy from limited samples.

[0074] The task adaptation module adopts a four-layer fully connected neural network structure. The input layer contains 8 neurons, which receive accuracy evaluation indicators, stability evaluation indicators, efficiency evaluation indicators and honeycomb core structural characteristics (such as wall thickness, unit density, etc.). The first hidden layer contains 64 neurons and uses the LeakyReLU activation function with an activation parameter of 0.2. The second hidden layer contains 32 neurons and also uses the LeakyReLU activation function. The output layer contains 4 neurons, corresponding to the normal coefficient, tangential coefficient, frequency adjustment factor and amplitude adjustment factor respectively. It uses the Sigmoid activation function and is scaled to obtain output values ​​in a suitable range. A Transformer encoder is also integrated in the task adaptation module to process historical cropping data sequences. The encoder contains 3 attention heads and a hidden dimension of 32, which is used to capture common features between different working conditions and realize knowledge transfer.

[0075] For example, when the input compensation optimization target value is 1.729, it is first normalized (divided by the preset normalization factor of 2.5) and then passed to the task adaptation module along with other features. After forward propagation, the output of the first hidden layer undergoes batch normalization, and 25% dropout is added to the output of the second hidden layer to prevent overfitting. Finally, the task adaptation module outputs a normal coefficient of 1.15 and a tangential coefficient of 0.85 as optimization parameters.

[0076] The policy optimization module is based on a dual-Q network architecture and consists of two neural networks with identical structures to reduce the bias in Q-value estimation. The module's input layer receives 16 features, including the output parameters of the task adaptation module and the current cutting state vector (including cutting force, vibration frequency, temperature, etc.). The first hidden layer contains 48 neurons and uses the ELU activation function. The second hidden layer contains 24 neurons and also uses the ELU activation function. The output layer generates the final compensation adjustment policy, including precise normal and tangential compensation coefficients. The policy optimization module also integrates a Monte Carlo sampling method. By adding small Gaussian noise to the network parameters and performing 10 forward propagations, it calculates the uncertainty of the decision and adjusts the exploration-exploitation balance parameters based on the degree of uncertainty.

[0077] The original normal compensation of 0.12mm was adjusted through the strategy optimization module, and after a coefficient of 1.15, the final compensation amount was 0.138mm. Similarly, the tangential compensation was adjusted from 0.08mm to 0.068mm after a coefficient of 0.85. When decision uncertainty is high (the coefficient of variation is greater than 0.15), the system will reduce the absolute value of the compensation coefficient to avoid over-adjustment.

[0078] The meta-learning network is trained using Model-Agnostic Meta-Learning (MAML). The training process consists of two phases: an inner loop and an outer loop. The inner loop uses a stochastic gradient descent optimizer for rapid adaptation to the specific cropping task, with a learning rate of 0.01 and a five-step gradient update using 8 samples at a time. The outer loop updates the meta-parameters using the Adam optimizer with a learning rate of 0.001, momentum parameters beta1 of 0.9, and beta2 of 0.999, with an outer loop update performed every 200 samples.

[0079] The loss function consists of three components: accuracy loss (the root mean square of the profile error), stability loss (the absolute value of the difference between the cutting force fluctuation and the threshold), and energy efficiency loss (the square of the difference between the energy consumption ratio and the target value). The weights of these three components are 0.5, 0.3, and 0.2, respectively. In addition, an L2 regularization term with a coefficient of 0.0001 is added to prevent overfitting.

[0080] The system maintains an experience replay buffer with a capacity of 10,000 samples, storing state-action-reward triplets. During each training session, a prioritized sampling strategy is used to select 128 samples from the buffer for batch update, with the priority calculated based on the temporal difference error. The network is initialized using the He initialization method, with all biases initialized to zero.

[0081] In actual application, when the normal and tangential compensation amounts were 0.138mm and 0.068mm, respectively, the contour after cutting was found to improve by 25% (from 0.08mm to 0.06mm), and the cutting force fluctuation was reduced by 20% (from 25N to 20N), but energy consumption increased by 3% (from 0.8 kWh to 0.824 kWh). The system calculated the total reward value as 25×0.5+20×0.3-3×0.2=12.5+6-0.6=17.9, and used this feedback to update the network parameters.

[0082] The system also implements an adaptive temperature parameter mechanism, initially using a higher temperature parameter (2.0) to encourage exploration, then gradually reducing it to 0.5 as learning progresses to enhance strategy utilization. When the cumulative number of samples reaches 5,000, the temperature parameter is reduced to 1.0, and when it reaches 8,000, it is further reduced to 0.5.

[0083] After multiple iterative optimizations, the meta-learning network adjusted the normal compensation to 0.142 mm and optimized the tangential compensation to 0.064 mm, which were used as the final compensation amounts for cutting control.

[0084] Traditional compensation methods primarily rely on offline simulation and fixed parameter control. These methods struggle to adapt to local property variations and complex boundary conditions when processing heterogeneous materials like honeycomb cores, leading to problems such as stress concentration, edge tearing, and dimensional deviation. The meta-learning network of this invention enables adaptive optimization of compensation strategies, enabling rapid adaptation to diverse processing conditions and avoiding the limitations of traditional fixed compensation strategies. A multi-dimensional evaluation system comprehensively considers accuracy, stability, and efficiency to ensure optimal compensation. An iterative learning mechanism enables the system to continuously improve from historical processing experience, improving the accuracy and robustness of compensation and offering greater adaptability than traditional compensation methods.

[0085] Figure 4 This diagram compares the effects of meta-learning network compensation optimization on six-axis ultrasonic cutting of honeycomb cores. In terms of surface profile improvement, traditional PID compensation achieved a baseline improvement of 60.0%, initial meta-learning compensation increased this to 70.0%, and after iterative optimization, meta-learning compensation achieved a significant improvement of 81.5%. This data directly demonstrates the outstanding performance of the meta-learning network in improving machining accuracy, especially when processing complex geometries such as honeycomb core structures.

[0086] The reduction rate of cutting force fluctuations also showed a gradual improvement: 59.0% for the traditional method, 67.4% for the initial meta-learning approach, and 76.0% after iterative optimization. This demonstrates that the optimized meta-learning network can more effectively reduce mechanical fluctuations during machining, improve machining stability, and effectively avoid potential tearing or deformation of the honeycomb core material during cutting.

[0087] The difference in energy efficiency improvement was particularly significant: while traditional PID compensation achieved only a 43.3% improvement, initial meta-learning compensation increased that to 50.2%, and after iterative optimization, meta-learning compensation reached 60.8%. This fully demonstrates the advantages of meta-learning networks in multi-objective optimization, enabling them to significantly reduce energy consumption while maintaining machining quality.

[0088] The chart clearly demonstrates the gradient improvement of the three compensation methods across all evaluation metrics, with the iteratively optimized meta-learning compensation achieving the best performance across all metrics. This deep learning-based compensation method adaptively adjusts machining parameters, optimizing toolpaths in real time based on the honeycomb core material's characteristics and machining requirements.

[0089] According to a second aspect of an embodiment of the present invention, an electronic device is provided, including: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0090] According to a third aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0091] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0092] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A control method for dynamic compensation and trajectory optimization of six-axis ultrasonic cutting tool paths for honeycomb cores, characterized in that: include: Obtain material characteristic parameters and cutting contour lines of the honeycomb core structure; Based on deep reinforcement learning, a tool-material interaction model is established, the material characteristic parameters and cutting contour line are input into the interaction model, the curvature change and tangent angle of the cutting contour line are analyzed, and the stress and strain prediction data of the honeycomb core structure at different contour feature positions are generated in combination with the arrangement of the honeycomb units. The entire processing area is divided into multiple feature areas based on the stress and strain prediction data, and differentiated initial processing parameters are set for different areas. The initial processing parameters include tool feed speed, spindle speed, ultrasonic amplitude and frequency; generating an initial tool path trajectory based on the area division result and the initial processing parameters; Establishing a digital twin model of the honeycomb core structure, collecting cutting force, vibration spectrum, and temperature field distribution data of the current processing area in real time, updating the processing state of the digital twin model, comparing the processing state with the stress and strain prediction data, calculating the deviation between the actual processing state and the predicted state, dynamically optimizing the processing parameters of the current position based on the deviation, compensating the initial tool path trajectory, and generating a compensated tool path trajectory; Processing is performed according to the compensated tool path trajectory, and actual processing data is fed back to the digital twin model in real time to continuously optimize processing parameters until the entire processing process is completed.

2. The method according to claim 1, characterized in that Based on deep reinforcement learning, an interaction model between a tool and a material is established, the material characteristic parameters and the cutting contour line are input into the interaction model, the curvature change and the tangent angle of the cutting contour line are analyzed, and the stress and strain prediction data of the honeycomb core structure at different contour feature positions are generated in combination with the arrangement of the honeycomb units. The entire processing area is divided into multiple feature areas according to the stress and strain prediction data, and differentiated initial processing parameters are set for different areas, wherein the initial processing parameters include tool feed speed, spindle speed, ultrasonic amplitude and frequency. The steps include: Calculating the curvature value and tangent angle of the cutting contour line to obtain the wall thickness distribution data of the honeycomb core structure and the arrangement data of the honeycomb units; Constructing a state space of the interaction model, wherein the state space includes a material state vector constructed based on the material characteristic parameters and a geometric characteristic vector constructed based on the curvature value and tangent angle, wall thickness distribution data, and arrangement data; Constructing an action space of the interaction model, wherein the action space includes a tool feed speed adjustment amount, a spindle speed adjustment amount, an ultrasonic amplitude adjustment amount, and a frequency adjustment amount; Inputting the material state vector and the geometric characteristic vector into the interactive model to predict stress and strain data; Calculating a stress gradient value based on the stress-strain data, and dividing the processing area into a plurality of characteristic areas according to the stress gradient value; for each characteristic area, calculating a tool feed speed based on the curvature value, calculating a spindle speed based on the arrangement data, calculating an ultrasonic amplitude based on the stress gradient value, and calculating the frequency based on the wall thickness distribution data and the arrangement data, to obtain differentiated initial processing parameters; A comprehensive reward value is calculated based on the cutting force, machining quality parameters, and machining time, and the comprehensive reward value is fed back to the interactive model for online learning to obtain optimized initial machining parameters.

3. The method according to claim 2, characterized in that The steps of calculating the tool feed rate based on the curvature value, calculating the spindle speed based on the arrangement data, calculating the ultrasonic amplitude based on the stress gradient value, and calculating the frequency based on the wall thickness distribution data and the arrangement data include: Calculating a curvature change rate of each point on the cutting contour line, and determining a discrete feed speed for each point based on the curvature change rate; when the curvature change rate is less than a first preset threshold, the discrete feed speed takes a maximum feed speed value; when the curvature change rate is greater than the first preset threshold and less than a second preset threshold, the discrete feed speed decreases according to a quadratic function with the curvature change rate; when the curvature change rate is greater than the second preset threshold, the discrete feed speed takes a minimum feed speed value; when the curvature gradient between adjacent points exceeds the second preset threshold, performing a smooth transition process on the discrete feed speed to obtain a continuous feed speed curve; Calculating the matching relationship between the ultrasonic period and the spindle speed based on the arrangement data of the honeycomb units, and taking the frequency multiplication value of the reference spindle speed and the arrangement data as the spindle speed; A stress segmentation value is calculated based on the maximum and minimum values ​​of the stress gradient value, the stress gradient value is divided into three segments based on the stress segmentation value, and the ultrasonic amplitude is adjusted to a corresponding preset value according to the position of the segment where the stress gradient value is located; The square root of the sum of the squares of the changes in the wall thickness distribution data in the horizontal and vertical directions is calculated to obtain the spatial variation rate, the arrangement period is determined according to the arrangement data, the natural frequency of the processing system is used as the reference frequency, and the frequency value is obtained based on the reference frequency, the spatial variation rate and the arrangement period.

4. The method according to claim 2, characterized in that The steps for calculating the comprehensive reward value based on cutting force, machining quality parameters and machining time include: The synthetic cutting force and cumulative stress gradient during the machining process are calculated. The ratio of the actual material removal volume to the theoretical material removal volume per unit time is calculated to obtain the time efficiency index. The weighted average of the local surface roughness and the local surface integrity score is obtained to obtain the machining quality. The comparison result between the synthetic cutting force and the force threshold, and the comparison result between the cumulative stress gradient and the stress threshold are used as the basic process constraint layer reward value, and when the corresponding thresholds are exceeded, the basic process constraint layer reward value is calculated based on the difference using an exponential function; Calculate the ratio of the product of feed speed and amplitude to its reference value, and the ratio of the product of frequency and amplitude to its reference value, and use the weighted exponential sum of the differences between the two ratios and the reference value as the parameter coupling layer reward value; The comparison result between the processing quality and the quality threshold, and the comparison result between the time efficiency index and the efficiency threshold are used as the quality efficiency layer reward value, and when the corresponding thresholds are exceeded, the quality efficiency layer reward value is calculated based on the difference using an exponential function; The basic process constraint layer reward value, the parameter coupling layer reward value, and the quality efficiency layer reward value are multiplied together to obtain a comprehensive reward value.

5. The method according to claim 1, characterized in that Based on the area division result and the initial processing parameters, the step of generating an initial tool path trajectory includes: Obtaining stress distribution data, geometric feature data, and processing difficulty data for each region in the region division result, and constructing the stress distribution data, geometric feature data, and processing difficulty data into a regional feature vector; analyzing the stress gradient, stress distribution, and curvature change in the regional feature vector to obtain a regional complexity evaluation result; A trajectory point distribution strategy is determined based on the regional complexity evaluation result, trajectory point arrangements are increased in the region where the stress gradient exceeds a preset gradient threshold, the tool posture is adjusted in the region where the curvature change exceeds a preset curvature change threshold, and the feed speed is optimized in the region where the stress distribution exceeds a preset stress distribution threshold; the trajectory point distribution strategy is combined with the initial machining parameters to form a trajectory point sequence that satisfies the six-axis motion constraints; the trajectory point sequence is converted into a continuous trajectory segment, and a transition trajectory is generated at the trajectory mutation position; the continuous trajectory segment and the transition trajectory are combined to form a complete initial tool path trajectory.

6. The method according to claim 1, wherein The steps of establishing a digital twin model of a honeycomb core structure, collecting cutting force, vibration spectrum, and temperature field distribution data of a current processing area in real time, updating the processing state of the digital twin model, comparing the processing state with the stress and strain prediction data, calculating the deviation between the actual processing state and the predicted state, dynamically optimizing the processing parameters of the current position according to the deviation, and compensating the initial tool path trajectory to generate a compensated tool path trajectory include: Generate the initial state of the digital twin model based on the geometric parameters of the honeycomb core structure and the material characteristic parameters; calculate the square root of the sum of the squares of the cutting force in three directions to obtain the synthetic cutting force, and calculate the inverse tangent of the cutting force in the Y direction and the X direction to obtain the cutting force direction angle; perform Fourier transform on the vibration spectrum to extract the main frequency band and bandwidth; reconstruct the temperature field distribution data based on the spatial basis function and the time response function; construct the synthetic cutting force, cutting force direction angle, main frequency band, bandwidth and reconstructed temperature field data into a real-time state vector; input the real-time state vector into a Kalman filter to obtain the processing state of the digital twin model; Calculating the Euclidean distance between the processing state and the stress-strain prediction data to obtain a deviation; An adaptive adjustment coefficient is calculated based on the deviation, a parameter adjustment amount is obtained based on the adaptive adjustment coefficient and the deviation, and a normal compensation amount and a tangential compensation amount are calculated based on the parameter adjustment amount; the current machining position on the initial tool path trajectory is offset by the normal compensation amount along the normal unit vector and by the tangential compensation amount along the tangential unit vector to generate a compensated tool path trajectory.

7. The method according to claim 6, characterized in that The steps of obtaining a parameter adjustment amount based on the adaptive adjustment coefficient and the deviation, and calculating a normal compensation amount and a tangential compensation amount based on the parameter adjustment amount include: Calculating the proportional term, integral term, and differential term based on the adaptive adjustment coefficient, and weighting the proportional term, integral term, and differential term in each direction to obtain a feed speed adjustment amount, a spindle speed adjustment amount, and a cutting depth adjustment amount; Converting the feed speed adjustment, spindle speed adjustment and cutting depth adjustment into normal compensation and tangential compensation; Collect machining surface profile data, cutting force fluctuation data and energy consumption data, take the deviation of the machining surface profile data from the target profile as the accuracy evaluation index, take the deviation of the cutting force fluctuation data from the set threshold as the stability evaluation index, and take the ratio of the energy consumption data to the benchmark energy consumption as the efficiency evaluation index; construct a compensation optimization target based on the accuracy evaluation index, stability evaluation index and efficiency evaluation index; construct a meta-learning network, the meta-learning network includes a task adaptation module for optimizing the compensation strategy and a strategy optimization module for online adjustment of the compensation amount; input the compensation optimization target into the task adaptation module to obtain the optimization parameters of the current machining condition, and adjust the normal compensation amount and the tangential compensation amount through the strategy optimization module based on the optimization parameters; adopt an iterative optimization method to update the parameters of the meta-learning network, and output the updated meta-learning network as the final normal compensation amount and the tangential compensation amount.

8. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 7.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images