An adaptive numerical control milling method and system based on roughness online prediction

Abstract

The application discloses a kind of self-adapting numerical control milling method and system based on roughness online prediction.The method comprises the following steps: collecting high-frequency three-dimensional vibration acceleration signals and three-phase stator current signals through sensors rigidly fixed to the spindle;Reading the machine tool processing parameters in the current numerical control system, which include the current kinematic data;Band-pass filtering the high-frequency three-dimensional vibration acceleration signals, and extracting time-domain and frequency-domain characteristic quantities to generate a transient feature vector;Substitute the transient feature vector and the current kinematic data into the polynomial support vector roughness prediction model to calculate the predicted surface roughness value;Construct a model predictive control objective function that includes a roughness tracking error penalty term, a cutting parameter control increment square penalty term, and a nonlinear chatter penalty term;Minimize the objective function and output the optimal milling feed speed and spindle speed correction instructions, and control the feed axis and spindle of the machine tool to perform milling on the workpiece according to the correction instructions.The present application can avoid chatter sensitive frequency, ensure cutting stability in harsh environments and reduce scrap rate.

Description

Technical Field

[0001] This application relates to the field of CNC milling process control and intelligent manufacturing equipment, and specifically to an adaptive CNC milling method and system based on online roughness prediction. Background Technology

[0002] In CNC milling of aerospace structural components and high-precision molds, workpiece surface roughness is a core indicator for measuring machining quality. Surface roughness is affected by a combination of time-varying factors, including tool wear, material hardness fluctuations, and machine tool vibration. Currently, surface roughness is typically obtained using offline contact stylus meters or white light interferometers. However, this offline inspection method suffers from significant lag; once roughness exceeds tolerances, the workpiece is often already scrap.

[0003] To achieve real-time monitoring during machining, some conventional technologies incorporate optical cameras to capture images of the machined surface online. However, in actual CNC milling environments, the spray of high-pressure cutting fluid and the scattering of metal chips directly obstruct the optical lenses, causing conventional vision measurement systems to malfunction. Furthermore, some solutions employ proportional-integral-differential algorithms to adjust the machine tool's feed rate based on a single measured cutting force signal, thereby indirectly controlling surface roughness. However, the relationship between surface roughness and cutting force is not a simple linear one. Simply adjusting the feed rate can easily disrupt the stability of the cutting process system, causing high-frequency chatter, which in turn leads to severe chatter marks on the machined surface, further deteriorating machining quality. Therefore, a closed-loop surface roughness optimization method that can adapt to harsh, obstructed environments and stably control dynamic quality is urgently needed. Summary of the Invention

[0004] This application provides an adaptive CNC milling method and system based on online roughness prediction.

[0005] This application provides an adaptive CNC milling method based on online roughness prediction, including:

[0006] The high-frequency triaxial vibration acceleration signal and the three-phase stator current signal are respectively collected by the acceleration sensor and the current sensor of the spindle driver, which are rigidly fixed to the spindle housing of the machine tool.

[0007] The machine tool machining parameters, which include current kinematic data, are read from the current CNC system with a machine tool low-level interpolator via an industrial protocol.

[0008] The high-frequency triaxial vibration acceleration signal is bandpass filtered using an edge computing gateway, and the time-domain and frequency-domain features of the filtered vibration signal and the three-phase stator current signal are extracted and spliced ​​to generate a transient feature vector.

[0009] The transient feature vector and the current kinematic data are substituted into the multinomial support vector roughness prediction model to calculate the predicted surface roughness value of the workpiece at the current moment.

[0010] A model predictive control objective function is constructed, which includes a roughness tracking error penalty term, a squared penalty term for the control increment of cutting parameters, and a nonlinear chatter penalty term based on the machine tool system stability lobe diagram.

[0011] In each discrete control cycle, the model predictive control objective function is minimized and the optimal milling feed rate and optimal spindle speed correction command are output and sent to the machine tool bottom interpolator. The machine tool's feed axis and spindle are then controlled to perform milling on the workpiece according to the correction command.

[0012] This application also provides an adaptive CNC milling system based on online roughness prediction, characterized by comprising: a CNC milling machine execution end, including a machine tool spindle housing, a spindle driver, and a machine tool bottom interpolator; a sensor acquisition end, including an acceleration sensor rigidly fixed to the machine tool spindle housing and a current sensor disposed at the power supply line of the output end of the spindle driver; and an edge computing control end, communicatively connected to the CNC milling machine execution end and the sensor acquisition end, the edge computing control end including an edge computing gateway; the edge computing gateway is configured to execute the above method to output optimal milling feed rate and optimal spindle speed correction commands to the machine tool bottom interpolator, and control the CNC milling machine execution end to perform adaptive milling machining on the workpiece.

[0013] The technical solution provided in this application brings the following beneficial effects: By collecting the underlying physical characteristics such as the three-dimensional high-frequency vibration signal and stator current signal of the machine tool spindle, and using a multinomial support vector regression model for composite mapping prediction, it eliminates the dependence of visual inspection on an unobstructed optical environment, and achieves stable and high-precision full-processing cycle prediction even in harsh scenarios with splashing cutting fluid. Furthermore, by utilizing a model predictive control algorithm with a control increment squared penalty term and a nonlinear chatter penalty term, while smoothly adjusting cutting parameters and approximating the target roughness, it actively imposes safety space constraints on control commands, avoiding chatter-sensitive frequencies of the machining system, ensuring the dynamic stability of the cutting process, and reducing the workpiece scrap rate. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the overall structure of the adaptive CNC milling system based on online roughness prediction provided in the embodiments of this application.

[0015] Figure 2 This is a block diagram of the logic structure of the control device provided in the embodiments of this application.

[0016] Figure 3 This is a flowchart of the adaptive CNC milling method based on online roughness prediction provided in the embodiments of this application.

[0017] Figure 4 This is a flowchart of the feature extraction steps provided in the embodiments of this application.

[0018] Explanation of reference numerals in the attached figures:

[0019] In the diagram: 101-Machine tool spindle housing, 102-Acceleration sensor, 103-Spindle driver, 104-Current sensor, 105-Machine tool bottom-level interpolator, 106-Edge computing gateway, 107-High-frequency A / D data acquisition card, 108-Anti-aliasing filter conditioning circuit module, 109-Wide voltage isolation power supply module, 201-Processor, 202-Memory, 203-Communication interface. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0021] like Figure 1 As shown, this application provides an adaptive CNC milling system based on online roughness prediction. This system forms the hardware network foundation for implementing the method of this application, mainly including a CNC milling machine execution end, a sensor acquisition end, and an edge computing control end. The CNC milling machine execution end includes a machine tool spindle housing 101, a spindle driver 103, and a machine tool bottom-level interpolator 105. To achieve high-precision capture of heterogeneous bottom-level physical signals, the sensor acquisition end is equipped with an accelerometer 102 and a current sensor 104. The accelerometer 102 is a piezoelectric triaxial accelerometer with a frequency response range covering 0.5Hz to 10kHz and an IP67 protection rating, enabling stable operation in harsh oil and moisture environments. The accelerometer 102 is rigidly fixed to the front labyrinth seal ring of the machine tool spindle housing 101 by bolts. The current sensor 104 is a through-hole Hall current sensor, directly inserted into the three-phase power supply line at the output end of the spindle driver 103.

[0022] On the edge computing control side, the system is equipped with an industrial gateway running a real-time operating system as the edge computing gateway 106. The edge computing gateway 106 is equipped with an Ethernet interface, enabling low-latency data handshake with the open CNC system containing the machine tool interpolator 105 via the industrial Ethernet protocol. Furthermore, to ensure that analog signals are converted into digital signals that the edge computing gateway 106 can interpret, a high-frequency A / D data acquisition card 107 and an anti-aliasing filtering and conditioning circuit module 108 are connected between the sensors and the gateway. The system is also equipped with a wide-voltage isolated power supply module 109 to provide a stable operating voltage for the edge computing gateway 106 and each sensor. It should be noted that the hardware models, communication protocols, and frequency response parameters described above are merely examples. Those skilled in the art can select components with equivalent functions according to the needs of the actual industrial site, as long as they can achieve the acquisition and real-time calculation of high-frequency physical signals.

[0023] like Figure 3 As shown, based on the aforementioned physical hardware architecture, this application embodiment provides an adaptive CNC milling method based on online roughness prediction, mainly executed by the edge computing gateway 106, and specifically includes the following steps:

[0024] In step S301, the high-frequency triaxial vibration acceleration signal and the three-phase stator current signal are respectively collected by the acceleration sensor and the current sensor of the spindle driver, which are rigidly fixed to the spindle housing of the machine tool.

[0025] During milling, an accelerometer 102 (i.e., a triaxial piezoelectric sensor) fixed to the machine tool spindle housing 101 along a three-dimensional orthogonal coordinate system continuously picks up the radial and axial vibrations of the spindle and outputs high-frequency triaxial vibration acceleration signals. Simultaneously, a current sensor 104 extracts the cutting load current of the spindle motor and generates a three-phase stator current signal. To ensure the temporal consistency of heterogeneous multi-source data, an edge computing gateway 106 triggers data alignment and acquisition between the triaxial piezoelectric sensor and the Hall current sensor by sending a hardware clock synchronization signal. By directly acquiring the underlying electromechanical physical signals near the cutting area, high-frequency dynamic changes can be captured at the moment of mechanical deformation and tool wear.

[0026] Step S302: Read the machine tool machining parameters in the current CNC system with the machine tool bottom interpolator through the industrial protocol. The machine tool machining parameters include the current kinematic data.

[0027] The edge computing gateway 106 establishes a network communication connection with the current CNC system of the CNC milling machine execution end via the industrial Ethernet protocol. It periodically reads the running register data inside the CNC system using a preset communication data block address. After address matching and data parsing, the edge computing gateway 106 extracts the nominal feed rate, nominal spindle speed, and depth of cut corresponding to the current timestamp, and merges and encapsulates these state variables as the current kinematic data. Direct access to the registers via the industrial protocol ensures that the acquisition of the reference kinematic parameters is not affected by external sensor errors.

[0028] Step S303: The high-frequency triaxial vibration acceleration signal is bandpass filtered using the edge computing gateway, and the time-domain and frequency-domain features of the filtered vibration signal and the three-phase stator current signal are extracted and spliced ​​to generate a transient feature vector.

[0029] like Figure 4 As shown, the specific process of feature dimensionality reduction and fusion includes:

[0030] In step S401, the edge computing gateway 106 acquires the current spindle speed and calculates the corresponding spindle rotation fundamental frequency constant based on the speed. The high-frequency triaxial vibration acceleration signal is then truncated based on this spindle fundamental frequency. The edge computing gateway 106 internally controls a bandpass filter to perform bandpass filtering within the first to second multiples of the spindle fundamental frequency band. The first multiple is configured as 1, and the second multiple as 15, thereby forcibly controlling the bandpass filter to retain only the vibration frequency components between 1 and 15 times the spindle fundamental frequency.

[0031] In step S402, for the filtered vibration signal and the three-phase stator current signal held in the time domain, the edge computing gateway 106 performs statistical extraction, calculating the root mean square (RMS) value, skewness feature, peak value of the frequency domain power spectral density, mean, and variance. During this extraction process, a sliding time window is used to segment the filtered vibration signal into frames. Within each sliding time window, the RMS value of the time-domain sample and the third-order central moment data (used as the skewness feature) are calculated independently. The time-domain samples within the sliding window are converted to the frequency domain using a fast Fourier transform, and the peak value of the frequency domain power spectral density is searched and extracted near the specific tooth-passing frequency.

[0032] Step S403 involves assembling and connecting the root mean square value, skewness feature, peak value, mean, and variance of the frequency domain power spectral density acquired from different sensing channels according to a preset dimensional order. This embodiment employs a rule prioritizing current channel features over vibration channel features to perform dimensionality reduction and combination, outputting a multidimensional array containing multiple row vectors as the transient feature vector. Through bandpass filtering and concatenation of multi-domain statistics, irregular electromagnetic background noise in CNC machine tools can be filtered out, extracting a high-density feature combination that represents the micro-cutting state of the cutting edge.

[0033] Step S304: Substitute the transient feature vector and the current kinematic data into the multinomial support vector roughness prediction model to calculate the predicted surface roughness value of the workpiece at the current moment.

[0034] The edge computing gateway 106 extracts the feed per tooth and tool parameters from the current kinematic data, and calculates and generates kinematic reference feature terms to characterize the macroscopic cutting geometry using kinematic formulas. Simultaneously, it extracts transient feature vectors to generate electromechanical incremental feature terms to characterize microscopic system fluctuations. Using a preset polynomial kernel function, the electromechanical incremental feature terms and preset reference state vectors are mapped to a high-dimensional feature space. Nonlinear support vector regression is then performed, and the error compensation term is obtained by solving the weighted combination product of the corresponding support vectors. The kinematic reference feature terms and error compensation term are then weighted and summed according to adaptive weight allocation to output the predicted surface roughness value.

[0035] Specifically, regarding the kernel operation mechanism of the polynomial support vector roughness prediction model, during the model training phase, a support vector machine classification hyperplane is constructed offline, and a grid search method is used for cross-validation to determine the degree constant, displacement term constant, and scale factor of the preset polynomial kernel function. During the real-time online prediction phase, the kernel mapping between the input electromechanical incremental feature terms and the preset baseline state vector is processed using the determined degree constant, displacement term constant, and scale factor. The obtained Lagrange multiplier vectors are numerically filtered, and training sample points corresponding to values ​​greater than zero are extracted as core support vectors. The edge computing gateway 106 calculates the mapping product between the input feature vector and each core support vector under the preset polynomial kernel function, sums it according to the preset bias constant, and outputs the error compensation term for the current control cycle.

[0036] Specifically, the system performs calculations on the polynomial support vector roughness prediction model:

[0037]

[0038] Among them, R a (t) represents the predicted surface roughness value at time t, with a reference range typically from 0.1 micrometers to 3.2 micrometers; f z(t) represents the current feed per tooth, with an example value of 0.12 mm / tooth; r e This indicates the radius of the tool tip arc; the example value is 0.8 mm. V i (t) represents the value of each element in the extracted transient feature vector; V base This indicates that the tool is in a wear-free and healthy state, and the pre-calibrated and extracted preset reference state vector is used. Represents a polynomial kernel function; represents the weight coefficient of the i-th feature obtained by support vector regression, and its value range is usually between -1.5 and 1.5; w1 and w2 are adaptively assigned weights, and the sum of the two is equal to 1, for example, 0.3 and 0.7 respectively.

[0039] Step S305: Construct the model predictive control objective function, which includes a roughness tracking error penalty term, a cutting parameter control increment squared penalty term, and a nonlinear chatter penalty term based on the machine tool system stability lobe diagram.

[0040] Inside the edge computing gateway 106, to ensure that control commands do not drastically change, the square weighted average of the deviation between the surface roughness setpoint (target roughness reference value, e.g., 0.8 micrometers) and the predicted surface roughness value output in step S304 is configured as a roughness tracking error penalty term. The square weighted average of the difference between the current CNC system output cutting parameters and the command parameters issued in the previous control cycle is configured as a cutting parameter control increment squared penalty term. Based on the coordinate positions of the cutting depth parameters and spindle speed parameters in the preset machine tool system stability lobe diagram, the relative distance between the current state point and the boundary line of the chatter instability region is calculated, and an inverse proportional function about this relative distance is constructed as a nonlinear chatter penalty term. The above three terms are superimposed and combined to form the model predictive control objective function used for optimization evaluation.

[0041] Furthermore, the specific construction and constraint logic of the nonlinear chatter penalty term is as follows: The system analyzes the current kinematic data in real time to obtain the instantaneous spindle speed and instantaneous cutting depth, and maps the instantaneous spindle speed and instantaneous cutting depth to a machine tool system stability leaf-shaped diagram with the spindle speed as the horizontal axis and the cutting depth as the vertical axis. The edge computing gateway 106 identifies the boundary envelope of the cutting stability domain in the machine tool system stability leaf-shaped diagram and calculates the shortest Euclidean distance from the current state point to the normal projection of the boundary envelope of the cutting stability domain. An exponential barrier function is used to construct the penalty weight coefficient, and the value of the penalty weight coefficient is configured to increase exponentially as the shortest Euclidean distance decreases. The penalty weight coefficient is multiplied by a preset chatter penalty constant to output the nonlinear chatter penalty term. By forcibly rejecting the solution space that approximates the boundary envelope of the cutting stability domain during the minimization process through the nonlinear chatter penalty term, it is ensured that the final output command is always within the cutting stability region.

[0042] Specifically, the system executes the calculation of the predictive control objective function of the model:

[0043]

[0044] Wherein, J(k) represents the objective function cost calculated in discrete control period k; P represents the rolling prediction step size, with a reference value range of 8 to 15 periods, preferably 10 in this embodiment; M represents the control step size, with a reference value range of 2 to 5 periods, preferably 3 in this embodiment; This represents the tracking error weighting coefficient used to constrain the roughness approximation; the example value is 0.85. R represents the roughness prediction estimate for the i-th future step at time k; ref Indicates the target roughness; and These represent the weighting coefficients for the feed and speed control increments, respectively, with example values ​​of 0.6 and 0.4. and These represent the control increments of the feed per tooth and the spindle speed over two consecutive control cycles, respectively. This represents the nonlinear chatter penalty term. According to the safety constraint logic, when the predicted spindle speed n(k) falls into the chatter instability region of the leaflet diagram, It is forced to be assigned a maximum constant that is exponentially amplified (e.g., 99999); when n(k) is in the absolutely stable region, The value is zero. By penalizing the rate of change of the square of the parameter and prohibiting crossing the chatter boundary line, this model can effectively prevent the algorithm from issuing aggressive commands that exceed the physical limits of the servo motor in order to quickly eliminate errors.

[0045] Step S306: Minimize the model predictive control objective function in each discrete control cycle, output the optimal milling feed rate and optimal spindle speed correction command and send it to the machine tool bottom interpolator, and control the machine tool's feed axis and spindle to perform milling on the workpiece according to the correction command.

[0046] The edge computing gateway 106 invokes its internal quadratic programming algorithm to iteratively search for a sequence of parameters that minimizes the model's predictive control objective function. Once optimization is complete, the first term of the sequence is extracted as the optimal milling feed rate and optimal spindle speed correction command for the current cycle. The edge computing gateway 106 then transmits this correction command back to the machine tool's bottom-level interpolator 105 via the fieldbus. The machine tool's bottom-level interpolator 105 subsequently adjusts the servo motor's operating cycle, completing the physical modification of the parameters. Thus, from physical signal perception to algorithm decision-making and hardware control execution, a complete and secure closed-loop physical control link for CNC milling is formed.

[0047] Based on the above structure and steps, the working principle of the system in this embodiment is summarized as follows: During the milling task performed by the machine tool, the wide-voltage isolated power supply module 109 continuously provides isolated power to each node. The accelerometer 102 and the current sensor 104 synchronously acquire electromechanical high-frequency signals under the trigger of a hardware clock. The raw signal is filtered for noise by the anti-aliasing filtering conditioning circuit module 108 and converted into a digital sequence by the high-frequency A / D data acquisition card 107 and transmitted to the edge computing gateway 106. The edge computing gateway 106, combined with the macroscopic cutting parameters provided by the machine tool's bottom interpolator 105, extracts a multi-dimensional matrix covering frequency domain features and sends it to the support vector regression network to predict the future roughness trend. Based on this error trend, the model prediction control module calls the cost function embedded with the leaf lobe diagram boundary for rolling optimization, smoothly issuing collaborative adjustment instructions to the machine tool driver, thus achieving efficient control of the workpiece surface roughness while ensuring cutting stability.

[0048] like Figure 2 As shown, the control device can be physically implemented as the edge computing gateway 106 in the above embodiment, which internally contains core hardware for data processing and resource scheduling. The control device includes a processor 201 and a memory 202 that establishes a high-speed communication connection with the processor 201. The control device also includes a communication interface 203 for connecting to external sensor networks and machine tool control systems. The memory 202, as a non-volatile or volatile storage medium, stores executable computer programs. When the processor 201 calls and executes the computer program stored in the memory 202, it can control the interfaces to work collaboratively, thereby fully realizing the adaptive CNC milling method based on online roughness prediction in the above method embodiment.

[0049] Furthermore, this application embodiment also provides a computer-readable storage medium storing a computer program thereon. When executed by processor 201, the computer program can implement the adaptive CNC milling method based on online roughness prediction in the above embodiments. By employing a control device with dedicated processing capabilities as an edge computing node, combined with the instructions in the storage medium, the computing power overhead of the CNC machine tool's main processor can be reduced, ensuring real-time response capability for high-frequency data acquisition and control solution cycles.

Claims

1. An adaptive CNC milling method based on online roughness prediction, characterized in that, include: The high-frequency triaxial vibration acceleration signal and the three-phase stator current signal are respectively collected by the acceleration sensor and the current sensor of the spindle driver, which are rigidly fixed to the spindle housing of the machine tool. The machine tool machining parameters, which include current kinematic data, are read from the current CNC system with a machine tool low-level interpolator via an industrial protocol. The high-frequency triaxial vibration acceleration signal is bandpass filtered using an edge computing gateway, and the time-domain and frequency-domain features of the filtered vibration signal and the three-phase stator current signal are extracted and spliced ​​to generate a transient feature vector. Substitute the transient feature vector and the current kinematic data into the multinomial support vector roughness prediction model to calculate the predicted surface roughness value of the workpiece at the current moment. A model predictive control objective function is constructed, which includes a roughness tracking error penalty term, a cutting parameter control increment squared penalty term, and a nonlinear chatter penalty term based on the machine tool system stability lobe diagram; In each discrete control cycle, the model predictive control objective function is minimized and the optimal milling feed rate and optimal spindle speed correction command are output and sent to the machine tool bottom interpolator. The machine tool's feed axis and spindle are then controlled to perform milling on the workpiece according to the correction command.

2. The method as described in claim 1, characterized in that, The process of using an edge computing gateway to perform bandpass filtering on the high-frequency triaxial vibration acceleration signal, extracting the time-domain and frequency-domain features of the filtered vibration signal and the three-phase stator current signal, and concatenating them to generate a transient feature vector includes: The high-frequency triaxial vibration acceleration signal is truncated according to the spindle fundamental frequency, and bandpass filtering is performed within the range of the first to second multiple of the spindle fundamental frequency band through a bandpass filter; Statistical extraction is performed on the filtered vibration signal and the three-phase stator current signal in the time domain, and the root mean square value, skewness characteristics, peak value, mean and variance of the frequency domain power spectral density are calculated respectively. The root mean square value, the skewness feature, the peak value of the frequency domain power spectral density, the mean, and the variance are assembled and connected according to a preset dimension order, and a multidimensional array is output as the transient feature vector.

3. The method as described in claim 2, characterized in that, The step of truncating the high-frequency triaxial vibration acceleration signal based on the spindle fundamental frequency and performing bandpass filtering within the first to second multiples of the spindle fundamental frequency range includes: Obtain the current spindle speed and calculate the corresponding spindle rotation fundamental frequency constant; Configure the first multiplier to the value 1 and the second multiplier to the value 15, and control the bandpass filter to retain the vibration frequency components between 1 times the spindle fundamental frequency and 15 times the spindle fundamental frequency; In the process of performing statistical extraction on the filtered vibration signal and the three-phase stator current signal in the time domain, the root mean square value, skewness feature, peak value of frequency domain power spectral density, mean and variance are calculated respectively. The filtered vibration signal is segmented into frames using a sliding time window. In each sliding time window, the root mean square value of the time domain sample and the third-order central moment data as the skewness feature are calculated independently. The time domain sample is then converted to the frequency domain by a fast Fourier transform to search and extract the peak value of the frequency domain power spectral density. The root mean square value, skewness feature, peak value of frequency domain power spectral density, mean, and variance obtained from different sensing channels are combined in a dimensionality reduction manner according to the rule that the current channel feature comes first and the vibration channel feature comes last, to generate the transient feature vector containing multiple row vectors.

4. The method as described in claim 1, characterized in that, The constructed model predictive control objective function includes a roughness tracking error penalty term, a cutting parameter control increment squared penalty term, and a nonlinear chatter penalty term based on the machine tool system stability lobe diagram, including: The square weighted average of the deviation between the surface roughness setpoint and the predicted surface roughness value is configured as the roughness tracking error penalty term; The square weighted sum of the difference between the current CNC system output cutting parameters and the command parameters issued in the previous control cycle is configured as the squared penalty term for the cutting parameter control increment. Based on the coordinate positions of the cutting depth parameters and spindle speed parameters in the preset machine tool system stability lobe diagram, the relative distance between the current state point and the boundary line of the chatter instability region is calculated, and an inverse proportional function about the relative distance is constructed as the nonlinear chatter penalty term. The roughness tracking error penalty term and the squared penalty term of the cutting parameter control increment are superimposed and combined to form the model predictive control objective function.

5. The method as described in claim 4, characterized in that, The step involves calculating the relative distance between the current state point and the boundary line of the chatter instability region based on the coordinate positions of the cutting depth parameters and spindle speed parameters in a preset machine tool system stability flap diagram, and constructing an inverse proportional function of the relative distance as the nonlinear chatter penalty term, including: The instantaneous spindle speed and instantaneous cutting depth are obtained by real-time analysis of the current kinematic data; The instantaneous spindle speed and the instantaneous cutting depth are mapped to the machine tool system stability lobe diagram with the spindle speed as the horizontal axis and the cutting depth as the vertical axis; Identify the boundary envelope of the cutting stability domain in the stability lobe diagram of the machine tool system, and calculate the shortest Euclidean distance from the current state point to the normal projection of the boundary envelope of the cutting stability domain. An exponential barrier function is used to construct the penalty weight coefficient, and the value of the penalty weight coefficient is configured to increase exponentially as the shortest Euclidean distance decreases; The nonlinear chatter penalty term is obtained by multiplying the penalty weight coefficient by a preset chatter penalty constant. By forcibly rejecting the solution space that approximates the boundary envelope of the cutting stability region through the nonlinear chatter penalty term during the minimization process, the optimal milling feed rate and the optimal spindle speed correction command located within the cutting stability region are output.

6. The method as described in claim 1, characterized in that, The polynomial support vector roughness prediction model includes kinematic baseline feature terms and electromechanical incremental feature terms. The step of substituting the transient feature vector and the current kinematic data into the polynomial support vector roughness prediction model to calculate the predicted surface roughness value of the workpiece at the current moment includes: Extract the feed per tooth and tool parameters from the current kinematic data, and generate the kinematic baseline feature terms by calculating the kinematic formulas. Extract the transient feature vector to generate the electromechanical incremental feature term; The electromechanical incremental feature terms and the preset reference state vector are mapped to a high-dimensional feature space using a preset polynomial kernel function to perform nonlinear support vector regression calculation, and the error compensation term is obtained by solving the weight combination product of the corresponding support vectors. The kinematic reference feature term and the error compensation term are weighted and summed according to the adaptive weight allocation to obtain the predicted surface roughness value.

7. The method as described in claim 6, characterized in that, The step involves mapping the electromechanical incremental feature terms and the preset reference state vector to a high-dimensional feature space using a preset polynomial kernel function for nonlinear support vector regression calculation. The error compensation term is obtained by solving the weight combination product of the corresponding support vectors, including: During the model training phase, a support vector machine classification hyperplane is constructed, and a grid search method is used for cross-validation to determine the degree constant, displacement term constant, and scale factor of the preset polynomial kernel function. During the online prediction phase, the kernel mapping between the input electromechanical incremental feature term and the preset reference state vector is processed using the determined degree constant, displacement term constant, and scale factor. The obtained Lagrange multiplier vectors are filtered to extract the training sample points with values ​​greater than zero as core support vectors. The product of the input feature vector and each of the core support vectors under the preset polynomial kernel function is calculated, and the product is summed according to the preset bias constant to output the error compensation term for the current control cycle.

8. The method as described in claim 1, characterized in that, The method involves acquiring high-frequency triaxial vibration acceleration signals and three-phase stator current signals respectively through an acceleration sensor rigidly fixed to the machine tool spindle housing and a current sensor of the spindle driver, including: A triaxial piezoelectric sensor is fixed as the acceleration sensor in the spindle housing of the machine tool along the three-dimensional orthogonal coordinate system direction by a threaded connection, and the radial and axial vibrations of the spindle during the machining process are used as the high-frequency triaxial vibration acceleration signal. A Hall current sensor is inserted into the three-phase power supply line at the output end of the spindle driver as the current sensor, and the cutting load current of the spindle motor is extracted as the three-phase stator current signal. The triaxial piezoelectric sensor and the Hall current sensor are triggered to perform data alignment and acquisition by a hardware clock synchronization signal.

9. An adaptive CNC milling system based on online roughness prediction, characterized in that, include: The CNC milling machine execution end includes a machine tool spindle housing (101), a spindle driver (103), and a machine tool bottom interpolator (105); the sensor acquisition end includes an acceleration sensor (102) rigidly fixed to the machine tool spindle housing (101), and a current sensor (104) disposed at the power supply line of the output end of the spindle driver (103). An edge computing control terminal is communicatively connected to the CNC milling machine execution terminal and the sensor acquisition terminal. The edge computing control terminal includes an edge computing gateway (106). The edge computing gateway (106) is configured to execute the method as described in any one of claims 1 to 8, to output the optimal milling feed rate and the optimal spindle speed correction command to the machine tool bottom interpolator (105), and to control the CNC milling machine execution terminal to perform adaptive milling on the workpiece.

10. The system as described in claim 9, characterized in that, The edge computing control terminal also includes a high-frequency A / D data acquisition card (107) and an anti-aliasing filter conditioning circuit module (108) connected between the sensor acquisition terminal and the edge computing gateway (106). The high-frequency triaxial vibration acceleration signal acquired by the accelerometer (102) and the three-phase stator current signal acquired by the current sensor (104) are filtered for noise by the anti-aliasing filter conditioning circuit module (108) and converted into digital sequences by the high-frequency A / D data acquisition card (107) before being transmitted to the edge computing gateway (106).

Images