Numerical control equipment active control method and system based on physical information neural network

Abstract

This invention discloses an active control method and system for CNC equipment based on a physical information neural network, comprising: acquiring the dynamic parameters of the CNC equipment's end effector; establishing a time-delay differential dynamic equation and converting it into physical residual terms; constructing a physical information neural network, the input of which is a multidimensional feature tensor containing a physical time independent variable at a fixed time, a state parameter of the end effector at the fixed time, and historical time-delay state parameters, and the output is a predicted vibration displacement corresponding to the fixed time; training the network using a loss function including a physical constraint loss function; during online operation, generating a high-dimensional time-series feature vector and assigning the time independent variable therein to the future target time, and outputting an advanced prediction result through forward inference of the network; determining whether chatter has occurred based on the advanced prediction result, and if so, outputting a speed modulation command. This application can realize the active prediction and suppression of chatter in CNC equipment cutting.

Description

Technical Field

[0001] This invention relates to the field of active control technology for industrial electromechanical equipment, and specifically to an active control method and system for CNC equipment based on physical information neural networks. More specifically, it relates to a servo closed-loop anti-vibration control method and system for CNC equipment based on edge computing and physical information neural networks. Background Technology

[0002] High-end CNC equipment is the core foundation of modern precision manufacturing. During the operation of complex CNC machining systems, especially in high-precision machining scenarios involving multi-tooth rotating envelopes and workpieces with large length-to-diameter ratios, such as leadscrew milling, the end effector of the machine tool is subjected to extremely complex time-varying nonlinear loads. Due to the "regenerative feedback time-delay" effect commonly present in CNC cutting kinematics caused by the overlap of spatially adjacent cutting trajectory surfaces, the underlying drive loop of the CNC system is highly susceptible to inducing strong self-excited high-frequency vibrations (i.e., regenerative chatter).

[0003] From the perspective of comprehensive energy efficiency and green manufacturing in CNC electromechanical systems, strong self-excited vibration not only causes high-frequency ripples on the workpiece surface, reduces transmission accuracy, and leads to material waste, but also causes drastic high-frequency alternation of the load torque of the spindle servo drive system. To maintain the commanded speed under nonlinear excitation force, the spindle drive motor of the CNC system must frequently output huge transient peak currents. This directly causes severe nonlinear copper losses in the motor stator (…). Heat generation and eddy current losses in the iron core lead to a significant waste of reactive power in CNC equipment during a single-piece machining cycle. Simultaneously, continuous high-frequency alternating loads can cause microscopic chipping and early thermal fatigue failure in high-value cutting tools such as cemented carbide or PCBN inserts, significantly increasing the emission of recalcitrant heavy metal industrial solid waste.

[0004] To suppress vibrations and reduce energy consumption in CNC machining, traditional CNC vibration damping strategies typically rely on passively reducing the machine tool spindle speed or critical depth of cut. This conservative strategy sacrifices material removal rate (MRR), lengthens the single-piece machining cycle, and significantly increases the long-term standby power consumption of peripheral auxiliary systems such as CNC machine tool hydraulic stations and high-pressure cooling pumps.

[0005] In recent years, the industry has attempted to introduce data-driven deep learning models for cutting state prediction and vibration control. However, these data-driven black-box AI models are detached from the underlying physical and mechanical laws of CNC machining. When faced with complex fine-tuning of cutting conditions in CNC equipment, they are prone to overfitting and distortion, generating erroneous control commands that violate the law of conservation of physical energy. Furthermore, traditional algorithms incur enormous numerical integration computation costs when solving differential dynamic equations involving large deformations and nonlinear time delays, failing to meet the sub-millisecond requirements of high-end CNC systems for controlling high-frequency self-excited vibrations. The stringent timeliness requirements of online real-time electromagnetic closed-loop intervention.

[0006] Therefore, how to provide a method and system that can predict and actively adjust high-frequency self-excited vibrations during CNC cutting is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] In view of the above problems, this invention is proposed to provide an active control method and system for CNC equipment based on physical information neural networks to overcome or at least partially solve the above problems. By deeply integrating the rigor of the underlying time-delay physical partial differential equations with the efficiency of the feedforward neural network algorithm, and by introducing an artificial intelligence reasoning mechanism constrained by physical mechanics laws into the CNC underlying drive loop, the alternating cutting load is smoothed from the source of servo electrical control. This achieves sub-millisecond-level advanced prediction and active servo adjustment of high-frequency self-excited vibration during CNC cutting, thereby reducing the power consumption of the spindle motor, extending the service life of core consumable components, and realizing energy saving, efficiency improvement, and green low-carbon CNC precision machining.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] In a first aspect, embodiments of the present invention provide an active control method for CNC equipment based on a physical information neural network, comprising the following steps: Steps for constructing and training a physical information neural network: The dynamic parameters of the end effector of the CNC equipment are obtained, a time-delay differential dynamic equation including the cutting regeneration effect is established, and the time-delay differential dynamic equation is converted into a physical residual term. A physical information neural network is constructed. The input of the physical information neural network includes a multidimensional feature tensor containing a fixed time as the physical time independent variable, the state parameter of the execution terminal at the fixed time, and the historical time delay state parameter. The output is the predicted vibration displacement corresponding to the fixed time. The physical information neural network is trained using a loss function to obtain the trained physical information neural network. The loss function includes a physical constraint loss function constructed using the physical residual term. Active control steps of CNC equipment: During the operation of the CNC equipment, the current state parameters and system time delay constant of the execution end are acquired in real time. Based on the system time delay constant, the time delay state parameters are extracted from the historical state parameters. The current physical time independent variable, the current state parameters and the time delay state parameters are combined to generate a high-dimensional time series feature vector. The current physical time independent variable in the high-dimensional time series feature vector is assigned the value of the current physical time plus the future target time after a preset look-ahead time window, and the current physical time is taken as the fixed time. The assigned high-dimensional time series feature vector is input into the trained physical information neural network for forward inference, and the predicted vibration displacement amplitude of the execution end at the future target time is output as the advanced prediction result. Closed-loop active adjustment step: Based on the advanced prediction result, determine whether the CNC equipment will experience chatter. If so, output a speed modulation command to the CNC equipment.

[0010] Preferably, in the physical information neural network construction and training step, the time-delay differential dynamic equation is converted into a physical residual term. for:

[0011] in, The cutting force coefficient is expressed as a unit. Radial cutting depth; Theoretical chip thickness; This represents the current transient vibration disturbance at the tool tip. The physical time-delay displacement characteristics of the historical period; The physical hysteresis time constant; Equivalent modal mass of CNC equipment; The damping coefficient of the CNC equipment structure; For equivalent dynamic stiffness, For a fixed time.

[0012] Preferably, in the physical information neural network construction and training step, the physical constraint loss function is constructed using the physical residual term as follows:

[0013] in, This represents the total number of unlabeled physical points randomly sampled within the time domain. For the first The moment of each physical point.

[0014] Preferably, the loss function further includes a data-driven loss function, which is the mean square error between the predicted vibration displacement and the sensor-measured vibration displacement; the step of training the physical information neural network using the loss function includes: The physical constraint loss function and the data-driven loss function are weighted and combined to construct the total loss function; The physical information neural network is trained by backpropagation based on the total loss function to update the network weight parameters.

[0015] Preferably, in the active control step of the CNC equipment, the high-dimensional temporal feature vector is used for online prediction. include:

[0016] in, For a fixed time; For forward-looking time windows; To control the end-effector rotation speed; To execute the end-effector offset; This represents the current vibration displacement; The current vibration velocity; This represents the historical time-delay displacement.

[0017] Preferably, in the physical information neural network construction and training step, after obtaining the trained physical information neural network, the step further includes: Perform model pruning and / or weight quantization operations on the trained physical information neural network to generate a lightweight forward inference engine; The lightweight forward inference engine is embedded in the non-volatile storage medium of the edge computing node.

[0018] Preferably, in the closed-loop active adjustment step, the step of determining whether the CNC equipment will experience chatter based on the advanced prediction result includes: The advanced prediction results are compared with a preset safety threshold; If the predicted result exceeds the preset safety threshold, it is determined that the CNC equipment will experience chatter. If the predicted result does not exceed the preset safety threshold, the CNC equipment is determined to be in a stable cutting state.

[0019] Preferably, the step of outputting a speed modulation command to the CNC equipment includes: The spindle speed variation algorithm is triggered to generate a speed modulation command containing a periodic nonlinear frequency modulation signal, which controls the actual spindle speed of the CNC equipment to fluctuate periodically and nonlinearly around the reference speed, so that the time delay constant responds to the chatter.

[0020] In a second aspect, embodiments of the present invention provide an active control system for CNC equipment based on a physical information neural network, used to implement the active control method for CNC equipment based on a physical information neural network described in the first aspect, comprising: The sensing module is used to acquire real-time vibration signals and real-time spindle speed signals from the end effector during the operation of CNC equipment. The edge inference module internally deploys a physical information neural network trained using a loss function, which includes a physical constraint loss function constructed with physical residual terms. The physical residual terms are derived from a time-delay differential dynamics equation containing a cutting regeneration effect. The edge inference module receives data output from the perception module, generates the high-dimensional time-series feature vector, assigns the current physical time variable in the generated high-dimensional time-series feature vector to the current physical time plus a preset look-ahead time window for the future target time, inputs the assigned high-dimensional time-series feature vector into the physical information neural network for forward inference, outputs the predicted vibration displacement amplitude of the execution terminal at the future target time as the advanced prediction result, and compares the advanced prediction result with a preset safety threshold to determine whether speed modulation is triggered. If triggered, a speed modulation command is generated. The servo execution module is used to receive the speed modulation command and output the speed modulation command to the spindle driver of the CNC equipment. Preferably, the edge reasoning module includes: A real-time data processing unit is used to construct a dynamic circular queue in a cache to continuously store historical vibration displacement sequences with timestamps, and to extract historical time-delay displacements from the dynamic circular queue by reverse indexing according to the system time delay constant. The lightweight inference engine unit internally loads the physical information neural network, which, after receiving the assigned high-dimensional temporal feature vector, performs forward computation through matrix multiplication and nonlinear activation mapping to output the advanced prediction result. The control decision unit is used to compare the advanced prediction result with the preset safety threshold, and determine whether to trigger speed modulation based on the comparison result. If triggered, a speed modulation command is generated.

[0021] This invention, through the interdisciplinary integration of underlying physical dynamics constraints and artificial intelligence computing, provides an active control architecture with time-varying delay compensation for electromechanical systems, resulting in significant electrical energy savings, system efficiency improvements, and environmental friendliness. Specific beneficial effects include at least: I. The physics-driven control logic possesses extremely high reliability and generalization ability. Relying on the precise underlying time-delay physical equations of the electromechanical system as hard penalty constraints, this invention, based on a physical information neural network forward inference engine, overcomes the "black box" overfitting risk of traditional pure data-driven control algorithms. Even when faced with extremely complex and variable time-varying loads and nonlinear operating condition disturbances in electromechanical systems, it can still maintain extremely high accuracy in physical state prediction and control command output performance, possessing industrial-grade field application and cross-condition generalization capabilities.

[0022] II. Direct Electrical Energy Saving and Significant Improvement in Power Efficiency of CNC Spindle Servo End. This invention is based on the underlying drive control of CNC machine tools. Through frequency domain stability leaflet diagram (SLD) analysis, it is confirmed that after introducing PINN-based spindle speed variation (SSV) advance intervention, the absolute stability boundary of the machine tool cutting system across the entire frequency band is significantly broadened. This closed-loop adjustment mechanism instantly smooths out the extremely nonlinear high-frequency alternating cutting resistance in CNC machining at the sub-millisecond level. The stable cutting state significantly suppresses the large reactive current pulses frequently output by the CNC machine tool spindle servo drive due to resistance to severe vibration, effectively reducing the nonlinear copper loss of the spindle motor stator winding (…). (Heat generation) and core eddy current losses. Through simulation verification using electromechanical system dynamics and force-electric coupling models, this CNC adjustment method significantly reduces the peak transient active power of the machine tool when performing high-intensity machining tasks, significantly reducing the overall power consumption of machining a single workpiece and fulfilling the direct electrical energy-saving requirements of high-end CNC equipment.

[0023] III. Breakthroughs in CNC Machining Efficiency and Source Reduction of Solid Waste from High-Value Cutting Tools. Thanks to the expanded absolute stability boundary of the CNC machine tool by the control system, theoretical dynamic models confirm that the machine tool effectively overcomes the conservative settings of traditional anti-vibration strategies that force a reduction in cutting depth or feed rate. Without self-excited chatter, the system allows for a considerable increase in the set critical cutting parameters, leading to a significant increase in material removal rate (MRR). Simultaneously, the physical blocking of high-frequency chatter by the CNC system significantly improves the micro-stress environment in the tool-chip contact area, greatly preventing micro-chipping and early thermal fatigue damage of carbide inserts and tools. The exponentially extended service life of high-value consumable tools not only ensures extremely high yield rates in machine tool processing but also substantially reduces the emission of recalcitrant, heavy metal-containing industrial solid waste at the source, fully aligning with the green CNC manufacturing and energy-saving and environmentally friendly industrial orientation. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0025] Figure 1 A flowchart illustrating the active control method for CNC equipment based on a physical information neural network provided in an embodiment of the present invention; Figure 2 This invention provides an architecture diagram of an active control system for CNC equipment based on a physical information neural network. Figure 3The simulation comparison diagram of the cutting chatter advance prediction response based on PINN is provided for the embodiments of the present invention; Figure 4 The simulation curve of the absolute prediction error in the entire time domain is provided for the embodiments of the present invention; Figure 5 This is a schematic diagram of the broadening of the stability leaflet diagram (SLD) boundary provided in an embodiment of the present invention. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] This invention specifically discloses an active control method for CNC equipment based on a physical information neural network, such as... Figure 1 As shown, it includes the following steps: Steps for constructing and training a physical information neural network: Obtain the dynamic parameters of the end effector of the CNC equipment, establish the time-delay differential dynamic equation including the cutting regeneration effect, and convert the time-delay differential dynamic equation into physical residual terms; A physical information neural network is constructed. The input of the physical information neural network includes: a multidimensional feature tensor with a fixed time as the physical time independent variable, the state parameter of the execution end at the fixed time, and the historical time delay state parameter at the fixed time. The output is the predicted vibration displacement corresponding to the fixed time. The physical information neural network is trained using a loss function to obtain the trained physical information neural network. The loss function includes a physical constraint loss function constructed with physical residual terms. Active control steps of CNC equipment: During the operation of CNC equipment, the current state parameters and system time delay constant of the execution end are acquired in real time. Based on the system time delay constant, the time delay state parameters are extracted from the historical state parameters. The current physical time independent variable, the current state parameters and the time delay state parameters are combined to generate a high-dimensional time series feature vector. The current physical time independent variable in the high-dimensional time series feature vector is assigned the value of the current physical time plus the future target time after a preset look-ahead time window, and the current physical time is taken as the fixed time. The assigned high-dimensional time series feature vector is input into the trained physical information neural network for forward inference, and the predicted vibration displacement amplitude of the execution end at the future target time is output as the advanced prediction result. Closed-loop active adjustment steps: Based on the advance prediction results, determine whether the CNC equipment will experience chatter. If so, output a speed modulation command to the CNC equipment.

[0028] Based on the same inventive concept, combined with Figure 2 As shown, this embodiment of the invention also provides an active control system for CNC equipment based on a physical information neural network (PIN) based active control method for CNC equipment. The technical features of this system are also applicable to the execution steps of the PIN based active control method for CNC equipment. The system includes: The sensing module is used to acquire real-time vibration signals and real-time spindle speed signals at the end of the actuator during the operation of CNC equipment using sensors. The edge inference module internally deploys a physical information neural network trained using a loss function, which includes a physical constraint loss function constructed with physical residuals. The physical residuals are derived from the time-delay differential dynamics equations that include the cutting regeneration effect. The edge inference module receives data output from the sensing module, generates a high-dimensional time-series feature vector, and assigns the physical time independent variable in the generated high-dimensional time-series feature vector to the current physical time plus a preset look-ahead time window for the future target time. The assigned high-dimensional time-series feature vector is then input into the physical information neural network for forward inference. The predicted vibration displacement amplitude at the execution end at the future target time is output as the advanced prediction result. The advanced prediction result is compared with a preset safety threshold to determine whether speed modulation is triggered. If triggered, a speed modulation command is generated. Among them, a dynamic circular queue can be maintained in the high-speed cache of the edge computing node to continuously store the historical vibration displacement sequence, generate physical residual terms using the dynamic parameters of CNC equipment calculated by the sensing module, and perform forward inference using the PINN inference engine. The servo execution module is used to receive the speed modulation command through the controller and output the speed modulation command to the spindle driver of the CNC equipment through the driver.

[0029] In one embodiment, the edge inference module includes: The real-time data processing unit is used to build a dynamic circular queue in the cache to continuously store historical vibration displacement sequences with timestamps, and to extract historical time-delay displacements from the dynamic circular queue by reverse indexing according to the system time delay constant. The lightweight inference engine unit internally loads a physical information neural network, which, after receiving the assigned high-dimensional temporal feature vector, performs forward computation through matrix multiplication and nonlinear activation mapping to output advanced prediction results. The control decision unit is used to compare the advanced prediction results with the preset safety threshold, and determine whether to trigger speed modulation based on the comparison results. If triggered, a speed modulation command is generated.

[0030] It should be noted that the embodiments of the present invention block the regenerative feedback time delay effect in multi-tooth cutting through forward inference and transient nonlinear modulation of the underlying servo. This method has broad applicability to CNC equipment. For example, this method / system is also applicable to side milling of large overhang end mills in conventional multi-axis linkage CNC machining centers: it can acquire signals through an accelerometer arranged in the spindle box, and use the number of cutting edges of the end mill as a parameter to input the time delay constant into a circular queue, and then issue a speed change command to the spindle driver through the OPC UA interface inside the CNC system before milling chatter occurs.

[0031] In order to demonstrate the superior control capability and significant electrical energy saving effect of the present invention when dealing with extremely complex, high-frequency alternating nonlinear loads, the following selected a lead screw cyclone milling system, which is very representative in CNC machine tool processing and has complex dynamic boundaries, as a preferred embodiment for detailed verification and explanation.

[0032] S1: Multimodal perception and edge control architecture construction: This embodiment constructs a time-delay closed-loop control system on a multi-axis CNC lathe equipped with a cyclone milling power head.

[0033] Hardware physical configuration: Select the GCr15 bearing steel ball screw blank with a large length-to-diameter ratio as the machining object, and evenly install multiple coated rotating cutting edges on the inner ring of the execution end (cyclone milling cutter head).

[0034] High-frequency sensing closed loop: A wideband piezoelectric accelerometer is rigidly mounted on the base shell of the cyclone milling cutter head and the support of the lead screw center frame. The sampling rate of the data acquisition card is set to [value missing]. It captures high-frequency alternating vibration electrical signals at the end of the system without distortion.

[0035] Communication and Command Bus: Sensor data, after analog-to-digital conversion, is transmitted in real-time via industrial fiber optic cable to an industrial edge AI computing node equipped with a GPU module and a real-time operating system. This edge node establishes a two-way handshake with the machine tool's CNC underlying servo ring via OPC UA and EtherCAT control bus protocols. It can not only read the current spindle speed in real time... It can control feed speed and coordinate position, and can also bypass conventional CNC logic to directly send control commands downwards. The hardware delay of sending commands to the driver is controlled within [a certain range]. Within.

[0036] S2: Real-world identification and hardware-level dynamic boundary solidification of physical parameters in cyclone milling: This step aims to identify the mechanical evolution process of cyclone milling, which includes multi-tooth envelope characteristics, online through the machine tool's own physical sensing hardware, and solidify it into the physical addressing register parameters at the bottom layer of the edge controller, providing a deterministic hardware boundary for the physical residual calculation of subsequent control logic.

[0037] S21: Hardware-driven online identification of static dynamic parameters: With the workpiece and cyclone milling power head assembled and the spindle stationary, the transfer function of the end effector of the machine tool spindle-tool system is first identified. A broadband pulse excitation force is applied in the main cutting direction of the tool tip using an impact hammer with a built-in piezoelectric sensor. The analog voltage excitation signal output by the impact hammer, and the time-domain vibration simulation response signal collected by a triaxial piezoelectric accelerometer rigidly mounted on the spindle bearing housing, are simultaneously connected to an external industrial edge computing node via a high-frequency shielded cable. A high-speed synchronous A / D acquisition card is used. Within the real-time operating system (RTOS) of the edge node, the onboard high-speed DSP chip performs a Fast Fourier Transform (FFT) on the digitized excitation and response signals to calculate the real and imaginary parts of the frequency response function (FRF) at the tool tip. Utilizing a multi-degree-of-freedom orthogonal polynomial fitting algorithm, the equivalent modal mass of the spindle-tool system under the principal vibration modes is accurately identified. Structural damping coefficient With equivalent dynamic stiffness After identification, the edge computing node writes the above parameters into the non-volatile memory (EEPROM) or specific physical addressing parameter register of the numerical control system (CNC).

[0038] S22: Bus demodulation of multi-tooth regenerative cutting force model and physical time delay constant: Cyclone milling exhibits typical multi-tooth regenerative cutting characteristics. This is because the inner ring of the forming cutter head is evenly distributed with... The time difference between two adjacent teeth cutting through the same spatial position of the workpiece under the combined action of high-speed rotation and axial feed of the PCBN insert and cutter head is the physical time delay constant. Industrial edge computing nodes utilize high-speed industrial Ethernet fieldbus (EtherCAT) to achieve sub-millisecond (ms) speeds. The polling cycle reads the encoder feedback register inside the spindle servo driver in real time, and demodulates the actual operating speed of the current tool turret spindle in real time. Based on physical formulas The unit is milliseconds. The hysteresis time constant of the machining system is defined as a time-varying coupled function determined by the spindle speed command and the number of teeth involved in cutting. This macroscopic dynamic law serves as the underlying hard constraint for subsequent control algorithms. The microprocessor of the edge controller calculates the corresponding physical hysteresis time constant in real time within the hardware arithmetic logic unit (ALU). Under this cycloidal envelope kinematic constraint, the nominal instantaneous chip thickness of a single PCBN tool is... The precise analytical formula is:

[0039] In the formula, This is the nominal feed rate per tooth. The axial feed rate register value is sent in real time by the interpolation algorithm of the CNC system. This is the real-time rotational phase angle calculated with the workpiece center as the reference after kinematic coordinate system transformation.

[0040] Furthermore, the actual physical displacement of the mechanical actuator end is incorporated into the cutting geometry to construct a multi-tooth regenerative dynamic cutting force. Mechanical mapping relationship at the servo motor end:

[0041] In the formula, The unit cutting force coefficient for a specific material (GCr15) stored in non-volatile memory; The radial depth of cut specified by the CNC machining program (G-code); The actual dynamic chip thickness is determined by the theoretical thickness. The current transient vibration disturbance of the tool tip is acquired by a high-frequency acquisition card and output through a hardware integration algorithm. And the physical time-delay displacement characteristics of historical periods extracted from the reverse index of the dynamic circular buffer queue. It is formed by the superposition of multiple nonlinearities.

[0042] S23: Physical residual hardware mapping of time-delay differential dynamics equations: Combining Newton's second law with the above dynamic cutting force model, the delay differential equation (DDE) for the end effector of the spindle-tool system is obtained:

[0043] In order to incorporate this physical equation as a hard penalty constraint into the subsequent edge controller calculations, the edge microprocessor rearranges terms of the equation and constructs physical residual terms in the control register. :

[0044] This residual equation constitutes the physical conservation boundary for the controller to determine the stability of the system's phase trajectory and calculate the active anti-vibration modulation. It represents the time-delay physical law of regenerative chatter caused by the overlapping surface ripples left on the workpiece surface by the two consecutive cuts. By monitoring the convergence and divergence of this physical residual, the controller directly assesses whether the current CNC machining system has deviated from the physically stable feasible region.

[0045] S3: PINN Network Architecture Design and Data Flow Process: To integrate the time-delay differential equation established in step S2 into the underlying computation of artificial intelligence, this embodiment constructs a dual-channel physical information neural network (PINN) based on a multilayer feedforward neural network (MLP). To ensure complete alignment between the offline training topology and the edge online inference topology, its network architecture design and joint loss convergence process are detailed below: S31: The topology and high-dimensional feature assembly of PINN networks: Input Layer: Composed of high-dimensional spatiotemporal feature tensors, it receives independent neuron nodes in six dimensions. Input Tensor The first three items are look-ahead spatiotemporal query parameters: The time coordinate is the independent variable, i.e., the time coordinate at a fixed moment. The rotational speed is the speed commanded by the tool turret. The eccentricity of the cutter head is set; the last three terms are local dynamic boundary conditions: To extract transient shifts from historical sequences, For transient vibration velocity, To determine the historical time-delay displacement of the regenerative interference effect, this high-dimensional assembly not only provides the time independent variable but also forcibly imposes the local physical boundaries required for solving the partial differential equations. It should be noted that the above tensor expression represents the sample data structure of the physical information neural network during the offline training phase. When the model is trained and solidified to the edge computing nodes to enter the online active control phase, to achieve sub-millisecond-level advance prediction, this input tensor is configured as a high-dimensional temporal feature vector. At this point, the time independent variable of the input layer changes from a fixed time during training. Explicit translation assignment to the future target time Historical sample boundary conditions ( 、 (etc.) are replaced with the real-time physical state parameters acquired by the sensor at the current physical moment. 、 wait).

[0046] Hidden Layers: Employs a 3-layer fully connected topology (64 neurons per layer). The network uses high-order smooth and differentiable functions such as the hyperbolic tangent (Tanh) as activation functions to ensure the smoothness of subsequent time-based activation. When performing automatic differentiation, the acceleration (second derivative) is continuously non-zero, avoiding the mathematical risk of partial differential equation calculation breakage caused by conventional ReLU.

[0047] Output Layer: The output of a single-dimensional neuron. Given the aforementioned high-dimensional feature tensor... Under the constraints of network weight mapping, the corresponding time coordinates are directly output forward. Electromechanical system execution end predicts relative vibration displacement .

[0048] S32: Data flow and dual-channel joint closed-loop optimization process: The network uses back propagation to converge the errors from both data-driven and physical constraint sources to update closed-loop parameters. First path: Data-Driven Loss: Using pre-constructed supervised sample pairs, calculate the forward predicted displacement of the network. Compared with the actual measured physical displacement Mean square error between :

[0049] In the formula: Mean squared error loss representing the data-driven channel; This represents the total number of data points in the real displacement monitoring samples obtained by high-frequency acquisition from sensors. The first representative sample set The specific time coordinates of each data point; Represents the neural network at time points The displacement value of the forward prediction output; Represents the corresponding data collected and recorded by the sensor. The objective physical displacement value at a given moment.

[0050] Second path: Physical residual constraint channel (PDE Loss): The Autograd automatic differentiation operator, based on a deep learning framework, is used to evaluate the network output. For input layer time independent variables Perform continuous partial derivative calculations to analytically determine the predicted velocity. With predicted acceleration Simultaneously, the historical predicted displacement is calculated by substituting the time-delay independent variable into the input. Substitute all the above network prediction variables into the time-delay dynamics equation solidified in step S2. In the process of calculating physical collocation points Mean square error of partial differential residual at point :

[0051] In the formula: The mean square error loss represents the physical residual constraint channel; This represents the total number of unlabeled collocation points randomly sampled within the time domain; Representing the The time coordinates of each physical point; These represent the system's equivalent modal mass, structural damping coefficient, and equivalent dynamic stiffness, respectively, as determined by the static identification in step S2. These represent the collocations of the neural network. The predicted displacement, predicted velocity, and predicted acceleration are obtained from forward propagation and automatic differentiation. These represent the unit cutting force coefficient, radial depth of cut, and phase angle of a specific material, respectively. The theoretical nominal chip thickness is determined; This represents the historical time delay prediction displacement output by the neural network after shifting the input timestamp forward by T milliseconds.

[0052] Combining the two constraint paths mentioned above, we can obtain the joint convergence mechanism: The network weights and combines the two into a total loss function. :

[0053] In the formula: This represents the total loss function value used during iterative optimization of the network's backpropagation. This represents the adaptive balance weight coefficients corresponding to the data-driven channel and the physical constraint channel.

[0054] Historical datasets containing parameters such as different reference rotation speeds and cutting depths are imported to jointly train the network. While ensuring that the network can fit the training data, the network weights are made to converge as much as possible to a reasonable physical and mechanically feasible region, thus avoiding the risk of overfitting and distortion of pure data-driven methods under varying cutting conditions.

[0055] The optimizer updates the hidden layer parameters through backpropagation of the computation graph gradient. A penalty term is added to ensure that the weight updates during neural network training are within the allowable range for multi-tooth regenerative chatter, preventing overfitting and data distortion caused by the black-box data model. Once the total loss falls below a set threshold, the microprocessor fixes these parameters and writes the final weight matrix into the onboard flash memory of the industrial edge AI node in binary instruction format. During online machining, all network weights are fixed, and data is only read online, not written, to ensure the reliability and safety of the control system.

[0056] This invention utilizes time extrapolation to test PINN.

[0057] Specifically, the system divides the timeline into two parts. This is the training period, during which only displacement data within this range is used for network anchoring training. To predict blind spots, real displacement input is stopped during this time period, and the network performs forward predictions on its own.

[0058] From the appendix Figure 3 It can be seen that, in Within the prediction blind zone, the prediction results obtained by PINN (dashed line) are almost identical to the actual dynamic response (solid line). Calculations show that the root mean square error (RMSE) of the prediction within this blind zone is as low as [insert value here]. The difference is orders of magnitude less than that. This indicates that the network has indeed learned the actual underlying cutting time delay, rather than merely fitting the data, and has good chatter advance warning performance.

[0059] S4: Edge-end Advanced Inference and Active Loop Closure: S41: Lightweight solidification and hardware acceleration of the PINN edge inference engine: To meet the extreme requirements for inference speed under complex and large deformation cutting in cyclone milling, the PINN model needs to be transformed into a lightweight inference engine suitable for industrial applications after offline training.

[0060] Model pruning and quantization: After completing the joint training with physical constraints, the hidden layer weights of the PINN model are pruned to remove unnecessary nodes. Based on this, acceleration tools such as TensorRT are used to transform the model weights from 32-bit floating-point numbers (FP32) to 16-bit half-precision numbers (FP16) or even 8-bit integers (INT8).

[0061] Edge memory direct read: The highly compressed, lightweight PINN model is directly deployed to the underlying bus of the CNC machine tool system and industrial edge computing nodes with GPU modules. The model resides in the GPU memory and does not require solving any partial differential equations during runtime. It only needs to perform simple network multilayer perceptron (MLP) matrix multiplication operations, eliminating the computing bottleneck.

[0062] S42: Sub-millisecond high-frequency feature assembly and phase space advance calculation: S421: Heterogeneous Data Acquisition from Multiple Sources and Dynamic Circular Queue Extraction: After starting the cyclone milling process (including the initial air feed phase), set the cutter head reference speed n. t0 To address the issue that the internal bus of the CNC system cannot acquire the high-frequency status of the mechanical end, a dual-channel heterogeneous data acquisition method is used in the system.

[0063] First, industrial edge AI computing nodes are deployed on high-frequency sensors at the cutting end, as well as on dedicated external high-frequency data acquisition cards (DAQ), optical fibers, and other equipment. They acquire the acceleration information at the cutting end in real time at a sampling rate of 20 kHz, and perform high-speed digital integration within the underlying FPGA to extract the current real physical vibration displacement. x ( t ) and speed x ( t The system creates and maintains a dynamic circular buffer queue in video memory, which stores the timestamped sequence of physical displacements (in...). t During the empty-cut cold start phase (<0), the initial value of the queue is filled with zero ambient noise).

[0064] Meanwhile, the edge computing node synchronously reads the real-time tool head speed of the underlying servo via the machine tool's high-speed fieldbus (EtherCAT). According to the formula The unit is milliseconds, accurately calculating the hysteresis time constant unique to the current period. Edge nodes are determined according to this time constant. The precise reverse indexing of the circular buffer queue, in actual physical operation, when the calculated time constant T is not an integer multiple of the discrete sampling period, uses linear interpolation or spline interpolation algorithms on the edge nodes to interpolate the displacement data of two adjacent sampling points in the dynamic circular buffer queue, thereby extracting the historical vibration displacement characteristics when the previous tooth passed through that spatial position. .exist to In the initial cutting stage (milliseconds), since historical ripple overlap has not yet physically formed, the extracted... Approaching zero, the network uses this initial physical boundary condition to predict system convergence; while crossing... Milliseconds later, the system formally integrates the nonlinear time-series characteristics containing historical ripple phases and compares them with the current state. , They are assembled into high-dimensional feature vectors.

[0065] In addition, in order to overcome the drift of physical and dynamic parameters caused by tool wear and machine tool thermal deformation during long-term cutting, the edge computing node has a built-in slow time-varying boundary periodic recalibration mechanism. It uses the non-cutting no-load stage to re-identify the equivalent dynamic stiffness and damping parameters of the system online and dynamically refresh them to the physical address register to offset the long-term slow time-varying cumulative error of the system.

[0066] S422: Rapid Forward Reasoning Across Window-of-Lookout: After completing the synchronous extraction and time-delay state retrieval of the aforementioned multi-source heterogeneous data, the industrial edge AI computing node assembles a high-dimensional feature matrix from the local dynamic boundary conditions established offline and the spatiotemporal query parameters. Specifically, the system uses the current physical time... Explicitly shift the preset look-ahead time window backward It can take the value of Construct target prediction timestamp Subsequently, the target timestamp Real-time cutter head speed using fieldbus data Cutter head installation eccentricity The current end physical displacement demodulated in the circular buffer queue Vibration velocity and historical time-delay displacement characteristics Unified encapsulation and combination to form a high-dimensional input feature tensor It is then instantly fed into the lightweight PINN forward inference engine. Within a sub-millisecond-level extremely short hardware computation cycle, it spans the time coordinate and outputs the predicted envelope amplitude of the relative vibration displacement of the machine tool tip under the future preset look-ahead time window, while simultaneously quantifying and evaluating the stability margin of the current CNC machining system in the dynamic phase space.

[0067] Figure 3 Comparison of PINN-based cutting chatter advance prediction response, and Figure 4 As shown in the full-time domain absolute prediction error curve, the inference engine embedded in this invention can still maintain extremely high advanced inference accuracy after entering the prediction blind zone in the offline training area of ​​0.15s. Its prediction trajectory is highly consistent with the actual dynamic response, providing reliable data support for subsequent active intervention.

[0068] It should be noted that, in order to verify the effectiveness of the PINN inference engine described in this invention, such as... Figure 3 Comparison of PINN-based cutting chatter advance prediction response, and Figure 4 The full-time-domain absolute prediction error curve is shown in the figure. The benchmark anchoring data used to verify the model's prediction accuracy is a numerical simulation dataset generated in advance based on rigorous time-delay dynamic equations. In the actual industrial machine tool deployment stage, the system described in this invention also supports and can directly access real cutting physics experimental data collected by high-frequency sensors for model training. This embodiment uses a numerical simulation dataset to conduct qualitative and quantitative principle verification of the active control system. The results show that after entering the prediction blind zone, the prediction trajectory of the inference engine highly matches the benchmark dynamic response, proving the extremely high advanced inference accuracy and feasibility of this method.

[0069] Traditional online solutions to time-delay differential equations (DDEs) rely on extremely time-consuming numerical discrete iterative integration (such as the history-dependent Runge-Kutta method), with accumulated computational overhead and time delays typically reaching milliseconds or even tens of milliseconds, failing to meet the requirements for active defense against self-excited chatter in CNC systems. However, the Physical Information Neural Network (PINN) embedded in the method of this invention has already deeply learned and implicitly solidified the dynamic evolution laws throughout the entire continuous time domain and the solution space of the operating conditions during the offline training phase. Therefore, in the online execution state at the edge, the inference engine bypasses the complex numerical integration process of differential equations. When the lightweight inference engine receives the feature tensor... After that, its underlying hardware only needs to execute a resident GPU function based on the network weight matrix. With bias vector The multilayer feedforward perceptron linear matrix multiplication and hyperbolic tangent (Tanh) nonlinear activation mapping are used. The pure hardware computation cycle of the entire forward propagation can be achieved within 0.7ms under hardware acceleration. This is achieved by directly performing future timestamps on the time coordinates of the independent variables at the input. The explicit assignment of values ​​allows the network to transcend the current physical timeline, achieving a one-time, non-progressive cross-temporal mapping and directly outputting future moments in advance. Predicted envelope amplitude of relative vibration displacement of the tool tip under ( ) This rapid forward reasoning process provides a crucial time window for the underlying servo system of CNC machine tools to actively intervene electromagnetically before the actual outbreak of self-excited vibration.

[0070] S423: Phase space convergence synchronization determination: The edge AI control module outputs the predicted future amplitude. Simultaneously, the transient convergence state of the current cyclone milling dynamics system in the phase space trajectory is quantitatively evaluated. The decision logic is as follows: the controller continuously transmits the predicted future look-ahead amplitude at microsecond clock cycles. With respect to the system's preset vibration safety threshold Perform high-frequency threshold polling and comparison. If the predicted amplitude remains within the safety threshold... If the condition is within a certain range, it is determined that the phase trajectory of the machining system in the dynamic phase space is in a convergent state, and the machine tool maintains the conventional PLC loop logic; conversely, if the future is affected by a sudden change in the lead screw blank allowance or the impact of hard particles, it will be considered that the machining system is in a convergent state. The predicted amplitude within the range shows a diverging trend and approaches the safety threshold. The system then simultaneously determines that the phase trajectory of the machining system is about to diverge, that is, it determines that self-excited regenerative vibration is about to occur. The outbreak occurred later. This is due to the system's look-ahead time window. ( Significantly greater than the hardware latency of issuing control commands from the CNC underlying bus. This provides an extremely valuable physical control time margin for the subsequent triggering of the spindle speed variation (SSV) algorithm and the forced destruction of regenerative coherence.

[0071] S43: Physical Destruction of Transient Closed-Loop and Regenerative Coherence Based on SSV: High-frequency threshold polling and driver takeover: The edge control module continuously polls the predicted amplitude with a microsecond clock cycle. When a change in the leadscrew machining allowance causes a sudden increase in the depth of cut, the predicted amplitude approaches the set system safety threshold. At that time, the control module determines that the phase space trajectory is about to diverge. The CNC system's underlying controller interrupts the conventional PLC loop logic. Within an ultra-short control cycle, the highest control authority of the cyclone milling cutter head servo driver is forcibly acquired.

[0072] Nonlinear speed modulation execution: The driver instantaneously triggers the spindle speed variation (SSV) algorithm, injecting a sinusoidal frequency modulation signal into the cutter head motor. The commanded cutter head speed is 10% of the set reference speed with an amplitude ratio, and... The modulation frequency exhibits nonlinear fluctuations.

[0073] Forced phase misalignment: The continuous nonlinear transient fluctuations in the cutterhead rotation speed cause... The time interval between adjacent teeth of the forming blade cutting through the same position in the workpiece space. The constantness of the cycloidal period in cyclone milling is no longer constant. This forced disruption of the constantness of the cycloidal period in cyclone milling from the underlying kinematics causes high-frequency perturbations in the system's time delay constant, disrupting the regenerative interference phase of the multi-tooth envelope cutting thickness (i.e., preventing the in-phase superposition of historical ripple energy), instantly suppressing and dissipating the self-excited chatter of the leadscrew machining in its infancy. At the same time, the stable cutting state smooths out the nonlinear load torque fluctuations of the spindle servo motor, reduces the reactive current pulses and copper loss heating of the motor stator, and achieves active noise reduction and energy efficiency optimization in CNC machining.

[0074] S5: Empirical Evidence on Environmental Performance Assessment and Stability Boundary Expansion of Control Systems: Through discrete domain dynamic analysis of electromechanical systems and deduction using a force-electric coupling dynamic model, a multi-dimensional comprehensive verification of the energy-saving and environmental protection efficiency of CNC equipment after implementing advanced closed-loop control intervention was conducted.

[0075] This invention achieves breakthrough of discrete domain dynamic boundary and improvement of macroscopic operation efficiency.

[0076] Firstly, regarding the analysis of the control domain boundary extension, the system analyzes the theoretical stability boundary of the electromechanical system based on the time-delay dynamic equation, such as... Figure 5 The diagram illustrates the broadening of the Stability Leaf Diagram (SLD). The dashed lines represent the stability boundary under traditional conservative settings, while the solid lines represent the boundary of this invention after incorporating a Physical Information Neural Network (PINN) and edge-closed-loop active control. The analytical results clearly show that after implementing closed-loop active adjustment, the critical chatter-free physical boundary (limit radial cutting depth) of the CNC machine tool across the entire frequency band has been significantly and uniformly shifted upwards. The gray shaded area in the diagram represents the broadened absolute machining safety margin of the system.

[0077] Secondly, in terms of macro-level efficiency improvement and energy saving, the expansion of the absolute stability boundary means that the CNC control system has broken free from the traditional constraint of only being able to conservatively set low-load parameters for vibration prevention. The machine tool is allowed to operate continuously and stably with feed process parameters far exceeding those of the norm, resulting in a leapfrog increase in material removal rate (MRR). The system compresses the single-piece lead screw machining cycle, reducing the basic power consumption of the large peripheral auxiliary equipment of the CNC machine tool (such as high-power high-pressure cooling pumps, chip conveyor motors, centralized hydraulic stations, and servo standby modules) during long-term operation, thus realizing high equipment specific energy efficiency from the macro-level process planning perspective.

[0078] This invention achieves electrical energy-saving effects based on servo-based force-electric weighting.

[0079] Firstly, the control effect can be directly evaluated through the internal data of the CNC system, eliminating the need for expensive external power analysis hardware. In practical industrial applications, this system utilizes the OPC UA industrial interface of the CNC system to perform closed-loop energy consumption monitoring based on the underlying feedback current waveform of the high-frequency gripping spindle servo driver. The force-electric coupling model shows that because the control system forcibly eliminates the interference phase relationship between multiple gears at the underlying kinematic level, the spatially drastically changing nonlinear cutting force peaks are instantly and physically smoothed out.

[0080] Secondly, after the cutting vibration is suppressed, the servo motor does not need to continuously output a large current to counteract the high-frequency excitation force. The effective value (RMS) current of the drive motor is greatly reduced, thereby reducing the nonlinear copper loss (I2R heating) of the spindle motor stator winding and the eddy current loss of the iron core. This helps to save the power consumption of high-end CNC machine tools and improve the overall efficiency of the entire system.

[0081] This invention enables full life-cycle management of core consumables and source reduction of solid waste.

[0082] First, chatter suppression effectively prevents high-frequency vibration marks from appearing on the workpiece surface, improving the surface quality and increasing the first-pass yield. This reduces energy waste and carbon emissions caused by recasting or secondary cutting of defective products.

[0083] Secondly, this invention reduces the generation of heavy metal solid waste. The stable, non-alternating-impact servo operation effectively improves the micro-stress state at the tool-chip contact point, significantly reducing the probability of micro-chipping and early thermal fatigue fracture of expensive PCBN forming turning tools, thus multiplying tool life. This not only reduces the flow loss of valuable tools in the workshop but also fundamentally reduces the emission of non-biodegradable industrial solid waste containing heavy metals, aligning with the national direction of green and low-carbon production and energy-saving and emission-reduction industrial development.

[0084] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0085] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for active control of CNC equipment based on physical information neural networks, characterized in that, Includes the following steps: Steps for constructing and training a physical information neural network: The dynamic parameters of the end effector of the CNC equipment are obtained, a time-delay differential dynamic equation including the cutting regeneration effect is established, and the time-delay differential dynamic equation is converted into a physical residual term. A physical information neural network is constructed. The input of the physical information neural network includes a multidimensional feature tensor containing a fixed time as the physical time independent variable, the state parameter of the execution terminal at the fixed time, and the historical time delay state parameter. The output is the predicted vibration displacement corresponding to the fixed time. The physical information neural network is trained using a loss function to obtain the trained physical information neural network. The loss function includes a physical constraint loss function constructed using the physical residual term. Active control steps of CNC equipment: During the operation of the CNC equipment, the current state parameters and system time delay constant of the execution end are acquired in real time. Based on the system time delay constant, the time delay state parameters are extracted from the historical state parameters. The current physical time independent variable, the current state parameters and the time delay state parameters are combined to generate a high-dimensional time series feature vector. The current physical time independent variable in the high-dimensional time series feature vector is assigned the value of the current physical time plus the future target time after a preset look-ahead time window, and the current physical time is taken as the fixed time. The assigned high-dimensional time series feature vector is input into the trained physical information neural network for forward inference, and the predicted vibration displacement amplitude of the execution end at the future target time is output as the advanced prediction result. Closed-loop active adjustment step: Based on the advanced prediction result, determine whether the CNC equipment will experience chatter. If so, output a speed modulation command to the CNC equipment.

2. The active control method for CNC equipment based on physical information neural networks as described in claim 1, characterized in that, In the physical information neural network construction and training steps, the time-delay differential dynamic equation is converted into a physical residual term. for: in, The cutting force coefficient is expressed as a unit. Radial cutting depth; Theoretical chip thickness; This represents the current transient vibration disturbance at the tool tip. The physical time-delay displacement characteristics of the historical period; The physical hysteresis time constant; Equivalent modal mass of CNC equipment; The damping coefficient of the CNC equipment structure; For equivalent dynamic stiffness, For a fixed time.

3. The active control method for CNC equipment based on physical information neural networks as described in claim 2, characterized in that, In the physical information neural network construction and training steps, the physical constraint loss function is constructed using the physical residual term as follows: in, This represents the total number of unlabeled physical points randomly sampled within the time domain. For the first The moment of each physical point.

4. The active control method for CNC equipment based on physical information neural networks as described in claim 1, characterized in that, The loss function further includes a data-driven loss function, which is the mean square error between the predicted vibration displacement and the sensor-measured vibration displacement; the step of training the physical information neural network using the loss function includes: The physical constraint loss function and the data-driven loss function are weighted and combined to construct the total loss function; The physical information neural network is trained by backpropagation based on the total loss function to update the network weight parameters.

5. The active control method for CNC equipment based on physical information neural networks as described in claim 1, characterized in that, In the active control step of the CNC equipment, the high-dimensional temporal feature vector used for online prediction include: in, For a fixed time; For forward-looking time windows; To control the end-effector rotation speed; To execute the end-effector offset; This represents the current vibration displacement; The current vibration velocity; This represents the historical time-delay displacement.

6. The active control method for CNC equipment based on physical information neural networks as described in claim 1, characterized in that, The physical information neural network construction and training steps, after obtaining the trained physical information neural network, also include: Perform model pruning and / or weight quantization operations on the trained physical information neural network to generate a lightweight forward inference engine; The lightweight forward inference engine is embedded in the non-volatile storage medium of the edge computing node.

7. The active control method for CNC equipment based on physical information neural networks as described in claim 1, characterized in that, The closed-loop active adjustment step, in which the step of determining whether the CNC equipment will experience chatter based on the advanced prediction result includes: The advanced prediction results are compared with a preset safety threshold; If the predicted result exceeds the preset safety threshold, it is determined that the CNC equipment will experience chatter. If the predicted result does not exceed the preset safety threshold, the CNC equipment is determined to be in a stable cutting state.

8. The active control method for CNC equipment based on physical information neural networks as described in claim 1, characterized in that, The step of outputting a speed modulation command to the CNC equipment includes: The spindle speed variation algorithm is triggered to generate a speed modulation command containing a periodic nonlinear frequency modulation signal, which controls the actual spindle speed of the CNC equipment to fluctuate periodically and nonlinearly around the reference speed, so that the time delay constant responds to the chatter.

9. A CNC equipment active control system based on a physical information neural network, used to implement the CNC equipment active control method based on a physical information neural network as described in any one of claims 1 to 8, characterized in that, include: The sensing module is used to acquire real-time vibration signals and real-time spindle speed signals from the end effector during the operation of CNC equipment. The edge inference module internally deploys a physical information neural network trained using a loss function, which includes a physical constraint loss function constructed with physical residual terms. The physical residual terms are derived from a time-delay differential dynamics equation containing a cutting regeneration effect. The edge inference module receives data output from the perception module, generates the high-dimensional time-series feature vector, assigns the current physical time variable in the generated high-dimensional time-series feature vector to the current physical time plus a preset look-ahead time window for the future target time, inputs the assigned high-dimensional time-series feature vector into the physical information neural network for forward inference, outputs the predicted vibration displacement amplitude of the execution terminal at the future target time as the advanced prediction result, and compares the advanced prediction result with a preset safety threshold to determine whether speed modulation is triggered. If triggered, a speed modulation command is generated. The servo execution module is used to receive the speed modulation command and output the speed modulation command to the spindle driver of the CNC equipment.

10. The active control system for CNC equipment based on a physical information neural network as described in claim 9, characterized in that, The edge reasoning module includes: A real-time data processing unit is used to construct a dynamic circular queue in a cache to continuously store historical vibration displacement sequences with timestamps, and to extract historical time-delay displacements from the dynamic circular queue by reverse indexing according to the system time delay constant. The lightweight inference engine unit internally loads the physical information neural network, which, after receiving the assigned high-dimensional temporal feature vector, performs forward computation through matrix multiplication and nonlinear activation mapping to output the advanced prediction result. The control decision unit is used to compare the advanced prediction result with the preset safety threshold, and determine whether to trigger speed modulation based on the comparison result. If triggered, a speed modulation command is generated.

Images