A machining process optimization method based on digital twinning

Abstract

The application relates to the field of intelligent manufacturing and digital twin technology, and particularly discloses a machining process optimization method based on digital twin. The method constructs a multi-physical field digital twin body combining a pulse neural network and a fluid-structure coupling theory, collects signals in real time through a high-sampling-rate vibration sensor and carries out time sparse coding; when a chatter precursor is identified, a microsecond-level pulse trigger signal is generated to drive a cutting fluid spraying mechanism to form a dynamic fluid damping barrier in a tool-workpiece contact area, and actively inhibit the propagation of chatter. Through the above technical scheme, millisecond-level state synchronization and microsecond-level response control can be realized, the cutting fluid is converted into an active damping unit, and the machining stability, precision and tool life are improved.

Description

Technical Field

[0001] This invention belongs to the field of intelligent manufacturing and digital twin technology, specifically relating to a method for optimizing machining processes based on digital twins. Background Technology

[0002] With the profound transformation of intelligent manufacturing technology, digital twin technology is increasingly being applied in the field of machining process monitoring and quality prediction, becoming a core means to achieve precise control in digital workshops. By constructing a virtual image of machining equipment and processes, the system can achieve real-time mapping and parallel simulation of cutting parameters, tool status, and workpiece quality, providing decision support for optimizing machining strategies. In high-feed, high-speed machining tasks of precision parts, the system's real-time perception and recognition of physical entities and its process optimization capabilities directly determine machining accuracy, surface integrity, and the service life of production equipment.

[0003] Suppressing high-frequency chatter during machining is a crucial step in ensuring machining stability. Current digital twin optimization methods focus on using multi-source sensing data to model the machining environment in real time. By introducing fluid dynamics analysis and neural network algorithms, they aim to accurately simulate the evolution of the physical field during cutting, constructing a closed-loop feedback control mechanism between the physical entity and the virtual model. This method attempts to achieve real-time compensation and proactive intervention for cutting vibrations by dynamically adjusting process parameters or auxiliary cooling flow fields, thereby protecting the tool and improving machining quality.

[0004] Traditional digital twin systems based on physical spring damping models suffer from significant computational lag when dealing with high-frequency chatter, making it difficult to capture transient changes during the cutting process. Because traditional control loops rely on complex cloud computing and linear logic analysis, their communication delays prevent the predictive feedback speed from meeting the real-time suppression requirements of high-frequency vibrations, easily leading to machining instability. Existing models lack a deep understanding of multidimensional fluid-structure interaction effects, resulting in insufficient adaptive adjustment of nonlinear vibration characteristics when facing different material properties or complex cutting fluid environments, making it difficult to achieve microsecond-level stress responses. Traditional artificial neural networks consume high energy and have limited time accuracy when processing large-scale time-series sensor data, failing to accurately correlate the dynamic relationship between the flow field and workpiece vibration. Therefore, a digital twin-based optimization method for machining processes is desired. Summary of the Invention

[0005] The purpose of this invention is to provide a machining process optimization method based on digital twins, which can effectively solve the problems mentioned in the background. Traditional digital twin systems suffer from computational lag and communication delay when dealing with high-frequency chatter, making it difficult to effectively suppress transient vibrations. Existing physical models lack adaptability to complex material properties and changes in the cutting fluid environment, failing to form an active vibration reduction mechanism with microsecond-level response capabilities. This invention constructs a "neuro-fluid" twin that integrates pulse neural networks and fluid-structure interaction theory, achieving high-precision sensing, low-latency prediction, and real-time generation of fluid damping barriers for machining chatter.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: a method for optimizing machining processes based on digital twins, comprising the following specific steps: Step 1: Construct a multiphysics digital twin that includes processing equipment, workpiece, cutting tool and cutting fluid system. The digital twin integrates a fluid-structure interaction model and a spiking neural network model to synchronously map the structural dynamics behavior and flow field dynamics of the physical entity. Step 2: Using high sampling rate vibration sensors arranged on the workpiece and the cutting tool, multi-dimensional vibration signals during the machining process are collected in real time, and the vibration signals are input into the pulse neural network model for time sparse encoding processing; Step 3: Based on fluid-structure interaction theory, establish a dynamic mapping relationship between the flow state parameters of the cutting fluid and the vibration response of the workpiece surface. The flow state parameters include velocity distribution, pressure gradient and fluid action direction. Step 4: When the pulse neural network model identifies the flutter precursor features, it immediately generates a pulse trigger signal with time precision. The pulse trigger signal directly drives the cutting fluid injection actuator to adjust the injection frequency and injection angle. Step 5: By adjusting the cutting fluid jet flow field, a dynamic fluid damping barrier is formed in the contact area between the workpiece and the tool, which suppresses the propagation and amplification of chatter energy and maintains the stability of the machining process.

[0007] Preferably, the multiphysics digital twin constructed in step 1 adopts a modular architecture, including a structural dynamics sub-model, a flow field evolution sub-model, and a neural response sub-model. The three interact with each other through a unified time synchronization mechanism to ensure that the virtual model and the physical entity maintain state consistency on a millisecond time scale.

[0008] Preferably, in step 2, the sampling frequency of the high sampling rate vibration sensor is not lower than a preset threshold to ensure that the rising edge and peak characteristics of the high-frequency flutter signal can be fully captured; the pulse neural network model adopts a neuron unit based on the leakage integration firing mechanism, and its membrane potential dynamic evolution process is determined by the external input current and the internal attenuation constant to realize the time encoding of the vibration signal.

[0009] Preferably, in step 3, the fluid-structure interaction model is discretized using the finite volume method to solve the coupled system of the Navier-Stokes equations and the structural dynamics equations. The interface between the fluid domain and the solid domain is processed using dynamic mesh technology to accurately reflect the instantaneous impact force and damping effect of the cutting fluid on the workpiece surface.

[0010] Preferably, the generation logic of the pulse trigger signal in step 4 is based on the synchronous firing behavior of the neuron cluster in the spiking neural network. When the energy spectral density of the input vibration signal exceeds a preset threshold in the frequency band, the neuron cluster immediately outputs a high-intensity pulse sequence with a time delay of no more than microseconds.

[0011] Preferably, in step 5, the cutting fluid injection actuator includes an array of multiple independently controllable micro-nozzles, each nozzle equipped with a piezoelectric actuator, which can adjust the injection angle and flow rate within a predetermined time after receiving a pulse trigger signal, forming a directional fluid damping barrier in a local area of ​​the workpiece.

[0012] Preferably, the formation location and strength of the fluid damping barrier are dynamically determined based on the spatial distribution characteristics of the current flutter mode. By analyzing the spatial coherence of the vibration signal, the region where the flutter energy is most concentrated is identified, and high-density fluid jets are preferentially deployed in this concentrated region.

[0013] Preferably, the spiking neural network model employs a combination of offline simulation data and online measured data for supervised learning during the training phase. The offline data originates from a high-fidelity multiphysics simulation platform and is used to cover vibration modes under typical machining conditions. The online data is used to continuously fine-tune the network weights to adapt to nonlinear disturbances caused by material batch differences and tool wear in the actual machining environment.

[0014] Preferably, the digital twin and the physical processing system use an edge computing architecture for data interaction. All vibration signal processing, flutter recognition, and injection control command generation are completed on the local edge node, avoiding the communication delay introduced by the traditional cloud loop and ensuring that the response time of the entire control closed loop is within the microsecond range.

[0015] Preferably, the method further includes a real-time feedback correction mechanism for the cutting fluid spraying effect. By monitoring the attenuation rate of the workpiece vibration signal after spraying, the threshold parameter for the next pulse trigger is dynamically adjusted to achieve adaptive optimization of chatter suppression strategies at different processing stages.

[0016] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention deeply integrates spiking neural networks with fluid-structure interaction theory to construct a "neuro-fluid" digital twin possessing the stress response characteristics of a biological nervous system, solving the problems of computational lag and poor environmental adaptability in high-frequency chatter suppression of traditional digital twin systems. The spiking neural network utilizes a time-sparse coding mechanism to improve the speed and accuracy of identifying chatter precursor signals, with a microsecond-level response capability far exceeding that of traditional artificial neural networks. 2. This invention transforms the cutting fluid from a passive cooling medium into an active vibration damping actuator, enabling the system to form a targeted fluid damping barrier at the initial stage of chatter, blocking the propagation path of vibration energy. This method does not rely on complex cloud-based computational closed loops; all control logic is completed at the edge, reducing the impact of communication latency on control performance. 3. The fluid-structure interaction model of this invention can adaptively learn the vibration characteristics under different material and cutting fluid combinations, enabling the system to maintain a stable suppression effect even in complex and variable machining environments, improving machining accuracy, tool life, and production efficiency. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention; Figure 2 This is a schematic diagram of the core principle framework of the "neural-fluid" digital twin that integrates spiking neural networks and fluid-structure interaction theory in this invention; Figure 3 This is a flowchart illustrating the logical flow of the invention based on time-sparse coding of multidimensional vibration signals and flutter precursor identification. Figure 4 This is a schematic diagram of the multi-level interaction relationship and data flow between the physical processing system, edge computing nodes, and digital twin model in this invention; Figure 5 This is a flowchart illustrating the logical process of active generation and adaptive feedback correction of the dynamic fluid damping barrier in this invention. Detailed Implementation

[0018] Example 1: Please refer to the appendix Figure 1 To be continued Figure 5 To make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments.

[0019] In the aforementioned method for optimizing machining processes based on digital twins, step 1 specifically includes constructing a multiphysics digital twin comprising machining equipment, workpiece, cutting tool, and cutting fluid system. During execution, the digital twin integrates structural dynamics sub-models, flow field evolution sub-models, and neural response sub-models through a modular architecture.

[0020] The structural dynamics sub-model pre-stores the overall stiffness matrix of the machining equipment, the equivalent mass parameters of each feed axis, and the rotational dynamic characteristics data of the spindle system. The physical properties of the workpiece are digitized into a voxelized mesh model containing three-dimensional spatial coordinates, material elastic modulus, Poisson's ratio, and density distribution. The geometric features of the cutting tool are modeled with high precision based on the actual tool tip radius, rake angle, clearance angle, and helix angle, and correlated with its material hardness and coefficient of thermal expansion.

[0021] The flow field evolution sub-model is specifically designed to simulate the fluid dynamics of the cutting fluid system at the pipes and nozzles, and internally defines the dynamic viscosity, density, and physical parameters of the cutting fluid as a function of temperature. The neural response sub-model is built on a programmable logic array and simulates the interconnection structure of a cluster of neurons, with a pre-defined synaptic weight matrix.

[0022] The structural dynamics sub-model, flow field evolution sub-model, and neural response sub-model interact with each other at high frequency through a unified time synchronization mechanism. The system master clock sends synchronization pulses to each sub-model with a base period of 100 microseconds. Upon receiving the synchronization pulse, each sub-model immediately aligns its internal state variables. The state variables include the six-degree-of-freedom coordinates of the blade tip in virtual space, the pressure vectors of each node in the fluid domain, and the current instantaneous value of the neuron membrane potential.

[0023] In the construction process of step 1, the multiphysics digital twin achieves synchronous mapping between the structural dynamics of the physical entity and the dynamic characteristics of the flow field. The structural dynamics include the mode shapes and nonlinear drift of the natural frequencies of the mechanical structure as the machining allowance decreases; the dynamic characteristics of the flow field encompass the impact pressure distribution of the cutting fluid jet on the workpiece surface and the turbulence intensity formed in the cutting area. By allocating a high-speed cache in memory, the system compares the measured data fed back by the physical sensors with the predicted data of the virtual model in real time, and dynamically fine-tunes the boundary conditions of the virtual model using a residual correction algorithm.

[0024] In step 2, high-sampling-rate vibration sensors deployed on the workpiece and cutting tool are used to acquire multi-dimensional vibration signals during the machining process in real time. These high-sampling-rate vibration sensors include piezoelectric accelerometers and non-contact laser vibrometers, with a sampling frequency set to no less than 100 kHz to ensure complete capture of the rising edge and peak characteristics of high-frequency chatter signals above 20 kHz. The multi-dimensional vibration signals encompass acceleration vector data in the spindle radial, axial, and feed directions. The acquired raw analog electrical signals are preprocessed by a front-end charge amplifier and then converted into a 16-bit wide digital signal stream by a high-bit-to-analog converter.

[0025] After acquiring the multidimensional vibration signal, the system inputs the signal into a spiking neural network model for time-sparse encoding processing. The spiking neural network model employs neuronal units based on a leaky integrated firing mechanism, whose internal logic defines the dynamic evolution of the membrane potential. The current value of the neuron's membrane potential depends on the combined effects of the residual membrane potential from the previous moment, the contribution of the external input current, and the internal decay constant.

[0026] The external input current contribution term is obtained by weighting the amplitude of the acquired vibration signal, while the internal attenuation constant determines the leakage rate of the membrane potential in the absence of external excitation. When the cumulative value of the neuron's membrane potential increases to a preset threshold voltage, the neuron immediately generates a pulse output signal, and then the membrane potential is forcibly reset to the reference potential and enters a preset absolute refractory period.

[0027] In this process, continuous vibration signals are converted into a series of pulse sequences discretely distributed along the time axis. This encoding method utilizes the time interval between pulses to carry the instantaneous change information of vibration energy, not only filtering out low-amplitude random noise interference, but also preserving the weak impact characteristics at the initial stage of flutter.

[0028] In step 3, based on fluid-structure interaction theory, a dynamic mapping relationship is established between the flow state parameters of the cutting fluid and the vibration response of the workpiece surface. The fluid-structure interaction model is discretized using the finite volume method to solve the coupled system of the Navier-Stokes equations and the structural dynamics equations. Within the fluid computational domain, the system decomposes the flow state parameters into a velocity distribution vector field, a pressure gradient scalar field, and a fluid action direction tensor. The momentum change of fluid particles is described as the algebraic sum of the pressure term, the viscosity term, and the external constraint force, thereby calculating the instantaneous impact force of the cutting fluid on the workpiece surface.

[0029] Within the solid-state computational domain, the vibration response of the workpiece surface is simplified as a multi-degree-of-freedom damped forced vibration system, with its displacement, velocity, and acceleration following Newton's second law. The interface between the fluid and solid domains is processed using dynamic meshing technology, meaning that when the workpiece surface experiences micron-level vibrational displacement, the node coordinates of the fluid mesh are updated synchronously, ensuring that stress continuity and velocity matching conditions are met at the physical interface. Through iterative calculations, the system derives the suppression function of the cutting fluid on the workpiece vibration amplitude under different flow rates and injection positions. This mapping relationship is stored as a multidimensional lookup table or an offline regression model, enabling the system to quickly retrieve the fluid parameter combination that maximizes the damping effect based on the currently observed vibration response values.

[0030] In step 4, when the spiking neural network model identifies flutter precursor features, it immediately generates a time-precise pulse trigger signal. The identification logic for flutter precursor features is based on the synchronized firing behavior of neuron clusters in the spiking neural network. The system monitors the firing frequency of neuron pulses within the frequency band in real time. When the energy spectral density of the input vibration signal continuously increases within the preset flutter-sensitive frequency band, and the corresponding neuron cluster generates a highly synchronized pulse firing sequence within a very short time window, the logic determines that the system is currently in the flutter nascent stage.

[0031] The spiking neural network immediately outputs a sequence of high-intensity pulses as a trigger signal. The generation of this pulse trigger signal does not involve traditional convolution operations or complex floating-point recursive calculations; instead, it is achieved through direct logical mapping of the pulse stream. The time delay from feature recognition to control signal output is strictly controlled to the microsecond level, typically not exceeding 10 microseconds. This pulse trigger signal directly acts on the drive circuit of the coolant injection actuator.

[0032] In step 5, a dynamic fluid damping barrier is formed in the workpiece-tool contact area by adjusting the cutting fluid jet flow field. The cutting fluid jet actuator includes multiple independently controllable micro-nozzle arrays, which are arranged in a ring or fan shape around the tool. Each nozzle is equipped with an independent piezoelectric actuator, which can complete its extension and retraction within 50 microseconds after receiving a pulse trigger signal, thereby driving a micro-piston to change the nozzle's throttling area or jet deflection angle.

[0033] By rapidly adjusting the injection frequency and angle, multiple high-pressure cutting fluid jets converge and interfere near the workpiece-tool cutting interface. The adjusted injection frequency maintains a phase relationship with the dominant chatter frequency, typically set with a phase difference close to 180 degrees, to achieve energy interference cancellation. The generated dynamic fluid damping barrier utilizes the incompressibility and dynamic pressure effect of the fluid to form a high-stiffness liquid film on the workpiece surface. This liquid film increases the system's modal damping ratio, suppressing the propagation and amplification of chatter energy within the workpiece.

[0034] In the subsequent processing of step 5, the system continuously monitors the vibration signal after the fluid damping barrier is formed. By analyzing the decay rate of the vibration amplitude after the injection action is executed, the data is fed back to the pulse neural network model. If the decay rate is lower than the preset target value, the system will automatically increase the current intensity of the next pulse trigger and dynamically fine-tune the injection directionality of the nozzle array.

[0035] Example 2: In another preferred embodiment, the digital twin-based machining process optimization method is specifically configured for milling milling conditions of thin-walled aerospace parts. In this application scenario, the local stiffness of the workpiece is extremely low and fluctuates drastically with changes in the machining trajectory.

[0036] In step 1, a real-time material removal model is introduced into the constructed multiphysics digital twin. The system updates the geometry of the virtual workpiece in real time according to the interpolation instructions from the CNC system, and recalculates the natural frequencies of each part of the workpiece accordingly. The neuron weights in the neural response sub-model are initialized to modes specific to the vibration characteristics of the thin-walled component. The structural dynamics sub-model employs finite element method reduction technology, transforming large-scale degree-of-freedom matrix operations into modal superposition operations based on eigenvectors, thereby improving the computational efficiency of the virtual model and ensuring its state consistency is maintained within milliseconds.

[0037] During step 2, a high sampling rate vibration sensor is mounted on the back of the thin-walled workpiece and on the side of the worktable. To filter out periodic cutting force interference caused by spindle rotation, the spiking neural network model incorporates a lateral inhibition mechanism. Specifically, when the input signal contains regular pulses consistent with a multiple of the spindle rotation speed, inhibitory neurons fire negative pulses, weakening the influence of this signal on the membrane potential and enabling the model to more sensitively capture asynchronous, randomly diverging early chatter signals.

[0038] In performing step 3, the fluid-structure interaction model specifically considers the acoustic contribution of the cutting fluid on the surface of the thin-walled part. The fluid damping effect includes not only mechanical damping but also the attenuation of acoustic-structure interaction formed by the coupling between the fluid and the workpiece surface. Cavitation margin monitoring was incorporated into the flow state parameters to ensure that cavitation does not occur during high-speed injection, maintaining the continuity of the damping barrier. Through experimental calibration, the system obtained a linear relationship between the cutting fluid pressure gradient and the workpiece bending mode suppression rate at different cutting depths.

[0039] When performing step 4, the pulse trigger signal generation logic is designed as a multi-level threshold response. When vibration energy accumulates in the low-frequency band, a long-period pulse sequence is generated to control the cutting fluid to provide steady-state damping in a continuous flow form; when a high-frequency explosive chatter precursor is detected, it immediately switches to an ultra-high-frequency pulse sequence, driving the piezoelectric actuator into a high-frequency pulsating injection mode. This pulsating injection can generate a huge dynamic impact force with minimal water consumption, precisely targeting the chatter peak.

[0040] During step 5, the micro-nozzle array automatically selects the number of nozzles to be activated based on the current processing point's location. For high-stiffness areas near the fixture, low-pressure spraying is used to conserve resources; for cantilevered thin-walled areas far from the fixture, all nozzle resources are utilized to form a high-density fluid barrier. The formation location of the fluid damping barrier is dynamically reconstructed based on the spatial coherence analysis results of the vibration signals. Specifically, the system calculates the cross-correlation function between the signals from each sensor, identifies the antinodes with the largest vibration displacement, and aligns the spray center with these antinodes to achieve precise point-to-point vibration suppression.

[0041] Example 3: In another preferred embodiment, the method of the present invention is applied to the ultra-precision machining process of difficult-to-machine materials. In this process, the tool wear condition is extremely sensitive to chatter; therefore, the digital twin integrates a tool wear prediction algorithm in step 1.

[0042] The tool wear prediction algorithm takes cutting time, cumulative cutting force, and spindle load changes as inputs to dynamically adjust the geometric parameters in the tool sub-model. When the tool flank wear increases, the system automatically reduces the threshold voltage required to generate pulses in the spiking neural network. This adjustment means that the system can enter monitoring mode more sensitively in the early stages of tool performance degradation, and predict self-excited vibrations induced by increased cutting forces in advance.

[0043] During signal encoding in step 2, the spiking neural network employs a multi-timescale encoding strategy. The first layer of neurons is responsible for capturing millisecond-level macroscopic vibration trends, while the second layer of neurons captures microsecond-level transient changes through a smaller leakage constant. This hierarchical architecture can distinguish between accidental impacts caused by hard particles within the workpiece material and continuous self-excited flutter.

[0044] In performing the modeling in step 3, the fluid-structure interaction model incorporates a nonlinear model of the cutting fluid lubrication performance. The cutting fluid not only acts as a damping medium, but its flow state in the tool tip interference zone also alters the friction coefficient. The system establishes the relationship between friction fluctuations and the self-excited vibration feedback loop. Logically, this can be described as follows: when the pressure in the cutting region increases, leading to a surge in friction, the fluid-structure interaction model calculates the required fluid lubrication flow rate and converts it into an increase in the strength of the damping barrier.

[0045] During step 4, the pulse trigger signal is assigned a spatial vector attribute. Each pulse carries not only a timestamp but also the target nozzle's identification number. This allows the actuator to immediately perform synchronous or alternating injection according to a predetermined topology upon receiving the signal.

[0046] In step 5, to address the stringent surface quality requirements of ultra-precision machining, the system incorporates a refined analysis of residual vibrations in the workpiece after coolant spraying. The feedback correction mechanism fine-tunes the synaptic weights of the pulse neural network using a self-learning algorithm based on the frequency distribution of the residual vibrations. This fine-tuning process is performed after each machining cycle, allowing the digital twin to continuously evolve and adapt to the microscopic hardness differences between different material batches.

[0047] To ensure microsecond-level response times under large data volumes, all computational processes involved in each step are completed within edge computing nodes. These edge computing nodes are directly connected to the machine tool controller's bus and utilize field-programmable gate arrays (FPGAs) as accelerators. This architecture completely eliminates network jitter and latency caused by data uploads to the cloud, ensuring that the end-to-end latency from signal acquisition to fluid barrier generation remains stable within a preset safety range.

[0048] In any of the above embodiments, the method further includes monitoring the cutting fluid recovery and filtration status. When filter clogging causes the flow velocity distribution parameters to deviate from a preset range, the digital twin automatically compensates for the pressure loss through a flow field evolution sub-model, instructing the pump group to increase its output power to ensure that the strength of the fluid damping barrier is not affected by external hardware losses. This self-healing mechanism ensures the robustness of the machining optimization process during long-term operation.

[0049] The strength of the fluid damping barrier is quantitatively described as the ratio between the dynamic pressure force generated by the fluid on the workpiece surface under injection pressure and the original vibration excitation force of the workpiece. The system aims to maintain this ratio between 1.5 and 2.0 times during the initial stage of flutter by dynamically adjusting it, ensuring that the vibration energy is rapidly dissipated. All logical judgments, parameter comparisons, and numerical conversions are achieved by switching the register states within the computing unit, without relying on any external manual intervention.

[0050] The aforementioned multiphysics digital twin also integrates an ambient temperature compensation function. When the ambient temperature in the machining workshop fluctuates, the flow field evolution sub-model automatically corrects the viscosity coefficient of the cutting fluid. A decrease in the viscosity coefficient leads to a decrease in the fluid damping coefficient. At this time, the system automatically instructs an increase in the injection frequency, compensating for the damping loss caused by changes in physical properties by increasing the injection time density. This logic ensures that the system maintains consistent chatter suppression capability in different seasons and different working periods.

[0051] When generating the pulse trigger signal, the system also incorporates a redundancy check mechanism. Upon detecting a flutter precursor, the pulse neural network simultaneously calculates the trigger timing in two independent parallel logic channels. Only when the pulse time deviation between the two channels is less than 5 microseconds will a formal command be issued to the actuator. This design prevents false triggering due to electromagnetic interference, protecting the high-precision piezoelectric actuator from damage by abnormal signals.

[0052] The proposed digital twin-based machining process optimization method combines the sparse coding characteristics of computational neuroscience with the coupling inhibition principle of fluid mechanics to construct a high-speed response protective barrier between physical and virtual spaces. This barrier enables the machining system to react instinctively to external stimuli, much like a living organism, fundamentally changing the traditional control model that relies on post-processing feedback compensation. In each machining cycle, the system continuously accumulates data and, through online fine-tuning of network weights, achieves deep learning and adaptation to complex nonlinear conditions.

[0053] This method has applications beyond metal cutting, extending to areas such as composite material grinding and ultrasonic machining of ceramics. When dealing with materials of different properties, it is only necessary to replace the corresponding material property library in the digital twin and adjust the sensitivity parameters of the spiking neural network. Its core "neuro-fluid" interaction logic remains consistent: using high-speed acquired vibration pulses to drive the fluid dynamics response, implementing precise interference at the source of the energy propagation chain.

[0054] When adjusting the cutting fluid jet flow field in real time, the system also considers the splashing behavior of the fluid within the enclosed machining space. Through predictions using a flow field evolution sub-model, excessive obstruction of the observation window or sensors during interference by the jet flow is avoided. The selection of the jet angle considers not only damping effects but also auxiliary chip removal. The system uses chip removal efficiency and vibration suppression as dual objectives, searching for the optimal intersection solution in a lookup table to ensure that the machining process remains stable and clean.

[0055] The interaction protocol between the multiphysics digital twin and the physical processing system adopts low-latency real-time industrial Ethernet. All status packets include high-precision time stamps in their headers, and the system automatically compensates for minute phase shifts caused by network transmission during data fusion. This precise control of time is the core guarantee for achieving microsecond-level chatter suppression. The entire system operates with rigorous logic, and each functional module supports the others, forming a complete, self-evolving closed-loop intelligent control system.

[0056] Example 4: This example further details the specific implementation of the spiking neural network model during the training phase. This process employs a deep supervised learning approach that combines offline simulation data with online experimental data.

[0057] In the offline phase, a massive number of machining condition samples are generated using a high-fidelity multiphysics simulation platform. This platform exhaustively simulates all typical combinations of spindle speeds from 1,000 to 30,000 revolutions per minute, feed rates from 0.01 mm to 0.5 mm per tooth, and depths of cut from 0.1 mm to 5 mm. Under each combination, perturbation signals of varying intensities are artificially introduced to simulate extreme conditions such as tool breakage and uneven material hardness. The vibration waveform data output by the simulation platform is converted into corresponding pulse time series, serving as the initial training set for a spiking neural network.

[0058] During offline training, the system adjusts the weights of synaptic connections between neurons to enable the model to accurately classify three modes: "steady state," "flutter budding," and "violent flutter." The weight adjustments follow a plasticity rule similar to that of biological nervous systems: if two neurons frequently fire impulses sequentially, the connection strength between them increases. The training objective is to maximize the model's early warning time for flutter precursors, that is, the synchronized firing characteristics of the neuron cluster before any visible change in vibration amplitude occurs.

[0059] Upon entering the online machining phase, the system activates an online fine-tuning mechanism. Since batch variations in materials and nonlinear disturbances caused by tool wear in the actual machining environment cannot be fully covered by offline simulation, the system captures real-time vibration data during the machining process. When a physical entity experiences a slight flutter that the digital twin fails to predict in time, this measured data segment is immediately marked as a learning sample. The system calculates the deviation between the pulse sequence of this learning sample and the preset template in the background and uses the idle computing power of edge computing nodes to perform micro-iterations on the local weights. This incremental learning mode allows the system's recognition accuracy to monotonically increase with machining time, gradually approaching its performance limit.

[0060] Example 5: This example describes in detail the control logic of the micro-nozzle array in the cutting fluid injection actuator and its coupling relationship with the pulse signal.

[0061] Each nozzle in the micro-nozzle array is equipped with an independent digital drive unit. When the pulse trigger signal generated by the spiking neural network reaches the execution layer, it first enters the pulse distributor. The pulse distributor distributes the trigger command to different nozzle combinations according to the currently identified flutter mode.

[0062] For first-order bending chatter, since its displacement vector is mainly concentrated in the radial plane, the distributor instructs the nozzles on both sides of this radial plane to inject cutting fluid in alternating pulses. Each pulse command precisely corresponds to one reciprocating motion of the piezoelectric actuator. This control method is similar to the fuel injection control of an internal combustion engine, but with a higher frequency and faster response. By adjusting the pulse duty cycle, the system can continuously adjust the fluid jet intensity. Logically, as the pulse width increases, the nozzle opening time lengthens, and the resulting instantaneous damping force increases linearly; as the pulse frequency increases, the fluid barrier renewal rate accelerates, enabling it to cope with higher frequency vibration fluctuations.

[0063] During injection, the sensor also monitors pressure fluctuations at the nozzle tip to ensure that the data acquired by the flow field evolution sub-model is consistent with the actual pressure value. If a nozzle is detected to have a response delay or insufficient output pressure, the system will immediately activate fault-tolerant logic, assign the monitoring task of that nozzle to an adjacent backup nozzle, and compensate for the spatial difference by increasing the deflection of the injection angle.

[0064] Example 6: This example details the real-time feedback correction mechanism for the cutting fluid spray effect. This real-time feedback correction mechanism aims to ensure that the chatter suppression strategy remains in an optimal state under dynamic conditions through closed-loop optimization.

[0065] After the fluid damping barrier is formed in step 5, the system monitors the attenuation of vibration energy in real time using auxiliary sensors installed on the edge of the workpiece. The feedback correction module calculates the time difference from the issuance of the pulse trigger signal to the decrease of the vibration amplitude below the safety threshold. If this time difference exceeds the preset response period, the current damping strength is determined to be insufficient.

[0066] The feedback correction module sends a negative feedback signal to the spiking neural network. This negative feedback signal causes a slight decrease in the neuron firing threshold in the next cycle, accelerating the triggering of the next pulse. Simultaneously, the system instructs the fluid system to increase the reference oil supply pressure, thereby increasing the mass of cutting fluid ejected per unit pulse. This result-oriented adaptive optimization enables the system to effectively cope with long-term drift factors such as changes in cutting fluid concentration and aging of the pumping system.

[0067] The correction mechanism also analyzes the drift of the vibration signal on the frequency axis. If the dominant frequency of chatter shifts during vibration suppression, the fluid-structure interaction model will recalculate the optimal alignment position of the injection angle and fine-tune the pointing vector of the nozzle array. Through this dual correction of frequency tracking and spatial alignment, the system achieves comprehensive and continuous active control over machining chatter.

[0068] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A method for optimizing machining processes based on digital twins, characterized in that, Includes the following steps: Step 1: Construct a multi-physics digital twin that includes processing equipment, workpiece, cutting tool and cutting fluid system. The multi-physics digital twin integrates a fluid-structure interaction model and a spiking neural network model to synchronously map the structural dynamics behavior and flow field dynamics of the physical entity. Step 2: Using high sampling rate vibration sensors arranged on the workpiece and the cutting tool, multidimensional vibration signals during the machining process are collected in real time, and the multidimensional vibration signals are input into the pulse neural network model for time sparse encoding processing. Step 3: Based on fluid-structure interaction theory, establish a dynamic mapping relationship between the flow state parameters of the cutting fluid and the vibration response of the workpiece surface. The flow state parameters of the cutting fluid include velocity distribution, pressure gradient and fluid action direction. Step 4: When the pulse neural network model identifies the flutter precursor features, it generates a pulse trigger signal with time accuracy. The pulse trigger signal drives the cutting fluid injection actuator to adjust the injection frequency and injection angle. Step 5: By adjusting the cutting fluid jet flow field, a dynamic fluid damping barrier is formed in the contact area between the workpiece and the tool, suppressing the propagation and amplification of chatter energy.

2. The method for optimizing machining processes based on digital twins according to claim 1, characterized in that, The multiphysics digital twin constructed in step 1 adopts a modular architecture, including a structural dynamics sub-model, a flow field evolution sub-model, and a neural response sub-model. The structural dynamics sub-model stores the overall stiffness matrix of the processing equipment, the equivalent mass parameters of the feed axis, and the rotational dynamic characteristics data of the spindle system. The physical properties of the workpiece are digitized into a voxelized mesh model that includes three-dimensional spatial coordinates, material elastic modulus, Poisson's ratio, and density distribution; the geometric features of the cutting tool are modeled based on the actual tip radius, rake angle, clearance angle, and helix angle, and are associated with its material hardness and coefficient of thermal expansion. The flow field evolution sub-model simulates the fluid dynamics of the cutting fluid system at the pipes and nozzles, and defines the dynamic viscosity, density, and physical parameters of the cutting fluid as a function of temperature. The neural response sub-model is built on a programmable logic array, and the synaptic weight matrix is ​​set by simulating the interconnection structure of a cluster of neurons; The structural dynamics sub-model, flow field evolution sub-model, and neural response sub-model interact with each other through a unified time synchronization mechanism. The system master clock sends synchronization pulses to each sub-model based on a preset time base, and each sub-model aligns its internal state variables after receiving the synchronization pulse.

3. The method for optimizing machining processes based on digital twins according to claim 2, characterized in that, The process of synchronously mapping the structural dynamics and flow dynamics of physical entities using a multiphysics digital twin in step 1 includes: the structural dynamics encompasses the mode shapes of mechanical structures and the nonlinear drift of natural frequencies as machining allowance decreases; The dynamic characteristics of the flow field encompass the impact pressure distribution of the cutting fluid jet on the workpiece surface and the turbulence intensity formed in the cutting area. By allocating a high-speed cache area in memory, the system compares the measured data fed back by the high sampling rate vibration sensor with the predicted data of the multiphysics digital twin in real time, and uses a residual correction algorithm to dynamically fine-tune the boundary conditions of the multiphysics digital twin.

4. The method for optimizing machining processes based on digital twins according to claim 3, characterized in that, The multidimensional vibration signal acquisition and time sparse coding process in step 2 includes: the high sampling rate vibration sensor includes a piezoelectric accelerometer and a non-contact laser vibrometer, whose sampling frequency is set to be higher than a preset frequency threshold in order to capture the rising edge and peak characteristics of the high-frequency flutter signal. The multidimensional vibration signal encompasses acceleration vector data in the radial, axial, and feed directions of the spindle; The spiking neural network model uses neuron units based on a leaky integrated firing mechanism. The current value of the neuron membrane potential depends on the residual value of the membrane potential at the previous moment, the contribution of the external input current, and the internal decay constant. The external input current contribution term is obtained by weighting the amplitude of the acquired multidimensional vibration signal, and the internal attenuation constant determines the leakage rate of the membrane potential without external excitation. When the cumulative value of the neuron's membrane potential increases to a preset threshold voltage, the neuron generates a pulse output signal. Subsequently, the membrane potential is forcibly reset to the reference potential and enters a preset absolute refractory period, converting the continuous vibration signal into a pulse sequence discretely distributed on the time axis.

5. The method for optimizing machining processes based on digital twins according to claim 4, characterized in that, The process of establishing a dynamic mapping relationship based on fluid-structure interaction theory in step 3 includes: the fluid-structure interaction model is discretized and the coupled system of Navier-Stokes equations and structural dynamics equations is solved by the finite volume method. Within the fluid computational domain, the cutting fluid flow state parameters are decomposed into a velocity distribution vector field, a pressure gradient scalar field, and a fluid action direction tensor. The momentum change of fluid particles is described as the algebraic sum of the pressure term, viscosity term, and external constraint force, and the instantaneous impact force of the cutting fluid on the workpiece surface is calculated accordingly. In the solid computation domain, the vibration response of the workpiece surface is simplified to a multi-degree-of-freedom damped forced vibration system, and the variation of its displacement, velocity and acceleration follows Newton's second law. The interface between the fluid domain and the solid domain is treated with dynamic mesh technology. When the workpiece surface vibrates and displaces, the node coordinates of the fluid mesh are updated synchronously to ensure that the stress continuity and velocity matching conditions are met at the physical interface. The suppression function of cutting fluid on workpiece vibration amplitude under different flow outputs and different injection positions was obtained by iterative calculation and stored as a multidimensional lookup table.

6. The method for optimizing machining processes based on digital twins according to claim 5, characterized in that, The process of identifying flutter precursor features and generating pulse trigger signals in step 4 includes: the identification of flutter precursor features is based on the synchronous firing behavior of neuron clusters in the spiking neural network model; The system monitors the firing frequency of neuronal pulses within the frequency band in real time. When the energy spectral density of the multidimensional vibration signal continues to increase within the preset flutter sensitive frequency band, and the corresponding neuronal cluster generates a synchronous pulse firing sequence within the preset time window, the logic determines that the system is currently in the flutter bud stage. The pulse neural network model outputs a set of high-intensity pulse sequences as the pulse trigger signal. The generation process is completed through direct logical mapping of the pulse flow. The time delay from feature recognition to control signal output is on the order of microseconds. The pulse trigger signal directly acts on the drive circuit of the cutting fluid injection actuator.

7. The method for optimizing machining processes based on digital twins according to claim 6, characterized in that, The process of forming a dynamic fluid damping barrier in step 5 includes: the cutting fluid injection actuator includes multiple independently controlled micro-nozzle arrays, which are arranged in a ring or fan shape around the tool; Each nozzle is equipped with an independent piezoelectric actuator, which drives a miniature piston to change the nozzle's throttling area or spray deflection angle after receiving the pulse trigger signal; By adjusting the injection frequency and injection angle, multiple high-pressure cutting fluid jets converge and interfere near the cutting interface between the workpiece and the tool. The adjusted injection frequency maintains a preset phase relationship with the main frequency of the flutter, and the phase relationship is set to a phase difference of approximately 180 degrees to achieve energy interference cancellation. The dynamic fluid damping barrier utilizes the incompressibility of fluids and the dynamic pressure effect to form a liquid film with high stiffness on the surface of the workpiece, thereby increasing the modal damping ratio of the system.

8. The method for optimizing machining processes based on digital twins according to claim 7, characterized in that, It also includes a real-time feedback correction mechanism for the cutting fluid spraying effect: after the dynamic fluid damping barrier is formed, the attenuation of vibration energy is monitored in real time by an auxiliary sensor installed on the edge of the workpiece. The feedback correction module calculates the time difference from the issuance of the pulse trigger signal to the decrease in vibration amplitude below the safety threshold; When the time difference exceeds the preset response period, the feedback correction module sends a negative feedback signal to the spiking neural network model, causing the neuron firing threshold of the next cycle to decrease, and instructing the fluid system to increase the reference oil supply pressure. The feedback correction mechanism analyzes the drift phenomenon of the vibration signal on the frequency axis, recalculates the optimal alignment position of the injection angle using the fluid-structure interaction model, and fine-tunes the pointing vector of the micro-nozzle array.

9. The method for optimizing machining processes based on digital twins according to claim 8, characterized in that, The multiphysics digital twin and the physical processing system interact with each other using an edge computing architecture. All signal processing, flutter recognition and control command generation are completed on the local edge node. The spiking neural network model employs supervised learning during the training phase by combining offline simulation data with online experimental data. In the offline phase, a high-fidelity multiphysics simulation platform is used to generate processing condition samples. By adjusting the weights of synaptic connections between neurons, the model is classified into stable states, flutter budding, and severe flutter modes. During the online processing phase, the system captures real vibration data and marks it as learning samples, and uses the computing power of the edge nodes to perform micro-iterations on the local weights.

10. The method for optimizing machining processes based on digital twins according to claim 9, characterized in that, For thin-walled part processing, the multiphysics digital twin introduces a real-time material removal model, which updates the geometry of the virtual workpiece in real time and recalculates the natural frequency according to the interpolation instructions of the CNC system. The spiking neural network model introduces a lateral inhibition mechanism. When the input signal contains regular pulses that are consistent with the frequency multiple of the spindle rotation speed, the inhibitory neurons fire negative pulses to weaken the influence of some signals on the membrane potential. The formation position of the dynamic fluid damping barrier is dynamically reconstructed based on the spatial coherence analysis results of the vibration signal. The position of the antinode with the largest vibration displacement is identified by calculating the cross-correlation function between the signals of each sensor, and the injection center is aligned with the antinode position. The system uses the flow field evolution sub-model to correct the viscosity coefficient of the cutting fluid based on fluctuations in ambient temperature, and automatically adjusts the injection frequency to compensate for the damping loss caused by viscosity changes.

Images