Period-trend decoupling prediction control method and system of an industrial large model

Abstract

This invention relates to a periodic-trend decoupled predictive control method and system for large-scale industrial models, belonging to the field of machine tool digital control technology. The method includes constructing a multi-source heterogeneous data stream; generating a periodic feature tensor and a trend residual sequence; processing the periodic feature tensor through a depthwise separable convolutional network to extract the tool wear coefficient and spindle thermal drift coefficient, generating a machining surface quality index and a tool remaining life prediction range; solving a finite-time domain optimization problem with feed rate limiting and spindle torque constraints, generating the feed axis speed loop feedforward compensation and spindle speed correction; and finally outputting the result to the cyclic synchronous position mode control word of the servo driver through a safe shutdown path. This invention achieves periodic-trend dual-scale decoupling through angle domain resampling and adaptive variational mode decomposition, enabling safe and reliable execution of predictive control commands for spindle load in a five-axis machining center.

Description

Technical Field

[0001] This invention relates to the field of machine tool digital control technology, and more specifically, to a periodic-trend decoupling predictive control method and system for a large industrial model. Background Technology

[0002] As the high-end equipment manufacturing industry develops towards intelligence and precision, five-axis machining centers, as key manufacturing equipment in aerospace, automotive molds, and energy equipment fields, directly determine product quality and production costs through their machining accuracy and efficiency. Currently, the control technology of five-axis machining centers has evolved from traditional digital control (NC) to adaptive control and intelligent monitoring. Mainstream solutions include trajectory planning based on the PLCopen motion control library, cutting parameter optimization based on expert systems, and tool life prediction based on cloud-based big data.

[0003] In recent years, industrial large-scale models, as a cutting-edge direction for the deep integration of the industrial internet and artificial intelligence, are gradually penetrating from general fields to vertical industrial scenarios. Industrial large-scale models typically refer to deep learning models with hundreds of millions of parameters, possessing cross-device knowledge transfer and multi-source heterogeneous data fusion capabilities. Their typical architecture adopts a cloud-edge collaborative paradigm of "cloud pre-training - edge lightweight deployment": the cloud completes the model's pre-training and parameter initialization based on massive amounts of industrial data across machine models and operating conditions, forming a base model with strong generalization capabilities; the edge side uses technologies such as knowledge distillation, network pruning, and quantization compression to lightweight the hundreds of millions of parameters to the millions (…). The parameters, while satisfying millisecond-level real-time inference constraints, retain the core decision-making capabilities of the base model. This type of model achieves predictive analysis and optimized control of complex industrial processes by integrating equipment operating data, process parameters, and physical mechanisms.

[0004] However, existing technologies generally adopt a centralized architecture of "sensing-transmission-cloud analysis-feedback control," which suffers from problems such as high data transmission latency (hundreds of milliseconds), insufficient utilization of edge computing power, and a disconnect between prediction and control. Specifically, existing solutions fail to effectively utilize the edge lightweight deployment capabilities of large industrial models, resulting in the inability of the prediction capabilities of cloud-based large models to be implemented in real time into the control closed loop. Regarding the periodic fluctuations in spindle load and the trend drift of tool wear, existing technologies mostly employ spectrum analysis on a single time scale or offline calibration of fixed models, which is difficult to adapt to the variable working conditions such as speed fluctuations and material changes during machining. This leads to low prediction accuracy and control lag, ultimately manifesting as unstable surface quality and unexpected tool failure.

[0005] Furthermore, the application of existing large-scale industrial models in five-axis machining centers still faces the following bottlenecks: First, the computing power gap between cloud-based pre-trained models and edge controllers has not been effectively bridged, and there is a lack of model compression and inference acceleration solutions for millisecond-level control cycles; Second, the fusion mechanism of multimodal industrial data (such as current, vibration, temperature, and position signals) is not deeply coupled with the physical mechanism of machining, resulting in a lack of interpretability and security in model prediction; Third, the mapping relationship between the prediction output of large-scale models and motion control commands is unclear, and the "prediction" and "control" links are independent of each other, failing to form a closed-loop architecture of "prediction as control". Summary of the Invention

[0006] The purpose of this invention is to provide a periodic-trend decoupled predictive control method and system for large-scale industrial models to improve the aforementioned problems. To achieve this objective, the technical solution adopted by this invention is as follows: In a first aspect, this application provides a periodic-trend decoupled predictive control method for large-scale industrial models, deployed on a heterogeneous computing architecture of edge intelligent controllers, comprising the following steps: The fieldbus master station of the edge intelligent controller polls and collects the instantaneous value of the three-phase current of the electric spindle, the nanometer-level position feedback of the grating ruler, the vibration signal of the accelerometer, and the thermal radiation signal of the infrared temperature sensor through a distributed clock synchronization mechanism to construct a multi-source heterogeneous data stream; the multi-source heterogeneous data stream is then timestamped by the FPGA hardware and input into the real-time data pool of the programmable logic controller during operation. Based on the multi-source heterogeneous data stream in the real-time data pool, differential operation of grating ruler position feedback is performed to obtain the instantaneous spindle rotation speed. Then, the non-uniform time series is converted into a periodic stable sequence with a fixed rotation phase through angle domain resampling. Subsequently, it is input into an adaptive variational mode decomposition algorithm to separate the high-frequency cutting vibration component and the low-frequency tool wear trend, generating a periodic feature tensor and a trend residual sequence. The fixed phase step size of the angle domain resampling corresponds to a 1-degree spindle rotation, and the number of modes in the variational mode decomposition is adaptively determined by the spectral kurtosis index. The periodic feature tensor and the trend residual sequence are fed into the lightweight time series model deployed by the edge intelligent controller. The periodic feature tensor is processed by a depthwise separable convolutional network to extract the multi-scale vibration mode of the cutting cycle. The trend residual sequence is processed by a recursive least squares estimator to extract the tool wear coefficient and the spindle thermal drift coefficient. The dual-channel output is weighted and spliced ​​by the feature fusion layer to generate the machining surface quality index and the tool remaining life prediction range. Based on the machining surface quality index and the tool remaining life prediction range, the multi-axis collaborative controller of the programmable logic controller soft motion kernel is called. After evaluation by the weighted composite cost function of the cutting trajectory tracking term and the machining quality maintenance term, the finite time domain optimization problem with feed rate limit and spindle torque constraint is solved, and the feed axis speed loop feedforward compensation and spindle speed correction are generated. The speed loop feedforward compensation and spindle speed correction are transmitted to the dual-channel redundant comparator of the edge intelligent controller to perform fault detection and consistency verification. The output is mapped to the process data object of the fieldbus slave station through the safety shutdown path, and finally output to the cyclic synchronous position mode control word of the servo drive to complete the safe and reliable execution of the predictive control command.

[0007] Preferably, the heterogeneous computing architecture includes: The high-performance application processor core of the edge intelligent controller executes the adaptive variational mode decomposition algorithm and lightweight temporal model inference, the real-time coprocessor core executes the finite-time domain optimization solution of the multi-axis collaborative controller, the FPGA core executes hardware timestamp marking and preprocessing caching of multi-source heterogeneous data streams, and the neural network processing unit executes feature extraction of multi-scale dilated convolutional networks.

[0008] Preferably, the angle domain resampling includes: The position feedback of the grating ruler is used to obtain the instantaneous spindle speed through differential operation. Based on this, a time-rotation phase mapping function is constructed through numerical integration. The non-uniform time domain sampling points are converted into a uniform rotation phase sequence with a fixed phase interval by linear interpolation, generating a periodically stable sequence. The angle domain resampling eliminates the cutting force waveform distortion caused by spindle speed fluctuations and converts the time-varying speed signal into a standard cutting waveform with an invariant phase.

[0009] Preferably, the adaptive variational mode decomposition algorithm includes: The periodic stable sequence is fed into the adaptive variational mode decomposition algorithm. The variational constraint model is solved iteratively by the alternating direction multiplier method. The eigenmode functions with compact spectrum are extracted. The eigenmode function with the lowest center frequency is output as the trend residual sequence to characterize the asymptotic accumulation of tool wear and thermal drift. The remaining eigenmode functions are superimposed to generate a periodic feature tensor to characterize the periodic vibration impact of the cutting process.

[0010] Preferably, the periodic feature tensor is processed by a depthwise separable convolutional network including: The periodic feature tensor is fed into a multi-scale dilated convolutional network. After factorization operations of depthwise convolution and pointwise convolution, the multi-scale vibration mode of the cutting cycle is extracted to identify early signs of cutting chatter. The dilation rate of the multi-scale dilated convolutional network increases with the rotation phase, the convolution kernel size is configured according to the rotation phase, and the dilation factor is distributed in a power-law sequence of 2.

[0011] Preferably, the recursive least squares estimator includes: The trend residual sequence is input into the recursive least squares estimator, and a physical constraint loss function is constructed by combining the Taylor wear formula. The tool wear coefficient and spindle thermal drift coefficient with adaptive forgetting factor are identified online.

[0012] Preferably, the weighted composite cost function includes: The surface quality index and the predicted remaining tool life range are fed into a multi-axis collaborative controller. Through hierarchical iterative solutions by the cutting trajectory tracking sub-optimizer, the machining quality preservation sub-optimizer, and the tool axis vector smoothing sub-optimizer, the feed axis speed loop feedforward compensation, torque feedforward compensation, and spindle speed correction are generated. The cutting trajectory tracking sub-optimizer minimizes the tool path tracking error, the machining quality preservation sub-optimizer addresses the cutting vibration suppression constraints, and the tool axis vector smoothing optimizer addresses the penalty constraints on the tool axis rate of change. The prediction time domain is adaptively adjusted based on the tool tip curvature; in high curvature regions, the control time domain is shortened to 5-10 cutting cycles, while in stable cutting regions, the prediction time domain is extended to 20-50 cutting cycles.

[0013] Preferably, the dual-channel redundant comparator includes: speed loop feedforward compensation, torque feedforward compensation, and spindle speed correction transmitted to the dual-channel redundant comparator; the main channel uses a predicted compensation based on a lightweight timing model inference; the backup channel uses an empirical lookup table compensation based on cutting time; and Bayesian inference is performed to evaluate the redundancy output confidence based on the posterior probability distribution of the two channel outputs. When the confidence is higher than a threshold, a weighted fusion output is used; when the confidence is lower than the threshold, a degradation mode is triggered, forcibly reducing the feed rate to a safe feed rate. Once the confidence recovers to above the threshold, the system automatically returns to normal control mode. The posterior probability distribution is dynamically updated based on the model prediction uncertainty and the timeliness of the lookup table data.

[0014] Preferably, the safety shutdown path output includes: the output of the dual-channel redundant comparator performs dynamic reconstruction of the process data object mapping through the fieldbus controller, and immediately reconstructs the process data object mapping when an anomaly is detected, switching control to the backup programmable logic controller logic; the dynamic reconstruction of the process data object mapping utilizes the fieldbus hot-connect function, wherein the fault response time is less than 50μs.

[0015] Secondly, this application also provides a period-trend decoupled predictive control system for a large-scale industrial model, comprising: The data acquisition module: The fieldbus master station of the edge intelligent controller uses a distributed clock synchronization mechanism to poll and acquire the instantaneous value of the three-phase current of the electric spindle, the nanometer-level position feedback of the grating ruler, the vibration signal of the accelerometer, and the thermal radiation signal of the infrared temperature sensor to construct a multi-source heterogeneous data stream; the multi-source heterogeneous data stream is timestamped by the FPGA hardware and then input into the real-time data pool of the programmable logic controller during operation. The decomposition module is used to perform differential calculations on the position feedback of the grating ruler based on the multi-source heterogeneous data stream in the real-time data pool to obtain the instantaneous spindle speed. Then, it converts the non-uniform time series into a periodic stable sequence with a fixed rotation phase through angle domain resampling. Subsequently, it inputs the adaptive variational mode decomposition algorithm to separate the high-frequency cutting vibration component from the low-frequency tool wear trend, generating a periodic feature tensor and a trend residual sequence. The fixed phase step size of the angle domain resampling corresponds to a 1-degree spindle rotation, and the number of modes in the variational mode decomposition is adaptively determined by the spectral kurtosis index. Extraction module: used to feed the periodic feature tensor and the trend residual sequence into the lightweight time series model deployed by the edge intelligent controller respectively; the periodic feature tensor is processed by a depthwise separable convolutional network to extract the multi-scale vibration mode of the cutting cycle, and the trend residual sequence is processed by a recursive least squares estimator to extract the tool wear coefficient and spindle thermal drift coefficient. The dual-channel output is weighted and spliced ​​by the feature fusion layer to generate the machining surface quality index and the tool remaining life prediction range. Optimization module: Based on the machining surface quality index and the tool remaining life prediction range, it calls the multi-axis collaborative controller of the programmable logic controller soft motion kernel, evaluates the weighted composite cost function of cutting trajectory tracking and machining quality maintenance, solves the finite-time domain optimization problem with feed rate limit and spindle torque constraint, and generates the feed axis speed loop feedforward compensation and spindle speed correction. The safety execution module is used to transmit the speed loop feedforward compensation and spindle speed correction to the dual-channel redundant comparator of the edge intelligent controller, perform fault detection and consistency verification, output through the safety shutdown path, map to the process data object of the fieldbus slave station, and finally output to the cyclic synchronous position mode control word of the servo drive to complete the safe and reliable execution of predictive control commands.

[0016] Thirdly, this application also provides a period-trend decoupling predictive control device for a large-scale industrial model, comprising: Memory, used to store computer programs; A processor is used to implement the periodic-trend decoupled predictive control method of the industrial large model when executing the computer program.

[0017] Fourthly, this application also provides a readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described periodic-trend decoupling predictive control method based on a large industrial model.

[0018] The beneficial effects of this invention are as follows: This invention converts the non-uniform time series of time-varying rotation speed into a fixed rotational phase series by angle domain resampling, eliminating the influence of rotation speed fluctuations on periodic analysis; it adaptively determines the number of variational mode decomposition modes by using spectral kurtosis index, separates high-frequency cutting vibration components from low-frequency tool wear trends, and generates structured periodic feature tensors and trend residual sequences, providing high-quality feature inputs for subsequent lightweight model inference.

[0019] This invention achieves lightweight deployment and millisecond-level real-time inference of large industrial models at the edge by using task collaborative allocation in a heterogeneous computing architecture, combined with depthwise and pointwise convolution factorization operations of multi-scale dilated convolutional networks, and recursive estimation of Taylor wear formula physical constraints. This breakthrough overcomes the computing power bottleneck and ensures the stability of parameter estimation under non-steady-state conditions.

[0020] This invention achieves multi-objective collaborative optimization of trajectory accuracy, surface quality, and motion stability by using a weighted composite cost function of cutting trajectory tracking, machining quality maintenance, and tool axis vector smoothing, combined with the nonlinear decay of dynamic weights and adaptive adjustment in the prediction time domain. By generating feedforward compensation through finite time domain optimization, it realizes real-time mapping from predicted load to control commands.

[0021] This invention achieves dynamic switching between high-confidence weighted fusion and low-confidence degradation modes through dual-channel redundant confidence assessment using Bayesian inference; and achieves seamless switching of control during faults through dynamic reconstruction of process data object mapping and fieldbus hot-connection function, ensuring the safety integrity level requirements of predictive control commands.

[0022] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing embodiments of the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings. Attached Figure Description

[0023] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a schematic diagram of the cycle-trend decoupling predictive control method for the industrial large model described in this embodiment of the invention. Figure 2 This is a schematic diagram of the periodic-trend decoupled predictive control system structure of the industrial large model described in this embodiment of the invention; Figure 3 This is a schematic diagram of the structure of the periodic-trend decoupling predictive control device for the industrial large-scale model described in this embodiment of the invention.

[0025] In the diagram: 701, Acquisition Module; 702, Decomposition Module; 703, Extraction Module; 704, Optimization Module; 705, Safe Execution Module; 800, Period-Trend Decoupling Predictive Control Device for Industrial Large Model; 801, Processor; 802, Memory; 803, Multimedia Component; 804, I / O Interface; 805, Communication Component. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0027] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this invention, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0028] Example 1:

[0029] This embodiment provides a period-trend decoupling predictive control method for large industrial models.

[0030] See Figure 1 The figure shows that the method includes steps S100, S200, S300, S400 and S500.

[0031] The S100 and the fieldbus master station of the edge intelligent controller poll and collect the instantaneous value of the three-phase current of the electric spindle, the nanometer-level position feedback of the grating ruler, the vibration signal of the accelerometer, and the thermal radiation signal of the infrared temperature sensor through a distributed clock synchronization mechanism to construct a multi-source heterogeneous data stream. After being timestamped by the FPGA hardware, the multi-source heterogeneous data stream is input into the real-time data pool of the programmable logic controller during operation.

[0032] It is understandable that the heterogeneous computing architecture in this step S100 includes: The high-performance application processor core of the edge intelligent controller executes the adaptive variational mode decomposition algorithm and lightweight temporal model inference, the real-time coprocessor core executes the finite-time domain optimization solution of the multi-axis collaborative controller, the FPGA core executes hardware timestamp marking and preprocessing caching of multi-source heterogeneous data streams, and the neural network processing unit executes feature extraction of multi-scale dilated convolutional networks.

[0033] It should be noted that this embodiment employs an edge intelligent controller with a heterogeneous computing architecture to achieve lightweight deployment and real-time inference of large industrial models at the edge. The controller integrates an FPGA programmable logic unit, an ARM application processor, an ARM real-time coprocessor, and a neural network processing unit (NPU), forming a complete hardware closed loop of "acquisition-inference-control-execution" through an on-chip high-speed interconnect bus. The FPGA core undertakes the preprocessing of periodic-trend decoupling for multi-source heterogeneous data streams. Through EtherCAT distributed clock synchronization, it achieves 250μs periodic synchronous acquisition and 64-bit timestamp marking of electric spindle current, grating ruler position, vibration acceleration, and infrared temperature signals. The FPGA internally implements an angle domain resampling hardware acceleration engine to convert the non-uniform time series of time-varying rotation speed into a periodically stable sequence with a fixed rotation phase, providing structured data input for subsequent variational mode decomposition (VMD).

[0034] The ARM Cortex-A53 application processor is used to deploy a lightweight dual-channel inference architecture for a large industrial model. Periodic feature tensors are processed by an NPU-accelerated multi-scale dilated convolutional network to extract multi-scale vibration modes of the cutting cycle; trend residual sequences are processed by a physically constrained recursive least squares estimator to extract tool wear coefficients and spindle thermal drift coefficients. The dual-channel outputs are weighted and concatenated by a feature fusion layer to generate the machining surface quality index and the tool remaining life prediction range. The number of model parameters is compressed to 10% of the original large model, and the inference latency is controlled within 5ms.

[0035] Among them, the ARM Cortex-R5 real-time coprocessor performs rolling time-domain optimization of model predictive control (MPC), constructs a three-term weighted composite cost function based on the prediction interval for cutting trajectory tracking, machining quality maintenance, and tool axis vector smoothing, solves the finite-time-domain optimization problem with feed rate limiting, spindle torque constraint, and tool axis rate of change constraint, and generates speed loop feedforward compensation, torque feedforward compensation, and spindle speed correction.

[0036] The safe execution channel is implemented through a dual-channel redundant comparator integrated in the FPGA. The main channel uses a lightweight timing model inference predictive compensation, while the backup channel uses an empirical lookup table compensation based on cutting time. Based on Bayesian inference, the redundancy output confidence assessment is performed. Through the process data object mapping dynamic reconstruction mechanism of the fieldbus controller, the output is sent to the cyclic synchronous position mode control word of the servo driver to complete the safe and reliable execution of predictive control commands.

[0037] Multi-source heterogeneous data streams are timestamped by the FPGA hardware and then input into the real-time data pool of the programmable logic controller during runtime. The periodic stable sequences in the data pool are adaptively decomposed by VMD to separate the high-frequency cutting vibration components and the low-frequency tool wear trend, generating periodic feature tensors and trend residual sequences, which are then fed into the dual-channel inference branches of the lightweight timing model.

[0038] The periodic branch employs a cascaded structure of depthwise separable convolution and dilated convolution, with the dilation rate distributed as a power of 2 with respect to the rotation phase. The input sequence has a length of 1024 sampling points, extracting multi-scale periodic patterns. The trend branch uses a recursive least squares estimator, combined with the Taylor wear formula to construct a physical constraint loss function, and an adaptive forgetting factor that decays nonlinearly according to the processed surface quality index, achieving online identification of the wear coefficient and real-time compensation for thermal drift.

[0039] The dual-channel outputs are weighted and spliced ​​by a feature fusion layer. The weight coefficients are dynamically adjusted through a gating mechanism to generate a predicted range for the machining surface quality index and the remaining tool life. The predicted range is input into the MPC kernel and evaluated by a weighted composite cost function of the cycle tracking term and the trend maintenance term. The dynamic weight coefficients decay exponentially according to the machining surface quality index. The prediction time domain is adaptively adjusted according to the curvature of the tool tip, shortening the high curvature region to 5-10 cutting cycles and extending the stable cutting region to 20-50 cutting cycles.

[0040] The feedforward compensation value generated by the optimized solution is evaluated using Bayesian confidence level by a dual-channel redundant comparator. When the confidence level is above a threshold, a weighted fusion output is used; when the confidence level is below the threshold, a degradation mode is triggered, forcibly reducing the feed rate to a safe feed rate. Once the confidence level recovers, the system automatically returns to normal control mode. The final output is dynamically reconstructed via process data object mapping and switched to backup safety logic via fieldbus hot-connection. The fault response time is less than 50μs, meeting the safety integrity level requirements of predictive control.

[0041] S200: Based on the multi-source heterogeneous data stream in the real-time data pool, perform differential operation on the grating ruler position feedback to obtain the instantaneous spindle rotation speed. Then, convert the non-uniform time series into a periodic stable sequence with a fixed rotation phase through angle domain resampling. Subsequently, input the adaptive variational mode decomposition algorithm to separate the high-frequency cutting vibration component and the low-frequency tool wear trend, and generate a periodic feature tensor and a trend residual sequence. The fixed phase step size of the angle domain resampling corresponds to a 1-degree spindle rotation, and the number of modes in the variational mode decomposition is adaptively determined by the spectral kurtosis index.

[0042] It is understandable that angle domain resampling in this step S200 includes: The position feedback of the grating ruler is used to obtain the instantaneous spindle speed through differential operation. Based on this, a time-rotation phase mapping function is constructed through numerical integration. The non-uniform time domain sampling points are converted into a uniform rotation phase sequence with a fixed phase interval by linear interpolation, generating a periodically stable sequence. The angle domain resampling eliminates the cutting force waveform distortion caused by spindle speed fluctuations and converts the time-varying speed signal into a standard cutting waveform with an invariant phase.

[0043] It should be noted that angle-domain resampling technology is used to convert the non-uniform time series of time-varying rotational speeds into a uniform angle series with a fixed rotational phase, thus eliminating the influence of rotational speed fluctuations on period analysis. The specific implementation process is as follows: The position feedback value p(t) of the grating ruler is used to calculate the instantaneous angular velocity ω(t) through numerical differentiation. To avoid amplification of differential noise, a five-point center difference formula is used. The angular displacement is then calculated through numerical integration and accumulated using the trapezoidal rule. The calculation formula is as follows:

[0044] In the formula, Let be the cumulative rotation angle of the principal axis at time t. Initial time Initial value of the spindle rotation angle. Let t be the starting time of integration, t be the current time, and ω(τ) be the instantaneous angular velocity of the principal axis at time τ. Let N be the time differential element, and N be the total number of sampling points within the integration interval. Let be the instantaneous angular velocity of the principal axis at the i-th sampling moment. Let be the instantaneous angular velocity of the principal axis at the (i-1)th sampling time. The time interval between the i-th sampling point and the (i-1)-th sampling point.

[0045] The initial value for integration is set to the starting phase (0 degrees) of the current rotation cycle of the spindle. During integration, θ automatically returns to zero when it exceeds 360 degrees, achieving continuous tracking of the rotation cycle. Angle-domain resampling establishes a uniform phase grid in the θ domain, with a phase step size... The corresponding time index is found by using the inverse function t(θ) of the time-phase mapping function, and then the signal value of that phase is calculated by linear interpolation.

[0046] The adaptive variational mode decomposition algorithm in this step includes: The periodically stable sequence is fed into an adaptive variational mode decomposition algorithm. The variational constraint model is solved iteratively using the alternating direction multiplier method. Eigenmode functions with compact spectra are extracted. The eigenmode function with the lowest center frequency is output as the trend residual sequence, representing the asymptotic accumulation of tool wear and thermal drift. The remaining eigenmode functions are superimposed to generate a periodic feature tensor, representing the periodic vibration and impact of the cutting process. The variational constraint model is executed according to the following optimization objective:

[0047] In the formula, Let k be the intrinsic mode function. To correspond to the center frequency, For data fidelity weighting coefficients, For the Dirac function, The imaginary unit, It is a periodically stable sequence. For convolution operation, k is the total number of intrinsic mode functions. Let t be the time differential operator, t be the time variable, and e be the base of the natural logarithm.

[0048] An adaptive variational mode decomposition algorithm is used to input a periodically stable sequence. The variational constraint model is solved iteratively using the alternating direction multiplier method to extract eigenmode functions with compact spectra. The eigenmode function with the lowest center frequency is output as the trend residual sequence, representing the asymptotic accumulation of tool wear and thermal drift. The remaining eigenmode functions are superimposed to generate a periodic feature tensor, representing the periodic vibration and impact of the cutting process.

[0049] The data fidelity weighting coefficients and modal frequency interval penalty coefficients of the variational constraint model are adaptively adjusted according to the spectral kurtosis index. This enhances data fidelity when the signal is stable and enhances modal separation when the signal is highly impulsive. The iterative convergence threshold is set to... It typically converges after 20-30 iterations. Angle domain resampling eliminates spectral leakage caused by rotation speed fluctuations, and VMD adaptive decomposition achieves precise decoupling of periodic and trend components, providing structured feature inputs for edge inference of lightweight industrial large models. The accuracy of period recognition is improved by 40%, and modal aliasing is reduced by 60%.

[0050] S300 feeds the periodic feature tensor and the trend residual sequence into the lightweight time series model deployed by the edge intelligent controller. The periodic feature tensor is processed by a depthwise separable convolutional network to extract the multi-scale vibration mode of the cutting cycle. The trend residual sequence is processed by a recursive least squares estimator to extract the tool wear coefficient and the spindle thermal drift coefficient. The dual-channel output is weighted and spliced ​​by the feature fusion layer to generate the machining surface quality index and the tool remaining life prediction range.

[0051] It is understandable that the periodic feature tensor processed by the depthwise separable convolutional network in this step S300 includes: The periodic feature tensor is fed into a multi-scale dilated convolutional network. After factorization operations of depthwise convolution and pointwise convolution, the multi-scale vibration mode of the cutting cycle is extracted to identify early signs of cutting chatter. The dilation rate of the multi-scale dilated convolutional network increases with the rotation phase, the convolution kernel size is configured according to the rotation phase, and the dilation factor is distributed in a power-law sequence of 2.

[0052] It should be noted that the periodic feature tensor is fed into a multi-scale dilated convolutional network, and the multi-scale vibration modes of the cutting cycle are extracted through factorization operations of depthwise convolution and pointwise convolution. The network contains 4 layers of depthwise separable convolutions, the dilation rate is distributed in a power-of-2 sequence with respect to the rotation phase, the convolution kernel size is configured according to the rotation phase, the number of parameters is reduced by 85%-90% compared to standard convolution, and the inference computation is reduced to 10%-15% of the number of floating-point operations.

[0053] The network consists of four depthwise separable convolutional layers, with the dilation rate distributed in a power-law sequence (1, 2, 4, 8) with respect to the rotation phase. The input sequence has a length of 1024 sampling points. Each layer is followed by a batch normalization layer and gated linear unit activation, and gradient vanishing is mitigated through residual connections. Finally, a 128-dimensional feature vector is output after global average pooling, representing the multi-scale vibration modes of the cutting cycle.

[0054] The network training employs a knowledge distillation strategy: a large, cloud-deployed model (100M parameters) serves as the teacher model, while a lightweight, edge-based model (8M parameters) acts as the student model. Joint optimization through soft-label distillation and feature distillation enables the student model to learn the decision boundaries and intermediate layer feature distributions of the teacher model. The training data consists of 200,000 sets of cutting experiment data, covering typical cutting conditions for various materials such as aluminum alloys, titanium alloys, and high-temperature alloys. The quantization stage uses 8-bit weighted quantization, and the activation values ​​employ dynamic range quantization, with inference latency controlled within 5ms.

[0055] The residual trend sequence is fed into a recursive least squares estimator to perform online identification of the tool wear coefficient and spindle thermal drift coefficient. Traditional recursive least squares is purely data-driven, and parameter estimation diverges during unsteady cutting. This embodiment introduces the physical constraint of Taylor's wear formula, integrating data-driven approaches with physical laws.

[0056] Taylor's wear formula describes the relationship between the tool flank wear rate and cutting parameters. It is expressed as: wear rate equals the material wear coefficient multiplied by a power of the cutting speed, then multiplied by a power of the feed rate. The material wear coefficient, speed exponent, and feed rate exponent in this formula are pre-calibrated through cutting experiments, with parameter libraries established for titanium alloys, aluminum alloys, and high-temperature alloys.

[0057] It should be noted that the physical constraint is implemented through an augmented loss function: the total loss equals the data fitting loss of recursive least squares plus the physical constraint loss. The physical constraint loss is the squared Euclidean distance between the estimated wear rate and the rate predicted by Taylor formula. This constraint is embedded as a soft constraint in the recursive update formula, and the parameter estimation direction is adjusted through gradient descent to ensure that the identification results conform to the physical laws of cutting.

[0058] Understandably, the adaptive forgetting factor decays non-linearly according to the processed surface quality index: a high quality index results in a large forgetting factor, enhancing the stability of the estimator; a low quality index results in a small forgetting factor, enhancing the ability to track the wear acceleration trend. The adaptive physical constraint weights are dynamically adjusted according to the processed surface quality index: a high quality index results in a small physical constraint weight, trusting data-driven decisions; a low quality index results in a large physical constraint weight, strengthening the constraints of physical laws.

[0059] Furthermore, the recursive least squares estimator includes: The trend residual sequence is input into a recursive least squares estimator, and a physical constraint loss function is constructed by combining it with the Taylor wear formula. Online identification of the tool wear coefficient and spindle thermal drift coefficient with an adaptive forgetting factor is then performed. The physical constraint term of the Taylor wear formula is as follows:

[0060] In the formula, For tool wear rate, Where v is the material wear coefficient and v is the cutting speed. is the cutting speed exponent, f is the feed rate, and b is the feed rate exponent; The formula for calculating the physical constraint loss function is as follows:

[0061] In the formula, The physical constraint loss function characterizes the deviation between the estimated wear rate and the Taylor formula prediction. To estimate the tool wear rate, The tool wear rate predicted by Taylor's formula. The square of the Euclidean distance; The recursive update formula for the tool wear coefficient is as follows:

[0062] In the formula, Let be the vector of tool wear coefficient and thermal drift coefficient at time t. Let be the vector of tool wear coefficient and thermal drift coefficient at time t-1. Let be the covariance matrix at time t-1. For the regression vector, An adaptive forgetting factor. This is the transpose of the regression vector. The residual sequence value of the current trend. For adaptive physical constraint weights, The gradient of the physical constraint loss function with respect to the wear coefficient vector, and the adaptive physical constraint weights. The value is dynamically adjusted according to the surface quality index, with a range of 0.01-0.1.

[0063] The dual-channel outputs are weighted and concatenated by a feature fusion layer, with the weight coefficients dynamically adjusted via a gating mechanism to generate the machining surface quality index and the tool remaining life prediction range. The 128-dimensional feature vector from the periodic channel output and the wear coefficient vector from the trend channel output are weighted and concatenated by the feature fusion layer. The fusion layer uses a gating mechanism to dynamically adjust the dual-channel weights: the periodic channel has a higher weight when the cutting process is stable, and the trend channel has a higher weight when there is chatter or abnormality. The fusion result is mapped through a two-layer fully connected network to generate the machining surface quality index and the tool remaining life prediction range. The machining surface quality index is a continuous value between 0 and 1, with values ​​above 0.90 corresponding to a surface roughness Ra of less than 0.8 micrometers for precision machining quality. The tool remaining life prediction range is a statistical confidence interval, calibrated using a conformal prediction mechanism: utilizing the residual distribution of historical machining data to generate upper and lower bounds with 95% confidence, rather than simple point estimation.

[0064] In other words, knowledge distillation and quantization compression enable lightweight deployment of large industrial models at the edge, physical constraint embedding solves the drift problem of pure data-driven models during unsteady cutting, and dual-channel fusion enables collaborative prediction of machining quality and tool life, with prediction accuracy improved by more than 30% compared to pure data-driven models.

[0065] S400, based on the machining surface quality index and the tool remaining life prediction range, calls the multi-axis collaborative controller of the programmable logic controller soft motion kernel, evaluates the weighted composite cost function of cutting trajectory tracking term and machining quality maintenance term, solves the finite time domain optimization problem with feed rate limit and spindle torque constraint, and generates the feed axis speed loop feedforward compensation amount and spindle speed correction amount.

[0066] It is understood that the weighted composite cost function described in S400 in this step includes: The surface quality index and the predicted remaining tool life range are fed into a multi-axis collaborative controller. Through hierarchical iterative solutions by the cutting trajectory tracking sub-optimizer, the machining quality preservation sub-optimizer, and the tool axis vector smoothing sub-optimizer, the feed axis speed loop feedforward compensation, torque feedforward compensation, and spindle speed correction are generated. The cutting trajectory tracking sub-optimizer minimizes the tool path tracking error, the machining quality preservation sub-optimizer addresses the cutting vibration suppression constraints, and the tool axis vector smoothing optimizer addresses the penalty constraints on the tool axis rate of change. The prediction time domain is adaptively adjusted based on the tool tip curvature; in high curvature regions, the control time domain is shortened to 5-10 cutting cycles, while in stable cutting regions, the prediction time domain is extended to 20-50 cutting cycles. The weighted composite cost function is calculated using the following formula:

[0067] In the formula, For the weighted composite cost function, For dynamic weighting coefficients, For cutting trajectory tracking, For processing quality maintenance items, This is the tool axis smoothing weighting coefficient. For the tool axis vector smoothing term, For adaptive control in the time domain, t is the current time. To control the sampling point index in the time domain, Let be the predicted value at the current time t with respect to the future time t+c. This represents the predicted control increment at time t with respect to a future time t+c. To control the increment, S is the weight matrix. It is the square of the Euclidean distance.

[0068] It should be noted that the cutting trajectory tracking item penalizes the deviation between the actual tool path and the reference path, using the sum of squared Euclidean distances between the predicted and reference positions. The machining quality maintenance item penalizes the degree to which the machined surface quality index deviates from the target value, with the target value set at 0.90 to correspond to precision machining requirements. The tool axis vector smoothing item penalizes the rate of change of the unit vector of the tool axis to avoid sudden changes in cutting force caused by abrupt changes in the tool axis.

[0069] The three terms are weighted and summed using dynamic weighting coefficients. The total cost function is calculated by multiplying the trajectory tracking term by the dynamic weighting coefficient, adding the quality maintenance term multiplied by the complement of the dynamic weights, adding the tool axis smoothing term multiplied by the smoothing weighting coefficient, and finally adding a weighted penalty term for controlling the increment. The dynamic weighting coefficient decays exponentially according to the machining surface quality index: a high quality index results in a high weight for trajectory tracking, prioritizing efficiency; a low quality index results in a high weight for quality maintenance, prioritizing accuracy. The smoothing weighting coefficient is adaptively adjusted according to the amplitude of the tool axis change rate: when the change rate exceeds a threshold, the weight is increased to suppress sudden changes in the tool axis.

[0070] The prediction time domain is adaptively adjusted based on the tool tip curvature. The tool tip curvature is obtained through parsing the tool position file pre-read by the CAM software, or calculated in real time based on the position feedback of the grating ruler. Curvature greater than 0.1 mm is considered a high curvature region (such as mold corners, abrupt changes in surface), and the prediction time domain is shortened to 5 to 10 cutting cycles to improve response speed; curvature less than 0.01 mm is considered a smooth cutting region (such as planes, straight walls), and the prediction time domain is extended to 20 to 50 cutting cycles to improve optimization quality; linear interpolation is used in the intermediate region. The control time domain is always one-third to one-fifth of the prediction time domain to reduce the number of optimization variables and accelerate the solution. Constraints include: feed rate limit of ±20% of the rated feed to prevent abrupt changes in cutting load; spindle torque limit of 90% of the rated torque to retain a safety margin; and tool axis rate of change limit of 0.5 radians per second to suppress abrupt changes in tool axis. The interior point method is used to handle constraints, and inequality constraints are embedded into the cost function through a barrier function, automatically maintaining constraint satisfaction during iteration.

[0071] Furthermore, the optimization process generates feedforward compensation for the feed axis's speed loop, torque feedforward compensation, and spindle speed correction. The speed loop feedforward compensation is directly superimposed on the servo driver's speed command, achieving proactive compensation for load disturbances; the torque feedforward compensation is mapped to the current loop torque command, offsetting cutting force fluctuations; and the spindle speed correction finely adjusts the spindle speed to adapt to changes in material hardness. These three compensation values ​​are output synchronously via EtherCAT cycles, enabling real-time execution of predictive control commands. The three weighted composite cost functions achieve multi-objective collaborative optimization of trajectory accuracy, machining quality, and motion smoothness. This reduces cutting force fluctuations caused by sudden tool axis changes during five-axis simultaneous machining of complex surfaces, and lowers surface roughness Ra by 1 to 2 levels. The adaptive prediction time domain improves response speed by half in high curvature regions and optimizes quality by 25% in stable regions. Model pre-compensation directly converts prediction results into control commands, achieving a closed loop of "prediction as control," rather than the traditional open loop of "prediction-alarm-manual intervention."

[0072] The S500 transmits the speed loop feedforward compensation and spindle speed correction to the dual-channel redundant comparator of the edge intelligent controller to perform fault detection and consistency verification. The output is mapped to the process data object of the fieldbus slave station through the safety shutdown path, and finally output to the cyclic synchronous position mode control word of the servo drive to complete the safe and reliable execution of the predictive control command.

[0073] It is understood that in this step, the dual-channel redundant comparator in S500 includes: speed loop feedforward compensation, torque feedforward compensation, and spindle speed correction, which are transmitted to the dual-channel redundant comparator. The main channel uses a predicted compensation amount based on a lightweight timing model inference, and the backup channel uses an empirical lookup table compensation amount based on cutting time. Based on the posterior probability distribution of the two channel outputs, Bayesian inference is performed to evaluate the confidence level of the redundant output. When the confidence level is higher than the threshold, a weighted fusion output is used. When the confidence level is lower than the threshold, a degradation mode is triggered, forcibly reducing the feed rate to a safe feed rate. After the confidence level recovers to above the threshold, it automatically recovers to the normal control mode. The posterior probability distribution is dynamically updated according to the model prediction uncertainty and the timeliness of the lookup table data.

[0074] It should be noted that the dual-channel redundant comparator receives the model prediction compensation from the main channel and the empirical lookup compensation from the backup channel. The uncertainty of the main channel is estimated using the Monte Carlo temporary retirement method: 20 forward inferences are performed, with a portion of neurons randomly retired each time; the mean of the output is used as the predicted value, and the standard deviation as the uncertainty. The timeliness of the backup channel increases linearly with the time since the last calibration; the longer the time, the worse the timeliness. Based on the Gaussian distribution assumption, the posterior probabilities of the two channels are calculated: the posterior probability of the model prediction channel is proportional to the reciprocal of the model prediction uncertainty, and the posterior probability of the empirical lookup channel is proportional to the reciprocal of the timeliness of the lookup data. After normalizing the posterior probabilities, the confidence level is defined as the maximum of the two channel posterior probabilities. When the confidence level is higher than the threshold of 0.85, a weighted fusion output is used, with the weight being the posterior probability; when the confidence level is lower than the threshold, a degradation mode is triggered. In degradation mode, the feed rate is forcibly reduced to half of the rated feed, preserving cutting capability while reducing risk, and an attempt to recover is made every 5 seconds. During recovery, the confidence level is reassessed. If the quality index rises or the model uncertainty decreases, causing the confidence level to recover above the threshold, the system will automatically return to normal control mode. If the confidence level remains low, a safety shutdown will be triggered.

[0075] The output of the dual-channel redundant comparator is dynamically reconfigured via the fieldbus controller to perform process data object mapping. In normal mode, control resides with the master controller, and the process data object mapping is configured to output model prediction compensation values ​​to the servo driver. Upon detecting an anomaly, the master station writes a new process data object mapping configuration to the slave station via the CoE protocol, triggering the slave station to reconfigure and switch control to the backup safety programmable logic controller (PLC) logic. Dynamic reconfiguration utilizes the fieldbus hot-connect function, with configuration parameters including checking if the idle flag is true and setting the restart flag to true when invalid, ensuring the reliability of the reconfiguration. The fault response time from the detection of an anomaly to the servo enable signal cutoff is a measured maximum of 32 microseconds, meeting the safety integrity level requirement of less than 50 microseconds.

[0076] It should be noted that the industrial big data platform interface of the edge intelligent controller receives machine tool digital twin parameters from the cloud through a unified architecture connection service. The digital twin includes a geometric model, kinematic model, dynamic model, and material property library, with material property coefficients of the Taylor wear formula calibrated for aluminum alloys, titanium alloys, and high-temperature alloys respectively. The inner loop of the meta-learning performs 5 to 10 rounds of gradient descent fine-tuning on the current processing batch data, updating the convolution kernel weights of the periodic feature tensor; the outer loop performs meta-parameter optimization across processing batches, with the learning rate scaling inversely proportional to the number of local samples at the edge, increasing the learning rate to accelerate convergence when samples are scarce. The adapted model inherits the periodic-trend decoupled prior knowledge shared across batches, while adapting to the differences in cutting characteristics of the current material.

[0077] In other words, the dual-channel soft redundancy architecture based on Bayesian inference achieves the quantification and dynamic fusion of model uncertainty, reducing the false downtime rate by 80% compared to a fixed threshold, while maintaining the required safety integrity level. The degradation mode and automatic recovery mechanism resolve the "safety-availability" contradiction, preventing frequent downtime from affecting processing continuity.

[0078] The safety shutdown path output includes: the output of the dual-channel redundant comparator performs dynamic reconstruction of the process data object mapping through the fieldbus controller; when an anomaly is detected, the process data object mapping is immediately reconstructed, and control is switched to the backup programmable logic controller logic; the dynamic reconstruction of the process data object mapping utilizes the fieldbus hot-connection function, wherein the fault response time is less than 50μs.

[0079] This invention deploys a heterogeneous computing architecture on an edge intelligent controller. It achieves decoupling of periodic and trend scales through angle domain resampling and adaptive variational mode decomposition. It enables collaborative prediction of machining quality and tool life through dual-channel edge inference of a lightweight industrial large model. It achieves real-time mapping from predicted load to control commands through rolling time-domain optimization of model predictive control. It achieves safe and reliable execution through dual-channel redundancy comparison and dynamic reconstruction of process data object mapping using Bayesian inference. This invention pushes the predictive capabilities of the industrial large model down to the edge, realizing a closed loop of "prediction as control". It solves the technical problems of high latency, wasted edge computing power, and disconnect between prediction and control in traditional centralized architectures, while meeting the stringent requirements of high-end equipment manufacturing for security and availability.

[0080] Example 2:

[0081] like Figure 2 As shown, this embodiment provides a period-trend decoupled predictive control system for a large-scale industrial model. See [link to relevant documentation]. Figure 2 The system includes: Acquisition module 701: The fieldbus master station of the edge intelligent controller polls and acquires the instantaneous value of the three-phase current of the electric spindle, the nanometer-level position feedback of the grating ruler, the vibration signal of the accelerometer, and the thermal radiation signal of the infrared temperature sensor through a distributed clock synchronization mechanism to construct a multi-source heterogeneous data stream; the multi-source heterogeneous data stream is timestamped by the FPGA hardware and then input into the real-time data pool of the programmable logic controller during operation. Decomposition module 702: Based on the multi-source heterogeneous data stream in the real-time data pool, it performs differential operation on the grating ruler position feedback to obtain the instantaneous spindle rotation speed. Then, it converts the non-uniform time series into a periodic stable sequence with a fixed rotation phase through angle domain resampling. Subsequently, it inputs an adaptive variational mode decomposition algorithm to separate the high-frequency cutting vibration component from the low-frequency tool wear trend, generating a periodic feature tensor and a trend residual sequence. The fixed phase step size of the angle domain resampling corresponds to a 1-degree spindle rotation, and the number of modes in the variational mode decomposition is adaptively determined by the spectral kurtosis index. Extraction module 703: used to feed the periodic feature tensor and the trend residual sequence into the lightweight time series model deployed by the edge intelligent controller respectively; the periodic feature tensor is processed by a depthwise separable convolutional network to extract the multi-scale vibration mode of the cutting cycle, and the trend residual sequence is processed by a recursive least squares estimator to extract the tool wear coefficient and the spindle thermal drift coefficient. The dual-channel output is weighted and spliced ​​by the feature fusion layer to generate the machining surface quality index and the tool remaining life prediction range. Optimization module 704: Based on the machining surface quality index and the tool remaining life prediction range, it calls the multi-axis collaborative controller of the programmable logic controller soft motion kernel, evaluates the weighted composite cost function of cutting trajectory tracking and machining quality maintenance, solves the finite time domain optimization problem with feed rate limit and spindle torque constraint, and generates the feed axis speed loop feedforward compensation and spindle speed correction. Safety Execution Module 705: This module transmits the speed loop feedforward compensation and spindle speed correction to the dual-channel redundant comparator of the edge intelligent controller, performs fault detection and consistency verification, outputs through the safety shutdown path, maps to the process data object of the fieldbus slave station, and finally outputs to the cyclic synchronous position mode control word of the servo drive to complete the safe and reliable execution of predictive control commands.

[0082] Specifically, this embodiment completes the overall machine deployment verification on a five-axis gantry machining center. The machine tool has a travel of 3 meters on the X-axis, 2 meters on the Y-axis, and 1.5 meters on the Z-axis. The A / C oscillating head travel is ±120 degrees, the maximum spindle speed is 24,000 rpm, and the maximum power is 45 kW. The control system uses the edge intelligent controller of this patent to replace the original CNC system of the machine tool, retaining the servo drive and the mechanical body to realize a closed loop of "predictive control-servo execution". The edge intelligent controller is installed in the machine tool electrical cabinet, fixed by DIN rail, with a fanless heat dissipation design and an IP54 protection rating. The fieldbus network uses shielded twisted pair cable, with a linear topology, a total length of 80 meters, and includes 6 servo drives and 4 remote input / output modules. Multi-source sensors are deployed according to the aforementioned scheme: current sensors are installed at the output end of the frequency converter, grating rulers are installed on each linear axis and the oscillating head axis, acceleration sensors are magnetically attached to the spindle housing and the oscillating head bearing seat, and temperature sensors are infrared-aimed at the outer ring of the spindle bearing.

[0083] Specifically, the verification of typical processing conditions is as follows: Working Condition 1: Machining of Titanium Alloy Impeller Surface The impeller blades are complex free-form surfaces with dramatic curvature changes, and the material is a difficult-to-machine titanium alloy. Traditional CNC systems use a fixed feed rate, resulting in overcutting at corners and noticeable blade surface vibration marks. This embodiment's system adaptively adjusts the prediction time domain based on the tool tip curvature, shortening the high curvature region to 8 cycles and increasing the tool axis smoothing weight to 0.25, achieving a smooth tool axis transition at corners. The measured surface roughness Ra decreased from 1.6 micrometers to 0.6 micrometers, and the machining time was reduced by 15%.

[0084] Working Condition 2: Deep Cavity Machining of Aluminum Alloy Molds The deep cavity of the mold contains a large number of uncut and intermittent cuts, leading to frequent false alarms from traditional data-driven models. This embodiment uses physical constraint embedding to ensure that the wear rate is zero during the uncutting period, preventing parameter estimation divergence; the constraints are automatically released during the cutting period to track actual wear. The measured tool life prediction error has been reduced from ±30% to ±8%, and tool failure is warned 2 hours in advance, preventing workpiece scrap.

[0085] Working Condition 3: Machining of High-Temperature Alloy Turbine Disks The turbine disk is made of a nickel-based high-temperature alloy, subject to high cutting forces and temperatures, posing stringent safety and integrity requirements. This embodiment utilizes a dual-channel redundant comparator with Bayesian confidence assessment to automatically trigger a degradation mode when the spindle bearing temperature abnormally rises. The feed rate is reduced to a safe value, and the system dynamically reconstructs and switches to backup safety logic via process data object mapping, safely shutting down after completing the current cutting cycle. The measured fault response time is 28 microseconds, meeting the safety requirements for aerospace component machining.

[0086] It should be noted that the specific methods by which each module performs operations in the system described in the above embodiments have been described in detail in the embodiments related to the method, and will not be elaborated here.

[0087] Example 3:

[0088] Corresponding to the above method embodiments, this embodiment also provides a periodic-trend decoupling predictive control device for a large industrial model. The periodic-trend decoupling predictive control device for a large industrial model described below can be referred to in correspondence with the periodic-trend decoupling predictive control method for a large industrial model described above.

[0089] Figure 3 This is a block diagram illustrating a cycle-trend decoupling predictive control device 800 for a large-scale industrial model, according to an exemplary embodiment. Figure 3As shown, the industrial large-scale model's cycle-trend decoupled predictive control device 800 includes a processor 801 and a memory 802. The industrial large-scale model's cycle-trend decoupled predictive control device 800 also includes one or more of a multimedia component 803, an I / O interface 804, and a communication component 805.

[0090] The processor 801 controls the overall operation of the industrial large-scale model's periodic-trend decoupling predictive control device 800 to complete all or part of the steps in the aforementioned industrial large-scale model's periodic-trend decoupling predictive control method. The memory 802 stores various types of data to support the operation of the industrial large-scale model's periodic-trend decoupling predictive control device 800. This data may include, for example, instructions for any application or method operating on the industrial large-scale model's periodic-trend decoupling predictive control device 800, as well as application-related data such as contact data, sent and received messages, images, audio, video, etc. The memory 802 can be implemented using any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The multimedia component 803 may include a screen and an audio component. The screen may be, for example, a touchscreen, and the audio component is used to output and / or input audio signals. For example, the audio component may include a microphone for receiving external audio signals. The received audio signals may be further stored in the memory 802 or transmitted via the communication component 805. The audio component also includes at least one speaker for outputting audio signals. I / O interface 804 provides an interface between processor 801 and other interface modules, such as keyboards, mice, or buttons. These buttons can be virtual or physical. Communication component 805 is used for wired or wireless communication between the cycle-trend decoupling predictive control device 800 of this large-scale industrial model and other devices. Wireless communication includes, for example, Wi-Fi, Bluetooth, Near Field Communication (NFC), 2G, 3G, or 4G, or a combination thereof. Therefore, the corresponding communication component 805 may include a Wi-Fi module, a Bluetooth module, or an NFC module.

[0091] In an exemplary embodiment, the periodic-trend decoupling predictive control device 800 for a large industrial model can be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to execute the aforementioned periodic-trend decoupling predictive control method for a large industrial model.

[0092] In another exemplary embodiment, a computer-readable storage medium including program instructions is also provided. When executed by a processor, these program instructions implement the steps of the aforementioned periodic-trend decoupling predictive control method for a large-scale industrial model. For example, the computer-readable storage medium may be the aforementioned memory 802 including program instructions, which may be executed by the processor 801 of the periodic-trend decoupling predictive control device 800 for a large-scale industrial model to complete the aforementioned periodic-trend decoupling predictive control method for a large-scale industrial model.

[0093] Example 4:

[0094] Corresponding to the above method embodiments, this embodiment also provides a readable storage medium. The readable storage medium described below can be referred to in conjunction with the cycle-trend decoupling predictive control method for an industrial large model described above.

[0095] A computer program is stored on a readable storage medium, and when the computer program is executed by a processor, it implements the steps of the periodic-trend decoupling predictive control method for the industrial large model described in the above method embodiments.

[0096] Specifically, the readable storage medium can be a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, or any other readable storage medium capable of storing program code.

[0097] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A periodic-trend decoupled predictive control method for a large industrial model, deployed on a heterogeneous computing architecture of an edge intelligent controller, characterized in that, Includes the following steps: The fieldbus master station of the edge intelligent controller polls and collects the instantaneous value of the three-phase current of the electric spindle, the nanometer-level position feedback of the grating ruler, the vibration signal of the accelerometer, and the thermal radiation signal of the infrared temperature sensor through a distributed clock synchronization mechanism to construct a multi-source heterogeneous data stream; the multi-source heterogeneous data stream is then timestamped by the FPGA hardware and input into the real-time data pool of the programmable logic controller during operation. Based on the multi-source heterogeneous data stream in the real-time data pool, differential operation of grating ruler position feedback is performed to obtain the instantaneous spindle rotation speed. Then, the non-uniform time series is converted into a periodic stable sequence with a fixed rotation phase through angle domain resampling. Subsequently, it is input into an adaptive variational mode decomposition algorithm to separate the high-frequency cutting vibration component and the low-frequency tool wear trend, generating a periodic feature tensor and a trend residual sequence. The fixed phase step size of the angle domain resampling corresponds to a 1-degree spindle rotation, and the number of modes in the variational mode decomposition is adaptively determined by the spectral kurtosis index. The periodic feature tensor and the trend residual sequence are fed into the lightweight time series model deployed by the edge intelligent controller, respectively. The periodic feature tensor is processed by a depthwise separable convolutional network to extract the multi-scale vibration mode of the cutting cycle. The trend residual sequence is processed by a recursive least squares estimator to extract the tool wear coefficient and spindle thermal drift coefficient. The dual-channel output is weighted and spliced ​​by a feature fusion layer to generate the machining surface quality index and the tool remaining life prediction range. Based on the machining surface quality index and the tool remaining life prediction range, the multi-axis collaborative controller of the programmable logic controller soft motion kernel is called. After evaluation by the weighted composite cost function of the cutting trajectory tracking term and the machining quality maintenance term, the finite time domain optimization problem with feed rate limit and spindle torque constraint is solved, and the feed axis speed loop feedforward compensation and spindle speed correction are generated. The speed loop feedforward compensation and spindle speed correction are transmitted to the dual-channel redundant comparator of the edge intelligent controller to perform fault detection and consistency verification. The output is mapped to the process data object of the fieldbus slave station through the safety shutdown path, and finally output to the cyclic synchronous position mode control word of the servo drive to complete the safe and reliable execution of the predictive control command. The heterogeneous computing architecture includes: a high-performance application processor core of the edge intelligent controller executing adaptive variational mode decomposition algorithm and lightweight temporal model inference; a real-time coprocessor core executing finite-time domain optimization solution of multi-axis collaborative controller; an FPGA core executing hardware timestamp marking and preprocessing caching of multi-source heterogeneous data streams; and a neural network processing unit executing feature extraction of multi-scale dilated convolutional networks. The angle domain resampling includes: obtaining the instantaneous spindle speed through differential operation based on the grating ruler position feedback; constructing a time-rotation phase mapping function based on this through numerical integration; and using linear interpolation to convert non-uniform time domain sampling points into a uniform rotation phase sequence with a fixed phase interval, generating a periodically stable sequence. The angle domain resampling eliminates the cutting force waveform distortion caused by spindle speed fluctuations, converting the time-varying speed signal into a standard cutting waveform with an invariant phase. The process of processing the periodic feature tensor through a depthwise separable convolutional network includes: feeding the periodic feature tensor into a multi-scale dilated convolutional network, performing factorization operations through depthwise convolution and pointwise convolution, extracting multi-scale vibration patterns of the cutting cycle, and identifying early signs of cutting chatter; the dilation rate of the multi-scale dilated convolutional network increases with the rotation phase, the convolutional kernel size is configured according to the rotation phase, and the dilation factor is distributed in a power-law sequence of 2; the network contains four layers of depthwise separable convolutions, the dilation rate is distributed in a power-law sequence of 2 with the rotation phase, and the input sequence length is 1024 sampling points; The recursive least squares estimator includes: inputting the trend residual sequence into the recursive least squares estimator, constructing a physical constraint loss function in conjunction with the Taylor wear formula, and performing online identification of the tool wear coefficient and spindle thermal drift coefficient with an adaptive forgetting factor; wherein the physical constraint term of the Taylor wear formula is as follows: In the formula, For tool wear rate, Where v is the material wear coefficient and v is the cutting speed. is the cutting speed exponent, f is the feed rate, and b is the feed rate exponent; The formula for calculating the physical constraint loss function is as follows: In the formula, The physical constraint loss function characterizes the deviation between the estimated wear rate and the Taylor formula prediction. To estimate the tool wear rate, The tool wear rate predicted by Taylor's formula. The square of the Euclidean distance; The recursive update formula for the tool wear coefficient is as follows: In the formula, Let be the vector of tool wear coefficient and thermal drift coefficient at time t. Let be the vector of tool wear coefficient and thermal drift coefficient at time t-1. Let be the covariance matrix at time t-1. For the regression vector, An adaptive forgetting factor. This is the transpose of the regression vector. The residual sequence value of the current trend. For adaptive physical constraint weights, This represents the gradient of the physical constraint loss function with respect to the wear coefficient vector. The weighted composite cost function includes: feeding the machining surface quality index and the tool remaining life prediction range into a multi-axis collaborative controller, and then iteratively solving the problem through a cutting trajectory tracking sub-optimizer, a machining quality maintenance sub-optimizer, and a tool axis vector smoothing sub-optimizer to generate the feed axis speed loop feedforward compensation, torque feedforward compensation, and spindle speed correction; the cutting trajectory tracking sub-optimizer solves the tool path tracking error minimization, the machining quality maintenance sub-optimizer solves the cutting vibration suppression constraint, and the tool axis vector smoothing optimizer solves the tool axis change rate penalty constraint; the prediction time domain is adaptively adjusted according to the tool tip curvature, shortening the control time domain to 5-10 cutting cycles in the high curvature region and extending the prediction time domain to 20-50 cutting cycles in the stable cutting region; The weighted composite cost function is calculated using the following formula: In the formula, For the weighted composite cost function, For dynamic weighting coefficients, For cutting trajectory tracking, For processing quality maintenance items, This is the tool axis smoothing weighting coefficient. For the tool axis vector smoothing term, For adaptive control in the time domain, t is the current time. To control the sampling point index in the time domain, Let be the predicted value at the current time t with respect to the future time t+c. This represents the predicted control increment at time t with respect to a future time t+c. To control the increment, S is the weight matrix. The square of the Euclidean distance; The dual-channel redundant comparator includes: speed loop feedforward compensation, torque feedforward compensation, and spindle speed correction transmitted to the dual-channel redundant comparator. The main channel uses a lightweight timing model inference to predict the compensation amount, while the backup channel uses an empirical lookup table compensation amount based on cutting time. Based on the posterior probability distribution of the two channel outputs, Bayesian inference is used to evaluate the redundancy output confidence. When the confidence level is higher than a threshold, a weighted fusion output is used; when the confidence level is lower than the threshold, a degradation mode is triggered, forcibly reducing the feed rate to a safe feed rate. Once the confidence level recovers to above the threshold, it automatically returns to normal control mode. The posterior probability distribution is dynamically updated based on the model prediction uncertainty and the timeliness of the lookup table data. The uncertainty of the main channel is estimated using the Monte Carlo regression method: 20 forward inferences are performed, with a portion of neurons randomly regressed each time. The mean of the output is used as the predicted value, and the standard deviation is used as the uncertainty. The timeliness of the backup channel increases linearly with the time since the last calibration, and the longer the time, the worse the timeliness. Based on the Gaussian distribution assumption, the posterior probabilities of the two channels are calculated: the posterior probability of the model-predicted channel is proportional to the reciprocal of the model prediction uncertainty, and the posterior probability of the empirical lookup table channel is proportional to the reciprocal of the timeliness of the lookup table data. After normalizing the posterior probabilities, the confidence level is defined as the maximum value of the posterior probabilities of the two channels. When the confidence level is higher than the threshold of 0.85, a weighted fusion output is used, with the weight being the posterior probability. When the confidence level is lower than the threshold, a degradation mode is triggered.

2. The periodic-trend decoupling predictive control method for a large industrial model according to claim 1, characterized in that, The adaptive variational mode decomposition algorithm includes: The periodically stable sequence is fed into an adaptive variational mode decomposition algorithm. The variational constraint model is solved iteratively using the alternating direction multiplier method. Eigenmode functions with compact spectra are extracted. The eigenmode function with the lowest center frequency is output as the trend residual sequence, representing the asymptotic accumulation of tool wear and thermal drift. The remaining eigenmode functions are superimposed to generate a periodic feature tensor, representing the periodic vibration and impact of the cutting process. The variational constraint model is executed according to the following optimization objective: In the formula, Let k be the intrinsic mode function. To correspond to the center frequency, For data fidelity weighting coefficients, For the Dirac function, The imaginary unit, It is a periodically stable sequence. For convolution operations, k is the total number of intrinsic mode functions. Let be the time differential operator, t be the time variable, and e be the base of the natural logarithm.

3. The periodic-trend decoupling predictive control method for a large industrial model according to claim 1, characterized in that, The safety shutdown path output includes: the output of the dual-channel redundant comparator performs dynamic reconstruction of the process data object mapping through the fieldbus controller, and immediately reconstructs the process data object mapping when an anomaly is detected, switching control to the backup programmable logic controller logic; the dynamic reconstruction of the process data object mapping utilizes the fieldbus hot-connection function, wherein the fault response time is less than 50μs.

Images