Self-adaptive cutting control method based on continuous feeding scene of cutting machine

Abstract

The invention relates to the technical field of industrial control, in particular to a self-adaptive cutting control method based on a continuous feeding scene of a cutting machine, which comprises the following steps: acquiring process input parameters of a material sample of the cutting machine, process constraint parameters under different processing tracks and corresponding actual process output parameters; respectively constructing a first training data set and a second training data set, and training to obtain a constraint process model and a state process model; a process input parameter flow of a to-be-processed material area is obtained online and input to the constraint process model, and a dynamic process constraint is generated; creating a candidate control trajectory based on the dynamic process constraint; inputting the candidate control trajectory into the state process model, outputting a prediction performance index, and selecting and determining an optimal control trajectory; and executing the optimal control track to realize self-adaptive cutting control, and carrying out online updating on the constraint process model and the state process model to realize optimal control on the machining quality and efficiency of the cutting machine.

Description

Technical Field

[0001] This invention relates to the field of industrial control technology, specifically to an adaptive cutting control method based on a continuous feeding scenario of a cutting machine. Background Technology

[0002] In the application of continuous feeding operations in cutting machines, existing industrial control systems typically use programmable logic controllers to drive servo actuators, coordinating the continuous feeding and processing of materials through pre-programmed, fixed control parameters and open-loop program instructions.

[0003] However, the control logic of such industrial control systems is static. They lack the ability to identify the online parameters of the controlled object (i.e., the material to be processed) and cannot establish mathematical models capable of predicting process dynamics. Therefore, their fixed programs cannot optimize their control laws online or adaptively adjust control outputs based on real-time changes in material properties (such as thickness and density inhomogeneities) or equipment state drift. When faced with dynamic operating conditions, such industrial control systems struggle to maintain optimal performance, directly leading to poor consistency in processing results and impacting equipment energy efficiency.

[0004] Therefore, this invention proposes an adaptive cutting control method based on continuous feeding scenarios of cutting machines. Summary of the Invention

[0005] The purpose of this invention is to provide an adaptive cutting control method for continuous feeding scenarios in cutting machines, aiming to solve the problem that existing industrial control systems cannot adapt to dynamic working conditions due to their rigid control logic. To achieve the above objective, this invention combines an online model for predicting the local safe processing boundary of materials with a model for forward-looking multi-strategy simulation optimization, and supplements this with a closed-loop online update mechanism for the two models. This approach improves production efficiency and equipment energy efficiency while ensuring consistent processing quality.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] An adaptive cutting control method based on continuous feeding scenarios of a cutting machine includes:

[0008] The process input parameters for obtaining material samples from the cutting machine, the process constraint parameters under different processing trajectories, and the corresponding actual process output parameters are obtained.

[0009] A first training dataset is constructed using the process input parameters and process constraint parameters, and a constrained process model is trained; a second training dataset is constructed using the processing trajectory and actual process output parameters, and a state process model is trained.

[0010] obtaining a process input parameter stream of a region of material to be processed, inputting the process input parameter stream into the constraint process model, generating a dynamic process constraint representing a safe operating boundary of the region;

[0011] based on the dynamic process constraint, creating a candidate control trajectory, inputting the candidate control trajectory into the state process model, outputting a predicted performance index, selecting and determining an optimal control trajectory;

[0012] implementing the optimal control trajectory to achieve adaptive cutting control, obtaining actual process output parameters, and using the actual process output parameters to update the constraint process model and the state process model online, respectively.

[0013] Preferably, the constraint process model comprises: measuring a plurality of material samples by a non-contact sensor to obtain initial process input parameters; executing a processing trajectory on the material samples and collecting high-frequency acoustic signals in real time during the processing; determining a control parameter whose energy satisfies a preset mutation condition as a critical constraint of process abnormalities, and recording the control parameter corresponding to the critical constraint as a process constraint parameter; inputting data points containing the initial processing trajectory and the corresponding process constraint parameter into an optimization algorithm; the optimization algorithm calculates and recommends a next processing trajectory according to the principle of maximizing information gain; repeating the steps of executing the processing trajectory recommended by the optimization algorithm, determining the process constraint parameter, and calculating the next optimal processing trajectory until the entire safe processing boundary is accurately calibrated; pairing the process input parameters with the corresponding process constraint parameters to form a first training data set; using the first training data set, a constraint process model is trained by a machine learning method, which can predict the damage degree according to the real-time process input parameters of the material and determine the safe processing boundary.

[0014] Preferably, the state process model comprises: recording a plurality of processing trajectories and corresponding actual process output parameters, including vibration and torque feedback signals during the actual processing, in the actual processing; pairing the processing trajectories with the corresponding actual process output parameters to form a second training data set; using the second training data set, a state process model is trained by a machine learning method, which can predict the processing result according to a given processing trajectory.

[0015] Preferably, the process input parameter stream is inputted into the constraint process model to generate dynamic process constraints representing the safe operation boundary of the region, including: collecting a process input parameter stream of a region of material to be processed in real time during continuous feeding of the production process; inputting the real-time process input parameter stream into the constraint process model; the constraint process model performing a forward propagation calculation on the input data stream, and outputting a set of multi-dimensional control parameters as the dynamic process constraints, the control parameter set containing process failure thresholds of the region under different pressing speeds and forces.

[0016] Preferably, the candidate control trajectories are created based on the dynamic process constraints, including: in a preset multi-dimensional control parameter space composed of force and motion speed, taking the dynamic process constraints as boundary conditions, generating a set of discretized candidate control trajectories in the safe region defined by the boundary through a preset search algorithm, wherein each candidate control trajectory is a time sequence containing force and motion speed commands.

[0017] Preferably, the predicted performance indicators are outputted, including: the state process model performing a forward propagation calculation on each candidate control trajectory and outputting a multi-dimensional predicted performance vector containing multiple performance dimension parameters as the predicted performance indicators in parallel; the predicted performance indicators containing predicted processing quality parameters, total energy consumption parameters of the processing process, and processing completion time parameters; the processing quality parameters being calculated by a control law simulating the dynamic response of a device inside the state process model according to the control trajectory; the total energy consumption parameters of the processing process being calculated by time series integration of the force and speed commands contained in the control trajectory; and the processing completion time parameters being determined according to the time series length of the control trajectory.

[0018] Preferably, the optimal control trajectory is selected and determined, including: setting preset weight coefficients for the processing quality parameters, the total energy consumption parameters of the processing process, and the processing completion time parameters in the predicted performance indicators; for each candidate control trajectory, multiplying each predicted performance indicator by the corresponding weight coefficient, and performing algebraic summation on the multiplication results to calculate a control objective function value representing the comprehensive performance of the trajectory; comparing the control objective function values generated for all candidate control trajectories, and selecting the candidate control trajectory with the optimal function value as the optimal control trajectory.

[0019] Preferably, the online updating of the constraint process model and the state process model by using the actual process output parameters respectively comprises: acquiring edge micro-morphology data of the workpiece after processing as the actual process output parameters; calculating the difference between the actual process output parameters and the predicted performance index, generating a predicted process correction signal for updating the state process model, and adjusting the internal parameters of the state process model by using the predicted process correction signal; determining the correction direction of the dynamic process constraint by evaluating the advantages and disadvantages of the actual process output parameters, generating a damage prediction correction signal for updating the constraint process model, and adjusting the internal parameters of the constraint process model by using the damage prediction correction signal.

[0020] Compared with the prior art, the present application has the following advantages:

[0021] 1. The present application dynamically generates a safe operation boundary by online acquisition of the process input parameter flow of the material to be processed, so that the cutting control can adapt to the regional differences of the material in real time, thereby ensuring that the precision and quality of cutting at each point in the continuous feeding process remain highly consistent, and solving the problem of poor consistency of processing results caused by the fixed control logic in traditional methods.

[0022] 2. The present application creates multiple candidate control trajectories within a safe processing boundary by introducing a state process model, and simulates and predicts the processing quality, total energy consumption and completion time of each trajectory, so that an optimal control trajectory can be selected in advance to be executed through a weighted comprehensive scoring mechanism. This method changes passive execution to active optimization, which can maximize the reduction of equipment energy consumption and improve production rhythm under the premise of ensuring processing quality, and solves the disadvantages of affecting equipment energy efficiency in the prior art.

[0023] 3. The present application compares the actual process output parameters with the predicted results after each cutting is completed, and corrects the constraint process model and the state process model in reverse, so that the control system can continuously improve itself through this closed-loop online learning mechanism, automatically compensates for the influence of factors such as tool wear and material batch differences, and ensures the long-term accuracy of the control model and the long-term stability of the cutting operation. BRIEF DESCRIPTION OF DRAWINGS

[0024] Figure 1 It is a schematic diagram of the overall architecture of the adaptive cutting control method based on the continuous feeding scene of the cutting machine of the present application.

[0025] Figure 2 It is a flowchart of the adaptive cutting control method based on the continuous feeding scene of the cutting machine of the present application.

[0026] Figure 3A schematic diagram of an active learning training process of a constraint process model of an embodiment of the present application;

[0027] Figure 4 A schematic diagram of an online decision and optimization process of an embodiment of the present application. DETAILED DESCRIPTION

[0028] The technical solutions of the present application will be described in detail below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can more clearly understand the core concept of the present application. It should be emphasized that these embodiments are only examples of several preferred implementations of the present application, and are not an exhaustive enumeration of all possible implementations of the present application. Therefore, any modification, replacement or variation that does not deviate from the core technical idea and scope of the present application, which can be easily thought of by those skilled in the art according to the content disclosed herein, should be considered to fall within the protection scope of the present application.

[0029] Embodiment 1:

[0030] Referring to Figure 1 A schematic diagram of the overall architecture to which the adaptive cutting control method described in this embodiment is applied. The method is applied to a gantry cutting machine equipped with a servo motor driven continuous feeding device and a hydraulic stamping execution mechanism. As shown in the figure, the system integrates multiple sensor modules, including a non-contact sensor group for online acquisition of material process input parameters, a high-frequency acoustic / vibration sensor for acquisition of training and correction signals, and a machine vision system for acquisition of actual cutting effect. The data acquisition of all these modules and the control of the feeding device and the cutting machine are all coordinated by the adaptive control system, which is the core of the present application.

[0031] Referring to Figure 1 and Figure 2 The adaptive cutting control method based on the continuous feeding scenario of the cutting machine of this embodiment includes:

[0032] Acquiring process input parameters of the material sample of the cutting machine, process constraint parameters under different processing trajectories, and corresponding actual process output parameters;

[0033] Using the process input parameters and the process constraint parameters to construct a first training data set, training to obtain a constraint process model; using the processing trajectory and the actual process output parameter to construct a second training data set, training to obtain a state process model;

[0034] Online acquisition of process input parameter flow of the material region to be processed, input of the process input parameter flow to the constraint process model, generation of dynamic process constraints representing the safe operation boundary of the region;

[0035] Based on the dynamic process constraint, a candidate control trajectory is created; the candidate control trajectory is input into the state process model, and a predicted performance index is output, and an optimal control trajectory is selected and determined;

[0036] The optimal control trajectory is executed to achieve adaptive cutting control, actual process output parameters are obtained, and the constraint process model and the state process model are updated online respectively using the actual process output parameters.

[0037] Further, with reference to Figure 3 , the constraint process model comprises: measuring a plurality of material samples by a non-contact sensor to obtain initial process input parameters; executing a machining trajectory on the material sample and collecting high-frequency acoustic signals in real time during the machining process; determining a control parameter satisfying a preset mutation condition of the energy of the acoustic signal as a critical constraint of process abnormality, and recording the control parameter corresponding to the critical constraint as a process constraint parameter; inputting data points containing the initial machining trajectory and the corresponding process constraint parameter into an optimization algorithm; the optimization algorithm calculates and recommends a next machining trajectory according to the principle of maximizing information gain; the steps of executing the machining trajectory recommended by the optimization algorithm, determining the process constraint parameter and calculating the next optimal machining trajectory are repeated until the entire safe machining boundary is accurately calibrated; pairing the process input parameters and the corresponding process constraint parameters to form a first training data set; using the first training data set, a constraint process model is trained by a machine learning method, and the constraint process model can predict the damage degree according to the real-time process input parameters of the material and determine the safe machining boundary.

[0038] With reference to Figure 3 , the constraint process model is an active learning process that interactsively communicates between multiple functional roles, namely "optimization algorithm", "experimental platform", "signal processor", "training database" and "ML engine", and its core lies in an iterative process (loop area in the figure) led by the "optimization algorithm", which intelligently explores the safe machining boundary through continuous "recommendation-execution-detection-learning" until the model converges. Specifically:

[0039] The process is performed on an experimental platform capable of accurately reproducing the cutting machine processing action. For a non-metallic material sample, the initial process input parameters of the material, including the density and elastic modulus of the material, are measured by a non-contact sensor. Then, an initial processing trajectory is performed on the material sample, and high-frequency acoustic signals during the processing are collected in real time using an acoustic emission sensor installed on the experimental platform. The program control system processes the high-frequency acoustic signals, first filters out the known device background noise through a digital filter, and then extracts the feature vector containing the root mean square energy, peak amplitude and frequency distribution from the signal. The system compares the continuous changes of the feature vector with the pre-set material damage acoustic feature mode library, identifies the signal precursor that characterizes the impending process abnormality, determines the time point when the precursor first appears as the critical constraint of the process abnormality, and records the control parameters corresponding to the critical constraint as the process constraint parameters under the trajectory.

[0040] Next, the data points containing the initial processing trajectory and the corresponding process constraint parameters are input into a Bayesian optimization algorithm. According to the principle of maximizing information gain, the optimization algorithm calculates and recommends the next most valuable processing parameter sampling point, which clearly defines the target values of the stamping force and movement speed to be used in the next experiment. To convert the static parameter sampling point into an executable dynamic processing trajectory, the trajectory generator built into the control system will automatically generate a standard trapezoidal acceleration and deceleration speed curve as the next processing trajectory based on the target values. The duration of the acceleration and deceleration phases of the trajectory is a pre-set fixed value that matches the actual device dynamic response capability, for example, 0.05 seconds, while the speed and applied force during the uniform speed running phase are exactly equal to the target values defined by the sampling point.

[0041] The system repeats the steps of performing the processing trajectory recommended by the optimization algorithm, determining the process constraint parameters, and calculating the next recommended processing trajectory until the entire safe processing boundary is accurately calibrated by the model with a pre-set confidence level. Finally, all the collected process input parameters and calibrated process constraint parameters are paired to form the first training data set, and the constraint process model is trained using the training data set.

[0042] This embodiment introduces an iterative process of active learning and exploration to construct the most accurate fault prediction model with the least number of experiments and the highest efficiency. This method avoids the large number of material samples and time consumption required by traditional grid testing, significantly reducing the research and development cost and cycle for establishing a control model for new materials.

[0043] Further, the state process model comprises: recording a plurality of machining trajectories and corresponding actual process output parameters in an actual machining process, including vibration and torque feedback signals in the machining process; pairing the machining trajectories and the corresponding actual process output parameters to form a second training data set; and training the state process model by using the second training data set and a machine learning method, so that the state process model can predict a machining result according to a given machining trajectory.

[0044] In the actual operation process of the cutting machine, the program control system records a plurality of machining trajectories containing different force and speed combinations, and synchronously collects actual process output parameters corresponding to each trajectory, the process output parameters including vibration signals collected by an accelerometer and torque signals fed back by a servo driver; the system timestamps and pairs the machining trajectories and the corresponding actual process output parameters to form a second training data set; and the state process model is trained by using the second training data set, so that the model can predict a system state in a machining process according to a given machining trajectory.

[0045] The embodiment identifies the state prediction model by using data in an actual production process, ensures that the model can highly accurately reflect real dynamic characteristics of a physical device, and makes subsequent simulation results based on the model extremely close to physical reality, thereby providing a reliable basis for selecting a truly optimal control trajectory.

[0046] Further, with reference to Figure 4 The process input parameter stream is input to the constraint process model to generate dynamic process constraints representing a safe operation boundary of the region, comprising: collecting a process input parameter stream of a region to be machined in real time in a continuous feeding production process; inputting the real-time process input parameter stream to the constraint process model; the constraint process model performs a forward propagation calculation on the input data stream, and outputs a multi-dimensional control parameter set as the dynamic process constraints, the control parameter set containing process failure thresholds of the region under different punching speeds and forces.

[0047] In the continuous feeding production process of the cutting machine, a non-contact sensor collects a process input parameter stream of a region to be machined in real time; a program control system inputs the real-time process input parameter stream to the constraint process model, and the constraint process model performs a forward propagation calculation on the input data stream, and outputs a multi-dimensional control parameter set as the dynamic process constraints, the control parameter set containing process failure thresholds of the region under different punching speeds and forces.

[0048] The embodiment can make the control system predict the safe operation boundary of the current material area with high precision before the machining action occurs. This forward-looking perception ability with high confidence is the premise of fundamentally avoiding machining abnormalities and quality defects, and changes the passive response mode of traditional control to the active prevention mode.

[0049] Further, with reference to Figure 4 , the creating candidate control trajectories based on the dynamic process constraints comprises: in a preset multi-dimensional control parameter space composed of force and motion speed, taking the dynamic process constraints as boundary conditions, generating a set of discrete candidate control trajectories in a safe region defined by the boundary through a preset search algorithm, wherein each candidate control trajectory is a time sequence containing force and motion speed commands.

[0050] The program control system generates a set of discrete and diverse candidate control trajectories in a safe region defined by the boundary through an A* search algorithm, wherein each candidate control trajectory is a time sequence containing force and motion speed commands.

[0051] Specifically, the multi-dimensional control parameter space is discretized into a two-dimensional grid, wherein each grid point represents a {force, motion speed} combination and serves as a node in the search graph. Each node is only allowed to move to adjacent nodes that are still within the safe operation boundary, and the cost of movement is defined as the time or energy consumed. The heuristic function of the A* search algorithm is designed as the weighted Euclidean distance between the current node state and a preset ideal target state point representing high efficiency and high quality, which can guide the search direction to preferentially move towards regions with better comprehensive performance.

[0052] The creating candidate control trajectories further comprises:

[0053] A historical optimal trajectory database is established and maintained to store the correspondence between historical process input parameters and control trajectories that are executed and verified as optimal. Before creating candidate control trajectories, the process input parameters of the current material area to be machined are matched with the records in the historical optimal trajectory database in terms of similarity. When the similarity of the matched records is higher than a preset threshold, the corresponding historical optimal control trajectory and its variants in the neighborhood are directly added to the candidate control trajectory set generated by the search algorithm as high-priority candidate trajectories to participate in subsequent prediction and selection. This enhanced generation mechanism based on historical experience greatly improves the decision-making efficiency and makes the generated trajectories more likely to approach the global optimal solution.

[0054] The embodiment ensures that all candidate strategies entering the decision-making link are essentially safe by creating control trajectories within the pre-judged safety boundary, which greatly reduces the search space, improves decision-making efficiency, and reduces dangerous operations that may cause damage to equipment or products from the source.

[0055] Further, with reference to Figure 4 , the output prediction performance indicators include: the state process model performs a forward propagation calculation on each candidate control trajectory and outputs a multi-dimensional prediction performance vector containing multiple performance dimension parameters as the prediction performance indicators in parallel; the prediction performance indicators include predicted processing quality parameters, total energy consumption parameters of the processing process, and processing completion time parameters; the processing quality parameters are calculated by a control law simulating the dynamic response of the equipment inside the state process model according to the control trajectory; the total energy consumption parameters of the processing process are obtained by time series integration of the force and speed instructions contained in the control trajectory; and the processing completion time parameters are determined according to the time series length of the control trajectory.

[0056] The embodiment enables the system to "rehearse" the consequences of multiple processing strategies at low cost and high efficiency without any physical processing through forward-looking process simulation. This virtual trial-and-error capability of rehearsing consequences without physical processing is the basis for achieving multi-objective optimization decision-making.

[0057] Further, with reference to Figure 4 , the selection and determination of an optimal control trajectory include: setting preset weight coefficients for the processing quality parameters, total energy consumption parameters of the processing process, and processing completion time parameters in the prediction performance indicators; for each candidate control trajectory, multiplying each prediction performance indicator by the corresponding weight coefficient, and performing algebraic summation on the results of the multiplication operation to calculate a control objective function value representing the comprehensive performance of the trajectory; the program control system compares the control objective function values generated by all candidate control trajectories and selects the candidate control trajectory with the optimal function value as the optimal control trajectory.

[0058] As a preferred mode of the embodiment, the selection and determination of an optimal control trajectory further include:

[0059] When generating the dynamic process constraints, the constraint process model also outputs a boundary confidence level associated with the dynamic process constraints. For each candidate control trajectory, the minimum safe distance between it and the boundary defined by the dynamic process constraints is calculated. Based on the boundary confidence level and the minimum safe distance, a risk penalty term is generated, wherein the risk penalty term is inversely proportional to the boundary confidence level and directly proportional to the reciprocal of the minimum safe distance. The control objective function value of the candidate control trajectory is calculated with the risk penalty term to generate a risk-adjusted final score. The candidate control trajectory with the highest risk-adjusted final score is selected as the optimal control trajectory. This method elevates control decision-making from a simple "performance optimization" to a cognitive level of "performance optimization under risk perception." It cleverly couples the outputs of the constraint prediction model and the state prediction model at the decision layer, constructing a more robust decision system.

[0060] Furthermore, to better integrate the cutting machine of this invention into modern intelligent manufacturing systems, its optimization objective is dynamically configurable. At least two production modes are predefined in the system, each corresponding to a different set of weighting coefficients. These production modes include: quality-priority mode, efficiency-priority mode, and energy-priority mode. A communication interface for a Manufacturing Execution System (MES) is provided to receive the currently active production mode instructions. When calculating the control objective function value, the corresponding set of weighting coefficients is automatically invoked based on the received production mode instructions. This design endows the control system with extremely high decision-making flexibility, enabling it to transform from a simple execution unit into an intelligent terminal capable of deeply understanding and executing upper-level business intentions.

[0061] The weighting coefficients are normalized and their sum is 1. In this embodiment, to achieve different production goals, at least the following three weighting configuration modes are preset:

[0062] (1) Quality-first mode: Set the weights as {processing quality: 0.7, total energy consumption: 0.15, completion time: 0.15};

[0063] (2) Efficiency-first mode: Set the weights as {processing quality: 0.2, total energy consumption: 0.2, completion time: 0.6};

[0064] (3) Balanced mode: Set the weights as {processing quality: 0.4, total energy consumption: 0.3, completion time: 0.3}.

[0065] Operators can manually select the currently active mode based on production tasks, or the upper-level MES system can automatically specify the mode in effect.

[0066] This embodiment introduces a configurable control objective function, giving the control system extremely high decision-making flexibility. Operators can change the machine's optimization objective by adjusting the weighting coefficients according to different production tasks, allowing it to flexibly switch between different modes such as pursuing ultimate quality or pursuing maximum efficiency.

[0067] Furthermore, the online updating of the constraint process model and the state process model using the actual process output parameters includes: acquiring the edge micro-morphology data of the workpiece after processing as the actual process output parameters; calculating the difference between the actual process output parameters and the predicted performance index, generating a predicted process correction signal for updating the state process model, and adjusting the internal parameters of the state process model using the predicted process correction signal; determining the correction direction for the dynamic process constraints by evaluating the quality of the actual process output parameters, generating a damage prediction correction signal for updating the constraint process model, and adjusting the internal parameters of the constraint process model using the damage prediction correction signal.

[0068] After the cutting machine executes the optimal control trajectory, it acquires the edge micro-morphology data of the processed workpiece through a high-resolution machine vision system and uses it as the actual process output parameter. The program control system calculates the difference between the actual process output parameter and the predicted performance index, generates a predicted process correction signal for updating the state process model, and uses the predicted process correction signal to adjust the internal parameters of the state process model. Specifically, the adjustment process adopts an incremental learning method. Each newly acquired data pair containing {actual execution trajectory, actual output parameter} is regarded as a mini-batch training sample. The system uses this sample to perform a weight fine-tuning update on the neural network serving as the state process model through a backpropagation algorithm. The learning rate is set to a small value, such as 1e-5, to ensure that the model maintains the stability of historical learning while adapting to new data. A similar method is used for the correction of the constraint process model, incrementally updating the model with new boundary exploration results.

[0069] Simultaneously, the system compares the actual process output parameters (i.e., edge micro-morphology) with a preset quality benchmark to evaluate the quality of the current processing. If the actual morphology is worse than the benchmark, it indicates that the optimal control trajectory executed this time has actually touched or approached the real damage boundary. Accordingly, the system generates a damage prediction correction signal for updating the constraint process model. This signal enables the dynamic process constraints to be adjusted in a more conservative direction within the parameter space region corresponding to the optimal control trajectory executed this time.

[0070] This embodiment establishes a dual-feedback correction loop, enabling the control system to possess continuous, two-dimensional self-evolution capabilities. It can not only learn how to more accurately predict the consequences of its behavior, but also how to more deeply understand the physical nature of materials, thereby continuously improving its intelligence level during long-term operation.

[0071] This embodiment constructs a constrained process model that maps material physical properties to their safe processing boundaries through active learning and exploration. In actual operation, this model is used to perform online safety predictions for each area of ​​material to be processed. Based on these predictions, the system determines a control trajectory that achieves optimal energy efficiency and effectiveness without exceeding safety boundaries through forward-looking process simulation and multi-objective optimization. After completing processing using this trajectory, the system further updates and evolves the two core models online using actual processing results through a dual feedback loop. This method enables the automated cutting machine to perceive local material changes in real time, actively avoid processing anomalies, and optimize the best control strategy within the safety boundaries. It also continuously learns and optimizes itself during production. In continuous feeding scenarios, it can adaptively adjust point-by-point to address material inhomogeneity, thereby significantly improving the quality consistency of the cutting edges and product qualification rate without sacrificing processing speed, while reducing abnormal tool wear and equipment energy consumption caused by parameter mismatch.

[0072] Example 2:

[0073] This embodiment provides a specific implementation of an adaptive cutting control method based on a continuous feeding scenario of a cutting machine. The application object of this embodiment is a gantry mobile cutting machine, which is equipped with a hydraulic punch head controlled by a program control system that can move in the XY plane, and a fixed worktable for placing whole pieces of leather material.

[0074] The method is first implemented using a dedicated experimental setup that can accurately reproduce the stamping action of the cutting machine, specifically:

[0075] Leather samples of varying thicknesses and regions were collected. For each sample, its thickness was measured using a laser displacement sensor, and its acoustic impedance was measured using an acoustic sensor. These measurements were combined to form the initial process input parameters for that sample. Subsequently, a series of processing trajectories were executed on the sample, each containing a specific time series of impact force and speed. During the execution of each processing trajectory, high-frequency acoustic signals were acquired in real time using an acoustic emission sensor mounted on the experimental apparatus. The program control system determined the time point at which the energy amplitude of the acoustic signal showed a preset steep increase as the critical constraint for microscopic damage to the leather sample under that processing trajectory, and recorded the processing parameters corresponding to this critical constraint as process constraint parameters. Simultaneously, the edge morphology of the leather sample after each processing was captured and quantified using a high-resolution industrial camera as the actual process output parameters. By repeating the above experiments, a first set of reference data containing the correspondence between "process input parameters and process constraint parameters" and a second set of reference data containing the correspondence between "processing trajectory and actual process output parameters" were obtained. Finally, a constrained process model was trained using the first set of reference data, and a state process model was trained using the second set of reference data.

[0076] The online adaptive control step of this method is executed on the gantry-type mobile cutting machine, specifically as follows:

[0077] A complete piece of leather material is placed on a fixed workbench, and the program control system loads a layout file containing the position coordinates of multiple parts to be punched. For the first target processing position in the layout file, the system first drives the gantry mechanism to move the hydraulic punch head to that position. At that position, the non-contact sensor acquires the real-time process input parameter stream of the leather area below online. The real-time process input parameter stream is input to the constraint process model, and the model calculates and generates a dynamic process constraint that characterizes the safe operating boundary of the area.

[0078] Based on the dynamic process constraints, the program control system creates a set of candidate control trajectories within a multi-dimensional control parameter space consisting of forces and motion velocities. Each candidate control trajectory is input into the state process model for process simulation, and the model outputs a set of predicted performance indicators for each trajectory. The system compares the predicted performance indicators of all candidate control trajectories according to a preset control objective function and selects and determines an optimal control trajectory.

[0079] Subsequently, the program control system sends the optimal control trajectory to the hydraulic system and servo drive to precisely execute the stamping action; after processing is completed, the machine vision system acquires the actual process output parameters of the stamping part; finally, the program control system uses the actual process output parameters to update the internal parameters of the constraint process model and the state process model online.

[0080] After completing the machining and model correction of the first target position, the program control system drives the gantry mechanism to move the stamping head to the next target machining position in the layout file, and repeats all the steps from online acquisition of process input parameters to online model update until the machining of all positions in the layout file is completed.

[0081] It should be noted that the foregoing embodiments are merely preferred implementations of the present invention, intended to help those skilled in the art understand the core concept of the present invention. These specific examples should not be construed as any limitation on the scope of protection of the present invention. Any modifications, equivalent substitutions, or improvements made to the technical solutions of the present invention within the spirit and principles of the present invention, as long as they do not depart from the spirit and scope of the present invention—such as using components with the same function but different structures, or adjusting the order of non-core steps—should be considered as included within the scope of protection of the present invention. The final scope of protection of the present invention shall be determined by the content defined in the appended claims.

Claims

1. An adaptive cutting control method based on a continuous feeding scenario of a cutting machine, characterized in that, include: The process input parameters for obtaining material samples from the cutting machine, the process constraint parameters under different processing trajectories, and the corresponding actual process output parameters are obtained. A first training dataset is constructed using the process input parameters and process constraint parameters, and a constrained process model is trained; a second training dataset is constructed using the processing trajectory and actual process output parameters, and a state process model is trained. The process input parameter stream of the material to be processed region is acquired online, and the process input parameter stream is input into the constrained process model to generate dynamic process constraints that characterize the safe operation boundary of the region. Based on the dynamic process constraints, candidate control trajectories are created; the candidate control trajectories are input into the state process model, the predicted performance index is output, and an optimal control trajectory is selected and determined. The optimal control trajectory is executed to achieve adaptive judgment control, the actual process output parameters are obtained, and the actual process output parameters are used to update the constraint process model and the state process model online.

2. The adaptive cutting control method based on continuous feeding scenario of a cutting machine according to claim 1, characterized in that, The constrained process model includes: measuring various material samples using non-contact sensors to obtain initial process input parameters; executing a processing trajectory on the material samples and acquiring high-frequency acoustic signals during the processing in real time; determining the control parameters whose energy of the acoustic signals meets preset mutation conditions as critical constraints for process anomalies, and recording the control parameters corresponding to the critical constraints as process constraint parameters; inputting data points containing the initial processing trajectory and corresponding process constraint parameters into an optimization algorithm; the optimization algorithm calculating and recommending the next processing trajectory based on the principle of maximizing information gain; repeatedly executing the steps of the processing trajectory recommended by the optimization algorithm, determining the process constraint parameters, and calculating the next optimal processing trajectory until the entire safe processing boundary is accurately calibrated; pairing the process input parameters with the corresponding process constraint parameters to form a first training dataset; using the first training dataset, training a constrained process model using machine learning methods, wherein the constrained process model can predict the degree of damage and determine the safe processing boundary based on the real-time process input parameters of the material.

3. The adaptive cutting control method based on continuous feeding scenario of a cutting machine according to claim 1, characterized in that, The state process model includes: recording multiple processing trajectories and corresponding actual process output parameters during actual processing, including vibration and torque feedback signals during processing; pairing the processing trajectories with the corresponding actual process output parameters to form a second training dataset; and using the second training dataset, training the state process model using machine learning methods, wherein the state process model can predict the processing result based on a given processing trajectory.

4. The adaptive cutting control method based on continuous feeding scenario of a cutting machine according to claim 1, characterized in that, The process input parameter stream is input into the constrained process model to generate dynamic process constraints characterizing the safe operating boundary of the region. This includes: during continuous feeding production, real-time acquisition of the process input parameter stream of the material to be processed region; inputting the real-time process input parameter stream into the constrained process model; the constrained process model performing a forward propagation calculation on the input data stream and outputting a multi-dimensional control parameter set as the dynamic process constraint, wherein the control parameter set contains the process failure threshold of the region under different stamping speeds and forces.

5. The adaptive cutting control method based on continuous feeding scenario of a cutting machine according to claim 1, characterized in that, The step of creating candidate control trajectories based on the dynamic process constraints includes: within a preset multi-dimensional control parameter space consisting of force and velocity, using the dynamic process constraints as boundary conditions, and through a preset search algorithm, generating a set of discretized candidate control trajectories within the safe area defined by the boundary, wherein each candidate control trajectory is a time series containing force and velocity commands.

6. The adaptive cutting control method based on continuous feeding scenario of a cutting machine according to claim 1, characterized in that, The output predicted performance index includes: the state process model performs a forward propagation calculation for each candidate control trajectory and outputs a multi-dimensional predicted performance vector containing multiple performance dimension parameters as the predicted performance index; the predicted performance index includes predicted processing quality parameters, total processing energy consumption parameters, and processing completion time parameters; the processing quality parameters are calculated based on the control trajectory by a control law simulating the dynamic response of a device within the state process model; the total processing energy consumption parameters are obtained by integrating the force and velocity commands contained in the control trajectory over time; and the processing completion time parameters are determined based on the time series length of the control trajectory.

7. The adaptive cutting control method based on continuous feeding scenario of a cutting machine according to claim 1, characterized in that, The step of selecting and determining an optimal control trajectory includes: setting preset weight coefficients for the processing quality parameter, total energy consumption parameter, and processing completion time parameter in the predicted performance indicators; for each candidate control trajectory, multiplying each predicted performance indicator with its corresponding weight coefficient, and summing the results of the multiplication algebraically to calculate a control objective function value characterizing the comprehensive performance of the trajectory; comparing the control objective function values ​​generated by all candidate control trajectories, and selecting the candidate control trajectory with the optimal function value as the optimal control trajectory.

8. The adaptive cutting control method based on continuous feeding scenario of a cutting machine according to claim 1, characterized in that, The step of updating the constraint process model and the state process model online using the actual process output parameters includes: acquiring the edge micro-morphology data of the workpiece after processing as the actual process output parameters; calculating the difference between the actual process output parameters and the predicted performance index, generating a predicted process correction signal for updating the state process model, and adjusting the internal parameters of the state process model using the predicted process correction signal; determining the correction direction for the dynamic process constraints by evaluating the quality of the actual process output parameters, generating a damage prediction correction signal for updating the constraint process model, and adjusting the internal parameters of the constraint process model using the damage prediction correction signal.

Images