Metal processing machine system processing parameter optimization method based on reinforcement learning

Abstract

The present application relates to the field of metal cutting and intelligent control technology, in particular to a metal processing mechanical system processing parameter optimization method based on reinforcement learning; comprising: obtaining vibration, acoustic emission, cutting force and spindle current physical sensing data, determining vibration energy, acoustic emission energy, cutting force step rate and spindle load deviation rate in the time sequence sampling window, and determining the chatter risk index and the impact risk index; combined with the spindle speed, the feed speed, the axial cutting depth and the radial cutting width to construct the processing state parameter group, and determine the candidate processing parameter through reinforcement learning; according to the simulation prediction value and the actual cutting state, the reliability is determined and the constraint parameter is updated; the variable safety limit is constructed to correct the candidate processing parameter, and the final processing instruction is obtained by weighted fusion; the present application improves the cutting stability and reduces the tool damage probability.

Description

Technical Field

[0001] This invention relates to the field of metal cutting and intelligent control technology, specifically to a method for optimizing machining parameters in a metal processing mechanical system based on reinforcement learning. Background Technology

[0002] During the cutting process, metal processing machinery systems typically require adjustments to machining parameters such as spindle speed, feed rate, axial depth of cut, and radial width of cut to balance material removal efficiency, machining quality, and equipment operational stability. Furthermore, they can monitor and assess the machining status by incorporating information on vibration, cutting force, acoustic emission, and spindle load, thus possessing strong operational condition characterization capabilities.

[0003] With the development of intelligent manufacturing and adaptive machining control technology, the optimization of machining parameters has gradually evolved from fixed empirical settings to dynamic adjustment based on model prediction, sensor feedback and intelligent algorithms. However, in complex metal cutting scenarios, existing technologies are still easily affected by factors such as chatter, impact, tool wear and changes in the credibility of simulation models, making it difficult to effectively coordinate the parameter optimization results with the actual executable safety range, making it difficult to simultaneously guarantee machining efficiency, surface quality and stability. Summary of the Invention

[0004] The purpose of this invention is to provide a method for optimizing machining parameters in metalworking machinery systems based on reinforcement learning. This method avoids the difficulty in coordinating parameter optimization with actual safety limits due to factors such as chatter or impact, and simultaneously ensures machining efficiency, surface quality, and stability. Specifically, the technical solution of this invention includes:

[0005] Step 1: Acquire physical sensing data of the metal processing machinery system, extract instantaneous risk indicators within the time-series sampling window, and determine the flutter risk index and impact risk index accordingly;

[0006] Step 2: Construct a machining state parameter group based on chatter risk index, impact risk index and machine tool process parameters, and output candidate machining parameters through reinforcement learning optimization unit. The machine tool process parameters include at least spindle speed, feed rate, axial depth of cut and radial width of cut.

[0007] Step 3: Determine the reliability parameters of the simulation model based on the cutting state prediction error output by the machining simulation unit; and update the constraint parameters, including the safety clearance shrinkage coefficient, safety function convergence coefficient, and smoothing coefficient, in conjunction with the chatter risk index and impact risk index.

[0008] Step 4: Construct variable safety limits based on the updated constraint parameters. Under the constraints of the upper and lower bounds of parameter execution and the maximum allowable change, modify the candidate processing parameters to obtain safe processing parameters.

[0009] Step 5: Based on the simulation model's credibility parameters, flutter risk index, and impact risk index, dynamically configure the adjustment weights of the safe processing parameters, the simulation recommended parameters output by the processing simulation unit, and the feedback adjustment parameters generated by the feedback control unit, and perform weighted fusion to generate the final processing command for execution.

[0010] Preferably, the physical sensing data includes vibration sensing data, acoustic emission sensing data, cutting force sensing data, and spindle current sensing data;

[0011] Instantaneous risk indicators include vibration energy, acoustic emission energy, cutting force step rate, and spindle load deviation rate of spindle current relative to the historical steady-state machining current mean obtained within the time-series sampling window.

[0012] Preferably, the chatter risk index is determined by vibration energy, acoustic emission energy, cutting force step rate, and spindle load deviation rate;

[0013] When a certain instantaneous risk indicator meets the preset flutter event judgment conditions, the index coefficient corresponding to the instantaneous risk indicator is amplified by a preset multiple.

[0014] Preferably, the impact risk index is determined by the cutting force step rate, the spindle load deviation rate, and the vibration peak value and acoustic emission burst characteristics extracted based on physical sensing data.

[0015] Preferably, the machining status parameter group includes statistical values ​​of chatter risk index, impact risk index, machine tool spindle speed, machine tool feed rate, machine tool axial depth of cut, machine tool radial width of cut, cutting force, spindle current, and surface roughness and tool wear estimates derived from physical sensor data and machine tool process parameters within a preset statistical period.

[0016] Preferably, the updating of the constraint parameters includes:

[0017] Update the safety clearance contraction coefficient based on the simulation model's credibility parameters, flutter risk index, and shock risk index;

[0018] Based on the simulation model's credibility parameters and flutter risk index, and combined with the preset basic safety function convergence coefficients, the safety function convergence coefficients are updated.

[0019] Based on the impact risk index and the difference between the candidate processing parameters and the parameters executed in the previous cycle, the smoothing coefficient is updated in conjunction with the preset basic smoothing coefficient.

[0020] Preferably, the construction of the variable safety clearance includes:

[0021] A safety state function is constructed based on the allowable upper limit values ​​corresponding to the chatter risk index, impact risk index, estimated tool wear, and spindle current monitoring value.

[0022] Based on the safety state function, the convergence coefficient of the safety function, and the sensitivity coefficient of the safety state function with respect to each processing parameter, safety constraints are constructed to form a variable safety limit.

[0023] The upper and lower bounds of the variable safety limit are updated by the safety limit contraction coefficient.

[0024] Preferably, the determination of safe processing parameters includes: under the conditions of meeting safety constraints, upper and lower bounds of parameter execution, and the maximum allowable change within a preset control cycle, optimizing the calculation based on candidate processing parameters, smoothing coefficients, and parameters executed in the previous cycle to obtain safe processing parameters.

[0025] Preferably, the adjustment weights include learning control weights, simulation correction weights, and fast feedback control weights;

[0026] The determination of adjustment weights includes:

[0027] When the simulation model's credibility parameter is greater than the first credibility threshold, the flutter risk index is less than the first risk threshold, and the impact risk index is less than or equal to the first impact threshold, the learning control weights will be set to values ​​within the first preset weight range.

[0028] When the simulation model's credibility parameter is greater than the first credibility threshold, the flutter risk index is greater than or equal to the first risk threshold, and the impact risk index is less than or equal to the first impact threshold, the simulation correction weight is set to a value within the second preset weight range.

[0029] When the confidence parameter of the simulation model is less than or equal to the first confidence threshold or the impact risk index is greater than the first impact threshold, the fast feedback control weight will be set to a value within the third preset weight range.

[0030] Preferably, steps one through five are performed in the first control cycle, the second control cycle, and the third control cycle, respectively, wherein the determination of the instantaneous risk index and the generation of feedback adjustment parameters by the feedback control unit are performed in the first control cycle.

[0031] The updating of constraint parameters and the determination of the corrected machining parameters are carried out in the second control cycle;

[0032] The output of candidate processing parameters and the calibration of simulation-recommended parameters are performed in the third control cycle;

[0033] The duration of the first control cycle is shorter than the duration of the second control cycle, and the duration of the second control cycle is shorter than the duration of the third control cycle.

[0034] Compared with the prior art, the present invention has the following beneficial effects:

[0035] 1. This invention continuously collects vibration sensing data, acoustic emission sensing data, cutting force sensing data, and spindle current sensing data, and extracts instantaneous risk indicators such as vibration energy, acoustic emission energy, cutting force step rate, and spindle load deviation rate within a time-series sampling window. It establishes a joint characterization mechanism for chatter risk index and impact risk index, enabling the differentiation and identification of abnormal cutting stability and transient impact anomalies. This can more accurately reflect the actual working condition changes in complex metal cutting scenarios, thereby enhancing the timeliness and pertinence of machining status judgment.

[0036] 2. This invention constructs a machining state parameter set by combining chatter risk index, impact risk index, spindle speed, feed rate, axial depth of cut, radial width of cut, cutting force statistics, spindle current statistics, surface roughness estimate, and tool wear estimate. Candidate machining parameters are then output by a reinforcement learning optimization unit. This establishes a parameter optimization path that takes into account material removal efficiency, surface quality, and stability risk. It avoids the problem of difficulty in distinguishing between stable and efficient working conditions and near-instability working conditions when adjusting only a single process parameter, thereby improving the adaptability of parameter optimization results. Attached Figure Description

[0037] To more clearly illustrate the technical solutions in the embodiments of this application and the prior art, the following is a brief introduction to the drawings used in the description of the embodiments and the prior art:

[0038] Figure 1 This is a schematic diagram of the steps of a method for optimizing processing parameters in a metal processing machinery system based on reinforcement learning, according to the present invention. Detailed Implementation

[0039] To make the above and other objects, features and advantages of the present invention more apparent and understandable, preferred embodiments will be described in detail below with reference to the accompanying drawings.

[0040] A method for optimizing machining parameters in a metalworking machinery system based on reinforcement learning includes the following steps:

[0041] Step 1: Acquire physical sensing data of the metal processing machinery system, extract instantaneous risk indicators within the time-series sampling window, and determine the flutter risk index and impact risk index accordingly;

[0042] Step 2: Construct a machining state parameter group based on chatter risk index, impact risk index and machine tool process parameters, and output candidate machining parameters through reinforcement learning optimization unit. The machine tool process parameters include at least spindle speed, feed rate, axial depth of cut and radial width of cut.

[0043] Step 3: Determine the reliability parameters of the simulation model based on the cutting state prediction error output by the machining simulation unit; and update the constraint parameters, including the safety clearance shrinkage coefficient, safety function convergence coefficient, and smoothing coefficient, in conjunction with the chatter risk index and impact risk index.

[0044] Step 4: Construct variable safety limits based on the updated constraint parameters. Under the constraints of the upper and lower bounds of parameter execution and the maximum allowable change, modify the candidate processing parameters to obtain safe processing parameters.

[0045] Step 5: Based on the simulation model's credibility parameters, flutter risk index, and impact risk index, dynamically configure the adjustment weights of the safe processing parameters, the simulation recommended parameters output by the processing simulation unit, and the feedback adjustment parameters generated by the feedback control unit, and perform weighted fusion to generate the final processing command for execution;

[0046] Physical sensing data includes vibration sensing data, acoustic emission sensing data, cutting force sensing data, and spindle current sensing data;

[0047] Instantaneous risk indicators include vibration energy, acoustic emission energy, cutting force step rate, and spindle load deviation rate of spindle current relative to the historical steady-state machining current mean obtained within the time sampling window.

[0048] The chatter risk index is determined by vibration energy, acoustic emission energy, cutting force step rate, and spindle load deviation rate;

[0049] When a certain instantaneous risk indicator meets the preset flutter event judgment conditions, the index coefficient corresponding to the instantaneous risk indicator is amplified by a preset multiple.

[0050] The impact risk index is determined by the cutting force step rate, spindle load deviation rate, and vibration peak value and acoustic emission burst characteristics extracted based on physical sensing data.

[0051] The machining status parameter group includes statistical values ​​of chatter risk index, impact risk index, machine tool spindle speed, machine tool feed rate, machine tool axial depth of cut, machine tool radial width of cut, cutting force, spindle current, and surface roughness and tool wear estimates derived from physical sensor data and machine tool process parameters within a preset statistical period.

[0052] This implementation method applies to a type of metalworking mechanical system that has a spindle drive unit, a feed drive unit, a CNC execution unit, and an edge computing node, such as... Figure 1 As shown;

[0053] The edge computing node is connected to the vibration sensor, acoustic emission sensor, cutting force sensor and spindle current acquisition interface. The CNC execution unit receives the machining instructions issued by the edge computing node and adjusts one or more of the spindle speed, feed rate, axial depth of cut and radial width of cut accordingly.

[0054] In this embodiment, physical sensing data refers to data that directly reflects the objective mechanical state and electrical load state of the cutting process. Specifically, vibration sensing data corresponds to the vibration amplitude changes of the tool and spindle system during cutting, acoustic emission sensing data corresponds to the high-frequency acoustic emission characteristics generated when the tool contacts the workpiece, cutting force sensing data corresponds to the force changes during the cutting process, and spindle current sensing data corresponds to the spindle load changes. All of the above data are sent to the edge computing node by the sensing acquisition module without human intervention.

[0055] At the beginning of the process, the sensor acquisition module first acquires the aforementioned physical sensing data within a continuously scrolling time-series sampling window; within each time-series sampling window, the edge computing node accumulates the energy of the vibration sensing data to obtain the vibration energy; and accumulates the energy of the acoustic emission sensing data to obtain the acoustic emission energy.

[0056] The cutting force step rate is obtained by subtracting the cutting force statistics of adjacent windows and converting them according to the window length; the spindle load deviation rate is obtained by comparing the current average spindle current with the historical steady-state machining current average.

[0057] Calculate the absolute value of the difference between the current window spindle current average and the historical steady-state machining current average, and then divide it by the historical steady-state machining current average to transform the original high-frequency sensing data into an instantaneous risk indicator.

[0058] Among them, the historical steady-state machining current average value is taken from the statistical results of stable cutting under the same tool, the same workpiece material, and the same combination of process parameters; since the spindle current varies with the speed and feed, by using the historical steady-state machining current average value as a comparison benchmark, the current spindle load state can be accurately assessed and the precursors of shock or chatter can be identified;

[0059] When determining the chatter risk index, the edge computing nodes normalize the vibration energy, acoustic emission energy, cutting force step rate and spindle load deviation rate respectively, and then sum them up according to the preset index coefficients.

[0060] Specifically, the normalization process uses a minimax conversion method to uniformly map all indicators to the [0,1] interval. The normalized value of vibration energy is equal to the vibration energy calculated in the current window minus the minimum vibration energy benchmark value collected under historical steady-state processing, and then divided by the difference between the maximum and minimum vibration energy benchmark values ​​under historical extreme flutter conditions. When the maximum and minimum benchmark values ​​are equal, the normalized value is directly set to 0.

[0061] Acoustic emission energy, cutting force step rate and spindle load deviation rate are normalized and converted according to the same logic; the preset chatter event judgment condition is that a certain instantaneous risk indicator exceeds its corresponding warning threshold within a certain number of consecutive sampling windows, or a certain indicator does not exceed the limit continuously but the rising slope exceeds the preset change threshold.

[0062] The basis for determining the warning threshold is the distribution statistics of the historical time series data of steady-state machining of the same tool under the same type of workpiece material. The steady-state mean plus three times the standard deviation is taken as the warning threshold of the corresponding indicator.

[0063] If the judgment condition is met, the coefficient of the instantaneous risk indicator will be switched to a higher coefficient. For example, the original preset coefficient of the indicator will be increased by a preset multiple, and the coefficients of the other indicators will be reduced at the same time to keep the sum of all indicator coefficients normalized to 1, so as to increase the proportion of the indicator in the flutter risk index.

[0064] When vibration or acoustic emission is abnormal, the risk index is accelerated by triggering the above coefficient switching mechanism, thus triggering the tightening action of safety constraints in advance.

[0065] When determining the impact risk index, the edge computing nodes adopt a different combination of features than the chatter risk index. Specifically, the cutting force step rate is used to represent the instantaneous force change, the vibration peak value is used to represent the instantaneous excitation of the structural component, the spindle load deviation rate is used to represent the degree of change on the load side, and the acoustic emission burst characteristics are used to represent the precursors of abnormal friction or micro-chipping of the tool edge.

[0066] The specific method for obtaining the acoustic emission burst characteristics is as follows: count the number of pulses or ringings whose acoustic emission signal amplitude exceeds a preset normal fluctuation threshold within the current time-series sampling window, and use these counts as the acoustic emission burst characteristic values ​​for that window.

[0067] After the edge computing nodes normalize the above four features by performing interval mapping according to the same minimax conversion rule, they are combined according to the corresponding weights to obtain the impact risk index. In view of the physical characteristics of impacts, which are usually manifested as short-term spikes, this step uses vibration peak value and acoustic emission sudden characteristics to accurately reflect the real transient impact conditions.

[0068] In step two, the edge computing node combines the flutter risk index, the shock risk index, and the machine tool process parameters into a processing state parameter set; the processing state parameter set is the input parameter actually read by the reinforcement learning optimization unit in the current statistical period;

[0069] In this embodiment, the parameter set includes at least: the statistical values ​​of chatter risk index, impact risk index, spindle speed, feed rate, axial depth of cut, radial width of cut, cutting force, spindle current, surface roughness estimate, and tool wear estimate obtained within a preset statistical period.

[0070] Among them, the statistical values ​​of chatter risk index and impact risk index can be the mean, maximum or weighted statistical values ​​within a preset statistical period; the statistical values ​​of cutting force and spindle current can be the average values ​​within the same statistical period; the estimated surface roughness value... Based on the empirical formula for the exponent:

[0071]

[0072] in, Main spindle speed For feed rate, The axial cutting depth, The workpiece material constants are pre-fitted through calibration experiments; the estimated tool wear value is... The control cycle is updated recursively using a Taylor extension model based on cutting work.

[0073]

[0074] in, This is the estimated tool wear value for the current cycle; This is the estimated tool wear value from the previous cycle; and These are the times of the current control cycle and the previous control cycle, respectively; For integration time; The wear coefficient of the material; This is instantaneous cutting force sensing data; This refers to the instantaneous cutting linear velocity;

[0075] By incorporating the aforementioned multidimensional parameters into the state parameter group, the reinforcement learning optimization unit can accurately distinguish between the stable cutting state under high material removal rate and the two technical conditions approaching the instability boundary.

[0076] After receiving the machining state parameter set, the reinforcement learning optimization unit outputs candidate machining parameters. Specifically, the reinforcement learning optimization unit can use a deep deterministic policy gradient algorithm to construct the machining state parameter set into an accurate 10-dimensional state feature vector, which corresponds to the statistical values ​​of chatter risk index, impact risk index, spindle speed, feed rate, axial depth of cut, radial width of cut, cutting force, spindle current, surface roughness estimate, and tool wear estimate, and inputs them into the trained action network.

[0077] The action network employs a fully connected neural network structure with two hidden layers, with the number of nodes in the hidden layers set to 64 and 32 respectively, and uses a modified linear unit activation function; the output layer uses a hyperbolic tangent activation function to limit the output extrema.

[0078] The deep deterministic policy gradient algorithm also includes an evaluation network for assessing action value. This evaluation network concatenates a 10-dimensional state feature vector with a 4-dimensional continuous vector output by the action network as input layer features, which then pass through two hidden layers with 128 and 64 nodes respectively. The hidden layers use a modified linear unit activation function, and the output layer is a single-node linear output without an activation function, used to predict the action value assessment value of the state-action pair.

[0079] During model training and updates, the evaluation network uses mean squared error as the loss function, and its objective value is the sum of the action value assessment value of the next state and the current reward; the action network calculates the deterministic policy gradient by maximizing the action value assessment value output by the evaluation network, and updates the action network weights accordingly.

[0080] The motion network ultimately outputs a precise 4-dimensional continuous vector, representing the current optimization increment of the normalized spindle speed, feed rate, axial depth of cut, and radial width of cut; the candidate machining parameters are the target parameters to be executed in the next control cycle, which are obtained by adding the parameters executed in the previous cycle to the current optimization increment after inverse normalization;

[0081] To train this action network, a specific reward function was defined, with the following calculation logic: the material removal rate of the current cycle is set as a positive reward; the portion of the surface roughness estimate exceeding the preset roughness requirement is set as the first penalty; and the sum of the flutter risk index and the impact risk index is set as the second penalty. The specific mathematical mapping is as follows:

[0082]

[0083] in, This is the reward function value for the current period; This represents the material removal rate for the current cycle. It is a positive extreme value; This is an estimate of the surface roughness. The target roughness is the baseline. For flutter risk index; To impact the risk index; , , The positive weight coefficients are preset; the above formula is used to convert each term into a dimensionless pure number and multiply it by the corresponding weight for linear combination, so that the network optimization can take into account efficiency, quality and stability on the same order of magnitude;

[0084] The candidate parameters are not sent directly to the CNC execution unit, but instead enter the subsequent constraint correction process. The above processing decouples the optimization action and the safe execution action, avoiding the direct interference of the low-frequency parameter optimization results with the transient cutting state.

[0085] In step three, the machining simulation unit outputs the cutting state prediction value based on the current state parameter group and candidate machining parameters; the cutting state prediction value can include at least the predicted cutting force, predicted spindle current, predicted tool wear trend, and predicted stability-related quantities.

[0086] Meanwhile, the machining simulation unit generates a stability limit diagram based on the preset machine tool dynamics model and cutting mechanics model. This diagram divides the stability domain and chatter domain with the feedforward spindle speed as the horizontal axis and the limit axial depth of cut as the vertical axis.

[0087] The edge computing node extracts the maximum depth of cut envelope without chatter from the limit map as the search space. With the material removal rate of the current cycle as the lower limit, the sequential iterative search algorithm is used to optimize along the envelope. The spindle speed and maximum axial depth of cut corresponding to the peak position on the envelope are selected. The feed rate and radial width are calculated according to the equal volume removal rate. The combination of process parameters that maximizes the cutting depth is used as the simulation recommended parameters output.

[0088] The edge computing nodes simultaneously read the actual cutting status and compare the actual values ​​with the predicted values ​​one by one to form the prediction error. If the prediction error is less than the preset error threshold, it means that the simulation model is consistent with the actual cutting under the current working condition. If the prediction error is greater than or equal to the preset error threshold, it means that the simulation model has drifted and the constraint effect of the current simulation conclusion is canceled.

[0089] The confidence parameter of the simulation model can be obtained from the prediction error through a monotonically decreasing relationship based on exponential decay; specifically, the confidence parameter of the simulation model is equal to the negative power of the natural constant, and the exponent of the negative power is the product of the preset decay coefficient and the absolute value of the current prediction error.

[0090] This explicit quantitative calculation method ensures that the smaller the prediction error, the higher the simulation model's credibility parameter, which approaches 1; the larger the prediction error, the lower the simulation model's credibility parameter, which approaches 0; the edge computing nodes then update the constraint parameters based on the simulation model's credibility parameter, flutter risk index, and shock risk index.

[0091] The update logic is as follows: when the credibility of the simulation model decreases, the safety clearance shrinkage coefficient is increased, which reduces the executable range; when the flutter risk index increases, the safety function convergence coefficient is increased, which makes the safe state return to the safe side faster; when the impact risk index increases or the difference between the candidate processing parameter and the parameter executed in the previous cycle exceeds the preset difference threshold, the smoothing coefficient is increased, which controls the parameter change rate within the preset change rate threshold.

[0092] In step four, the edge computing nodes construct variable safety limits using the updated constraint parameters. The variable safety limits are configured as an executable range that is dynamically updated with the current simulation confidence level, risk level, wear state, and load state. The limits are simultaneously constrained by three aspects: first, the upper and lower bounds of parameter execution; second, the maximum allowable change within the preset control period; and third, the safety constraints corresponding to the safety state function.

[0093] Among them, the upper and lower bounds of parameter execution are derived from the intersection range of machine tool process limits, tool recommended process window, and stable machining database; the maximum allowable change within the preset control cycle is derived from the change capability that the actuator can complete within one control cycle;

[0094] Under the above constraints, the edge computing node optimizes and corrects the candidate processing parameters so that the corrected processing parameters are as close as possible to the candidate processing parameters, while maintaining the continuity of changes with the parameters executed in the previous cycle; the output result is the corrected processing parameters, which is the safe processing parameters.

[0095] In step five, the edge computing node further determines the adjustment weights of three types of parameters: the learning control weights corresponding to the safe machining parameters, the simulation correction weights corresponding to the simulation recommendation parameters, and the fast feedback control weights corresponding to the feedback adjustment parameters. Then, the weights are weighted and fused to obtain the final machining instructions, which are then converted by the CNC execution unit into control quantities for the spindle speed, feed rate, axial depth of cut, and radial width of cut.

[0096] In a typical machining scenario, when the local material cutting resistance of the workpiece suddenly increases, the cutting force step rate and spindle load deviation rise first, and the edge computing node then increases the impact risk index; if the vibration energy and acoustic emission energy continue to increase, the chatter risk index increases synchronously.

[0097] At this point, although the reinforcement learning optimization unit may still provide candidate machining parameters to increase the feed rate, the safety limits for variation will be tightened due to the combined effect of the simulation model's credibility and risk indicators. The corrected machining parameters are restricted to a more conservative range, and the final machining command is reflected in a smaller feed adjustment range and a smoother change in rotational speed. In this way, the output results directly affect the objective physical quantities of the machine tool, reducing the possibility of tool breakage and instability.

[0098] Furthermore, the updating of constraint parameters includes:

[0099] Update the safety clearance contraction coefficient based on the simulation model's credibility parameters, flutter risk index, and shock risk index;

[0100] Based on the simulation model's credibility parameters and flutter risk index, and combined with the preset basic safety function convergence coefficients, the safety function convergence coefficients are updated.

[0101] Based on the impact risk index and the difference between the candidate processing parameters and the parameters executed in the previous cycle, the smoothing coefficient is updated in conjunction with the preset basic smoothing coefficient.

[0102] The construction of variable safety limits includes:

[0103] A safety state function is constructed based on the allowable upper limit values ​​corresponding to the chatter risk index, impact risk index, estimated tool wear, and spindle current monitoring value.

[0104] Based on the safety state function, the convergence coefficient of the safety function, and the sensitivity coefficient of the safety state function with respect to each processing parameter, safety constraints are constructed to form a variable safety limit; and the upper and lower bounds of the variable safety limit parameters are updated through the safety limit contraction coefficient.

[0105] The determination of safe processing parameters includes:

[0106] Under the conditions of meeting safety constraints, upper and lower bounds of parameter execution, and the maximum allowable change within the preset control cycle, the safe processing parameters are obtained by optimization calculation based on the candidate processing parameters, smoothing coefficient, and parameters executed in the previous cycle.

[0107] This implementation focuses on the process of updating constraint parameters, constructing variable safety limits, and obtaining the corrected machining parameters; the execution entity is still the edge computing node, which is communicatively connected to the machining simulation unit, the sensor acquisition module, and the CNC execution unit.

[0108] In this embodiment, the safety clearance contraction coefficient is a quantity used to adjust the width of the parameter execution range; when its value tends to be conservative, the deviation of the candidate processing parameter from the current process center is reduced; when its value tends to be relaxed, the exploration space allowed for the candidate processing parameter is increased.

[0109] The convergence coefficient of the safety function is the constraint strength for the regression of the safe state to the safe side, and is used to control the degree of conservatism near the safety boundary; the smoothing coefficient is used to constrain the rate of change of the correction parameter in the current cycle relative to the execution parameter in the previous cycle, and to prevent the machine tool execution chain from being subjected to additional shocks due to excessively rapid parameter changes;

[0110] When updating constraint parameters, the edge computing node first receives three inputs: the simulation model confidence parameter, the flutter risk index, and the shock risk index; simultaneously, it reads the difference between the candidate processing parameters and the parameters executed in the previous cycle; the update order can be performed according to the following logic:

[0111] First, update the safety clearance shrinkage coefficient. If the simulation model's reliability parameter decreases, it indicates that the simulation model's description of the current cutting state is less reliable, triggering the shrinkage of the execution space. If the chatter risk index increases, it indicates that the cutting stability is deteriorating, triggering the shrinkage of the execution space. If the impact risk index increases, it indicates that there is a transient force anomaly, simultaneously triggering the shrinkage of the execution space. The safety clearance shrinkage coefficient is calculated by co-updating the simulation reliability and the two types of risk indices.

[0112] The safety function convergence coefficient is then updated; this coefficient is based on the fundamental safety function convergence coefficient and is corrected by combining the simulation model's credibility parameters and the flutter risk index; the safety function convergence coefficient is used to constrain the flutter boundary of the system principle.

[0113] When the credibility of the simulation model decreases or the flutter risk index increases, the edge computing node increases this coefficient to enhance the constraint strength of subsequent safety constraints on candidate actions; when the credibility of the simulation model decreases or the flutter risk index increases, the edge computing node increases this coefficient to make the subsequent safety constraints on candidate actions stronger.

[0114] Update the smoothing coefficient; this coefficient is based on the basic smoothing coefficient, and is superimposed with the impact risk index and the difference between the candidate machining parameters and the parameters executed in the previous cycle; if the impact risk index is high, it indicates that there is a significant transient impact in the cutting process. In this case, even if the direction of the candidate machining parameters is correct, it is not advisable to adjust it too much in one step.

[0115] If the difference between the candidate processing parameter and the parameter executed in the previous cycle exceeds the preset jump threshold, the smoothing coefficient is increased; the smoothing process limits the difference between the correction parameter and the parameter executed in the previous cycle to meet the preset execution continuity requirements.

[0116] When constructing the variable safety limit, the edge computing node first establishes a safety state function; this safety state function is composed of the chatter risk index, the impact risk index, the estimated value of tool wear, and the monitoring value of spindle current, and each quantity is compared with its respective allowable upper limit value;

[0117] Specifically, the safety state function The expression is constructed using a normalized weighted quadratic form of each index relative to its allowable upper limit, to ensure mathematical continuity and differentiability:

[0118]

[0119] in, for The safety state function value at any given time; For flutter risk index, Set its maximum allowed value; To impact the risk index, Set its maximum allowed value; This is an estimate of the tool wear. Set its maximum allowed value; The main spindle current monitoring value, Set its maximum allowed value; to The positive sensitivity coefficients are pre-calibrated; and when any variable approaches its upper limit, the penalty term is rapidly amplified quadratically.

[0120] As long as the actual values ​​of the above items are within the allowable upper limit, the safety state function remains in the safety domain; if any item is close to the allowable upper limit, the safety state function approaches the safety boundary; if any item exceeds the allowable upper limit, the safety state function enters the out-of-bounds state.

[0121] The above allowable upper limits are set based on specific processing objects and process constraints; the chatter risk upper limit is taken from the risk statistical interval before the continuous amplitude increase occurs in the stable processing test; the impact risk upper limit is taken from the cutting force step statistical interval before chipping occurs.

[0122] The upper limit of tool wear is taken from the allowable wear amount in the tool life standard; the upper limit of spindle current is taken from the spindle drive capability and the allowable range of thermal load; after these upper limits are uniformly introduced into the safety state function, risk, wear and load are no longer isolated monitoring quantities, but jointly participate in the subsequent constraint calculation.

[0123] After the safety state function is established, the edge computing nodes combine the convergence coefficient of the safety function and the sensitivity coefficient of the safety state function with respect to each processing parameter to construct safety constraints.

[0124] The specific construction logic is as follows: the predicted value of the safety state function in the next period, that is, the sum of the current safety state function value plus the product of the change of each processing parameter and the corresponding sensitivity coefficient, should not exceed the safety regression target after combining the current safety state function value with the attenuation increment.

[0125] Specifically, the safety regression target is derived through the following quantitative decomposition: the current safety state function value is subtracted from the allowable upper limit value of the safety state function to obtain the deviation margin; the deviation margin is then multiplied by the convergence coefficient of the safety function to obtain the decay increment; finally, the current safety state function value is added to the decay increment as the limit boundary for the next period.

[0126] The sensitivity coefficient characterizes the trend of change in the safety state function when the spindle speed, feed rate, axial depth of cut, or radial width of cut change independently near the current working condition.

[0127] The sensitivity coefficient can be provided by the processing simulation unit; when the reliability of the simulation model is insufficient, the edge computing node will perform data-driven recursive estimation based on the actual measurement data in the most recent control cycle.

[0128] The specific structured calculation rule is as follows: extract the difference between the actual execution values ​​of each processing parameter in the current control cycle and the previous cycle to construct a multi-dimensional input change vector, and at the same time extract the difference between the actual calculated values ​​of the safety state function in these two cycles as the output change scalar;

[0129] Considering the mathematical coupling effect caused by the simultaneous change of multiple parameters, if the absolute value of the input change of only a single processing parameter is greater than the preset dead zone threshold, the output change scalar is directly divided by the input change of that parameter to obtain the instantaneous gradient.

[0130] If the absolute values ​​of the input changes of multiple processing parameters are all greater than the preset dead zone threshold, a recursive least squares algorithm with a forgetting factor is adopted. The multidimensional input change vector is used as the input feature, and the output change scalar is used as the observation benchmark. The sensitivity vector of the current cycle is jointly and decoupled in the multidimensional space, thereby avoiding pseudo gradient distortion caused by direct division in a single dimension.

[0131] Then, the moving average filtering algorithm is used to perform a weighted average of the currently obtained instantaneous gradient or sensitivity vector with the sensitivity estimate of the previous period, and the current sensitivity coefficient is updated.

[0132] If the absolute value of all input changes is less than or equal to the preset dead zone threshold, the estimated value of the previous cycle remains unchanged; the edge computing node determines the trend vector of the system safety state function when the processing parameters increase based on the sensitivity coefficient;

[0133] After obtaining the safety constraints, the edge computing node forms a variable safety bound; in this embodiment, the variable safety bound is configured as an executable constraint range that is updated in real time with risk, wear, load and model confidence; the edge computing node then uses the safety bound shrinkage coefficient to update the upper and lower bounds of the parameters of this range.

[0134] If the safety clearance shrinkage coefficient increases, the execution upper and lower bounds will converge towards the current process center; if the safety clearance shrinkage coefficient decreases, the execution upper and lower bounds will be appropriately widened; thus, the executable parameter boundaries are determined by the real-time safety state function.

[0135] During the modified processing parameter acquisition stage, the edge computing node simultaneously sends the candidate processing parameters, smoothing coefficients, and parameters executed in the previous cycle into the optimization calculation process.

[0136] The optimization process must satisfy at least three conditions: First, it must satisfy the above-mentioned safety constraints; second, it must satisfy the upper and lower bounds of the updated parameter execution; and third, it must satisfy the maximum allowable change within the preset control cycle. The maximum allowable change is limited by the mechanical adjustment limit of the machine tool actuator within a control cycle, which is used to limit abnormal parameter jumps.

[0137] The optimization calculation uses a clear objective function for solution, which can be summarized into two items: the first item is the sum of squares of the differences between the corrected processing parameters and the candidate processing parameters, which is used to make the corrected processing parameters as close as possible to the candidate processing parameters and maintain the optimization direction; the second item is the sum of squares of the differences between the corrected processing parameters and the parameters executed in the previous cycle, multiplied by a smoothing coefficient, which is used to constrain the deviation by the smoothing coefficient and avoid abrupt changes.

[0138] The edge computing node adds the above two items as the overall objective function, and uses convex optimization algorithms such as sequential quadratic programming to solve for the minimum value of the overall objective function within the constraint domain formed by the above three types of conditions. The solution obtained at this time is the corrected processing parameters. The corrected processing parameters output by the optimization solution satisfy the safety boundary conditions and the control cycle adjustment threshold of the actuator.

[0139] In a specific machining condition, when the estimated value of tool wear gradually increases and the value of spindle current monitoring is close to the upper limit of allowable limit, the safety state function will shrink toward the boundary. At this time, even if the chatter risk index and the shock risk index have not exceeded the limit significantly, the safety limit shrinkage coefficient will still increase due to the combined effect of the reliability of the simulation model and the load level, resulting in the narrowing of the upper and lower bounds of parameter execution.

[0140] If the candidate machining parameters output by the reinforcement learning optimization unit still result in an increase in feed rate, the optimization calculation will, under the combined effect of safety constraints and smoothing constraints, correct the candidate machining parameters to a smaller feed rate change, or only retain a small correction in rotational speed; thus, the machining parameters output to the CNC execution unit are truly executable and unlikely to cause further load increases.

[0141] Furthermore, the adjustment weights include: learning control weights, simulation correction weights, and fast feedback control weights;

[0142] The determination of the adjustment weights includes: when the simulation model confidence parameter is greater than the first confidence threshold, the flutter risk index is less than the first risk threshold, and the impact risk index is less than or equal to the first impact threshold, the learning control weights are set to values ​​within the first preset weight range.

[0143] When the simulation model's credibility parameter is greater than the first credibility threshold, the flutter risk index is greater than or equal to the first risk threshold, and the impact risk index is less than or equal to the first impact threshold, the simulation correction weight is set to a value within the second preset weight range.

[0144] When the confidence parameter of the simulation model is less than or equal to the first confidence threshold or the impact risk index is greater than the first impact threshold, the fast feedback control weight will be set to a value within the third preset weight range.

[0145] Steps one through five are executed in the first control cycle, the second control cycle, and the third control cycle, respectively. The determination of instantaneous risk indicators and the generation of feedback adjustment parameters by the feedback control unit are executed in the first control cycle.

[0146] The updating of constraint parameters and the determination of the corrected machining parameters are carried out in the second control cycle;

[0147] The output of candidate processing parameters and the calibration of simulation-recommended parameters are performed in the third control cycle;

[0148] The duration of the first control cycle is shorter than the duration of the second control cycle, and the duration of the second control cycle is shorter than the duration of the third control cycle.

[0149] This implementation focuses on explaining the adjustment weight arbitration logic and the operation mode under multiple control cycles; the execution entities include the sensor acquisition module, the feedback control unit, the reinforcement learning optimization unit, the machining simulation unit, and the numerical control execution unit in the edge computing node;

[0150] In this embodiment, the learning control weight is the proportion of safe processing parameters in the final processing command, used to reflect the degree of execution of the reinforcement learning optimization results; the simulation correction weight is the proportion of simulation recommended parameters in the final processing command, used to reflect the degree of conservative correction of the processing simulation unit to the current working condition; the fast feedback control weight is the proportion of feedback adjustment parameters in the final processing command, used to reflect the degree of suppression of transient anomalies by the feedback control unit; the three together participate in weighted fusion, and the sum remains unchanged;

[0151] Before determining the weights, the system first updates three basic input parameters: the simulation model credibility parameter, the chatter risk index, and the impact risk index. After all three parameters have been updated, the edge computing nodes perform weight allocation according to the following judgment conditions: The first case is when the simulation model credibility parameter is greater than the first credibility threshold and the chatter risk index is less than the first risk threshold. This indicates that the simulation model is consistent with the actual cutting, and the current cutting stability is within the preset safety threshold.

[0152] Edge computing nodes set the learning control weights to a preset high weight, such as a value between 0.7 and 0.9, and distribute the remaining weights equally or according to a preset ratio to the simulation correction weights and fast feedback control weights, so that the final processing command is closer to the safe processing parameters obtained by reinforcement learning.

[0153] By setting a higher preset weight to accelerate the optimization convergence rate within the safety boundary, parameters such as spindle speed and feed rate are guided to continuously adjust towards the optimization range;

[0154] The second scenario is when the simulation model's credibility parameter is greater than the first credibility threshold and the chatter risk index is greater than or equal to the first risk threshold; this indicates that the simulation model is still reliable, but the cutting process has approached or entered the chatter-sensitive area.

[0155] Edge computing nodes set the simulation correction weights to a preset high weight, such as a value between 0.6 and 0.8, while correspondingly reducing the learning control weights, so that the final processing instructions are more biased towards the simulation recommended parameters;

[0156] The simulation-recommended parameters represent the safety-side parameter boundaries that tend towards stable processing. Through the above weight configuration, the control commands are preferentially adjusted towards stability, while simultaneously suppressing the exploratory actions of reinforcement learning.

[0157] The third scenario is when the simulation model's credibility parameter is less than or equal to the first credibility threshold, or the impact risk index is greater than the first impact threshold. If either of these conditions is met, the edge computing node will set the fast feedback control weight to a preset high weight, for example, directly set it to above 0.8, and reduce the other two types of weights to a minimum value.

[0158] When the simulation model is not reliable enough or there are significant shock anomalies, there is a risk of control hysteresis based on reinforcement learning output or simulation recommended parameters; the system switches the weight configuration so that the feedback adjustment parameters dominate the final processing command;

[0159] The feedback control unit outputs more direct parameter corrections, such as forcibly reducing the feed rate and depth of cut, or adjusting the spindle speed to an empirically stable speed range, so as to produce a vibration suppression effect on the physical execution surface of the machine tool that is equivalent to increasing the system damping, thereby suppressing the current abnormality first, and then waiting for the subsequent cycle to recover and find the optimization again;

[0160] The first confidence threshold, the first risk threshold, and the first impact threshold are all preset judgment conditions; their setting principle is not a fixed single constant, but is determined based on the historical stable processing data corresponding to the target machine tool, cutting tool, and workpiece material; specifically, the first confidence threshold is set near the lower edge of the confidence interval that the simulation can serve as an effective reference.

[0161] The first risk threshold is set near the risk range where there is no sustained instability but a warning trend has already appeared; the first impact threshold is set near the impact statistics range before the blade breaks; the above threshold configuration ensures that the weight switching is strictly performed based on the objective physical state of the system.

[0162] When generating the final processing instruction, the edge computing node weights and fuses the safe processing parameters, simulation recommended parameters, and feedback adjustment parameters according to the above three weights to obtain the final processing instruction. In order to resolve the conflict between the timing synchronization and physical lag of parameter generation at different time granularities, the weighted fusion and instruction issuance actions are strictly executed with the shortest first control cycle as the timing trigger benchmark.

[0163] The specific physical execution chain is as follows: In each first control cycle, the system reads the latest generated feedback adjustment parameters at the current moment, and uses a zero-order hold mechanism to latch the safety processing parameters generated in the most recent second control cycle and the simulation recommended parameters generated in the most recent third control cycle;

[0164] Based on the adjustment weights synchronously refreshed in the current first control cycle, real-time weighted fusion is performed and sent to the CNC execution unit; this ensures that millisecond-level abnormal feedback signals will not be blocked by physical hysteresis due to waiting for a long period of reinforcement learning optimization or simulation calculation, thus realizing the time-series decoupling and asynchronous coordination of the fast and slow control loops;

[0165] After receiving the instruction, the CNC execution unit decomposes it into the corresponding parts of the spindle speed setting value, feed rate setting value, axial depth of cut correction amount, and radial width of cut correction amount, and controls the machine tool to execute it. Since the final machining instruction comes from the comprehensive result of three sources, different control emphases will be reflected in the stable period, risk warning period, and abnormal period.

[0166] Steps one through five in this embodiment are not run at the same rate, but are respectively allocated to the first control cycle, the second control cycle, and the third control cycle; here, the first control cycle is shorter than the second control cycle, and the second control cycle is shorter than the third control cycle; the multi-cycle architecture is used to adapt to the processing time requirements of different computing tasks.

[0167] During the first control cycle, the sensor acquisition module and the feedback control unit operate; specifically, the sensor acquisition module continuously samples vibration, acoustic emission, cutting force and spindle current in a short period of time, and the edge computing node determines the instantaneous risk index based on this, while the feedback control unit synchronously generates feedback adjustment parameters.

[0168] Since shock and flutter precursors usually manifest as short-cycle abrupt changes in sensor signals, these tasks are assigned to be executed at the shortest cycle to reflect changes in the field conditions earlier.

[0169] During the second control cycle, the edge computing nodes perform constraint parameter updates and determine the corrected processing parameters; the safety clearance shrinkage coefficient, safety function convergence coefficient, and smoothing coefficient are updated at low frequency during the second control cycle based on the risk parameters and confidence parameters updated in the previous stage.

[0170] The construction of variable safety limits and the correction of candidate processing parameters are completed within the same cycle; this arrangement can balance the amount of solution and the execution time, and avoid the impact on real-time performance by performing heavy optimization calculations in the shortest cycle.

[0171] During the third control cycle, the reinforcement learning optimization unit outputs candidate processing parameters, and the processing simulation unit completes the calibration of the simulation recommended parameters. The reinforcement learning inference and simulation calibration operations that provide trend parameter adjustments are executed through the third control cycle, and high-frequency anomalies are suppressed by the feedback control unit under the first control cycle.

[0172] When the three control cycles run in coordination, the data flow relationship is as follows: The first control cycle first generates instantaneous risk indicators and feedback adjustment parameters and triggers the fusion and distribution action; the second control cycle asynchronously reads these risk indicators and, at the same time, combines the candidate processing parameters and simulation results output by the previous third control cycle that are latched to complete the safety correction.

[0173] The third control cycle updates the candidate processing parameters and simulation recommended parameters for the next round based on the state parameter group summarized in the second control cycle; the final processing command is generated by reconstructing the processing results under multiple control cycles according to real-time weights.

[0174] In a high-load milling scenario, if local hard spots appear in the material during the cutting process, the cutting force step rate and acoustic emission sudden characteristics will increase in the first control cycle. The feedback control unit will generate feedback adjustment parameters to suppress the impact.

[0175] During the second control cycle, the edge computing node detects that the impact risk index exceeds the first impact threshold, so it increases the rapid feedback control weight and readjusts the candidate processing parameters. If the simulation model is still highly reliable and the flutter risk decreases during the third control cycle, the system will gradually restore the simulation correction weight or the learning control weight.

[0176] Therefore, the machine tool prioritizes stable operation during abnormal periods, and continues parameter optimization after the abnormality is resolved. The entire process affects objective machining parameters such as spindle speed, feed rate, and depth of cut and width of cut, thereby improving cutting stability and reducing the probability of tool damage.

[0177] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make several improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for optimizing machining parameters in a metalworking machinery system based on reinforcement learning, characterized in that, Includes the following steps: Step 1: Acquire physical sensing data of the metal processing machinery system, extract instantaneous risk indicators within the time-series sampling window, and determine the flutter risk index and impact risk index accordingly; Step 2: Based on the chatter risk index, the impact risk index, and the machine tool process parameters, construct a machining state parameter group, and output candidate machining parameters through a reinforcement learning optimization unit. The machine tool process parameters include at least the spindle speed, feed rate, axial depth of cut, and radial width of cut. Step 3: Determine the reliability parameters of the simulation model based on the cutting state prediction error output by the machining simulation unit; and update the constraint parameters, including the safety clearance shrinkage coefficient, safety function convergence coefficient, and smoothing coefficient, in conjunction with the chatter risk index and the impact risk index. Step 4: Construct a variable safety limit based on the updated constraint parameters. Under the constraints of the upper and lower bounds of parameter execution and the maximum allowable change, modify the candidate processing parameters to obtain safe processing parameters. Step 5: Based on the simulation model credibility parameters, the flutter risk index, and the impact risk index, dynamically configure the adjustment weights of the safe processing parameters, the simulation recommended parameters output by the processing simulation unit, and the feedback adjustment parameters generated by the feedback control unit, and perform weighted fusion to generate the final processing command for execution.

2. The method for optimizing machining parameters of a metalworking machinery system based on reinforcement learning according to claim 1, characterized in that, The physical sensing data includes vibration sensing data, acoustic emission sensing data, cutting force sensing data, and spindle current sensing data. The instantaneous risk indicators include vibration energy, acoustic emission energy, cutting force step rate, and spindle load deviation rate of spindle current relative to the historical steady-state machining current mean obtained within the time-series sampling window.

3. The method for optimizing machining parameters of a metalworking machinery system based on reinforcement learning according to claim 2, characterized in that, The chatter risk index is determined by the vibration energy, the acoustic emission energy, the cutting force step rate, and the spindle load deviation rate. When a certain instantaneous risk indicator meets the preset flutter event judgment conditions, the index coefficient corresponding to the instantaneous risk indicator is amplified by a preset multiple.

4. The method for optimizing machining parameters of a metalworking machinery system based on reinforcement learning according to claim 2, characterized in that, The impact risk index is determined by the cutting force step rate, spindle load deviation rate, and vibration peak value and acoustic emission burst characteristics extracted based on the physical sensing data.

5. The method for optimizing machining parameters of a metalworking machinery system based on reinforcement learning according to claim 1, characterized in that, The processing status parameter group includes statistical values ​​of chatter risk index, impact risk index, machine tool spindle speed, machine tool feed rate, machine tool axial depth of cut, machine tool radial width of cut, cutting force, spindle current, and surface roughness and tool wear estimates derived from physical sensor data and machine tool process parameters within a preset statistical period.

6. The method for optimizing machining parameters of a metalworking machinery system based on reinforcement learning according to claim 5, characterized in that, The updating of the constraint parameters includes: The safety clearance contraction coefficient is updated based on the simulation model credibility parameters, the flutter risk index, and the impact risk index. Based on the simulation model's credibility parameters and the flutter risk index, and combined with the preset basic safety function convergence coefficient, the safety function convergence coefficient is updated. Based on the impact risk index and the difference between the candidate processing parameters and the parameters executed in the previous cycle, the smoothing coefficient is updated in conjunction with the preset basic smoothing coefficient.

7. The method for optimizing machining parameters of a metalworking machinery system based on reinforcement learning according to claim 6, characterized in that, The construction of the variable safety limit includes: Based on the allowable upper limit values ​​corresponding to the chatter risk index, the impact risk index, the estimated tool wear, and the spindle current monitoring value, a safety state function is constructed. Based on the safety state function, the convergence coefficient of the safety function, and the sensitivity coefficient of the safety state function with respect to each processing parameter, safety constraints are constructed to form the variable safety limit. The upper and lower bounds of the variable safety limit are then updated using the safety limit contraction coefficient.

8. The method for optimizing machining parameters of a metalworking machinery system based on reinforcement learning according to claim 7, characterized in that, The determination of the safe processing parameters includes: Under the conditions of satisfying the safety constraints, the upper and lower bounds of the parameter execution, and the maximum allowable change within the preset control cycle, the safe processing parameters are obtained by optimization calculation based on the candidate processing parameters, the smoothing coefficient, and the parameters executed in the previous cycle.

9. The method for optimizing machining parameters of a metalworking machinery system based on reinforcement learning according to claim 1, characterized in that, The adjustment weights include learning control weights, simulation correction weights, and fast feedback control weights; The determination of the adjustment weights includes: When the confidence parameter of the simulation model is greater than the first confidence threshold, the flutter risk index is less than the first risk threshold, and the impact risk index is less than or equal to the first impact threshold, the learning control weight is set to a value within the first preset weight range. When the simulation model credibility parameter is greater than the first credibility threshold, the flutter risk index is greater than or equal to the first risk threshold, and the impact risk index is less than or equal to the first impact threshold, the simulation correction weight is set to a value within the second preset weight range. When the confidence parameter of the simulation model is less than or equal to the first confidence threshold or the impact risk index is greater than the first impact threshold, the fast feedback control weight is set to a value within the third preset weight range.

10. The method for optimizing machining parameters of a metalworking machinery system based on reinforcement learning according to claim 1, characterized in that, Steps one through five are executed in the first control cycle, the second control cycle, and the third control cycle, respectively. The determination of instantaneous risk indicators and the generation of feedback adjustment parameters by the feedback control unit are executed in the first control cycle. The updating of constraint parameters and the determination of the corrected machining parameters are carried out in the second control cycle; The output of candidate processing parameters and the calibration of simulation-recommended parameters are performed in the third control cycle; The duration of the first control cycle is less than the duration of the second control cycle, and the duration of the second control cycle is less than the duration of the third control cycle.

Images