Metal cutting process parameter optimization analysis method based on machine learning

Abstract

The invention discloses a metal cutting process parameter optimization analysis method based on machine learning, and particularly relates to the field of machine learning. Comprising multi-dimensional process parameter feature extraction and preprocessing, cutting state intelligent identification based on integrated learning, dynamic process parameter sensitivity analysis and weight calculation, process parameter intelligent optimization under a multi-target constraint condition, and adaptive parameter adjustment and real-time control strategy. According to the method, the interaction relationship between complex nonlinear features and process parameters in the cutting process is comprehensively captured, and accurate and intelligent recognition of different cutting states such as normal cutting, tool abrasion and abnormal flutter is achieved through a three-layer integrated learning architecture; the technical bottlenecks that an existing system lacks real-time self-adaptive adjustment capacity and is low in process optimization efficiency are overcome, pertinence and effectiveness of parameter adjustment are ensured, and the technical current situation that machining quality fluctuates and repeatability is poor due to traditional fixed parameters is changed.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and more specifically, to a method for optimizing and analyzing metal cutting process parameters based on machine learning. Background Technology

[0002] Traditional metal cutting process parameter optimization primarily relies on empirical formulas and single-objective optimization methods. In practical applications, process engineers first determine initial process parameters, including basic parameters such as cutting speed, feed rate, and depth of cut, by consulting cutting manuals or empirical formulas. Then, during actual machining, the cutting effect is evaluated by manually observing cutting phenomena, measuring surface roughness, and checking tool wear. When the machining quality is found to be unsatisfactory, process personnel manually adjust the process parameters based on experience, typically using single-factor experimental methods or orthogonal experimental methods for parameter optimization.

[0003] Existing automated cutting systems typically employ simple threshold-based control strategies, monitoring the cutting process by setting alarm thresholds for cutting force, vibration, or temperature. When a monitored parameter exceeds a preset threshold, the system issues an alarm signal or executes a pre-set protective action, such as reducing the cutting speed or stopping machining. Regarding parameter optimization, traditional methods primarily use single-objective optimization algorithms, such as genetic algorithms or particle swarm optimization, to optimize a single index like surface roughness or cutting force. The entire optimization process is usually conducted offline, requiring numerous trial cuts to verify the optimization results before the optimized parameters are applied to mass production.

[0004] Existing technologies suffer from several significant drawbacks, severely hindering the intelligent development of metal cutting processes. First, traditional experience-driven methods lack the ability to deeply analyze the complex nonlinear characteristics of the cutting process, failing to accurately capture the interactions and coupling relationships between process parameters. This results in parameter optimization often being limited to local optima, making it difficult to achieve globally optimal process configurations. Second, existing monitoring systems primarily rely on single sensor signals and fixed thresholds, failing to achieve intelligent identification and classification of cutting conditions and lacking early warning capabilities for gradual changes such as tool wear. More importantly, traditional optimization methods cannot handle process parameter optimization problems under multi-objective constraints. While pursuing surface quality, they often neglect other important indicators such as machining efficiency and tool life, resulting in poor overall machining efficiency. Existing systems lack real-time adaptive adjustment capabilities and cannot adjust process parameters in a timely manner according to the dynamic changes in the cutting process. When cutting conditions change, using fixed process parameters can easily lead to fluctuations in machining quality or equipment damage. In addition, traditional methods are insufficient in parameter sensitivity analysis, failing to quantify the degree of influence of different process parameters on cutting quality and lacking a scientific basis for parameter adjustment, making the process optimization process inefficient and with poor repeatability. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, this invention provides a machine learning-based method for optimizing and analyzing metal cutting process parameters, which addresses the problems mentioned in the background section through the following approach.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for optimizing and analyzing metal cutting process parameters based on machine learning, comprising: S1: Multi-dimensional process parameter feature extraction and preprocessing: Multi-dimensional physical signals during the metal cutting process are collected through a sensor array to establish a feature library relating process parameters and cutting state, including signal acquisition, noise filtering, feature dimension expansion, and data standardization. S2: Intelligent identification of cutting state based on ensemble learning: Using the multi-dimensional process parameter features extracted in S1 as input, an ensemble learning model integrating multiple algorithms is constructed to intelligently identify and classify metal cutting state, including base learner training, model fusion strategy design, state identification decision, and confidence assessment. S3: Dynamic process parameter sensitivity analysis and weight calculation: Based on the intelligent identification results of cutting state in S2, a parameter sensitivity weight matrix is ​​established to dynamically analyze and calculate the sensitivity of process parameters under different cutting states, including parameter sensitivity quantification, dynamic weight update, and interaction effect analysis. S4: Intelligent optimization of process parameters under multi-objective constraints: Based on the dynamic process parameter sensitivity analysis and weight calculation results of S3, a multi-objective constraint optimization model is constructed to intelligently optimize metal cutting process parameters, including the establishment of multi-objective functions, setting of constraint conditions, design of optimization algorithms, and selection of solution sets; S5: Adaptive parameter adjustment and real-time control strategy: Based on the intelligent optimization results of process parameters under the multi-objective constraints of S4, an adaptive parameter adjustment and real-time control strategy is designed, including real-time monitoring feedback, parameter deviation detection, adaptive adjustment algorithm, and control command generation.

[0007] The technical effects and advantages of this invention are as follows: This approach effectively addresses the limitations of existing technologies, such as the lack of in-depth analysis capabilities and the limitations of single-sensor monitoring, through multi-dimensional process parameter feature extraction and preprocessing in S1 and intelligent cutting state recognition based on ensemble learning in S2. By constructing a joint feature matrix in the time and frequency domains and a combination of nonlinear features, it can comprehensively capture the complex nonlinear features and interactions between process parameters during the cutting process. The three-layer ensemble learning architecture, combined with a multi-algorithm fusion strategy of random forest, support vector machine, and gradient boosting decision tree, achieves accurate and intelligent recognition of different cutting states, such as normal cutting, tool wear, and chatter abnormalities. The weighted average fusion strategy and the confidence evaluation mechanism based on entropy value overcome the limitations of traditional fixed threshold judgment and provide early warning capabilities for gradual changes, enabling the system to accurately identify subtle changes in the cutting state. This invention solves the technical problems of traditional optimization methods being unable to handle multi-objective constraints and lacking scientific basis for parameter adjustment by using dynamic process parameter sensitivity analysis and weight calculation in S3 and intelligent optimization of process parameters under multi-objective constraints in S4. The improved Morris sensitivity analysis method, combined with a dynamic weight update mechanism, achieves accurate quantitative analysis of the degree of influence of different process parameters, providing a scientific quantitative basis for parameter adjustment. The construction of the parameter interaction effect analysis matrix reveals the coupling relationship and mutual influence mechanism between process parameters. The multi-objective optimization function considers multiple key indicators such as surface roughness, cutting force, and tool life in a unified manner. Through the improved non-dominated sorting genetic algorithm NSGA-III and adaptive parameter adjustment mechanism, the search for the global optimal solution is achieved, effectively avoiding the problems of local optima and poor overall benefits caused by traditional single-objective optimization. This invention addresses the technical bottlenecks of existing systems, such as lack of real-time adaptive adjustment capabilities and low process optimization efficiency, through S5's adaptive parameter adjustment and real-time control strategy. The real-time monitoring and feedback system established by the sliding window data processing mechanism and statistical process control method can continuously monitor changes in the cutting process state and detect parameter deviations in a timely manner. The fuzzy PID adaptive adjustment algorithm, combining the technical advantages of fuzzy logic and PID control, achieves intelligent response to dynamic changes in the cutting process. The control command generation strategy based on cutting state identification results and parameter sensitivity weights ensures the targetedness and effectiveness of parameter adjustment. The control command priority queue mechanism determines the adjustment order according to the severity of the anomaly and parameter sensitivity, realizing closed-loop adaptive operation of the entire metal cutting process parameter optimization analysis and control method. This changes the technical status quo of machining quality fluctuations and poor repeatability caused by traditional fixed parameters. Attached Figure Description

[0008] Figure 1 This is a schematic diagram of the overall structure of the present invention.

[0009] Figure 2This is a schematic diagram of the S1 structure of the present invention.

[0010] Figure 3 This is a schematic diagram of the S2 structure of the present invention.

[0011] Figure 4 This is a schematic diagram of the S3 structure of the present invention.

[0012] Figure 5 This is a schematic diagram of the S4 structure of the present invention.

[0013] Figure 6 This is a schematic diagram of the S5 structure of the present invention. Detailed Implementation

[0014] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0015] refer to Figures 1-6 The method for optimizing and analyzing metal cutting process parameters based on machine learning, as shown, includes: S1: Multi-dimensional process parameter feature extraction and preprocessing: Multi-dimensional physical signals in the metal cutting process are collected through a sensor array to establish a feature library relating process parameters and cutting state, including signal acquisition, noise filtering, feature dimension expansion, and data standardization.

[0016] The signal acquisition specifically involves installing a triaxial accelerometer vibration sensor on the spindle housing of the target metal cutting equipment. The sensor is rigidly connected to the equipment using bolts. The sensor model is a piezoelectric accelerometer with a measurement range of ±50g and a frequency response range of 1Hz to 5000Hz. An acoustic emission sensor is installed at the tool clamping location, fixed with a magnetic base. This sensor operates at a frequency range of 100kHz to 1MHz and is used to capture high-frequency acoustic emission signals during the cutting process. A Hall current sensor is installed on the power supply line of the spindle motor, using a through-hole mounting method. Its measurement range is 0 to 100A, used to monitor changes in the spindle motor current. An infrared temperature sensor is installed near the tool, approximately 50mm from the cutting point, with a measurement range of 0 to 500℃, used to monitor temperature changes in the cutting area.

[0017] All sensors are connected to the data acquisition card via shielded cables. The data acquisition card uses 16-bit analog-to-digital conversion accuracy and is configured with four synchronous acquisition channels, each corresponding to a different sensor signal. The data acquisition system is set to a uniform sampling frequency of 1000Hz to ensure the time synchronization of all signals. The acquired digital signals are transmitted to an industrial computer via a USB interface for storage and processing. The industrial computer is configured with a quad-core processor and 8GB of memory, running a Windows operating system and LabVIEW data acquisition software.

[0018] A real-time display interface is established during data acquisition, allowing simultaneous observation of waveform changes from four sensor signals. Signal amplitude alarm thresholds can be set, and an alarm is automatically triggered when any sensor signal exceeds the preset range. The acquired raw data is stored in binary format on the local hard drive, and the files are named using timestamps for easy data retrieval and analysis.

[0019] The noise filtering process employs wavelet transform to denoise the acquired raw signals. First, the Daubechies wavelet basis function is selected as the decomposition basis function. This wavelet basis function has good time-frequency localization characteristics and can effectively preserve the important feature information of the signal. Each sensor signal is decomposed into three levels of wavelet decomposition, which decomposes the raw signal into low-frequency approximation coefficients and three levels of high-frequency detail coefficients. The low-frequency approximation coefficients contain the main energy and useful information of the signal, while the high-frequency detail coefficients mainly contain noise components.

[0020] In the wavelet decomposition process, the first level of decomposition divides the original signal into a low-frequency part from 0 to 250 Hz and a high-frequency part from 250 Hz to 500 Hz. The second level of decomposition further divides the low-frequency part into two frequency bands: 0 to 125 Hz and 125 Hz to 250 Hz. The third level of decomposition divides the lowest frequency band into two frequency bands: 0 to 62.5 Hz and 62.5 Hz to 125 Hz. By analyzing the energy distribution characteristics of each frequency band, it is determined that the noise is mainly concentrated in the high-frequency detail coefficients.

[0021] High-frequency detail coefficients are thresholded, and a soft thresholding method is used to remove noise components. The soft thresholding method can maintain the continuity of the signal while removing noise, avoiding the oscillation phenomenon that may occur in the hard thresholding method. The threshold selection adopts a heuristic threshold estimation method, which automatically determines the optimal threshold size based on the signal length and noise variance.

[0022] After thresholding, the low-frequency approximation coefficients and the processed high-frequency detail coefficients are retained, and the denoised signal is recovered by wavelet reconstruction algorithm. The reconstruction process is the reverse of the decomposition process, synthesizing the signals of each frequency band layer by layer, and finally obtaining the complete denoised signal. The denoised signal retains the original sampling frequency and data length to ensure compatibility with subsequent processing steps.

[0023] To verify the denoising effect, the improvement in signal-to-noise ratio before and after denoising was calculated, and the effectiveness of noise filtering was confirmed through comparative analysis. The denoised signal data were stored in the same file format to provide a high-quality data foundation for subsequent feature extraction and analysis.

[0024] The feature dimension expansion adds time-frequency domain features and statistical features to the original time-domain and frequency-domain features to construct a more comprehensive feature set. The time-frequency domain feature extraction adopts the short-time Fourier transform method, which divides the denoised signal into several time windows. The length of each window is set to 256 sampling points, and the overlap rate between windows is set to 50%. Fourier transform is performed on the signal in each time window to obtain the spectrum information in that time period. The spectrum results of all time windows are connected to form a time-spectrum graph.

[0025] Instantaneous frequency features are extracted from the time-spectrum graph to calculate the dominant frequency component at each moment, reflecting the change of signal frequency over time; spectral centroid features are extracted to calculate the centroid position of spectral energy within each time window, which reflects the concentration of signal frequency distribution; spectral bandwidth features are extracted to calculate the width of the effective bandwidth within each time window, reflecting the dispersion of signal frequency distribution.

[0026] Statistical feature extraction includes higher-order moment features and distribution shape features. The skewness feature of the signal is calculated, reflecting the degree of asymmetry of the signal amplitude distribution relative to the mean; positive skewness indicates a right-skewed distribution, and negative skewness indicates a left-skewed distribution. The kurtosis feature of the signal is calculated, reflecting the sharpness of the signal amplitude distribution; a larger kurtosis value indicates a sharper distribution, and a smaller kurtosis value indicates a flatter distribution.

[0027] Multi-scale feature extraction is enhanced by calculating the same feature parameters using different time window lengths. Three window lengths are set: 128 points, 256 points, and 512 points, corresponding to different temporal and frequency resolutions. Short windows provide higher temporal resolution, suitable for capturing transient feature changes, while long windows provide higher frequency resolution, suitable for analyzing steady-state frequency components.

[0028] Cross-feature construction involves calculating the correlation characteristics between different sensor signals, including the cross-correlation coefficients between vibration signals and current signals, and between acoustic emission signals and temperature signals. These cross-features can reflect the correlation between different physical quantities, providing richer information for tool condition identification.

[0029] Feature standardization employs a zero-mean unit variance standardization method, converting all feature values ​​into standardized values ​​with a mean of zero and a standard deviation of one. Standardization eliminates differences in dimensions and numerical ranges between different features, ensuring that each feature has the same weight in subsequent processing. The length of the expanded feature vector is increased to three times the original, providing more comprehensive feature information for machine learning algorithms.

[0030] S2: Intelligent identification of cutting state based on ensemble learning: Using the multi-dimensional process parameter features extracted in S1 as input, an ensemble learning model integrating multiple algorithms is constructed to intelligently identify and classify metal cutting states, including base learner training, model fusion strategy design, state identification decision, and confidence assessment.

[0031] The S2 design employs a three-layer ensemble learning architecture. The first layer includes three base learners: a random forest, a support vector machine, and a gradient boosting decision tree. The random forest uses the Bootstrap aggregation method to construct multiple decision trees. During training, each decision tree randomly draws the same number of samples with replacement from the original training set to form a subset of training sets. Simultaneously, at each node split, a subset of features is randomly selected from all features for optimal splitting. The number of decision trees is set to 100 to improve the model's stability and generalization ability. The maximum depth of each tree is limited to 15 layers to prevent overfitting. Gini impurity is used as a standard to measure the purity of the dataset during node splitting; the smaller the Gini impurity value, the purer the dataset. Splitting stops when the Gini impurity reaches its minimum value or the maximum depth.

[0032] Support Vector Machines (SVMs) achieve classification by finding the optimal separating hyperplane. They use radial basis functions as kernel functions to map low-dimensional, nonlinearly separable data to a high-dimensional space, making it linearly separable. The kernel parameter γ is set to 0.1, which controls the influence range of a single training sample. A smaller γ value results in a larger influence range and a relatively smooth decision boundary. The penalty parameter C is set to 100, which controls the degree of penalty for misclassified samples. A larger C value indicates a lower tolerance for misclassification and aims for higher training accuracy.

[0033] Gradient boosting decision trees use a forward stepwise algorithm to gradually build multiple weak learners. Each new decision tree is trained to correct the prediction errors of all previous trees. The learning rate is set to 0.1 to control the contribution of each tree to the final prediction result. A smaller learning rate requires more trees to achieve the same training effect but usually has better generalization performance. The subsampling rate is set to 0.8, which means that 80% of the training samples are randomly selected each time a new tree is trained. This random subsampling method can increase the randomness of the model and reduce the risk of overfitting, while improving the training speed.

[0034] The second layer employs a weighted average fusion strategy, dynamically allocating fusion weights by evaluating the performance of each base learner on the validation set. First, the training dataset is divided into a training set and a validation set in a 7:3 ratio. The three base learners of the first layer are trained using the training set, and then the classification accuracy of each base learner is tested on the validation set. The fusion weight of each base learner is calculated based on the validation set accuracy, with base learners that have higher accuracy receiving greater weights, ensuring that high-performing models play a more important role in the final decision.

[0035] The weight calculation employs a power function for nonlinear transformation, raising the validation accuracy of each base learner to the power of α before normalization. Specifically, the calculation process involves first calculating the α-th power of the accuracy of each base learner, where the adjustment parameter α is set to 2. This parameter controls the concentration of weight allocation; a larger α value amplifies accuracy differences, giving more prominent weights to superior models. Then, the sum of the α-th powers of the accuracy of all base learners is used as a normalization factor. Finally, the final fusion weight is obtained by dividing the α-th powers of the accuracy of each base learner by the normalization factor.

[0036] Let w_i be the weight of the i-th base learner, acc_i be its validation accuracy, and α be the adjustment parameter, then w_i = (acc_i)^α / Σ(acc_j)^α.

[0037] The dynamic weight allocation mechanism adjusts the contribution of each base learner in real time according to the changes in the performance of the validation set. When a base learner performs well on new validation data, its weight will increase accordingly, and when its performance declines, its weight will decrease. This dynamic adjustment mechanism enables the ensemble model to adaptively adapt to changes in data distribution, make full use of the advantages of each base learner, and fuse the prediction results of the three base learners through weighted averaging to obtain a more stable and accurate ensemble prediction result.

[0038] The third layer uses a meta-learner for final decision fusion. Logistic regression is selected as the meta-learner to further optimize the weighted fusion result of the second layer output. The training data of the meta-learner consists of the prediction probabilities of each base learner in the first layer on the validation set and the weighted fusion result of the second layer, forming a new feature vector that is input into the logistic regression model. By learning the nonlinear relationship between the prediction results of each layer, logistic regression automatically discovers the optimal fusion strategy and overcomes the limitations that may exist in simple weighted averaging.

[0039] The confidence level calculation uses the entropy method to quantitatively evaluate the reliability of the final prediction result. The specific calculation process is as follows: first, obtain the prediction probability distribution of each category output by the meta-learner, and then calculate the information entropy value of the probability distribution as the confidence index. The entropy value calculation formula is the sum of the negatives of the logarithmic probability products of each category probability. When the prediction probability distribution is more concentrated in a certain category, the smaller the entropy value, the higher the confidence level. When the probability distribution is more uniform, the larger the entropy value, the lower the confidence level.

[0040] Let p_i be the predicted probability of the i-th state, and n be the total number of state categories. Then confidence = 1 - (-Σp_i × log(p_i)) / (log(n)).

[0041] The meta-learner training process employs cross-validation to ensure the model's generalization ability. The original training data undergoes 5-fold cross-validation, with each fold using data from the other 4 folds to train the first-layer base learner. Predictions generated on the current fold are then used as training samples for the meta-learner. The logistic regression meta-learner uses L2 regularization to prevent overfitting, with the regularization parameter set to 0.01. Model complexity is constrained by adding a penalty term for the sum of squared weights to the loss function. The meta-learner learns the weight combination pattern of the predictions from each base learner in the first layer, combined with entropy confidence evaluation, and finally outputs an ensemble prediction result optimized by three layers, along with its reliability evaluation.

[0042] S3: Dynamic Process Parameter Sensitivity Analysis and Weight Calculation: Based on the intelligent identification results of cutting state in S2, a parameter sensitivity weight matrix is ​​established to dynamically analyze and calculate the sensitivity of process parameters under different cutting states, including parameter sensitivity quantification, dynamic weight update, and interaction effect analysis.

[0043] The parameter sensitivity quantification employs an improved Morris sensitivity analysis method to quantify three main process parameters: cutting speed, feed rate, and depth of cut. First, Morris sampling trajectories are constructed within the value range of each parameter. The cutting speed value range is divided into 10 equally spaced levels, and the feed rate and depth of cut are also divided into 10 levels, forming a three-dimensional parameter space grid structure. Fifty sampling trajectories are designed in the parameter space, with each trajectory containing 4 sampling points. Only one parameter changes between adjacent sampling points, and the change step size is one-tenth of the parameter range.

[0044] Cutting experiments or simulations are performed on the sampling points in each sampling trajectory to obtain the corresponding cutting quality index response values. The basic effect of each parameter is calculated, which is the ratio of the change in output response caused by the change of a single parameter to the change in parameter. The basic effect is calculated by dividing the response difference between adjacent sampling points by the change in the corresponding parameter. Each parameter produces a basic effect value in a trajectory.

[0045] The arithmetic mean of the absolute values ​​of the basic effects of each parameter in the 50 trajectories is obtained to obtain the average sensitivity index of the parameter. At the same time, the standard deviation of the basic effects of each parameter is calculated to measure the stability and nonlinearity of the parameter sensitivity. The average sensitivity index reflects the overall influence of the parameter on the output, and the standard deviation reflects the strength of the interaction between the parameter and other parameters.

[0046] Let μ_i be the average sensitivity of the i-th parameter, σ_i be the standard deviation of the sensitivity, EE_i be the basic effect, and r be the number of sampling trajectories set to 50. Then μ_i = (1 / r)Σ|EE_i|, σ_i = √[(1 / r)Σ(EE_i-μ_i)^2].

[0047] By comparing the average sensitivity index of each parameter, the primary and secondary relationships of the parameter influence are determined. The sensitivity index is then normalized so that the sum of the sensitivity indices of all parameters equals 1, forming a parameter sensitivity weight vector. This weight vector directly reflects the degree of influence of each process parameter on cutting quality, providing a quantitative basis for subsequent parameter optimization.

[0048] The dynamic weight update mechanism adjusts the parameter weights in real time based on the current cutting state and historical optimization results. First, a parameter optimization success rate statistics system is established to record the number of times each process parameter has been adjusted and the number of times the cutting quality has been successfully improved during the historical optimization process. The success rate is calculated by dividing the number of successful improvements by the total number of adjustments, thus forming the historical optimization effect evaluation index for each parameter.

[0049] The system maintains a sliding time window with a window length set to the last 100 optimization operations. It updates the optimization success rate of each parameter in real time. When the adjustment of a parameter can improve surface roughness, reduce cutting force, or extend tool life, the adjustment is recorded as a successful operation. The current optimization success rate of each parameter is normalized to obtain a relative success rate value, which is used for weight update calculation.

[0050] The weight update process uses a weighted average method to integrate historical weights with current sensitivity indicators and optimization success rates. The historical weight retention factor is set to 0.7 to ensure the smoothness and stability of weight changes. The sensitivity indicator balance factor is set to 0.6, and the optimization success rate balance factor is set to 0.4. The sum of the two is 1 to ensure the rationality of the new weight components.

[0051] Let w_i(t+1) be the updated weight, β be the historical weight retention factor of 0.7, S_i be the optimization success rate of parameter i, the maximum value of the optimization success rate of all parameters is S_max, and α_1 and α_2 be the balance factors of 0.6 and 0.4 respectively. Then the weight update formula is: w_i(t+1)=βw_i(t)+(1-β)[α_1μ_i+α_2(S_i / S_max)].

[0052] After each round of parameter optimization, the system automatically triggers a weight update calculation. The parameter weights from the previous time step are multiplied by the historical retention factor, plus the product of the current sensitivity index and its balance factor, and finally the product of the normalized optimization success rate and its balance factor, to obtain the updated parameter weights. The updated weight vector is then normalized to ensure that the sum of all parameter weights equals 1, forming a new dynamic weight allocation scheme to guide the next round of process parameter optimization and adjustment.

[0053] The interaction effect analysis uses a second-order interaction sensitivity index to quantify the strength of the interaction between parameters. First, during the Morris sampling process, the order of parameter changes and the corresponding response changes in each trajectory are recorded. When different parameters are changed in two consecutive steps in the trajectory, the joint effect of the two parameters is calculated. The joint effect is measured by the difference between the total response change caused by the two-step parameter change and the theoretical response change when each parameter is changed individually. The absolute value of the difference reflects the strength of the interaction between parameters.

[0054] Let S_ij be the interaction sensitivity of parameters i and j, Y be the cutting quality index, and X_i and X_j be the corresponding process parameters. Then S_ij = Var(E[Y|X_i, X_j]) / Var(Y), where Var represents the variance and E represents the expected value.

[0055] The system establishes a parameter interaction effect matrix, with rows and columns corresponding to the three parameters of cutting speed, feed rate, and depth of cut, respectively. Each element in the matrix represents the average interaction intensity between two corresponding parameters, which is obtained by averaging the absolute values ​​of the interaction effects of the parameter pair across all sampled trajectories. The diagonal elements of the matrix are set to zero because the parameter does not have an interaction effect with itself, while the off-diagonal elements reflect the degree of mutual influence between different parameter pairs.

[0056] In the calculation of the interaction effect, when parameter A is adjusted first and then parameter B is adjusted for a certain trajectory, the changes in the cutting quality index before and after the adjustment are recorded. The total amount of actual observed change is compared with the theoretical superposition value of parameter A changing alone and parameter B changing alone. The deviation between the two is the interaction effect amount. The average value of the interaction effect of the same parameter pair in all trajectories is taken to eliminate the influence of random error and obtain a stable interaction strength index.

[0057] Based on the numerical values ​​of the interaction effect matrix, parameter pairs with significant interactions are identified. Parameter pairs with interaction strength exceeding a set threshold are marked as strong interaction parameter combinations. These parameters need to be adjusted synchronously in subsequent optimization processes to avoid optimizing a single parameter while ignoring its synergistic effect with other parameters. The results of the interaction effect analysis are used to guide the formulation of parameter optimization strategies, ensuring that the mutual influence between parameters is fully considered when adjusting process parameters, thereby improving the stability and reliability of the optimization effect.

[0058] S4: Intelligent optimization of process parameters under multi-objective constraints: Based on the results of dynamic process parameter sensitivity analysis and weight calculation in S3, a multi-objective constraint optimization model is constructed to intelligently optimize metal cutting process parameters, including the establishment of multi-objective functions, setting of constraint conditions, design of optimization algorithms, and selection of solution sets.

[0059] The multi-objective function establishment first takes surface roughness Ra as the first optimization objective. By collecting roughness measurement data of the workpiece surface after processing, a functional relationship model between roughness and cutting speed v, feed rate f, and cutting depth ap is established. A roughness prediction model is fitted using regression analysis. This model can predict the corresponding surface roughness value based on the combination of input process parameters, with the goal of minimizing the predicted roughness value.

[0060] Next, the cutting force F is established as the second optimization objective. The three-dimensional cutting force data during the cutting process are collected in real time by force sensors installed on the machine tool spindle and workpiece fixture, and the magnitude of the combined cutting force is calculated. Similarly, the data fitting method is used to establish a prediction model between the cutting force and the three main process parameters. The optimization objective is to minimize the predicted cutting force value, thereby reducing the machine tool load and improving machining stability.

[0061] Then, the tool life T is set as the third optimization objective. By monitoring the relationship between tool wear and cutting time, a tool life prediction model is established. This model can predict the tool life based on the current process parameters. Unlike the first two objectives, the optimization objective of tool life is to find the maximum value, that is, to hope that the tool can obtain the longest service life under the given process parameters.

[0062] The three optimization objectives are expressed as functions: minf_1(x)=Ra(v, f, a_p) minf_2(x)=F(v, f, a_p) maxf_3(x)=T(v, f, a_p) Where x = [v, f, a_p] is the process parameter vector, minf_1(x) represents the combination of process parameters that minimizes the surface roughness Ra, minf_2(x) represents the combination of process parameters that minimizes the cutting force F, and maxf_3(x) represents the combination of process parameters that maximizes the tool life T.

[0063] After establishing three separate objective functions, the three objective functions are combined with the results of the sensitivity analysis and weight calculation of the dynamic process parameters obtained in S3. The first weight w_1 is multiplied by the surface roughness objective function, the second weight w_2 is multiplied by the cutting force objective function, and the third weight w_3 is multiplied by the negative value of the tool life objective function. The three are added together to form a comprehensive weighted objective function. Since the tool life objective function seeks to maximize the value while the other two objectives seek to minimize the value, the tool life objective function needs to be negatively represented during the weighted combination so that the entire comprehensive objective function becomes a unified form that seeks to minimize the value.

[0064] The weighted objective function is expressed as: F(x) = w_1f_1(x) + w_2f_2(x) + w_3(-f_3(x)).

[0065] This method of establishing a weighted objective function can unify multiple conflicting optimization objectives into a single mathematical framework for solving. At the same time, by dynamically adjusting the weight values ​​of each objective, the importance of different objectives can be balanced, thus effectively transforming a multi-objective optimization problem into a single-objective optimization problem.

[0066] The constraint conditions are first established by setting boundary constraints for process parameters. The minimum cutting speed v is directly obtained from the recommended minimum cutting speed value in the technical manual of the machining material. The maximum value is calculated by dividing the maximum spindle speed in the machine tool technical parameter table by the workpiece diameter and then multiplying by pi. At the same time, it is compared with the maximum allowable cutting speed of the tool material provided by the tool supplier and the smaller value is taken. The minimum feed rate f is obtained from the minimum feed rate parameter in the machine tool feed system technical specifications. The maximum value is directly adopted from the maximum feed rate value marked in the machine tool technical parameter table and compared with the maximum allowable feed rate corresponding to the workpiece surface roughness requirement and the smaller value is taken. The minimum cutting depth ap is set to twice the minimum cutting thickness of the tool to ensure cutting stability. The maximum value is directly adopted from the maximum cutting depth value marked in the tool technical parameters and does not exceed one-third of the workpiece clamping height.

[0067] The machining accuracy constraint is set by directly referring to the tolerance grade requirements marked in the workpiece drawing to determine the target accuracy value. Then, the repeatability and machining accuracy parameters of the machine tool are obtained from the machine tool accuracy grade certificate as the benchmark value of the accuracy constraint. The upper limit of the machining accuracy constraint is directly adopted from the maximum permissible deviation specified in the workpiece technical requirements, while the lower limit is determined according to the highest machining accuracy that the machine tool can achieve. This constraint is determined by comparing the accuracy value calculated by substituting the optimized process parameters into the machining accuracy prediction formula with the target accuracy range to determine whether the requirements are met.

[0068] The equipment capacity constraints are mainly determined by directly obtaining the upper limit of the power constraint from the rated power value marked on the machine tool nameplate. This value is then multiplied by a safety factor of 0.9 to obtain the actual constraint value to ensure safe operation of the equipment. The maximum value of the spindle torque constraint is obtained directly from the maximum output torque value in the machine tool technical parameter table. The maximum value of the feed force constraint is directly read from the maximum thrust parameter in the machine tool feed system technical specifications. The machine tool rigidity constraint is determined by the maximum cutting force bearing capacity value in the machine tool basic parameters. When the calculated cutting force exceeds this value, it is determined that the constraint conditions are not met.

[0069] The constraints include process parameter boundary constraints, machining accuracy constraints, and equipment capability constraints: v_min≤v≤v_max f_min≤f≤f_max a_p, min≤a_p≤a_p, max IT≤IT_target P≤P_max The optimization algorithm design first determines the basic parameter settings of the genetic algorithm. The population size is set to 100 individuals, and each individual represents a complete combination of process parameters, including three gene loci: cutting speed, feed rate, and cutting depth. The number of generations is set to 500 to ensure that the algorithm has enough iterations to find the global optimum. The crossover probability is set to 0.8, and the mutation probability is set to 0.02. These parameter values ​​are determined through pre-algorithm testing to maintain good population diversity and convergence performance.

[0070] Individual coding adopts real number coding method. The three gene positions of each individual correspond to the actual values ​​of cutting speed v, feed rate f and cutting depth ap respectively. The initial population is generated randomly, and the initial individuals are randomly generated within the constraints of each process parameter to ensure that the initial population has good diversity and wide search space coverage. After each individual is generated, the constraint conditions are checked immediately. Individuals that do not meet the constraints are regenerated until all constraints are met.

[0071] The fitness function directly adopts the weighted multi-objective function established in S4. Since this function is a minimum value problem, the fitness value is calculated using the reciprocal of the function value, so that individuals with smaller function values ​​have higher fitness. In the fitness calculation process, the process parameter values ​​of the individual are first substituted into the three single-objective functions to calculate the predicted values ​​of surface roughness, cutting force and tool life respectively. Then, the values ​​are weighted and summed according to the preset weights to obtain the comprehensive objective function value. Finally, the reciprocal is taken as the fitness value of the individual.

[0072] The selection operation uses a roulette wheel selection method, calculating the probability of each individual being selected based on its fitness value. Individuals with higher fitness have a greater probability of being selected for reproduction. The crossover operation uses an arithmetic crossover method, randomly pairing selected parent individuals and performing a linear combination of the corresponding gene loci of each pair of parent individuals to generate two offspring individuals. The combination coefficient during the crossover process is randomly selected between 0 and 1 to ensure that the gene values ​​of the offspring individuals are still within the range of the parent gene values.

[0073] The mutation operation uses the Gaussian mutation method, which adds a random perturbation that conforms to a normal distribution to each gene position of the selected individual to be mutated. The mutation intensity is dynamically adjusted according to the current generation. In the early stage, a larger mutation intensity is used to enhance the global search capability, and in the later stage, the mutation intensity is gradually reduced to improve the local search accuracy. The mutated individual also needs to be checked for constraints. Gene values ​​that exceed the constraint range are directly corrected to the nearest constraint boundary value.

[0074] The algorithm terminates when the maximum number of generations is reached or when the optimal solution has not improved for 50 consecutive generations. After each generation of evolution, the best individual in the current population and its fitness value are recorded, and convergence is judged. When the algorithm terminates, the final best individual is output as the optimized combination of process parameters, and the corresponding objective function prediction value is output as the performance index of the optimization result.

[0075] Let p_c,max=0.9 be the maximum crossover probability, p_c,min=0.6 be the minimum crossover probability, t be the current iteration number, and T_max be the maximum iteration number. Then the adaptive crossover probability formula is: p_c(t)=p_c,max-(p_c,max-p_c,min)×(t / T_max)^2.

[0076] The solution set selection process chooses the optimal combination of process parameters from the Pareto front solution set using a strategy based on dominance and crowding distance. First, the Pareto front solutions are sorted by dominance relationship. The number of other solutions a solution dominates is calculated based on its values ​​in the three objective functions. Solutions with fewer dominance values ​​are closer to the Pareto optimal front. Solutions with the same number of dominance values ​​are then sorted by crowding distance. Solutions with larger crowding distances are more dispersed in the objective space, representing a wider range of solutions.

[0077] By combining the two indices of dominance level and crowding distance, the solution with the lowest dominance level and the largest crowding distance is selected as the final output from the Pareto front solution set. This screening strategy can effectively ensure the convergence and dispersion of the output solution, avoid the occurrence of local optima, and obtain a Pareto optimal solution set with good representativeness.

[0078] The specific calculation process is as follows: For each solution in the Pareto pre-boiling solution set, calculate its dominance number, i.e. how many other solutions are dominated by that solution; Sort the solution set according to the number of dominances, with the solution having the fewer dominances listed first; For solutions within the same dominance level, calculate the crowding distance; the larger the crowding distance, the more dispersed the spatial distribution of the solutions. By combining the two indicators of dominance level and congestion distance, the solution with the lowest dominance level and the largest congestion distance is selected as the final output.

[0079] This method ensures that the output solution has good convergence and dispersion, and is an effective strategy for selecting solution sets in multi-objective optimization.

[0080] S5: Adaptive parameter adjustment and real-time control strategy: Based on the intelligent optimization results of process parameters under the multi-objective constraints of S4, an adaptive parameter adjustment and real-time control strategy is designed, including real-time monitoring feedback, parameter deviation detection, adaptive adjustment algorithm, and control command generation.

[0081] The real-time monitoring feedback adopts multi-sensor data fusion technology. Vibration sensors, acoustic emission sensors, current sensors and temperature sensors deployed on the metal cutting equipment continuously collect physical signal data during the cutting process at fixed time intervals of 100 milliseconds. The system establishes a data acquisition buffer, and each sensor channel is configured with an independent circular buffer. The buffer size is set to 1000 data points. When the buffer data reaches the set threshold, the data processing flow is automatically triggered.

[0082] The sliding window data processing mechanism maintains a time window with a length of 20 sampling points. The window updates data in a first-in-first-out manner. Each time new data arrives, the oldest data point is removed and the latest data point is added. The system calculates characteristic parameters, including the mean signal amplitude, standard deviation, peak factor, and waveform factor, within the sliding window. The changing trend of the cutting process is identified by comparing the changes in the values ​​of characteristic parameters between consecutive windows.

[0083] The characteristic parameter change trend analysis adopts the linear regression method to calculate the slope of the characteristic parameter change over time within the window. When the absolute value of the slope is greater than the set threshold, it is judged as a trend change. The anomaly calculation is based on the Mahalanobis distance method, which compares the feature vector of the current window with the feature vector of the normal cutting state and calculates the distance value in the multidimensional feature space as the anomaly index.

[0084] Feedback data transmission adopts the real-time Ethernet protocol to ensure that monitoring data can be transmitted from the sensor node to the control processing unit within 10 milliseconds. The system establishes a data timestamp mechanism to assign a precise time stamp to each sampled data to ensure the time synchronization of multi-sensor data and provide accurate real-time status information for subsequent parameter deviation detection and adaptive adjustment.

[0085] The parameter deviation detection is based on a multi-level control chart monitoring system established by statistical process control methods. Independent control charts are established for the three main process parameters: cutting speed, feed rate, and depth of cut. The system first collects historical data under normal cutting conditions, and establishes a control benchmark by calculating the mean and standard deviation of the historical data. The number of historical data samples is set to no less than 100 valid data points to ensure statistical reliability.

[0086] In establishing the upper and lower control limits, the system uses a moving average control chart to statistically analyze five consecutive sampling points as a subgroup. For each subgroup, the system calculates the average value of the data within the subgroup, and then calculates the population mean and standard deviation based on the average values ​​of multiple subgroups. The upper control limit is determined by adding three times the standard deviation to the population mean and dividing by the square root of the number of samples in the subgroup. The lower control limit is determined by subtracting the same value from the population mean.

[0087] Real-time deviation detection is achieved by comparing the currently monitored process parameter values ​​with the established control limits. When the real-time monitored value of a process parameter exceeds its corresponding upper or lower control limit, the system immediately marks the parameter as a deviation state. The system simultaneously monitors the change patterns of multiple consecutive data points. When seven consecutive data points are located on the same side of the center line or one of two consecutive data points exceeds twice the standard deviation range, the system determines it as a parameter trend deviation.

[0088] The severity of deviation is assessed using a deviation quantification method, which calculates the ratio of the distance between the real-time parameter value and the control limit. The greater the degree to which the parameter value exceeds the control limit, the higher the severity score of the deviation. The system determines the response speed and adjustment range of subsequent adaptive adjustment based on the severity of the deviation, providing quantified deviation information input for the fuzzy PID adaptive adjustment algorithm.

[0089] The formulas for calculating the upper control limit (UCL) and lower control limit (LCL) are as follows: UCL=μ+3σ / √N LCL=μ-3σ / √N Where μ is the historical data mean, σ is the standard deviation, and N is the number of subgroup samples, set to 5.

[0090] The adaptive adjustment algorithm adopts a fuzzy PID controller structure. It achieves online adjustment of PID parameters by establishing a fuzzy rule base. The input variables of the fuzzy controller include parameter deviation value and deviation change rate. The output variables are the proportional gain, integral time and derivative time correction values ​​of the PID controller. The fuzzification process converts the precise values ​​of deviation value and deviation change rate into fuzzy linguistic variables. The fuzzy set of deviation value is defined as seven levels: negative large, negative medium, negative small, zero, positive small, positive medium and positive large. The deviation change rate is also divided into the same seven levels of fuzzy set.

[0091] The fuzzy rule base is built based on expert experience and cutting process knowledge, and contains 49 fuzzy inference rules to cover all possible input combinations. When the deviation is negative and the rate of change of the deviation is negative, the system increases the proportional gain and decreases the integral time to speed up the response. When the deviation is close to zero and the rate of change of the deviation is zero, the system keeps the current PID parameters unchanged to maintain a stable state. The fuzzy inference adopts the Madanis type inference method, which synthesizes the antecedent of the rule through the minimum value operation and aggregates the consequent of the rule through the maximum value operation.

[0092] The defuzzification process uses the centroid method to convert the fuzzy inference results into precise PID parameter correction values. The system calculates the centroid position of the output fuzzy set as the final parameter adjustment amount. The centroid method determines the output value by performing an area-weighted average on the membership function. The PID parameter update adopts an incremental adjustment method. The new PID parameter value is equal to the current parameter value plus the correction value output by the fuzzy controller. At the same time, upper and lower limits on parameter changes are set to prevent over-adjustment.

[0093] The adaptive adjustment execution frequency is synchronized with the real-time monitoring system. The PID parameters are calculated and updated once in each monitoring cycle. The system maintains a parameter adjustment history record, recording the time of each adjustment, the parameter values ​​before and after the adjustment, and the reason for the adjustment, providing data support for subsequent algorithm optimization and parameter adjustment effect evaluation.

[0094] The formula for calculating the output of a PID controller is: u(t)=K_p(t)e(t)+K_i(t)∫e(t)dt+K_d(t)de(t) / dt in: K_p(t) = K_p0 + ΔK_p K_i(t) = K_i0 + ΔK_i K_d(t) = K_d0 + ΔK_d Where u(t) is the output signal of the PID controller at time t, e(t) is the system error at time t, K_p(t) is the proportional gain parameter, K_i(t) is the integral gain parameter, K_d(t) is the derivative gain parameter, K_p0 is the base value of the proportional gain, K_i0 is the base value of the integral gain, K_d0 is the base value of the derivative gain, ΔK_p is the adjustment amount of the proportional gain, ΔK_i is the adjustment amount of the integral gain, and ΔK_d is the adjustment amount of the derivative gain.

[0095] The control commands generate optimized parameter values ​​based on the output of an adaptive adjustment algorithm. These are then used to generate machine-executable control commands through the standard protocol format of the CNC system. The system first converts the optimized cutting speed, feed rate, and depth of cut parameters into corresponding S-codes, F-codes, and depth of cut values ​​in the CNC code. The S-code controls the spindle speed, and its value is obtained by dividing the optimized cutting speed by the workpiece diameter, multiplying by 1000, and then dividing by pi. The F-code controls the feed rate, directly using the optimized feed rate parameter value. The depth of cut is controlled through incremental values ​​in the Z-axis coordinate.

[0096] The instruction formatting process follows the ISO 6983 CNC programming standard. The generated control instructions contain complete address characters and parameter values. The system automatically adds necessary auxiliary function codes, including M-codes for coolant control and G-codes for tool compensation, ensuring that the generated instructions can be correctly parsed and executed by the CNC system. The instruction sequence is arranged according to the machining process requirements, prioritizing spindle speed adjustment instructions, followed by feed rate adjustment instructions, and finally coordinate adjustment instructions related to the depth of cut.

[0097] The command transmission uses a real-time communication protocol to establish a connection with the CNC system. The generated control commands are sent to the input buffer of the CNC system through an RS232 serial port or Ethernet interface. The system sets a priority mechanism for command transmission. Emergency adjustment commands have the highest priority and are sent immediately. Regular optimization and adjustment commands are sent in a queue according to time order. During the transmission process, the system monitors the response status of the CNC system and sends the next command only after confirming that each command has been successfully received and executed.

[0098] Command execution verification is achieved by reading the status feedback information of the CNC system. The system continuously monitors the actual spindle speed, actual feed rate, and tool position information, and compares the actual execution results with the sent control commands for verification. When the command execution deviation is detected to exceed the set threshold, the system automatically resends the correction command or triggers the abnormal handling process to ensure the accurate execution of control commands and the stability of the machining process.

[0099] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other. In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for optimizing and analyzing metal cutting process parameters based on machine learning, characterized in that, include: S1: Multi-dimensional process parameter feature extraction and preprocessing: Multi-dimensional physical signals during the metal cutting process are collected through a sensor array to establish a feature library relating process parameters and cutting state, including signal acquisition, noise filtering, feature dimension expansion, and data standardization. S2: Intelligent identification of cutting state based on ensemble learning: Using the multi-dimensional process parameter features extracted in S1 as input, an ensemble learning model integrating multiple algorithms is constructed to intelligently identify and classify metal cutting state, including base learner training, model fusion strategy design, state identification decision, and confidence assessment. S3: Dynamic process parameter sensitivity analysis and weight calculation: Based on the intelligent identification results of cutting state in S2, a parameter sensitivity weight matrix is ​​established to dynamically analyze and calculate the sensitivity of process parameters under different cutting states, including parameter sensitivity quantification, dynamic weight update, and interaction effect analysis. S4: Intelligent optimization of process parameters under multi-objective constraints: Based on the dynamic process parameter sensitivity analysis and weight calculation results of S3, a multi-objective constraint optimization model is constructed to intelligently optimize metal cutting process parameters, including the establishment of multi-objective functions, setting of constraint conditions, design of optimization algorithms, and selection of solution sets; S5: Adaptive parameter adjustment and real-time control strategy: Based on the intelligent optimization results of process parameters under the multi-objective constraints of S4, an adaptive parameter adjustment and real-time control strategy is designed, including real-time monitoring feedback, parameter deviation detection, adaptive adjustment algorithm, and control command generation.

2. The method according to claim 1, characterized in that, S1 includes: Vibration sensors, acoustic emission sensors, current sensors, and temperature sensors are deployed on the metal cutting equipment to synchronously acquire process data at a sampling frequency of 1000Hz. The acquired raw signals are subjected to wavelet denoising processing, and a three-level decomposition is performed using the Daubechies wavelet basis function to retain the main low-frequency components. The processed signal is expanded in terms of feature dimensions, and time-domain statistical features including mean, variance, skewness, and kurtosis are extracted, while frequency-domain features include power spectral density, frequency centroid, and spectral entropy. Construct a joint feature matrix in the time and frequency domains, and calculate the cross-correlation coefficients and phase difference characteristics between the signals from each sensor. A nonlinear feature combination is established, and quadratic and cubic cross feature terms are generated through polynomial feature cross to construct a 64-dimensional comprehensive feature vector. The Z-score standardization method is used to normalize all features to ensure the comparability of features with different dimensions.

3. The method according to claim 1, characterized in that, S2 includes: The design incorporates a three-layer ensemble learning architecture. The first layer consists of three base learners: a random forest, a support vector machine, and a gradient boosting decision tree. The random forest has 100 decision trees with a maximum depth of 15, using Gini impurity as the splitting criterion. The support vector machine uses the RBF kernel function with kernel parameter γ=0.1 and penalty parameter C=100. The gradient boosting decision tree has a learning rate of 0.1 and a subsampling rate of 0.

8. The second layer adopts a weighted average fusion strategy, which dynamically allocates weights based on the accuracy of each base learner on the validation set. The third layer outputs the cutting state identification results and confidence assessment, with the confidence level calculated using the entropy method.

4. The method according to claim 1, characterized in that, S3 includes: An improved Morris sensitivity analysis method was used to quantify the influence of process parameters. Cutting speed v, feed rate f, and depth of cut a_p were used as the main analysis parameters, and Morris sampling trajectories were constructed within the range of values ​​of each parameter. Establish a dynamic weight update mechanism to adjust parameter weights based on the current cutting status and historical optimization results; Construct a parameter interaction effect analysis matrix and calculate the second-order interaction sensitivity.

5. The method according to claim 1, characterized in that, S4 includes: A multi-objective optimization function is established, with surface roughness Ra, cutting force F, and tool life T as optimization objectives; The constraints include process parameter boundary constraints, machining accuracy constraints, and equipment capability constraints.

6. The method according to claim 5, characterized in that, S4 further includes: An improved non-dominated sorting genetic algorithm, NSGA-III, was used for multi-objective optimization. The population size was set to 100, the crossover probability was 0.9, the mutation probability was 0.1, and the maximum number of iterations was 200. An optimal combination of process parameters is selected from the Pareto front solution set by employing a solution set selection strategy based on dominance relationship and crowding distance.

7. The method according to claim 1, characterized in that, S5 includes: A real-time monitoring and feedback system was established to continuously monitor the cutting process status with a sampling period of 100ms. A sliding window data processing mechanism was designed with a window length of 20 sampling points to calculate the changing trend and anomaly of the characteristic parameters within the window. A parameter deviation detection algorithm is constructed, and a control chart is established using statistical process control methods.

8. The method according to claim 7, characterized in that, The S5 also includes: A fuzzy PID adaptive adjustment algorithm is designed, combining the advantages of fuzzy logic and PID control. The fuzzy controller inputs are error and error change rate, and the output is the PID parameter adjustment amount. The fuzzy rule base contains 25 rules, and the membership function adopts a triangular function. Based on the intelligent identification results of cutting conditions in S2 and the sensitivity analysis and weight calculation of dynamic process parameters in S3, corresponding control commands are generated.

Images