Thread turning force prediction method based on mutation theory and domain adversarial

Abstract

The application discloses a thread turning specific main cutting force prediction method based on mutation theory and domain adversarial, and belongs to the technical field of machining cutting process modeling and prediction. The method first obtains 1252 groups of cutting parameter and specific main cutting force measured data and chip state through 6061-T6 aluminum alloy thread turning experiment, and then obtains 1002501 groups of theoretical data by discrete cusp type mutation regular balance surface; a random forest model of 100 decision trees is constructed to realize 100% classification of chip bifurcation state, a full connection neural network with two hidden layers is taken as a main stem, a domain adversarial neural network is fused, and cusp mutation geometric features are taken as prior constraints and embedded into a loss function. After independent standardization and non-random division of the measured and theoretical data and training of the dataset, the average absolute percentage error of the model in specific main cutting force prediction is 1.50%, and the maximum absolute error is 4.47%, which is significantly better than that of a transfer learning and support vector machine model. The method architecture is decoupled and easy to be deployed in engineering, provides a quantitative basis for active induction of chip bifurcation and realization of cutting energy saving, and is suitable for 6061-T6 aluminum alloy thread turning process.

Description

Technical Field

[0001] This invention belongs to the field of machining cutting process modeling and prediction technology, specifically involving a method for predicting the specific main cutting force in thread turning based on catastrophe theory and domain adversarial. It can simultaneously achieve high-precision classification of chip bifurcation state and accurate prediction of specific main cutting force during thread turning, and is applicable to the cutting force prediction and cutting energy-saving optimization of 6061-T6 aluminum alloy thread turning process. Background Technology

[0002] In thread turning, continuous changes in cutting parameters can trigger abrupt chip bifurcation, which can reduce the primary cutting force by up to 64.68%, providing an important potential avenue for energy conservation in the cutting process. The core prerequisite for achieving energy-saving utilization of this chip bifurcation abrupt change is establishing a model that can accurately predict the chip bifurcation state and the primary cutting force.

[0003] In existing technologies, modeling of cutting abrupt changes mainly relies on catastrophe theory. However, catastrophe theory is only a qualitative theory, and the analogy models built upon it lack quantitative prediction capabilities, resulting in persistently high prediction errors. For example, the maximum prediction error for the critical depth of cut during chip angle abrupt changes is 19.47%, and the maximum prediction error for the critical cutting thickness during chip bifurcation abrupt changes is 50.44%. Even if some models achieve 100% prediction of chip bifurcation, the maximum absolute prediction error compared to the main cutting force still reaches 22.07%, which cannot meet the energy-saving requirements of engineering.

[0004] Machine learning technology has been applied to cutting process modeling due to its powerful nonlinear fitting ability. Although pure data-driven models can improve fitting accuracy, their generalization ability is poor. Existing research combining machine learning with catastrophe theory has only verified the advantages of data-driven methods in chip state classification or single cutting parameter prediction. It has not yet achieved the collaborative prediction of chip bifurcation state and main cutting force, nor has it applied domain adversarial neural networks (DANN) to the field of cutting force prediction. There is a lack of effective methods to transform the qualitative geometric features of catastrophe theory into quantitative constraints that machine learning models can learn.

[0005] In addition, existing cutting prediction models mostly adopt a combination of classification and regression tasks, which is prone to mutual interference between tasks and makes engineering deployment difficult. Traditional process optimization relies on trial and error, which has high trial cutting costs and time overhead. There is an urgent need for a high-precision, highly generalizable, and easily deployable collaborative prediction method to provide a quantitative basis for actively inducing chip bifurcation to achieve cutting energy saving. Summary of the Invention

[0006] This invention addresses the problems of low modeling accuracy of catastrophe theory, poor generalization of pure data-driven models, lack of efficient collaborative prediction methods, and weak practicality of model engineering in existing prediction methods for chip bifurcation state and specific main cutting force in thread turning. It provides a prediction method for specific main cutting force in thread turning based on catastrophe theory and domain adversarial theory, which achieves high-precision collaborative prediction of chip bifurcation state and specific main cutting force, improves the generalization ability and engineering deployment of the model, and provides quantitative basis for energy-saving optimization of cutting process, reducing trial cutting costs and time overhead.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A method for predicting the main cutting force of thread turning based on catastrophe theory and domain adversarial approaches includes the following steps:

[0009] Step 1: Conduct thread turning experiments on 6061-T6 aluminum alloy to obtain basic cutting test data, chip bifurcation / non-bifurcation states, and 1252 sets of actual equilibrium point coordinate measured data. ,in For cutting width, For cutting thickness, The ratio of the main cutting force;

[0010] Step 2: Based on catastrophe theory, the bifurcation catastrophe in thread turning is determined to be a cusp catastrophe. The cusp catastrophe regular equilibrium surface is discretized to obtain the coordinate data of the theoretical equilibrium point. ),in For theoretical control parameters, These are theoretical state parameters;

[0011] Step 3: Construct a random forest classification model to... The chip bifurcation state is the input, and the chip bifurcation state is the output, thus achieving high-precision classification of chip bifurcation state;

[0012] Step 4: Construct a hybrid model that integrates a fully connected neural network and a domain adversarial neural network. Embed the folded geometric features of the cusp-type abrupt equilibrium surface as a prior constraint into the model training process to achieve the prediction of the main cutting force.

[0013] Step 5: After preprocessing the measured data and theoretical data respectively, divide the dataset, train and validate the RF model and FCN-DANN hybrid model, and collaboratively output the chip bifurcation state and the prediction result of the specific main cutting force to complete the prediction of the specific main cutting force for thread turning.

[0014] Step 6: Design ablation experiments and comparative experiments to verify the predictive performance and generalization ability of the model.

[0015] Furthermore, in step 1, the thread turning experiment was conducted on a CW6163E precision horizontal machine tool, using a Kistler 9257A three-dimensional force measurement system to record the dynamic main cutting force. The parameters of the high-speed steel linear double-edged symmetrical turning tool were: rake angle... -5°, blade tip angle , blade angle , blade inclination angle The cutting parameters are: initial cutting speed m / min, feed rate per revolution mm, depth of cut Set 7 levels: 0.05mm, 0.10mm, 0.15mm, 0.20mm, 0.25mm, 0.30mm, and 0.35mm, to control the cutting thickness. Based on geometric relationships The calculated cutting thicknesses are 0.032mm, 0.064mm, 0.096mm, 0.129mm, 0.160mm, 0.193mm, and 0.225mm.

[0016] Furthermore, in step 2, the equation of the cusp-type abrupt change canonical equilibrium surface is: The equilibrium surface is divided using a uniform grid method with a step size of 0.01 under control parameters. and Discretizing within the interval [−5,5] yields 1002501 sets of theoretical equilibrium point coordinate data. ).

[0017] Furthermore, in step 3, the RF model consists of 100 decision trees. The 1252 sets of measured data are divided into training and test sets in an 8:2 ratio using stratified sampling. The Gini coefficient is used as the decision tree splitting criterion. The model is trained by five-fold cross-validation. The prediction accuracy of the chip bifurcation state on both the training and test sets is 100%.

[0018] Furthermore, in step 4, the basic architecture of the FCN is: based on As input, compared to the predicted value of the main cutting force. For the output, two hidden layers are set, each containing 32 neurons. The activation function of the hidden layers is ReLU, and the output layer has linear activation. The Adam optimizer is used to update the parameters, and the initial loss function is the mean squared error between the predicted and measured values. .

[0019] Furthermore, in step 4, the DANN components include a source domain input layer, a target domain input layer, a source domain encoder, a target domain encoder, a domain classifier, and a gradient inversion layer; the source domain input layer receives standardized theoretical data ( The target domain input layer receives the standardized data. Both the source and target domain encoders are single-layer 128-node ReLU networks, mapping the source and target domain inputs to 64-dimensional features, respectively; the domain classifier is a single-layer 32-node Sigmoid network used for source / target domain classification; GRL reverses the gradient sign during backpropagation to achieve adversarial training.

[0020] Furthermore, in step 4, a domain adversarial loss is constructed. ,in The maximum mean difference is used to measure the distribution distance of features between the source and target domains; , which is the binary cross-entropy, used to measure the classification ability of a domain classifier; The weighting coefficients are dynamically calculated for GRL; the total loss function is constructed. ,in This is the domain adversarial loss weight coefficient, with a value of 0.1.

[0021] Furthermore, in step 5, the data preprocessing involves processing the measured data. Compared with theoretical data ( Z-score standardization is performed separately, and the standardization formula is as follows: , and These are the mean and standard deviation of each variable in its own dataset, and standardization is performed independently in the source and target domains.

[0022] Furthermore, in step 5, a non-random partitioning strategy is adopted to partition the measured data: 539 sets of measured data covering different cutting thickness levels are used as the test set, and the remaining 713 sets are used as the training and validation set; the theoretical data are not partitioned and are all used for the constrained training of DANN.

[0023] Furthermore, in step 5, the FCN-DANN hybrid model uses the Adam optimizer to update parameters, with mean absolute percentage error and coefficient of determination as evaluation metrics. The maximum number of training epochs is set to 100, an early stopping strategy is adopted with an early stopping patience value of 20, and the model's generalization ability is evaluated through five-fold cross-validation. The model's MAPE on the test set, compared to the prediction of the main cutting force, is 1.50%, and R... 2 The value is 0.9825, and the maximum absolute prediction error is 4.47%.

[0024] The thread turning ratio main cutting force prediction method based on catastrophe theory and domain adversarial in this invention has the following significant advantages compared with the prior art:

[0025] 1. Innovative Theoretical Guidance Mechanism Leads to Significantly Improved Prediction Accuracy: For the first time, the folded geometric features of the cusp catastrophe regular equilibrium surface are transformed into a quantitative constraint that can be learned by the neural network. Through independent standardization of the source / target domain and adversarial training of the domain, the organic integration of qualitative description of catastrophe theory and quantitative prediction by machine learning is achieved. Ablation experiments have confirmed that this constraint mechanism reduces the prediction error of the catastrophe critical region by 72.8%, and the maximum absolute prediction error of the model compared with the main cutting force is only 4.47%, and the MAPE of the test set is 1.50%, which is far superior to the traditional catastrophe theory model.

[0026] 2. Modular decoupled design with strong engineering practicality: The chip state classification (RF model) and specific main cutting force regression (FCN-DANN model) are decoupled to avoid mutual interference between classification and regression tasks. In actual cutting process, the chip bifurcation state can be quickly determined by the RF model first, and then the corresponding FCN-DANN model can be called to evaluate energy consumption. The model is convenient to deploy and has high computational efficiency.

[0027] 3. Excellent prediction accuracy and generalization ability, and strong stability: The RF model achieves 100% prediction accuracy for chip bifurcation states and maintains this accuracy on independent test sets; the FCN-DANN hybrid model achieves excellent R-values ​​on the test set. 2 The R-value reached 0.9825, significantly outperforming the TL model (R²). 2 =0.9418) and SVM model (R 2 =0.7174), and the five-fold cross-validation showed that the model had no overfitting and excellent generalization ability and stability.

[0028] 4. Provides direct support for energy saving in cutting and reduces process optimization costs: This invention achieves accurate prediction of chip bifurcation state and specific main cutting force, providing a quantitative basis for actively inducing chip bifurcation mutations and enabling the cutting system to work in a low specific main cutting force state; compared with traditional process optimization that relies on trial and error, this method can significantly reduce trial cutting costs and time expenditures, and has practical engineering application value. Attached Figure Description

[0029] Figure 1 This is a picture of the actual cutting tool used in the experiment;

[0030] Figure 2 For the experimental workpiece, see the actual photograph.

[0031] Figure 3 Images of actual machine tools and measuring instruments;

[0032] Figure 4 For the main cutting force With cutting width Cutting thickness Relationship curve diagram;

[0033] Figure 5 This is a cusp mutation regularized equilibrium surface;

[0034] Figure 6 This is a schematic diagram of the overall architecture of the hybrid RF model and FCN-DANN model.

[0035] Among them, 1. iPhone, 2. Shadowless lamp, 3. Workpiece, 4. Chuck, 5. Kistler 9257A force measurement system base, 6. Double-edged symmetrical turning tool, 7. Center. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0037] The thread turning ratio main cutting force prediction method based on catastrophe theory and domain adversarial theory provided by this invention specifically includes the following steps:

[0038] Step 1:

[0039] Measured data on the chip state and actual equilibrium point of thread turning were obtained using 6061-T6 aluminum alloy as the workpiece. Thread turning experiments were conducted on a CW6163E precision horizontal machine tool equipped with a frequency converter. The workpiece was a bar stock with one symmetrical conical end and one cylindrical end. The thread pitch to be cut was... mm; using front angle -5°, blade tip angle , blade angle , blade inclination angle A high-speed steel linear double-edged symmetrical turning tool, with the tool holder horizontally mounted and perpendicular to the workpiece's axis of rotation, ensuring the tool tip passes through the workpiece spindle's center of rotation; the cutting method is longitudinal turning and dry cutting, completed in three passes, with an initial cutting speed of... m / min, feed rate per revolution mm, the second cutter depth The cutting depth was set to seven levels: 0.05mm, 0.10mm, 0.15mm, 0.20mm, 0.25mm, 0.30mm, and 0.35mm. The corresponding basic data for the cutting experiment are shown in Table 1.

[0040] A Kistler 9257A three-dimensional force measurement system was used to record the dynamic main cutting force signal, and slow-motion video recording of the chip state was recorded via a mobile phone; cutting thickness was also recorded. Based on geometric relationships Calculations show that the main cutting force is... The cutting width is calculated by taking the average value of the stroke from the dynamic main cutting force signal. It is calculated by combining the cutting time with the tapered structure of the workpiece; from and , Points were taken at equal intervals with a step size of 0.01 mm in the relationship curve, and finally 1252 sets of actual equilibrium point coordinates were obtained. And the chip bifurcation / non-bifurcation state corresponding to each set of data.

[0041] Table 1. Data for turning 6061-T6 aluminum alloy tapered bars with high-speed steel double-edged symmetrical turning tools

[0042]

[0043] Step 2:

[0044] Obtaining equilibrium point data from the theory of cusp-type catastrophe regular equilibrium surface: Based on catastrophe theory, the control parameter for chip bifurcation catastrophe in thread turning is the cutting width. Cutting thickness The state parameter is the specific main cutting force. The mutation type was determined to be a cusp mutation, and its regular equilibrium surface equation is: ( For state parameters, (For control parameters).

[0045] Using a uniform grid method, the control parameters are adjusted in steps of 0.01. Discretizing the cusp-type abrupt change regular equilibrium surface in the interval [−5,5] yields 1002501 sets of theoretical equilibrium point coordinate data. This provides geometric prior knowledge of catastrophe theory for subsequent models.

[0046] Step 3:

[0047] Constructing a Random Forest (RF) model to classify chip bifurcation states based on cutting width. Cutting thickness The model is an RF classification model with chip bifurcation state (0 = no bifurcation, 1 = bifurcation) as input and chip bifurcation state as output. The model consists of 100 decision trees. The 1252 sets of measured data are divided into a training set (1001 sets) and a test set (251 sets) in an 8:2 ratio using stratified sampling to ensure that the ratio of chip state samples in the training set and the test set is consistent with the original data.

[0048] Using the Gini coefficient as the decision tree splitting criterion and the chip state prediction accuracy as the core evaluation index, the model was trained through five-fold cross-validation (801 training sets and 200 validation sets per fold). The convergence criterion was that the number of decision trees reached 100, achieving 100% classification accuracy for chip branching states.

[0049] Step 4:

[0050] A hybrid model guided by catastrophe theory, FCN-DANN, is constructed to predict the specific main cutting force. This model uses a fully connected neural network (FCN) as the backbone and a domain adversarial neural network (DANN) as constraints. The folded geometric features of the cusp catastrophe equilibrium surface are embedded as prior constraints in the training process to achieve high-precision prediction of the specific main cutting force. Specifically, this includes:

[0051] (I) FCN Infrastructure Design: Based on , As input, compared to the predicted value of the main cutting force. For the output, two hidden layers are set, each containing 32 neurons. The activation function of the hidden layers is ReLU, and the output layer has linear activation. The Adam optimizer is used to update the network parameters, and the initial loss function is the mean squared error between the predicted and measured values. ;

[0052] (II) DANN Component Design: The DANN component includes a source domain input layer, a target domain input layer, a source domain encoder, a target domain encoder, a domain classifier, and a gradient inversion layer (GRL); the source domain input layer receives standardized theoretical data ( The target domain input layer receives the standardized process parameters and the predicted value of the main cutting force. Both the source domain encoder and the target domain encoder are single-layer 128-node ReLU networks, mapping the source domain and target domain inputs to 64-dimensional features, respectively. , The domain classifier is a single-layer 32-node Sigmoid network. , Random mini-batch samples are used for source / target domain classification; GRL reverses the gradient sign during backpropagation to achieve adversarial training;

[0053] (III) Loss Function Construction: Constructing Domain Adversarial Loss ,in The maximum mean difference is used to measure the distribution distance between features in the source and target domains, thereby aligning feature distributions. Binary cross-entropy (GRL) is used to measure the classification ability of a domain classifier. It is maximized after inverting the gradient using GRL. (i.e., minimize) , (Weight coefficients dynamically calculated for GRL) to enhance feature discriminative power; set domain adversarial loss weight coefficients. =0.1, construct the total loss function .

[0054] Step 5:

[0055] Data preprocessing and dataset partitioning for the measured data in step 1 Compared with the theoretical data in step 2 ( Z-score standardization is performed separately, using the following formula: ( and These are the mean and standard deviation of each variable in its own dataset, respectively. Standardization is performed independently in the source and target domains to avoid distribution shifts caused by differences in units.

[0056] A non-random partitioning strategy was adopted to divide the measured data: 539 groups (43%) of measured data covering different cutting thickness levels were used as the test set, and the remaining 713 groups were used as the training and validation set; the theoretical data were not partitioned and were all used for the constrained training of DANN.

[0057] Step 6:

[0058] Model training and validation: The Adam optimizer was used to update the parameters of the FCN-DANN hybrid model, with mean absolute percentage error (MAPE) and coefficient of determination (R²). 2 The model uses a dual evaluation metric; the maximum number of training rounds is set to 100, and an early stopping strategy is adopted (patience=20, training is terminated if the performance on the validation set does not improve for 20 consecutive rounds). At the same time, the model's generalization ability is evaluated through five-fold cross-validation (570 training sets and 143 validation sets per fold). The chip bifurcation state classification results of the RF model and the specific main cutting force prediction results of the FCN-DANN hybrid model are output together to complete the prediction of the specific main cutting force for thread turning.

[0059] Step 7:

[0060] Model performance verification design includes ablation experiments and comparative experiments to verify model performance: The ablation experiments are set up with four control groups: only FCN, FCN+MMD, FCN+BCE, and FCN+DANN complete structures, to verify the necessity of the catastrophe theory constraint and the synergistic effect of MMD and BCE; The comparative experiments compare the established model with the transfer learning (TL) model and the support vector machine (SVM) model to verify the model's prediction accuracy and generalization ability.

[0061] The specific implementation of the present invention will be described in detail below with reference to specific experimental data, model parameters and accompanying tables. This embodiment is only used to explain the present invention and is not intended to limit the scope of protection of the present invention.

[0062] Example 1: Experimental Data Acquisition

[0063] Conduct the 6061-T6 aluminum alloy thread turning experiment according to the experimental plan in step 1, and measure the depth of cut. Each level corresponds to the cutting thickness. The chip thicknesses were 0.032 mm, 0.064 mm, 0.096 mm, 0.129 mm, 0.160 mm, 0.193 mm, and 0.225 mm, respectively. The basic experimental data are shown in the table. The chip condition for experiment number 1 was non-forked, and the chip condition for experiment number 2 was forked. The experiment number 3 was in... =1.33mm (cutting width from largest to smallest) A sudden chip bifurcation occurs at a cutting width of 1.34 mm (from smallest to largest), verifying that this mutation follows the Maxwell convention.

[0064] Dynamic main cutting force was acquired using a Kistler 9257A force measurement system, and chip status was recorded using slow-motion video from a mobile phone. Combined with geometric calculations and data processing of cutting parameters, 1252 sets of data were ultimately obtained. The measured data and corresponding chip states provide a complete data foundation for subsequent model training and validation.

[0065] Example 2: RF Model Training and Validation

[0066] Following step 3, construct the RF model. The 100 decision trees output the chip state classification results using independent voting. The final output of the model follows the majority voting principle. (in (This is an indicator function; it takes the value 1 if the condition is true, and 0 otherwise).

[0067] The model training results show that the prediction accuracy of the chip bifurcation state on each fold of the five-fold cross-validation is 100%, and the prediction accuracy on 251 independent test sets is still 100%, which verifies the high reliability, robustness and generalization ability of the RF model in the chip bifurcation state classification task.

[0068] Example 3: Training and Validation of the FCN-DANN Hybrid Model

[0069] Construct and train the FCN-DANN hybrid model according to steps 4-6, where Z-Score normalization is performed independently for both experimental and theoretical data, and the domain adversarial loss weight coefficients in the total loss function are... =0.1 is determined by the grid search method; the model training adopts an early stopping strategy, which achieves convergence before the preset 100 training rounds are reached, effectively avoiding overfitting;

[0070] Training results show that the model's performance on the training set is better than the predicted main cutting force (MAPE) of 0.20% and R. 2 =0.9999, MAPE=0.23% on the validation set, R 2 =0.9998, no overfitting characteristics; MAPE=1.50% on 539 independent test sets, R 2 =0.9825, compared to the maximum absolute prediction error of 4.47% for the main cutting force, achieving higher accuracy and stability prediction than the main cutting force.

[0071] Example 4: Model Performance Validation

[0072] To systematically verify the effect of catastrophe theory constraints on model performance and the advantages of the established model compared with existing models, ablation experiments and comparative experiments were designed. The experimental results are presented quantitatively in Tables 2, 3 and 4.

[0073] (a) Ablation experiment: To verify the contribution of each component of the catastrophe theory constraint, four sets of control experiments were designed. The experimental model structure, loss function and theoretical purpose are shown in Table 2. All experimental groups adopted the same network architecture and dataset division, with only the loss function composition differing, to ensure the reliability of the comparison results.

[0074] Table 2 Ablation Experimental Design

[0075]

[0076] Table 3 shows the comparison of the main cutting force prediction performance of the four models on 539 test sets. The results confirm that the complete model of this invention achieves the best performance in both prediction accuracy and goodness of fit, with MAPE reduced by 71.0% compared to the baseline model, and R...2 The value was improved from 0.7205 to 0.9825. The distribution alignment of MMD is the basis, and the adversarial training of BCE is the supplement. Only by working together can the unique geometric details of mutations be fully captured.

[0077] Table 3 Comparison of ablation test results

[0078]

[0079] (II) Comparative Experiments: The constructed model was compared with existing mainstream transfer learning (TL) models and support vector machine (SVM) models. All models were trained using the same dataset and partitioning method, and the results were compared with MAPE and R... 2 The accuracy of chip state prediction was used as a unified evaluation index, and the results are shown in Table 4.

[0080] Table 4 Modeling accuracy of different models

[0081]

[0082] The results show that the established model has the lowest MAPE (Marginal Cutting Force Prediction) and R0 (R0) on the training set, validation set, and test set. 2 All were the highest; in chip state prediction, the established model still maintained 100% accuracy on the test set, which was significantly better than the TL model and SVM model, fully verifying the high accuracy, strong generalization ability and stability of the established model.

[0083] Example 5: Engineering Application

[0084] In the actual optimization of the 6061-T6 aluminum alloy thread turning process, the model of this invention is deployed to the machine tool CNC system to achieve an integrated application of real-time chip state identification, accurate prediction of the main cutting force, and energy-saving process parameter optimization: firstly, the RF model is used to collect the cutting width in real time. Cutting thickness The system takes milliseconds as input to determine the chip bifurcation status. If the chip is determined to be non-bifurcation, the system automatically adjusts the cutting parameters to induce chip bifurcation. Then, the FCN-DANN model is used to predict the specific main cutting force after bifurcation, and the energy-saving effect is quantitatively evaluated.

[0085] Practical engineering applications show that this method can quickly achieve coordinated prediction of chip state and specific main cutting force, providing real-time quantitative basis for energy-saving optimization of cutting process. Compared with the traditional trial and error method, the trial cutting cost is reduced by more than 60%, and the process optimization efficiency is improved by more than 80%, which has significant engineering application value and economic value.

[0086] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A method for predicting the main cutting force in thread turning based on catastrophe theory and domain adversarial theory, characterized in that, Includes the following steps: Step 1: Conduct thread turning experiments on 6061-T6 aluminum alloy to obtain basic cutting test data, chip bifurcation / non-bifurcation states, and 1252 sets of actual equilibrium point coordinate measured data. ,in For cutting width, For cutting thickness, The ratio of main cutting force; Step 2: Based on catastrophe theory, the bifurcation catastrophe in thread turning is determined to be a cusp catastrophe. The cusp catastrophe regular equilibrium surface is discretized to obtain the coordinate data of the theoretical equilibrium point. ),in For theoretical control parameters, These are theoretical state parameters; Step 3: Construct a random forest classification model to... The chip bifurcation state is the input, and the chip bifurcation state is the output, thus achieving high-precision classification of chip bifurcation state; Step 4: Construct a hybrid model that integrates a fully connected neural network and a domain adversarial neural network. Embed the folded geometric features of the cusp-type abrupt equilibrium surface as a prior constraint into the model training process to achieve the prediction of the main cutting force. Step 5: After preprocessing the measured data and theoretical data respectively, divide the dataset, train and validate the RF model and FCN-DANN hybrid model, and collaboratively output the chip bifurcation state and the prediction result of the specific main cutting force to complete the prediction of the specific main cutting force for thread turning. Step 6: Design ablation experiments and comparative experiments to verify the predictive performance and generalization ability of the model.

2. The method for predicting the main cutting force of thread turning based on catastrophe theory and domain adversarial as described in claim 1, characterized in that, In step 1, the thread turning experiment was conducted on a CW6163E precision horizontal machine tool. A Kistler 9257A three-dimensional force measurement system was used to record the dynamic main cutting force. The parameters of the high-speed steel linear double-edged symmetrical turning tool were: rake angle... -5°, blade tip angle , blade angle , blade inclination angle The cutting parameters are: initial cutting speed m / min, feed rate per revolution mm, depth of cut Set 7 levels: 0.05mm, 0.10mm, 0.15mm, 0.20mm, 0.25mm, 0.30mm, and 0.35mm, to control the cutting thickness. Based on geometric relationships The calculated cutting thicknesses are 0.032mm, 0.064mm, 0.096mm, 0.129mm, 0.160mm, 0.193mm, and 0.225mm.

3. The method for predicting the main cutting force of thread turning based on catastrophe theory and domain adversarial as described in claim 1, characterized in that, In step 2, the equation of the cusp-type abrupt change canonical equilibrium surface is: The equilibrium surface is divided using a uniform grid method with a step size of 0.01 under control parameters. and Discretizing within the interval [−5,5] yields 1002501 sets of theoretical equilibrium point coordinate data. ).

4. The method for predicting the main cutting force of thread turning based on catastrophe theory and domain adversarial as described in claim 1, characterized in that, In step 3, the RF model consists of 100 decision trees. The 1252 sets of measured data are divided into training and test sets in an 8:2 ratio using stratified sampling. The Gini coefficient is used as the decision tree splitting criterion. The model is trained by five-fold cross-validation. The accuracy of the model in predicting the chip bifurcation state on both the training and test sets is 100%.

5. The method for predicting the main cutting force of thread turning based on catastrophe theory and domain adversarial as described in claim 1, characterized in that, In step 4, the basic architecture of the FCN is as follows: As input, compared to the predicted value of the main cutting force. For the output, two hidden layers are set, each containing 32 neurons. The activation function of the hidden layers is ReLU, and the output layer has linear activation. The Adam optimizer is used to update the parameters, and the initial loss function is the mean squared error between the predicted and measured values. .

6. The method for predicting the main cutting force of thread turning based on catastrophe theory and domain adversarial as described in claim 1, characterized in that, In step 4, the DANN components include a source domain input layer, a target domain input layer, a source domain encoder, a target domain encoder, a domain classifier, and a gradient inversion layer; the source domain input layer receives standardized theoretical data ( The target domain input layer receives the standardized data. Both the source domain encoder and the target domain encoder are single-layer 128-node ReLU networks, which map the source domain and target domain inputs into 64-dimensional features, respectively. The domain classifier is a single-layer 32-node Sigmoid network used for source / target domain classification; GRL reverses the gradient sign during backpropagation to achieve adversarial training.

7. The method for predicting the main cutting force of thread turning based on catastrophe theory and domain adversarial as described in claim 6, characterized in that, In step 4, construct the domain adversarial loss. ,in The maximum mean difference is used to measure the distribution distance of features between the source and target domains; , which is the binary cross-entropy, used to measure the classification ability of a domain classifier; The weighting coefficients are dynamically calculated for GRL; the total loss function is constructed. ,in This is the domain adversarial loss weight coefficient, with a value of 0.

1.

8. The method for predicting the main cutting force of thread turning based on catastrophe theory and domain adversarial as described in claim 1, characterized in that, In step 5, the data preprocessing involves processing the measured data. Compared with theoretical data ( Z-score standardization is performed separately, and the standardization formula is as follows: , and These are the mean and standard deviation of each variable in its own dataset, and standardization is performed independently in the source and target domains.

9. The method for predicting the main cutting force of thread turning based on catastrophe theory and domain adversarial as described in claim 1, characterized in that, In step 5, a non-random partitioning strategy is used to partition the measured data: 539 sets of measured data covering different cutting thickness levels are used as the test set, and the remaining 713 sets are used as the training and validation set; the theoretical data are not partitioned and are all used for the constraint training of DANN.

10. The method for predicting the main cutting force of thread turning based on catastrophe theory and domain adversarial as described in claim 1, characterized in that, In step 5, the FCN-DANN hybrid model uses the Adam optimizer to update parameters, with mean absolute percentage error (MAPE) and coefficient of determination (COD) as evaluation metrics. The maximum number of training epochs is set to 100, an early stopping strategy is employed with an early stopping patience value of 20, and the model's generalization ability is evaluated using five-fold cross-validation. The model's MAPE on the test set, compared to the prediction of the main cutting force, is 1.50%, and R0 is... 2 The value is 0.9825, and the maximum absolute prediction error is 4.47%.

Images