Micro-texture cutter alloy milling force multi-mechanism fusion prediction method under thermal assistance condition

Abstract

The invention relates to a micro-texture cutter alloy milling force multi-mechanism fusion prediction method under a heat-assisted condition, and belongs to the technical field of titanium alloy mechanical manufacturing. Comprising the following steps: S1, a milling vibration signal noise reduction method: optimizing a wavelet packet threshold noise reduction method of variational mode decomposition based on a dung beetle algorithm; and S2, real-time monitoring and prediction of the milling force: a regression analysis model based on Bayesian optimization, a convolutional neural network, a bidirectional long-short-term memory network and a multi-head attention mechanism. The DBO-VMD-WPT noise reduction method adopted by the invention has the advantages of high-precision noise reduction, high adaptability, wide applicability and the like, and is particularly suitable for processing non-stationary signals in a complex noise environment.

Description

Technical Field

[0001] The present invention relates to a multi-mechanism fusion prediction method for milling force, belonging to the technical field of titanium alloy machining manufacturing. Background Art

[0002] With the rapid development of manufacturing technologies in the "Three Navigations" (aviation, aerospace, and navigation) fields, the application demand for difficult-to-machine metal materials such as titanium alloys is increasing day by day, and the tool performance has become the key bottleneck restricting the development of the industry. Research shows that combining substrate texturing with coating technology can significantly improve the cutting performance of tools. However, when preparing micro-textures using laser processing technology, defects such as remelted layer stacking and micro-cracks often occur, affecting the service performance and service life of tools. To solve the above problems, various auxiliary processes have gradually been introduced in the field of laser material processing to optimize the preparation efficiency and improve tool performance. Among them, the thermal-assisted laser processing technology has become a research hotspot in the auxiliary processes due to its advantages such as high efficiency, wide applicability, and low cost, showing broad application prospects.

[0003] During the milling process of titanium alloys, the cutting force borne on the tool surface increases significantly. The dynamic change of the cutting force will trigger the dynamic response of the tool-workpiece system, and then induce the vibration of the mechanical system. When the fluctuation frequency of the cutting force is close to the natural frequency of the machine tool system, resonance will be triggered, resulting in a significant amplification of the vibration amplitude. Severe vibration will not only accelerate tool wear but may even cause tool fracture and failure. In addition, vibration will also have a negative impact on the machining surface quality, resulting in problems such as increased surface roughness, decreased dimensional accuracy, and increased form and position errors. Therefore, achieving accurate prediction of cutting force is of great significance for optimizing tool service life and improving machining quality.

[0004] For example: The publication number is CN116150910B, and the invention creation name is a micro-texture design method and parameter prediction method for a milling titanium alloy ball-end mill. Its technical solution discloses the micro-texture design of a titanium alloy ball-end mill and predicts milling parameters through a stepwise regression model, which is convenient to use and simple to operate. However, it has limitations due to the influence of specific structures, and it does not involve the processing of milling vibration signals, and is lacking in aspects such as model construction and data processing, and cannot handle complex signal situations.

[0005] In actual machining, the measurement of cutting force is usually directly measured using a quartz three-component dynamometer. This method has high accuracy and reliability. However, the dynamometer has limitations such as complex installation, high cost, and strict requirements for the machining environment in actual applications, which limit its wide application.

[0006] Therefore, it is urgent to propose a multi-mechanism fusion prediction method for the milling force of micro-texture tools alloy under thermal-assisted conditions to solve the above technical problems. Summary of the Invention

[0007] To solve the above problems, a multi - mechanism fusion prediction method for the milling force of micro - textured tool alloys under thermal - assisted conditions is provided. A brief overview of the present invention is given below to provide a basic understanding of certain aspects of the present invention. It should be understood that this overview is not an exhaustive overview of the present invention. It is not intended to identify the key or important parts of the present invention, nor is it intended to limit the scope of the present invention.

[0008] The technical solution of the present invention:

[0009] A multi - mechanism fusion prediction method for the milling force of micro - textured tool alloys under thermal - assisted conditions, comprising the following steps:

[0010] S1. Milling vibration signal noise reduction method:

[0011] Wavelet packet threshold noise reduction method based on dung beetle optimization variational mode decomposition (DBO - VMD - WPT);

[0012] S2. Real - time monitoring and prediction of milling force:

[0013] Regression analysis model based on Bayesian optimization, convolutional neural network, bidirectional long - short - term memory network and multi - head attention mechanism (BO - CNN - BiLSTM - Multihead - Attention).

[0014] Preferably: S1 includes the following steps:

[0015] S1.1. Build a test experimental platform and collect signals;

[0016] S1.2. Optimize VMD parameters;

[0017] S1.3. Signal decomposition and denoising.

[0018] Preferably: In S1.1, build a milling and performance test experimental platform and collect milling vibration signals;

[0019] Adopt sliding window average sampling to reduce the data volume while retaining signal characteristics;

[0020] In S1.2, use the DBO algorithm to globally search for the optimal parameter combination (K, α) of VMD, and use the minimum envelope entropy as the fitness function to achieve adaptive optimization of parameters;

[0021] S1.3 includes the following steps:

[0022] S1.3.1. Decompose the signal based on the optimized parameters to obtain the intrinsic mode functions (IMFs);

[0023] S1.3.2. Perform wavelet packet decomposition on each IMF for several layers using the db4 wavelet basis function to obtain the coefficients C and the length vector L;

[0024] S1.3.3. Apply soft thresholding to the wavelet coefficients to remove the noise components;

[0025] S1.3.4. Reconstruct the signal using the denoised wavelet coefficients.

[0026] Preferably: S1.2 includes the following steps:

[0027] S1.2.1. Set the DBO parameters: set the population size to 5 and the maximum number of iterations to 30; set the lower and upper bounds of the parameters, the range of α is [100, 8000], and the range of K is [2, 12];

[0028] S1.2.2. Find the optimal parameters and plot the convergence curve.

[0029] Preferably: in S1.3.2, perform three-layer wavelet packet decomposition on each IMF; in S1.3.3, the threshold is 0.1.

[0030] Preferably: S2 includes the following steps:

[0031] S2.1. Data preprocessing, including the following steps:

[0032] S2.1.1. Import the dataset, divide the dataset into a training set and a test set, set the proportion of the training set in the total dataset, the last column in the dataset is the output (milling force), use the randperm function to randomly shuffle the dataset, and calculate the number of training set samples and the input feature dimension;

[0033] S2.1.2. Use the mapminmax function to normalize the data of the training set and the test set to the interval [0, 1];

[0034] S2.1.3. Convert the normalized data to the cell array format;

[0035] S2.2. Define the fitness function fical to evaluate the performance of the model under different parameter combinations;

[0036] S2.3. Set the range of parameters to be optimized, including the number of hidden layer nodes (InitialLearnRate), the initial learning rate (NumOfUnits), and the L2 regularization coefficient (L2Regularization);

[0037] S2.4. Use the bayesopt function for Bayesian optimization to find the optimal parameter combination and extract the optimal parameter values from the optimization results;

[0038] S2.5. Determine the input feature dimension, output feature dimension, and convolutional kernel size;

[0039] S2.6. Construct a hybrid neural network model, including an input layer, a convolutional layer, a pooling layer, a bidirectional long short-term memory layer (BILSTM), an attention mechanism layer, a fully connected layer, and an output layer;

[0040] S2.7. Use the layerGraph function to combine each layer into a layer graph and connect the relevant attributes;

[0041] S2.8. Set the training options, including the optimization algorithm, the maximum number of training epochs, the batch size, and the learning rate scheduling strategy;

[0042] S2.9. Use the training set data to train the model to obtain a trained network and training information;

[0043] S2.10. Make predictions for the training set and the test set respectively, and denormalize the prediction results;

[0044] S2.11. Calculate the root mean square error (RMSE), coefficient of determination (R 2 ), mean absolute error (MAE), mean absolute percentage error (MAPE), mean bias error (MBE), and mean square error (MSE).

[0045] Preferably: In step 2.3, the optimization range of the initial learning rate is [0.001, 1], the range of the number of hidden layer nodes is set to [10, 45], and the optimization range of the L2 regularization coefficient is set to [0.0000000001, 0.01].

[0046] Preferably: In step 2.7, use the connectLayers function to connect the miniBatchSize attributes of the fold layer and the unfold layer.

[0047] Preferably: In step 2.8, the optimization algorithm selects the Adam gradient descent algorithm, the maximum number of training times is 100 times, the batch size is 64, and the learning rate decay factor is 0.1.

[0048] Preferably: S2 also includes S2.12. Visualization of the results;

[0049] Draw the loss curve and RMSE curve during the training process of the model structure, the comparison chart of the prediction results of the training set and the test set, the error chart, and the linear fitting chart.

[0050] The present invention has the following beneficial effects:

[0051] (1) The DBO-VMD-WPT noise reduction method adopted by the present invention has advantages such as high-precision noise reduction, strong self-adaptability, and wide applicability, and is particularly suitable for processing non-stationary signals in complex noise environments.

[0052] (2) In the prediction of milling force during the milling of titanium alloy with micro-textured tools using the BO-CNN-BiLSTM-Multihead-Attention regression analysis model adopted by the present invention, through multi-modal feature fusion, dynamic attention mechanism, and efficient hyperparameter optimization, the non-linear and multi-scale coupling problems difficult to handle by traditional methods are solved, and it has significant method innovation and engineering application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 It is a flowchart of the noise reduction algorithm of the present invention.

[0054] Figure 2 It is a curve graph of the DBO optimizing the VMD parameters.

[0055] Figure 3 It is a signal comparison graph of the noise reduction algorithm of the present invention.

[0056] Figure 4 It is a curve graph of the BO algorithm optimizing the hyperparameters of the BiLSTM network.

[0057] Figure 5 It is a BO-CNN-BiLSTM-Multihead-Attention prediction model.

[0058] Figure 6 It is a comparison graph of the prediction results of the training set.

[0059] Figure 7 It is an effect graph of the training set.

[0060] Figure 8 It is a comparison graph of the training results of the test set.

[0061] Figure 9 It is an effect graph of the test set;

[0062] Figure 10 It is a diagram of the equipment required for collecting milling vibration signals. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0063] To make the objectives, technical solutions, and advantages of the present invention clearer and more understandable, the present invention will be described below through specific embodiments shown in the drawings. However, it should be understood that these descriptions are only exemplary and do not intend to limit the scope of the present invention. In addition, in the following description, the descriptions of well-known structures and technologies are omitted to avoid unnecessarily confusing the concepts of the present invention.

[0064] DETAILED DESCRIPTION OF THE EMBODIMENT 1: In combination withFigure 1-10 For this embodiment, the multi - mechanism fusion prediction method for the milling force of a micro - textured tool alloy under thermal - assisted conditions in this embodiment includes the following steps:

[0065] S1. Milling vibration signal denoising method:

[0066] Wavelet packet threshold denoising method based on dung beetle optimization of variational mode decomposition (DBO - VMD - WPT);

[0067] In S1.1, the test experimental platform has a thermal imager, a force - measuring fixing device, a dynamometer, a fixture, a tool shank, a milling force signal receiver, a milling experimental platform with a ball - end mill, an X - ray powder diffractometer, a white - light interferometer, a computer, a super - depth - of - field microscope, a Rockwell indentation tester, a laser marking machine, and a Vickers hardness tester; it includes the following steps:

[0068] S1.1. Build a test experimental platform and collect signals;

[0069] Build a milling and performance test experimental platform and collect milling vibration signals;

[0070] Adopt sliding - window average sampling to reduce the data volume while retaining signal characteristics; greatly reduce the computational burden, provide reliable guarantee for subsequent efficient processing, and reduce the impact on model performance;

[0071] S1.2. Optimize VMD parameters;

[0072] Use the DBO algorithm to globally search for the optimal parameter combination (K, α) of VMD, with the minimum envelope entropy as the fitness function to achieve adaptive optimization of parameters; use DBO to optimize VMD parameters to adaptively determine the optimal decomposition parameters with the minimum envelope entropy. Compared with traditional fixed - parameter decomposition, it can more accurately reflect signal characteristics; S1.2 includes the following steps:

[0073] S1.2.1. Sequentially set DBO parameters: the population size is set to 5, and the maximum number of iterations is 30; set the lower and upper bounds of the parameters, the range of α is [100, 8000], and the range of K is [2, 12];

[0074] S1.2.2. Find the optimal parameters and draw the convergence curve;

[0075] S1.3. Signal decomposition and denoising; WPT further reduces noise on the basis of IMF, improves signal quality, and improves the quality of milling vibration signals;

[0076] S1.3 includes the following steps:

[0077] S1.3.1. Decompose the signal based on the optimized parameters to obtain the intrinsic mode functions (IMFs);

[0078] S1.3.2. Perform three-layer wavelet packet decomposition on each IMF using the db4 wavelet basis function to obtain the coefficient C and the length vector L;

[0079] S1.3.3. Apply soft thresholding to the wavelet coefficients with a threshold of 0.1 to effectively remove noise in specific frequency bands;

[0080] S1.3.4. Reconstruct the signal using the denoised wavelet coefficients;

[0081] S2. Real-time monitoring and prediction of milling force:

[0082] Based on the regression analysis model of Bayesian optimization, convolutional neural network, bidirectional long short-term memory network and multi-head attention mechanism (BO-CNN-BiLSTM-Multihead-Attention), which focuses on the optimized extraction and capture of key features such as hyperparameters, local features, and training information, improves the prediction accuracy and is applicable to dynamic data processing; S2 includes the following steps:

[0083] S2.1. Data preprocessing, including the following steps:

[0084] S2.1.1. Import the dataset, divide the dataset into a training set and a test set, set the proportion of the training set in the total dataset to 70%, the last column in the dataset is the output (milling force), use the randperm function to randomly shuffle the dataset, and calculate the number of training set samples and the input feature dimension;

[0085] S2.1.2. Divide the dataset into a training set and a test set according to the previously determined proportion, and at the same time obtain the number of samples in the training set and the test set. Use the mapminmax function to normalize the data in the training set and the test set to the interval [0, 1];

[0086] S2.1.3. Convert the normalized data into a cell array format;

[0087] S2.2. Define the fitness function fical to evaluate the performance of the model under different parameter combinations;

[0088] S2.3. Set the parameter ranges to be optimized, including the number of hidden layer nodes (InitialLearnRate), the initial learning rate (NumOfUnits), and the L2 regularization coefficient (L2Regularization);

[0089] The optimization range of the initial learning rate is [0.001, 1], the range of the number of hidden layer nodes is set to [10, 45], and the optimization range of the L2 regularization coefficient is set to [0.0000000001, 0.01];

[0090] S2.4. Use the bayesopt function for Bayesian optimization to find the optimal parameter combination, and extract the optimal parameter values from the optimization results;

[0091] S2.5. Determine the input feature dimension, output feature dimension, and convolution kernel size;

[0092] S2.6. Build a hybrid neural network model, including an input layer, a convolutional layer, a pooling layer, a bidirectional long short-term memory layer (BILSTM), an attention mechanism layer, a fully connected layer, and an output layer;

[0093] S2.7. Use the layerGraph function to combine each layer into a layer graph, and use the connectLayers function to connect the miniBatchSize attributes of the fold layer and the unfold layer;

[0094] S2.8. Set the training options, including the optimization algorithm, the maximum number of training epochs, the batch size, and the learning rate scheduling strategy; select the Adam gradient descent algorithm as the optimization algorithm, the maximum number of training times is 100, the batch size is 64, and the learning rate decay factor is 0.1;

[0095] S2.9. Use the training set data to train the model to obtain the trained network and training information;

[0096] S2.10. Make predictions on the training set and the test set respectively, and denormalize the prediction results;

[0097] S2.11. Calculate the root mean square error (RMSE), coefficient of determination (R 2 ), mean absolute error (MAE), mean absolute percentage error (MAPE), mean bias error (MBE), and mean square error (MSE) metrics to make predictions on the training set and the test set;

[0098] S2.12. Plot the loss curve and RMSE curve during the training process of the model structure, the comparison chart of the prediction results of the training set and the test set, the error chart, and the linear fitting chart, which greatly facilitates the comprehensive evaluation of the model training effect and prediction accuracy, is intuitive and easy to adjust and optimize, and ensures the effectiveness of the model in practical applications.

[0099] Example 1:

[0100] Step 1: The ball-end milling cutter is a cemented carbide tool with the material YG8; the milling metal is selected as titanium alloy with the material Ti6Al4V.

[0101] Step 2: Select two materials, AlSiTiN and AlCrN, for the composite coating preparation process on the surface of cemented carbide.

[0102] Step 3: Design the test parameters as shown in Table 1. A total of 28 cemented carbide tools are used, corresponding to different laser parameters, texture geometric parameters, and thermal assistance parameters respectively. Each tool cuts 15,000 mm, and the milling vibration signals are collected during the cutting process.

[0103] Step 4: Tool test parameters in Table 1

[0104]

[0105]

[0106] Step 5: Use a VDL-1000E three-axis vertical milling machine for processing. A Kistler 9257 dynamometer is equipped on the machine to collect milling force data. The milling force data are shown in Tables 2 - 5. In addition, vibration and noise sensors are installed to collect vibration and noise signals during the milling process. The equipment diagram is as Figure 1 shown.

[0107] Table 2 First milling data

[0108]

[0109] Table 3 Second milling data

[0110]

[0111] Table 4 Third milling data

[0112]

[0113] Table 5 Fourth milling data

[0114]

[0115] Step 6: The collected milling vibration signals include three directions: X, Y, and Z. After testing, the Z-axis signal is very helpful for the subsequent progress. Therefore, only the Z-axis milling vibration signals are preprocessed in this experiment.

[0116] Step 7: Perform DBO-VMD-WPT noise reduction processing on the collected milling vibration signals. The algorithm flow chart is as Figure 2 shown.

[0117] Step 8: Optimize the VMD decomposition mode number K and penalty factor α based on the DBO algorithm. The optimization curve graph is as Figure 3 shown.

[0118] Step 9: The VMD decomposes the signal to obtain multiple intrinsic mode functions (IMFs).

[0119] Step 10: As Figure 3As shown, it is a comparison chart of milling vibration signal noise reduction.

[0120] Step Eleven: Extract the eigenvalue of the noise-reduced signal.

[0121] Step Twelve: Use the Relief-F algorithm to screen the selected eigenvalues.

[0122] Step Thirteen: Take the 22 groups of screened eigenvalues as input values and the measured milling force as the output value. Collect 4 times of milling force signals for each tool to obtain 112 groups of data sets.

[0123] Step Fourteen: Randomly select 70% of the samples in the data set for training and the remaining 30% for testing. And perform normalization processing on each group of data to improve the performance and stability of the model.

[0124] Step Fifteen: The optimization process of the BO algorithm for the hyperparameters of the BiLSTM network is as Figure 4 shown. Among them, the blue curve represents the loss value of each iteration. From the trend of the curve, it can be seen that as the number of iterations increases, the loss value gradually decreases; the yellow curve represents the root mean square error value of each iteration, and it also shows a trend of continuous decrease in error as the number of iterations increases. It can be seen from this that the BO algorithm is quite effective in optimizing the hyperparameters of the BiLSTM network.

[0125] Step Sixteen: As Figure 5 shown, it is the neural network diagram of the BO-CNN-BiLSTM-Multihead-Attention regression analysis model.

[0126] Step Seventeen: From Figure 6 the comparison chart of the training set results shown, it can be clearly seen that the model shows a high prediction accuracy in the training stage. The RMSE value of the model is only 1.3149. This data intuitively shows that during the training process, the deviation between the model prediction value and the true value is extremely small, which strongly verifies the accuracy and reliability of the model.

[0127] Step Eighteen: It can be clearly seen from Figure 7 that in the evaluation of the training set, the prediction R 2 value of this model is as high as 0.9967. This data indicates that there is an almost perfect linear correlation between the model and the actual data. This means that the model can accurately capture the internal trend of the data, highly fit the change law of the actual data, and show extremely excellent performance in data trend analysis and prediction.

[0128] Step Nineteen: By comparing Figure 8 and Figure 9Through the analysis, it can be intuitively understood the results and effects of milling force prediction on the test set based on the same vibration eigenvalue. During the prediction process of the test set, the R 2 value of the model is 0.99194 and the RMSE value is 2.4113. Although these values show a certain degree of decline compared with the training set, this does not affect the excellent prediction performance of the model, and it can still play an outstanding role in milling force prediction.

[0129] It should be noted that in the above embodiments, as long as the technical solutions are not contradictory, they can be arranged and combined. Those skilled in the art can exhaust all possibilities based on the mathematical knowledge of permutation and combination. Therefore, the present invention will not describe the technical solutions after permutation and combination one by one, but it should be understood that the technical solutions after permutation and combination have been disclosed by the present invention.

[0130] The above are only the preferred embodiments of the present invention and are not used to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A multi - mechanism fusion prediction method for the milling force of micro - textured tool alloys under thermal - assisted conditions, characterized in that: It includes the following steps: S1. Milling vibration signal noise reduction method: Wavelet packet threshold noise reduction method based on optimizing variational mode decomposition by dung beetle algorithm; S2. Real-time monitoring and prediction of milling force: Regression analysis model based on Bayesian optimization, convolutional neural network, bidirectional long short-term memory network and multi-head attention mechanism.

2. The multi-mechanism fusion prediction method for the milling force of a micro-textured tool alloy under hot-assisted conditions according to claim 1, wherein: S1 includes the following steps: S1.

1. Build a test experiment platform and collect signals; S1.

2. Optimize VMD parameters; S1.

3. Decompose and denoise the signals.

3. The multi-mechanism fusion prediction method for the milling force of micro-textured tool alloys under thermally assisted conditions according to claim 2, wherein: In S1.1, build a milling and performance test experiment platform and collect milling vibration signals; Adopt sliding window average sampling to reduce the data volume while retaining signal characteristics; In S1.2, use the DBO algorithm to globally search for the optimal parameter combination (K, α) of VMD, use the minimum envelope entropy as the fitness function, and realize the adaptive optimization of parameters; S1.3 includes the following steps: S1.3.

1. Decompose the signal based on the optimized parameters to obtain the intrinsic mode functions; S1.3.

2. Perform several layers of wavelet packet decomposition on each IMF, use the db4 wavelet basis function to obtain the coefficient C and the length vector L; S1.3.

3. Apply soft thresholding to the wavelet coefficients to remove the noise components; S1.3.

4. Reconstruct the signal using the denoised wavelet coefficients.

4. The multi-mechanism fusion prediction method for the milling force of a micro-textured tool alloy under hot-assisted conditions according to claim 3, characterized in that: S1.2 includes the following steps: S1.2.

1. Set DBO parameters: set the population size to 5 and the maximum number of iterations to 30; set the lower and upper bounds of the parameter α range [100, 8000], and the K range [2, 12]; S1.2.

2. Find the optimal parameters and draw the convergence curve.

5. The multi-mechanism fusion prediction method for the milling force of the micro-textured tool alloy under hot-assisted conditions according to claim 4, wherein: In S1.3.2, perform three-layer wavelet packet decomposition on each IMF; in S1.3.3, the threshold is 0.

1.

6. The multi-mechanism fusion prediction method for the milling force of a micro-textured tool alloy under thermally assisted conditions according to claim 5, wherein: S2 includes the following steps: S2.

1. Data preprocessing, including the following steps: S2.1.

1. Import the dataset, divide the dataset into a training set and a test set, set the proportion of the training set in the total dataset, the last column in the dataset is the output, use the randperm function to randomly shuffle the dataset, and calculate the number of training set samples and the input feature dimension; S2.1.

2. Use the mapminmax function to normalize the data of the training set and the test set to the interval [0, 1]; S2.1.

3. Convert the normalized data into a cell array format; S2.

2. Define the fitness function fical to evaluate the performance of the model under different parameter combinations; S2.

3. Set the range of parameters to be optimized, including the number of hidden layer nodes, the initial learning rate, and the L2 regularization coefficient; S2.

4. Use the bayesopt function for Bayesian optimization to find the optimal parameter combination and extract the optimal parameter values from the optimization results; S2.

5. Determine the input feature dimension, output feature dimension, and convolutional kernel size; S2.

6. Build a hybrid neural network model, including an input layer, a convolutional layer, a pooling layer, a bidirectional long short-term memory layer, an attention mechanism layer, a fully connected layer, and an output layer; S2.

7. Use the layerGraph function to combine each layer into a layer graph and connect the relevant attributes; S2.

8. Set training options, including optimization algorithm, maximum number of training epochs, batch size, and learning rate scheduling strategy; S2.

9. Use the training set data to train the model to obtain a trained network and training information; S2.

10. Make predictions on the training set and the test set respectively, and denormalize the prediction results; S2.

11. Calculate the root mean square error (RMSE), coefficient of determination (R 2 ), mean absolute error (MAE), mean absolute percentage error (MAPE), mean bias error (MBE), and mean square error (MSE).

7. The multi-mechanism fusion prediction method for the milling force of a micro-textured tool alloy under thermally assisted conditions according to claim 6, wherein: In step 2.3, the optimization range of the initial learning rate is [0.001, 1], the range of the number of hidden layer nodes is set to [10, 45], and the optimization range of the L2 regularization coefficient is set to [0.0000000001, 0.01].

8. The multi-mechanism fusion prediction method for the milling force of a micro-textured tool alloy under hot-assisted conditions according to claim 7, wherein: In step 2.7, use the connectLayers function to connect the miniBatchSize attributes of the fold layer and the unfold layer.

9. The multi-mechanism fusion prediction method for the milling force of a micro-textured tool alloy under thermally assisted conditions according to claim 8, wherein: In step 2.8, select the Adam gradient descent algorithm as the optimization algorithm, the maximum number of training times is 100, the batch size is 64, and the learning rate decay factor is 0.

1.

10. The multi - mechanism fusion prediction method for the milling force of micro - textured tool alloy under thermal - assisted conditions according to claim 9, characterized in that: S2 also includes S2.

12. Visualization of results; Plot the loss curve and RMSE curve during the training process of the model structure, the comparison chart, error chart, and linear fitting chart of the prediction results of the training set and the test set.

Images