A cutting force prediction method based on physics-guided data-driven network

Abstract

The present invention discloses a cutting force prediction method based on a physics-guided data-driven network. The method is applicable to a variety of manufacturing and processing scenarios, especially in processing environments requiring high precision and high efficiency. The method comprises the following steps: 1. collecting a model training data set; 2. constructing a physics-guided data-driven network; 3. constructing a loss function for the physics-guided data-driven network; 4. training and obtaining a cutting force prediction model that meets a specified threshold; and 5. inputting processing information during the processing into the cutting force prediction model to obtain a predicted value and achieve visualization. The present invention addresses the limitations of purely physics-based methods and data-driven methods. By combining the mechanism analysis of the physical model with the learning ability of the data-driven model, the method can more accurately capture the complex dynamics of the cutting process and use physical principles to guide the learning process, thereby reducing the reliance on large amounts of data and improving the prediction accuracy, generalization, and interpretability of the cutting force.

Description

Technical Field

[0001] The present invention belongs to the technical field of cutting force prediction, and in particular is a cutting force prediction method based on a physics-guided data-driven network. Background Art

[0002] Cutting force prediction during intelligent machining plays a crucial role in improving production efficiency, reducing costs, ensuring machining quality, and enhancing the safety of equipment and operators. Overestimating cutting force predictions can lead to the selection of smaller cutting parameters, reducing machining efficiency. Inaccurate or uneven cutting force predictions can cause dimensional errors and reduced surface quality during machining, impacting product precision and performance. Furthermore, they can increase tool wear and shorten tool life.

[0003] Current cutting force prediction methods are primarily categorized as physics-based and data-driven. Physics-based methods include simplified models based on empirical formulas, mechanical principle models that deeply analyze cutting mechanisms, and simulations using computer simulation techniques. Data-driven methods include statistical learning methods through historical data mining, and complex nonlinear models constructed using machine learning and artificial intelligence techniques. Each of these methods has advantages and limitations. Empirical formula methods are easy to apply but have a limited scope of applicability and fail to reflect machining mechanisms. Mechanical principle models have broad applicability but are complex and influenced by multiple factors. Computer simulations provide detailed processes but consume a lot of resources. Statistical learning models rely on large amounts of data and have limited generalization capabilities. Machine learning methods, while offering powerful predictive capabilities, require extensive training data and suffer from poor model interpretability. Therefore, to overcome the limitations of purely physics-based and data-driven methods and leverage their complementary strengths, there is an urgent need to develop physics-guided, data-driven methods that reduce computational costs and improve model transparency, interpretability, and analytical capabilities. Summary of the Invention

[0004] In order to address the shortcomings of the above-mentioned physics-based methods and data-driven methods, the present invention proposes a cutting force prediction method based on a physics-guided data-driven network, in order to combine the interpretability of the physical model and the generalization ability of the data-driven model, thereby achieving more accurate cutting force prediction.

[0005] In order to achieve the above-mentioned object, the present invention adopts the following technical solutions:

[0006] The cutting force prediction method based on a physics-guided data-driven network of the present invention is characterized in that it is performed according to the following steps:

[0007] Step 1: Use a force sensor installed on the CNC machine tool workbench to collect the cutting force signal during the machining process online and use it as a label value; collect the machining information of the CNC machine tool as the input of the model, including: machining parameters, machining part material properties and cutting fluid conditions, and record the cutting time;

[0008] Step 2: Construct a physical guidance data-driven network, including: a data-driven module and a physical guidance module;

[0009] Step 2.1, the data driven module includes a feature extraction block and a head decision block;

[0010] The feature extraction block consists of M autoencoder blocks with the same structure and shared weights;

[0011] The head decision block is composed of an attention mechanism layer and a fully connected layer in series;

[0012] Step 2.1.1: Input the collected processing information into the data driving module, and process it in turn through the M autoencoders of the feature extraction block to obtain a processing feature vector;

[0013] Step 2.1.2: The attention mechanism layer of the head decision block extracts features from the processing feature vector to obtain processing attention features, which are then processed by the fully connected layer to obtain data-predicted cutting force;

[0014] Step 2.2, the physical guidance module includes an extended Kalman filter block and a mechanical model block;

[0015] Step 2.2.1, input the data-predicted cutting force into the extended Kalman filter block, process it through the extended Kalman filter method, and obtain the specific force coefficient and initial cutting angle required by the mechanical model block;

[0016] Step 2.2.2, inputting the specific force coefficient, the initial intrusion angle, and the machining parameters into the mechanical model block to obtain the physically predicted milling force;

[0017] Step 3: Use formula (1) to construct the loss function of the physical guided data driven network :

[0018] (1)

[0019] In formula (1), represents the data loss function, represents the physical term loss function; represents the coefficient of scale difference;

[0020] Step 4: Train the physical guidance data driven network by gradient descent method to obtain a trained milling force prediction model:

[0021] Step 4.1: Train the data-driven module and calculate the data loss function To update the parameters of the data driven module, when When it drops to the specified threshold, the pre-trained data-driven module is obtained;

[0022] Step 4.2: Train the pre-trained data-driven module and physical guidance module and calculate the loss function To update the parameters of the network, when When it drops to the specified threshold, the trained cutting force prediction model is obtained;

[0023] Step 5: Input the processing information of a section of the processing process into the trained cutting force prediction model, obtain the predicted value of the cutting force under the corresponding processing conditions and visualize it.

[0024] The cutting force prediction method based on a physics-guided data-driven network according to the present invention is also characterized in that step 2.2.1 includes:

[0025] Step a: Use the specific force coefficient and the initial intrusion angle as the estimated variables of the extended Kalman filter algorithm, use the data predicted cutting force as the observed variable in the extended Kalman filter, and use formula (2) to obtain the state estimation variable of the i+1 stage: :

[0026] { { i i + 1 } = [ P i ] { i i } + { u i } { F ^ i } = g i ( { i i } ) + { v i } (2)

[0027] In formula (2), [ P i ] is the state transition matrix of stage i, is the predicted cutting force at stage i, is the nonlinear function of stage i, is the process noise of stage i, is the measurement noise of stage i; is the state estimation variable of stage i, and { i i } = [ K tc i K rc i K you i K re i F xr i F ˙ xr i F yr i F ˙ yr i ϕ 0 i ] T ,in, They represent the tangential cutting force coefficient of stage i, the radial cutting force coefficient of stage i, the tangential cutting edge force coefficient of stage i, and the radial cutting edge force coefficient of stage i, respectively. are the horizontal and vertical jumping forces in stage i, They are derivatives, is the initial angle of stage i;

[0028] Step b: Use equations (3) and (4) to obtain the prior estimates of the state estimation variables of stage i: and the posterior estimate ;

[0029] { i ^ i } − = [ P i ] { i ^ i − 1 } + (3)

[0030] { i ^ i } + = { i ^ i } − + [ G i ] ( { F ^ i } − g i ( { i ^ i } − ) ) (4)

[0031] In formula (3) and formula (4), [ G i ] is the Kalman filter gain of stage i.

[0032] In step 2.2.2, the physical prediction milling force of stage i is obtained using formula (5): :

[0033] F p i = ∑ j = 1 N c [ ∫ z jl z yes T i K i h jz ( t i , ϕ 0 i , z ) db ( z )] (5)

[0034] In formula (5), is the rotation matrix of stage i, is the specific force coefficient matrix of stage i, for The uncut thickness of the jth blade on the tool at is the time of stage i, for The cutting depth of the tool is is the number of blades, is the upper limit of the cutting height of the jth blade on the tool during processing, It is the lower limit of the height of the cutting part of the j-th cutting edge on the tool during processing.

[0035] Furthermore, the step 3 includes:

[0036] Step 3.1: Use formula (6) to construct the data loss function :

[0037] (6)

[0038] In formula (6), The milling force predicted by the data of stage i, is the real milling force at stage i, and n represents the total number of stages;

[0039] Step 3.2: Use formula (7) to construct the physical term loss function :

[0040] (7)

[0041] In formula (7), represents the physical consistency function, and f : = F ^ i −  [ E ( F ^ i ) , l i , t i ] ,  [ ] is a nonlinear function operator, representing the mechanical model process, E [ ] represents the extended Kalman filter process, represents the processing parameters of stage i;

[0042] Step 3.3: Calculate the scale difference coefficient using formula (8) ;

[0043] (8)

[0044] In formula (8), epoch is the number of times all training data sets are used for training. Represents the activation function.

[0045] The electronic device of the present invention includes a memory and a processor, and is characterized in that the memory is used to store a program that supports the processor to execute the cutting force prediction method, and the processor is configured to execute the program stored in the memory.

[0046] The present invention provides a computer-readable storage medium, wherein a computer program is stored on the computer-readable storage medium, and the computer program executes the steps of the cutting force prediction method when the computer program is executed by a processor.

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] 1. The present invention can learn the implicit patterns from a large amount of data through a data-driven module, so that the prediction model is not only applicable to specific cutting conditions, but also can adapt to different materials, tools and processing parameters.

[0049] 2. The present invention embeds the prior knowledge of the cutting force mechanical model into the data-driven model by introducing the physical term loss function, and uses the physical model labels and real labels to jointly train the network, thereby reducing the dependence on large amounts of data, improving physical consistency, and enhancing the interpretability of traditional data-driven models.

[0050] 3. By combining the mechanism analysis of the physical model and the generalization ability of the data-driven model, the present invention can more accurately capture the complex nonlinear relationship in the cutting process, thereby improving the accuracy of cutting force prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 Flowchart of a cutting force prediction method based on a physics-guided data-driven network according to an embodiment of the present invention;

[0052] Figure 2 This is a structural diagram of a feature extraction block according to an embodiment of the present invention;

[0053] Figure 3 A comparison diagram of the predicted value and actual value of the cutting force in the x-direction according to an embodiment of the present invention;

[0054] Figure 4 2 is a comparison chart of the predicted value and actual value of the y-direction cutting force of an embodiment of the present invention. DETAILED DESCRIPTION

[0055] The present invention will be described in further detail below with reference to the accompanying drawings in the embodiments of the present invention. The embodiments are only examples and are not intended to limit the present invention.

[0056] In this embodiment, a cutting force prediction method based on a physics-guided data-driven network achieves high-precision and high-generalization cutting force prediction by combining physical models and data-driven models. Figure 1 This is a flow chart of a cutting force prediction method based on a physics-guided data-driven network according to an embodiment of the present invention. The specific steps are as follows:

[0057] Step 1: Use a force sensor mounted on the CNC machine tool worktable to collect cutting force signals during machining online and use them as label values. CNC machine tool processing information, including machining parameters, part material properties, and cutting fluid conditions, is collected as model input, and cutting time is recorded. In this embodiment, the CNC machine tool is a JDGR50A10H vertical milling machining center, which has multiple internal wiring ports and a sensor mounting bracket. A Kistler 9265B force sensor with a sampling frequency of 24 kHz is mounted on the sensor mounting bracket. During machining, a double-edged milling cutter is used to perform transverse cutting on an AISI 4340 workpiece. During this process, CNC machine tool processing information and cutting force signals are collected simultaneously. Data cleaning is performed on the collected processing information and cutting force signals, and the processing information data is normalized. Data cleaning includes removing redundant data, processing missing data, and addressing outliers.

[0058] Step 2: Construct a physical guidance data-driven network, including: a data-driven module and a physical guidance module;

[0059] Step 2.1, the data driven module includes the feature extraction block and the head decision block;

[0060] The feature extraction block consists of M autoencoder blocks with the same structure and shared weights;

[0061] The head decision block is composed of an attention mechanism layer and a fully connected layer in series;

[0062] Step 2.1.1: Input the collected processing information into the data driving module and process it in turn through the M autoencoders of the feature extraction block to obtain the processing feature vector. Figure 2 As shown, in this example, M is 3. is the input i-stage processing information vector, is the output feature vector of stage i processing.

[0063] Step 2.1.2: The attention mechanism layer of the head decision block extracts features from the processing feature vector to obtain processing attention features, which are then processed by the fully connected layer to obtain data to predict the cutting force.

[0064] Step 2.2, the physical guidance module includes the extended Kalman filter block and the mechanical model block;

[0065] Step 2.2.1. Input the data-predicted cutting force into the extended Kalman filter block and process it through the extended Kalman filter method to obtain the specific force coefficient and initial cutting angle required by the mechanical model block;

[0066] Step a: Use the specific force coefficient and the initial intrusion angle as the estimated variables of the extended Kalman filter algorithm, use the data predicted cutting force as the observed variable in the extended Kalman filter, and use formula (1) to obtain the state estimation variable of the i+1 stage: :

[0067] { { i i + 1 } = [ P i ] { i i } + { u i } { F ^ i } = g i ( { i i } , i ) + { v i } (1)

[0068] In formula (1), [ P i ] is the state transition matrix of stage i, is the predicted cutting force at stage i, is the nonlinear function of stage i, is the process noise of stage i, is the measurement noise of stage i; is the state estimation variable of stage i, and { i i } = [ K tc i K rc i K you i K re i F xr i F ˙ xr i F yr i F ˙ yr i ϕ 0 i ] T ,in, They represent the tangential cutting force coefficient of stage i, the radial cutting force coefficient of stage i, the tangential cutting edge force coefficient of stage i, and the radial cutting edge force coefficient of stage i, respectively. are the horizontal and vertical jumping forces in stage i, They are derivatives, is the initial angle of stage i.

[0069] When the sampling frequency is high, the nonlinear function can be linearized, so the state space can be expressed as shown in formula (2):

[0070] { F ^ i } = g ( { i ^ i } − , i ) + [ J i ] ( { i i } − { i ^ i } − ) + { v i } [ J i ] = [ ∂ g ∂ i ] i = { i ^ i } − (2)

[0071] In formula (2), is the Jacobian matrix of stage i;

[0072] Step b: Use equations (3) and (4) to obtain the prior estimates of the state estimation variables of stage i: and the posterior estimate ;

[0073] { i ^ i } − = [ P i ] { i ^ i − 1 } + (3)

[0074] { i ^ i } + = { i ^ i } − + [ G i ] ( { F ^ i } − g i ( { i ^ i } − ) ) (4)

[0075] In formula (3) and formula (4), [ G i ] is the Kalman filter gain of stage i, which is updated according to formula (5);

[0076] [ P i ] − = [ T i − 1 ] [ P i − 1 ] + [ T i − 1 ] T + [ Q ] [ G i ] = [ P i ] − [ J i ] T ( [ J i ] [ P i ] − [ J i ] T + [ R ] ) − 1 [ P i ] + = ( I − [ G i ] [ J i ] ) [ P i ] − (5)

[0077] In formula (5), [ P i ] − is the prior estimate of the covariance matrix of stage i, [ P i ] + is the posterior estimate of the covariance matrix of stage i, [ Q ] is the covariance matrix of process noise, [ R ] is the covariance matrix of the measurement noise.

[0078] Step 2.2.2, input the specific force coefficient, initial intrusion angle and machining parameters into the mechanical model block to obtain the physical predicted milling force;

[0079] Use formula (6) to obtain the physical prediction milling force of stage i :

[0080] F p i = ∑ j = 1 N c [ ∫ z jl z yes T i K i h jz ( t i , ϕ 0 i , z ) db ( z )] (6)

[0081] In formula (6), is the rotation matrix of stage i, is the specific force coefficient matrix of stage i, for The uncut thickness of the jth blade on the tool at is the time of stage i, for The cutting depth of the tool is is the number of blades, is the upper limit of the cutting height of the jth blade on the tool during processing, It is the lower limit of the height of the cutting part of the j-th cutting edge on the tool during processing.

[0082] Step 3: Use formula (7) to construct the loss function of the physical guided data driven network :

[0083] (7)

[0084] In formula (7), represents the data loss function, represents the physical term loss function; represents the coefficient of scale difference;

[0085] Step 3.1: Use formula (8) to construct the data loss function :

[0086] (8)

[0087] In formula (8), The milling force predicted by the data of stage i, is the real milling force at stage i, and n represents the total number of stages.

[0088] Step 3.2: Use formula (9) to construct the physical term loss function :

[0089] (9)

[0090] In formula (9), represents the physical consistency function, and f : = F ^ i −  [ E ( F ^ i ) , l i , t i ] ,  [ ] is a nonlinear function operator, representing the mechanical model process, E [ ] represents the extended Kalman filter process, Indicates the processing parameters of stage i.

[0091] Step 3.3: Calculate the scale difference coefficient using formula (10) ;

[0092] (10)

[0093] In formula (10), epoch is the number of times all training data sets are used for training. Represents the activation function.

[0094] Step 4: Train the physical guided data driven network through the gradient descent method. In this case, the gradient descent algorithm can be Adam, SGD, MSprop or Adadelta to obtain the trained milling force prediction model:

[0095] Step 4.1: Train the data-driven module and calculate the data loss function To update the parameters of the data driven module, when When it drops to the specified threshold, the pre-trained data-driven module is obtained;

[0096] Step 4.2: Train the pre-trained data-driven module and physical guidance module and calculate the loss function To update the parameters of the network, when When it drops to the specified threshold, the trained cutting force prediction model is obtained;

[0097] Step 5: Input the processing information of a section of the processing process into the trained cutting force prediction model, obtain the predicted value of the cutting force under the corresponding processing conditions and realize visualization. The visualization result of the comparison between the cutting force prediction and the actual value in this embodiment is as follows: Figure 3 and Figure 4 As shown, Figure 3 It is a visual diagram of the comparison of cutting forces in the x-direction. Figure 4 It is a visualization diagram of the cutting force comparison in the y direction.

[0098] In this embodiment, an electronic device includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the above method, and the processor is configured to execute the program stored in the memory.

[0099] In this embodiment, a computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the above method are executed.

[0100] In summary, the present invention combines the mechanism analysis of the physical model and the generalization ability of the data-driven model, and can predict the cutting force more accurately, solving the problem that the traditional cutting force prediction method may have insufficient accuracy due to over-reliance on empirical formulas or a single data-driven model; the present invention enhances the adaptability of the model to different cutting conditions by introducing physical principles, solving the problem that a single data-driven model may have insufficient generalization ability when facing new or unknown cutting conditions; the present invention combines the physical model to improve the interpretability of the model, making the prediction results easier to understand and accept, solving the problem that the pure data-driven model is regarded as a "black box" and its prediction results are difficult to explain; the present invention reduces the dependence on large amounts of data through a physically guided method, guides the training and prediction of the model through physical principles, and solves the problem that some data-driven models require a large amount of labeled data for training. This method has the advantages of high prediction accuracy, strong adaptability, good interpretability, and reduced dependence on large amounts of data.

Claims

1. A cutting force prediction method based on a physics-guided data-driven network, characterized in that: Follow these steps: Step 1: Use a force sensor installed on the CNC machine tool workbench to collect the cutting force signal during the machining process online and use it as a label value; collect the machining information of the CNC machine tool as the input of the model, including: machining parameters, machining part material properties and cutting fluid conditions, and record the cutting time; Step 2: Construct a physical guidance data-driven network, including: a data-driven module and a physical guidance module; Step 2.1, the data driven module includes a feature extraction block and a head decision block; The feature extraction block consists of M autoencoder blocks with the same structure and shared weights; The head decision block is composed of an attention mechanism layer and a fully connected layer in series; Step 2.1.1: Input the collected processing information into the data driving module, and process it in turn through the M autoencoders of the feature extraction block to obtain a processing feature vector; Step 2.1.2: The attention mechanism layer of the head decision block extracts features from the processing feature vector to obtain processing attention features, which are then processed by the fully connected layer to obtain data-predicted cutting force; Step 2.2, the physical guidance module includes an extended Kalman filter block and a mechanical model block; Step 2.2.1, input the data-predicted cutting force into the extended Kalman filter block, process it through the extended Kalman filter method, and obtain the specific force coefficient and initial cutting angle required by the mechanical model block; Step 2.2.2, inputting the specific force coefficient, the initial intrusion angle, and the machining parameters into the mechanical model block to obtain the physically predicted milling force; Step 3: Use formula (1) to construct the loss function of the physical guided data driven network : (1) In formula (1), represents the data loss function, represents the physical term loss function; represents the coefficient of scale difference; Step 4: Train the physical guidance data driven network by gradient descent method to obtain a trained milling force prediction model: Step 4.1: Train the data-driven module and calculate the data loss function To update the parameters of the data driven module, when When it drops to the specified threshold, the pre-trained data-driven module is obtained; Step 4.2: Train the pre-trained data-driven module and physical guidance module and calculate the loss function To update the parameters of the network, when When it drops to the specified threshold, the trained cutting force prediction model is obtained; Step 5: Input the processing information of a section of the processing process into the trained cutting force prediction model, obtain the predicted value of the cutting force under the corresponding processing conditions and visualize it.

2. The cutting force prediction method based on a physics-guided data-driven network according to claim 1, characterized in that: Step 2.2.1 includes: Step a: Use the specific force coefficient and the initial intrusion angle as the estimated variables of the extended Kalman filter algorithm, use the data predicted cutting force as the observed variable in the extended Kalman filter, and use formula (2) to obtain the state estimation variable of the i+1 stage: : (2) In formula (2), is the state transition matrix of stage i, is the predicted cutting force at stage i, is the nonlinear function of stage i, is the process noise of stage i, is the measurement noise of stage i; is the state estimation variable of stage i, and ,in, They represent the tangential cutting force coefficient of stage i, the radial cutting force coefficient of stage i, the tangential cutting edge force coefficient of stage i, and the radial cutting edge force coefficient of stage i, respectively. are the horizontal and vertical jumping forces in stage i, They are derivatives, is the initial angle of stage i; Step b: Use equations (3) and (4) to obtain the prior estimates of the state estimation variables of stage i: and the posterior estimate ; (3) (4) In formula (3) and formula (4), is the Kalman filter gain of stage i.

3. The cutting force prediction method based on a physics-guided data-driven network according to claim 1, characterized in that: In step 2.2.2, the physical prediction milling force of stage i is obtained using formula (5): : (5) In formula (5), is the rotation matrix of stage i, is the specific force coefficient matrix of stage i, for The uncut thickness of the jth blade on the tool at is the time of stage i, for The cutting depth of the tool is is the number of blades, is the upper limit of the cutting height of the jth blade on the tool during processing, It is the lower limit of the height of the cutting part of the j-th cutting edge on the tool during processing.

4. The cutting force prediction method based on a physics-guided data-driven network according to claim 1, characterized in that: The step 3 comprises: Step 3.1: Use formula (6) to construct the data loss function : (6) In formula (6), The milling force predicted by the data of stage i, is the real milling force at stage i, and n represents the total number of stages; Step 3.2: Use formula (7) to construct the physical term loss function : (7) In formula (7), represents the physical consistency function, and , is a nonlinear function operator, representing the mechanical model process, represents the extended Kalman filter process, represents the processing parameters of stage i; Step 3.3: Calculate the scale difference coefficient using formula (8) ; (8) In formula (8), epoch is the number of times all training data sets are used for training. Represents the activation function.

5. An electronic device comprising a memory and a processor, characterized in that: The memory is used to store a program that supports the processor to execute the cutting force prediction method according to any one of claims 1 to 4, and the processor is configured to execute the program stored in the memory.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the cutting force prediction method according to any one of claims 1 to 4 are executed.

Images