Variable working condition tool wear prediction method based on compound machine tool

Abstract

The invention relates to the technical field of compound machine tools, and discloses a variable working condition tool wear prediction method based on a compound machine tool, which comprises the following steps: s1, collecting dynamic signals and acquiring process parameters; s2, constructing a tool wear prediction model; and s3, multi-task training of the meta learning framework: obtaining updated Wy and by through meta learning. And deploying the tool wear prediction model in an actual machining system of the compound machine tool, and quickly outputting a tool wear prediction value at the current moment according to the received dynamic signal and process parameters. The method has the beneficial effects that the machine tool does not need to be shut down and the tool wear prediction accuracy is high.

Description

Technical Field

[0001] The invention relates to the technical field of composite machine tools, and in particular to a variable working condition tool wear prediction method based on composite machine tools. Background Art

[0002] The tool is an indispensable core component in CNC machining, and its status directly affects the machining accuracy, workpiece surface quality and production efficiency. During the machining process, tool wear gradually accumulates. If the wear status cannot be monitored and predicted in time, it may lead to workpiece size deviations, surface defects or tool breakage, thereby causing production interruptions and even equipment damage. Especially for compound machine tools, compound machine tools usually have multiple tool magazines, each of which stores multiple tools. Multiple tools are used during the machining process of compound machine tools. When any of the tools is worn beyond the preset value, the workpiece machining quality will not meet the standard. Carrying out tool wear detection and prediction research is of great practical significance for improving machining efficiency, ensuring product quality and reducing production costs.

[0003] Existing tool wear detection methods are mainly divided into two categories: direct method and indirect method. The direct method is represented by the image method, which uses microscopes, industrial cameras and other equipment to obtain images of the tool wear surface and analyze the wear area or wear amount. This method can directly observe the tool wear state, but it requires stopping the machine for image acquisition, cannot achieve online detection, and has high cost and low detection efficiency, which is not suitable for large-scale production applications. The indirect method is based on machining process parameters such as cutting force, spindle current, vibration signal, etc., and infers the tool wear state through data analysis and physical modeling. However, since the tool wear process is essentially a complex process with multi-factor coupling and dynamic nonlinearity, the physical modeling method is often difficult to obtain high-precision prediction results under variable working conditions, and the applicability and robustness of the model are poor. Summary of the invention

[0004] In order to solve the above-mentioned problems existing in the prior art, the present invention provides a variable working condition tool wear prediction method based on a composite machine tool. This prediction method does not require the machine tool to stop. Based on the combination of deep learning and physical mechanism, a dynamic working condition tool wear prediction method is constructed through multivariate data fusion and feature extraction, and meta-learning is used to effectively improve the prediction accuracy.

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

[0006] A variable working condition tool wear prediction method based on a composite machine tool comprises the following steps:

[0007] s1: Collect dynamic signals and obtain process parameters: Collect dynamic signals: cutting force (F c,t ), spindle current (I s,t ) and vibration signal (Vt );Obtain process parameters: cutting speed v c , feed rate: v f , cutting temperature T c , tool yield strength σ y ;

[0008] s2: Build a tool wear prediction model: s2-1: Based on the long short-term memory network model LSTM, build the input X of the tool wear prediction model i ; s2-2: Build the tool wear prediction loss function L enhanced , L enhanced = L LSTM + λ·L physics ; In the above formula, L physics is the physical model loss function, λ is the physical loss weight coefficient, used to adjust the weight of the physical constraint loss; L LSTM is the supervised loss of the long short-term memory network model LSTM; s2-3: Build the tool wear state vector h t , by training the long short-term memory network model LSTM, obtain the tool wear prediction value: In the above formula, represents the tool wear prediction value at the current time step t, h t is the tool wear state vector, W y is the output layer weight matrix, b y is the bias vector; During the training process of the short-term memory network model LSTM, according to the tool wear prediction loss function L enhanced established in s2-2, the training objective is to minimize the tool wear prediction loss function L enhanced , and obtain the initial W y and b y through training;

[0009] s3: Multi-task training of the meta-learning framework: Through meta-learning, obtain the updated W y and b y .

[0010] Preferably, in step s1, the cutting force (F c,t ), spindle current (I s,t ), vibration signal (V t ), cutting temperature T c are obtained by collecting dynamic data of the composite machine tool processing process through sensors; the cutting speed v c , feed rate: v f are obtained through the communication interface of the composite machine tool; the tool yield strength σ y is obtained by referring to materials.

[0011] Preferably, in step S2-1, the input X of the tool wear prediction model i includes two parts: input dynamic signals and process parameters: In the above formula, represents the dynamic signal, including cutting force (F c,t ), spindle current (I s,t ), vibration signal (V t ), cutting temperature T c ; represents the process parameters, including cutting speed v c , feed rate v f , yield strength σ y ; t represents the time step of a certain signal acquisition, and T represents the total number of signal acquisitions.

[0012] Preferably, the construction method of L physics in step S2-2 is as follows:

[0013] Construct a chip force model: F c = K c ·a p ·f z ; Construct a temperature rise model:

[0014] In the above formula, F c is the cutting force, K c is the material cutting force coefficient, a p is the cutting depth, f z is the feed per tooth; T c is the rising temperature, v c is the cutting speed, η c is the thermal efficiency, A c is the cutting contact area, and k is the material thermal conductivity;

[0015] Take the deviation between the model calculation result and the actual acquisition value as one of the loss functions of the tool wear prediction model, and add it to the optimization objective in the form of a regularization term, so that the LSTM model can learn in the direction that conforms to physical laws;

[0016] Construct a cutting force model loss function:

[0017] Construct a temperature rise constraint loss function:

[0018] In the above formula, is the cutting force obtained from actual acquisition, is the cutting force obtained from model calculation; is the temperature rise obtained from actual acquisition, is the temperature rise obtained from model calculation;

[0019] Define the physical model loss function L physics as follows:

[0020] L physics = λ1L force + λ2L temperature ;

[0021] In the above formula, λ1 and λ2 are weight factors used to balance the importance of different physical constraints.

[0022] Preferably, in step s2-3, the construction method of the tool wear state vector h t is as follows: In the LSTM model, the dynamic signal and process parameters are used as independent long short-term memory network channels to extract time series representations, which are specifically defined as follows:

[0023]

[0024] In the above formula, is the time series representation extracted using the dynamic signal as the long short-term memory network channel, is the time series representation extracted using the process parameters as the long short-term memory network channel; the outputs of the two channels are fused in the hidden layer to form the representation of the tool wear state vector h t :

[0025] Preferably, in step s2-3, the entire training process is optimized for parameters through backpropagation and gradient descent algorithms. The tool wear prediction loss function is differentiated with respect to all trainable parameters in the long short-term memory network model, and the gradient of each parameter is calculated, and each parameter is updated along the negative direction of the gradient:

[0026] In the above formula, η is the learning rate, and θ lstm is the set of all learnable parameters in the LSTM model, including W y and b y .

[0027] Preferably, the multi-task training method of the meta-learning framework in step s3 is as follows:

[0028] s3-1: Divide the samples into multiple tasks T i , and within each task, the data is divided into a training subset and a validation subset

[0029]

[0030] In the above formula, represents the training set of the i-th task in the LTSM model, Indicates the training set and validation set of the i-th task in the LTSM model;

[0031] s3-2: Randomly obtain the shared initialization parameter θ of meta-learning, and update it through the inner loop of meta-learning to obtain the personalized model parameter θ i ′; Based on the personalized model parameter θ i ′, update it through the outer loop of meta-learning to obtain the updated global shared parameter θ new ; Among them, θ, θ i ′, θ new are all sets of all learnable parameters in the LSTM model, and both include the parameters W y and b y .

[0032] Preferably, in step s3-2, the inner loop update is as follows:

[0033]

[0034] In the above formula, α is the inner loop learning rate, used for in-task gradient update, is the model supervision loss of the training data on task T i , is the physical model loss corresponding to task T i , and λ is the physical loss weight coefficient;

[0035] The outer loop update based on θ i ′ is as follows:

[0036]

[0037] In the above formula, β is the outer loop learning rate, used for global update; is the validation loss of task T i , used to guide the update of the global shared parameter θ; is the physical model loss, ensuring the generalization ability of θ in physical laws;

[0038] Obtain the updated global shared parameter θ through the outer loop new , obtain the updated W y and b y , and then obtain the tool wear prediction value after meta-learning

[0039] Preferably, it further includes the following steps:

[0040] s4: Fast fine-tuning and deployment of new working condition tasks:

[0041] s4-1: When the model encounters unseen new working condition tasks during actual application deployment, without changing the model structure, rapid fine-tuning is completed through gradient update; the global shared parameters θ learned in step s3 are used in the fine-tuning process new , and the following update is performed in combination with the small sample data of the new working condition task:

[0042]

[0043] In the above formula, θ′ new is the model parameter after rapid adaptation to the new working condition task, is the supervised loss of the new task training set, is the physical model loss term of the new task, and finally the fine-tuned model parameter θ′ new is the long short-term memory network model parameter θ for actual deployment and prediction, lstm , that is, the parameters W y and b y under the new working condition task are obtained;

[0044] s4-2: The tool wear prediction value model After multi-task meta-learning and rapid fine-tuning of the new working condition task, it is deployed in the actual machining system of the composite machine tool, and dynamically receives dynamic signals and process parameters. Based on the updated model parameter θ lstm , it quickly outputs the tool wear prediction value at the current moment

[0045] Therefore, the present invention has the following beneficial effects: (1) By introducing dynamic signals and process parameters, combining with the LSTM model, a tool wear prediction loss function is constructed, and the tool wear prediction value is obtained by training the LSTM model and the parameters W y and b y are updated through meta-learning to improve the prediction accuracy; (2) Multi-task training using the meta-learning framework, the backpropagation and gradient descent algorithms are used for parameter optimization during the meta-learning process, and combined with the inner loop and outer loop updates, the model parameters are continuously optimized to enable it to extract the common features between different working condition tasks and further improve the prediction accuracy; (3) For new working condition tasks, without changing the model structure, rapid fine-tuning is achieved through gradient update to obtain θ′ new , realizing rapid adaptation to the new working condition. Description of the Drawings

[0046] Figure 1 is the block diagram of the long short-term memory network model.

[0047] Figure 2 is the block diagram of meta-learning and meta-testing.

[0048] Figure 3 It is a comparison chart of the predicted value of tool wear and the actual value of tool wear. Specific implementation mode

[0049] In order to make the technical problems, technical solutions and beneficial technical effects to be solved by the present invention clearer and more understandable, the present invention will be further described in detail below in conjunction with the drawings and multiple exemplary embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention, rather than to limit the protection scope of the present invention.

[0050] It should be understood that the LSTM model involved in this embodiment is the abbreviation of the Long Short-Term Memory network model. The terms involved in the embodiment: Long Short-Term Memory network model, Long Short-Term Memory network model LSTM, and LSTM model all refer to the Long Short-Term Memory network model.

[0051] Embodiment 1: A variable-condition tool wear prediction method based on a composite machine tool, comprising the following steps:

[0052] s1: Collect dynamic signals and obtain process parameters: Collect dynamic signals: cutting force (F c,t ), spindle current (I s,t ), and vibration signal (V t ); Obtain process parameters: cutting speed v c , feed speed: v f , cutting temperature T c , tool yield strength σ y ;

[0053] Cutting force (F c,t ), spindle current (I s,t ), vibration signal (V t ), cutting temperature T c are obtained by collecting dynamic data in the processing state of the composite machine tool through sensors; cutting speed v c , feed speed: v f are obtained through the communication interface of the composite machine tool itself; tool yield strength σ y is obtained by referring to materials;

[0054] s2: Construct a tool wear prediction model, as Figure 1 shown:

[0055] s2-1: Based on the Long Short-Term Memory network model LSTM, construct the input X i of the tool wear prediction model; The input X i of the tool wear prediction model includes two parts: input dynamic signals and process parameters:

[0056]

[0057] In the above formula, represents the dynamic signal, including cutting force (F c,t ), spindle current (I s,t ), vibration signal (V t ), cutting temperature T c ; represents the process parameters, including cutting speed v c , feed rate v f , yield strength σ y ; t represents the time step of a certain signal acquisition, and T represents the total number of signal acquisitions;

[0058] S2-2: Construct the tool wear prediction loss function L enhanced , L enhanced = L LSTM + λ·L physics ; In the above formula, L physics is the physical model loss function, λ is the physical loss weight coefficient, used to adjust the weight of the physical constraint loss; L LSTM is the supervised loss of the long short-term memory network model LSTM;

[0059] The construction method of L physics in step S2-2 is as follows:

[0060] Construct the chip force model: F c = K c ·a p ·f z ; Construct the temperature rise model:

[0061] In the above formula, F c is the cutting force, K c is the material cutting force coefficient, a p is the cutting depth, f z is the feed per tooth; T c is the rising temperature, v c is the cutting speed, η c is the thermal efficiency, A c is the cutting contact area, k is the material thermal conductivity;

[0062] Take the deviation between the model calculation result and the actual acquisition value as one of the loss functions of the tool wear prediction model, and add it to the optimization objective in the form of a regularization term, so that the LSTM model can learn in the direction that conforms to the physical law;

[0063] Construct the cutting force model loss function:

[0064] Construct the temperature rise constraint loss function:

[0065] In the above formula, is the cutting force obtained from actual acquisition, is the cutting force obtained from model calculation; is the temperature rise obtained from actual acquisition, is the temperature rise obtained from model calculation;

[0066] Define the loss function L of the physical model physics as follows:

[0067] L physics = λ1L force + λ2L temperature ;

[0068] In the above formula, λ1 and λ2 are weight factors used to balance the importance of different physical constraints;

[0069] s2-3: Construct the tool wear state vector h t , and through training the long short-term memory network model LSTM, obtain the tool wear prediction value:

[0070] In the above formula, represents the tool wear prediction value at the current time step t (i.e., at time t), h t is the tool wear state vector, W y is the weight matrix of the output layer, b y is the bias vector; during the training process of the long short-term memory network model LSTM, according to the tool wear prediction loss function L enhanced established in s2-2, the training objective is to minimize the tool wear prediction loss function L enhanced , and obtain the initial W y and b y through training;

[0071] The construction method of the tool wear state vector h t is as follows: in the LSTM model, use the dynamic signal and process parameters as independent long short-term memory network channels to extract the time series representation, which is specifically defined as follows:

[0072]

[0073] In the above formula, is the time series representation extracted using the dynamic signal as the long short-term memory network channel, is the time series representation extracted using the process parameters as the long short-term memory network channel; fuse the outputs of the two channels in the hidden layer to form the representation of the tool wear state vector h t :

[0074] During the training process, according to the established tool wear prediction loss function L enhanced , the model not only minimizes the supervised loss L between the tool wear prediction value LSTM and the true label y of tool wear, but also combines the established physical model loss function L physics to perform physical consistency verification on the intermediate features output by the model. The entire training process optimizes the parameters through backpropagation and gradient descent algorithms, and the training objective is to minimize the tool wear prediction loss function L enhanced ;

[0075] Specifically, in each forward propagation, the long short-term memory network performs time series modeling on the dynamic signal and process parameters respectively, outputs the fused tool state feature vector, and obtains the wear prediction value at the current time step t through the fully connected layer; subsequently, the system calculates the error between the tool wear prediction value and the true label of tool wear, and constructs a physical consistency loss based on the deviation between the physical model prediction value and the measured data; the backpropagation process relies on the chain rule to calculate the derivative of the tool wear prediction loss function with respect to all trainable parameters in the long short-term memory network model, and calculates the gradient of each parameter; in order to ensure that the learning direction conforms to the optimization objective, a gradient descent algorithm is introduced, and each parameter is updated along the negative direction of the gradient:

[0076]

[0077] In the above formula, η is the learning rate, and θ lstm is the set of all learnable parameters in the LSTM model, including W y and b y ;

[0078] S3: Multi-task training of the meta-learning framework. Through meta-learning, the updated W y and b y are obtained, as shown in Figure 2 :

[0079] S3-1: Divide the samples into multiple tasks T i . Within each task, the data is divided into a training subset and a validation subset

[0080]

[0081] In the above formula, represents the training set of the i-th task in the LTSM model, represents the validation set of the i-th task in the LTSM model;

[0082] s3-2: Randomly obtain the shared initialization parameter θ of meta-learning, and update it through the inner loop of meta-learning to obtain the personalized model parameter θ i ′; The inner loop update is as follows:

[0083]

[0084] In the above formula, α is the inner loop learning rate, which is used for in-task gradient update, is the model supervision loss of the training data on task T i ; is the physical model loss corresponding to task T i , and λ is the physical loss weight coefficient;

[0085] Based on the personalized model parameter θ′ i , update it through the outer loop of meta-learning to obtain the updated global shared parameter θ new ; Based on θ i ′, the outer loop update is as follows:

[0086]

[0087] In the above formula, β is the outer loop learning rate, which is used for global update; is the validation loss of task T i , which is used to guide the update of the global shared parameter θ; is the physical model loss, which ensures the generalization ability of θ in physical laws;

[0088] Among them, θ, θ i ′, θ new are all sets of all learnable parameters in the LSTM model, and both include the parameters W y and b y ; Through the outer loop, the updated global shared parameter θ new is obtained, and the updated W y and b y are obtained, and then the tool wear prediction value after meta-learning is obtained The tool wear prediction value function updated by meta-learning: Deploy it in the actual machining system of the composite machine tool, receive dynamic signals and process parameters in real time, and based on the updated model parameter θ lstm , quickly output the tool wear prediction value at the current moment

[0089] Example 2: On the basis of the meta-learning in Example 1, when the model encounters unseen new working condition tasks in the actual application scenario, update it in the following way:

[0090] s4: Quick fine-tuning and deployment of new working condition tasks:

[0091] s4-1: When the model encounters unseen new working condition tasks during actual application deployment, without changing the model structure, rapid fine-tuning is completed through gradient update; the global shared parameters θ learned in step s3 are used in the fine-tuning process. new And the following update is performed in combination with the small sample data of the new working condition task:

[0092]

[0093] In the above formula, θ′ new is the model parameter after rapid adaptation to the new working condition task. is the supervised loss of the new task training set. is the physical model loss of the new task. Finally, the fine-tuned model parameter θ′ new is the long short-term memory network model parameter θ actually deployed for prediction. lstm , that is, the parameters W y and b y under the new working condition task are obtained;

[0094] s4-2: The tool wear prediction value model After multi-task meta-learning and rapid fine-tuning of the new working condition task, it is deployed in the actual machining system of the composite machine tool, and receives dynamic signals and process parameters in real time. Based on the updated model parameter θ lstm , the tool wear prediction value at the current moment is quickly output Thereby realizing the monitoring and early warning of the tool state, with high precision, strong robustness and good working condition adaptability.

[0095] Such as Figure 3 shown is the comparison chart of the tool wear prediction value (i.e., the physical model prediction value) and the true label y (wear width experimental value) of the tool wear under four different working conditions.

[0096] Although specific embodiments of the present invention are described in detail herein, they are given for purposes of explanation only and should not be considered as limiting the scope of the present invention. Various substitutions, changes and modifications can be conceived without departing from the spirit and scope of the present invention.

Claims

1. A variable working condition tool wear prediction method based on a composite machine tool, characterized in that: The following steps are involved: s1: Collect dynamic signals and obtain process parameters: Collect dynamic signals: cutting force (F c,t ), spindle current (I s,t ) and vibration signal (V t ); Get process parameters: cutting speed v c , Feed speed: v f , cutting temperature T c , tool yield strength σ y ; s2: Construct tool wear prediction model: s2-1: Construct tool wear prediction model based on long short-term memory network model LSTM input x i ; s2-2: Construct tool wear prediction loss function L enhanced , L enhanced =L LSTM +λ·L physics ; In the above formula, L physics is the physical model loss function, λ is the physical loss weight coefficient, which is used to adjust the weight of the physical constraint loss; L LSTM It is the supervision loss of the long short-term memory network model LSTM; s2-3: Construct tool wear state vector h t , by training the long short-term memory network model LSTM, the tool wear prediction value is obtained: In the above formula, represents the tool wear prediction value at the current time step t, h t is the tool wear state vector, W y is the output layer weight matrix, b y is the bias vector; in the short-term memory network model LSTM training process, according to the tool wear prediction loss function L established in s2-2 enhanced The training goal is to minimize the tool wear prediction loss function L enhanced , the initial W is obtained through training y and b y ; s3: Multi-task training of meta-learning framework: Through meta-learning, we obtain the updated W y and b y .

2. The variable working condition tool wear prediction method based on a composite machine tool according to claim 1 is characterized in that the steps In s1, cutting force (F c,t ), spindle current (I s,t ), vibration signal (V t ), cutting temperature T c It is obtained by collecting dynamic data of the composite machine tool during the machining process through sensors; cutting speed v c , Feed speed: v f It is obtained through the communication interface of the compound machine tool; the tool yield strength σ y Obtained by consulting the information.

3. The variable working condition tool wear prediction method based on a composite machine tool according to claim 1 is characterized in that: Step s2-1, tool wear prediction model input X i It includes two parts: input dynamic signal and process parameters: In the above formula, Represents dynamic signals, including cutting force (F c,t ), spindle current (I s,t ), vibration signal (V t ), cutting temperature T c ; represents the process parameters, including cutting speed v c , feed speed v f , yield strength σ y ; t represents the time step of a certain signal acquisition, and T represents the total number of times the signal is acquired.

4. The variable working condition tool wear prediction method based on a composite machine tool according to claim 3 is characterized in that: L of step s2-2 physics The construction method is as follows: Constructing the chip force model: F c =K c ·a p ·f z ; Construct temperature rise model: In the above formula, F c is the cutting force, K c is the material cutting force coefficient, a p is the cutting depth, f z is the feed per tooth; T c is the temperature rise, v c is the cutting speed, η c is the thermal efficiency, A c is the cutting contact area, k is the thermal conductivity of the material; The deviation between the model calculation results and the actual collected values ​​is used as one of the loss functions of the tool wear prediction model and added to the optimization target in the form of a regularization term, so that the LSTM model can learn in a direction that conforms to the laws of physics. Construct cutting force model loss function: Construct the temperature rise constraint loss function: In the above formula, For the actual collected cutting force, Cutting forces calculated for the model; is the actual collected temperature rise, is the temperature rise calculated by the model; Define the physical model loss function L physics as follows: L physics =λ1L force +λ2L temperature ; In the above formula, λ1 and λ2 are weight factors used to balance the importance of different physical constraints.

5. The variable working condition tool wear prediction method based on a composite machine tool according to claim 4 is characterized in that: In step s2-3, the tool wear state vector h t The construction method is as follows: In the LSTM model, dynamic signals and process parameters are used as independent long-term and short-term memory network channels to extract time series representations, which are specifically defined as follows: In the above formula, To extract time series representation using dynamic signals as LSTM channels, To extract time series representation using process parameters as channels of long short-term memory networks; The outputs of the two channels are feature fused in the hidden layer to form the tool wear state vector h t Representation:

6. A variable working condition tool wear prediction method based on a composite machine tool according to claim 1 or 5, characterized in that: In step s2-3, the entire training process is optimized through back propagation and gradient descent algorithms. The tool wear prediction loss function is derived with respect to all trainable parameters in the long short-term memory network model, the gradient of each parameter is calculated, and each parameter is updated along the negative direction of the gradient: In the above formula, η is the learning rate, θ lstm is the set of all learnable parameters in the LSTM model, including W y and b y .

7. The variable working condition tool wear prediction method based on a composite machine tool according to claim 1 is characterized in that: The multi-task training method of the meta-learning framework in step s3 is as follows: s3-1: Divide the sample into multiple tasks T i , within each task, the data is divided into training subsets and validation subset D i val : In the above formula, represents the training set of the i-th task in the LTSM model, Represents the training set and validation set of the i-th task in the LTSM model; s3-2: Randomly obtain the shared initialization parameter θ of meta-learning, and update it through the inner loop of meta-learning to obtain the personalized model parameter θ′ for this task i ; Based on the personalized model parameter θ′ i , through the meta-learning outer loop update, we get the updated global shared parameter θ new ; where θ, θ′ i ,θ new are all sets of all learnable parameters in the LSTM model, including the parameter W y and b y .

8. The variable working condition tool wear prediction method based on a composite machine tool according to claim 7 is characterized in that: In step s3-2, the inner loop is updated as follows: In the above formula, α is the inner loop learning rate, which is used for the gradient update within the task. For task T i The model supervision loss on the training data, For task T i The corresponding physical model loss, λ is the physical loss weight coefficient; Based on θ′ i The outer loop is updated as follows: In the above formula, β is the outer loop learning rate, which is used for global update; For task T i The verification loss is used to guide the update of the global shared parameter θ; The physical model loss ensures the generalization ability of θ in physical laws. The updated global shared parameter θ is obtained through the outer loop new , get the updated W y and b y , and then obtain the tool wear prediction value after meta-learning 9. The variable working condition tool wear prediction method based on a composite machine tool according to claim 1 is characterized in that: The following steps are also included: s4: Rapid fine-tuning and deployment of new working conditions and tasks: s4-1: When the model is deployed in an actual application scenario and encounters a new task, it can be quickly fine-tuned through gradient update without changing the model structure. The fine-tuning process uses the global shared parameter θ learned in step s3. new , and perform the following updates based on the small sample data of the new working condition task: In the above formula, θ′ new Model parameters that are quickly adapted to new working conditions and tasks, is the supervision loss of the new task training set, is the physical model loss term of the new task, and finally the fine-tuned model parameter θ′ is obtained new That is, the parameter θ of the long short-term memory network model actually deployed for prediction lstm , that is, to obtain the parameter W under the new working condition task y and b y ; s4-2: Tool wear prediction model After multi-task meta-learning and rapid fine-tuning of new working conditions, it is deployed in the actual processing system of the composite machine tool, receiving dynamic signals and process parameters in real time, and based on the updated model parameters θ lstm , quickly output the tool wear prediction value at the current moment

Images