A tool wear state monitoring method, system, model training method and product based on meta learning

Abstract

This invention belongs to the field of tool wear condition monitoring, and relates to a tool wear condition monitoring method, system, model training method, and product based on meta-learning. The training method is based on the MSCNN-BiLSTM-Attention model. Since the tool wear value VB reflects the tool wear condition, the data is divided into three stages: initial wear, normal wear, and excessive wear. After preprocessing the signals, a model-independent meta-learning training strategy is introduced to simultaneously train the inner and outer loops. During training, by setting reasonable parameters and network structure, the problems of insufficient training sample quantity and uneven distribution are solved, improving the convergence speed of the algorithm network model and achieving accurate prediction of tool wear values.

Description

Technical Field

[0001] This invention belongs to the field of tool wear condition monitoring, and more specifically, relates to a tool wear condition monitoring method, system, model training method and product based on meta-learning. Background Technology

[0002] As key equipment in enterprise production, intelligent CNC machine tools connect product development to production, and their performance directly impacts the upgrading of the entire enterprise and even the entire industrial chain. With the continuous improvement of automation and intelligence in the machining process of intelligent CNC machine tools, equipment condition monitoring technology is becoming a research hotspot in the field.

[0003] In intelligent CNC machine tool systems, cutting tools, as the most direct and core actuators involved in raw material removal and part forming, play a decisive role in the product's machining accuracy and surface quality through their real-time operating status. However, as the cutting process progresses, tool wear is inevitable, leading to reduced product dimensional accuracy, increased surface roughness, and in severe cases, even chipping or breakage. This not only affects product quality but can also damage the intelligent CNC machine tool, resulting in unplanned downtime and significantly reduced production efficiency. Statistics show that downtime caused by tool wear accounts for approximately 20% of total downtime for intelligent CNC machine tools. Cutting tools are characterized by short service life, rapid wear, and frequent replacement. Therefore, accurate monitoring of tool wear can effectively ensure product quality, significantly improve the utilization rate of intelligent CNC machine tools, and reduce unexpected downtime losses. Existing research indicates that by introducing intelligent tool monitoring technology, manufacturing enterprises can save approximately 30% of processing costs.

[0004] In traditional tool wear monitoring, judgment mainly relies on human subjective experience. Due to the inconsistency of processing environment and processing conditions, the judgment of tool wear is subject to certain subjectivity and limitations, which affects the tool replacement time. The tool may be replaced before it reaches the wear standard, or it may be replaced only after the tool has been severely worn.

[0005] To effectively address the aforementioned issues, this invention utilizes various signals collected during the machining process of intelligent CNC machine tools. By employing various signal processing techniques and constructing an artificial intelligence meta-learning network, it delves into the relationship between various signals and the evolution of tool wear, thereby building a tool wear state monitoring model. This allows for accurate monitoring of the tool wear state, enabling precise intervention and tool replacement decisions before tool failure. Summary of the Invention

[0006] In view of the above-mentioned defects or improvement needs of the prior art, the present invention provides a tool wear condition monitoring method, system, model training method and product based on meta-learning. The purpose is to solve the technical problem of unreasonable tool change time caused by the subjectivity and limitations of the judgment of tool wear condition in the prior art.

[0007] To achieve the above objectives, according to one aspect of the present invention, a training method for a tool wear state monitoring model based on meta-learning is provided, which is trained based on the MSCNN-BiLSTM-Attention model, and includes the following steps: The existing tool wear data is used to collect tool signals and then divide them according to the tool wear stage: the tool wear value VB reflects the tool wear condition and is divided into three stages: initial wear, normal wear, and excessive wear; then, the signals are divided into initial wear stage signals, normal wear stage signals, and excessive wear stage signals according to the truth value label. The signals corresponding to the classified tool wear samples are preprocessed: based on the classified signals, invalid signals collected during the tool feed and retraction stages are removed, and each valid signal is truncated. Then, all data in the truncated valid signals are standardized, and the maximum value of the actual tool wear is normalized so that its value is between (0,1). A meta-learning training strategy is introduced: preprocessed signal data is divided into a support set and a query set, and input into the MSCNN-BiLSTM-Attention network model for simultaneous training of the inner and outer loops. The MSCNN-BiLSTM-Attention network model sequentially includes an input layer, a multi-scale convolutional layer, a BiLSTM layer, an attention mechanism layer, a fully connected layer, and an output layer. The input layer is used to input data from the support set. The multi-scale convolutional layer extracts multi-scale local features from the input signal and performs feature fusion. The BiLSTM layer extracts temporal features from the fused multi-scale local features. The attention mechanism layer further filters key features from the extracted temporal features. The fully connected layer converts the filtered key features into a one-dimensional vector. Finally, the output layer performs weighted calculations on the one-dimensional vector to obtain the predicted tool wear value and outputs it. The inner loop consists of three loops, each corresponding to the training task of the support set for the three wear stages of the tool. In each inner loop, the support set first performs a gradient update on the model parameters, resulting in updated parameters. Then, based on the corresponding query set, the updated parameters are used to calculate the loss, and the gradient of the loss on the current model parameters is obtained. After all inner loops are completed, the losses of all inner loops are summed to obtain the outer loop loss value. This outer loop loss value is used as the input to the outer loop to perform gradient descent updates on the initial parameters of the MSCNN-BiLSTM-Attention model. The updated, independent parameters are then used as the input to the next inner loop, thus completing one round of training. The inner and outer loop processes are then repeated until a preset termination condition is met, resulting in the tool wear state monitoring model.

[0008] Furthermore, force signals in the X direction from three tools were selected from the existing tool wear dataset. Force signal in the Y direction Force signal in the Z direction Vibration signal in the X direction Vibration signal in the Y direction Vibration signal in the Z direction A total of 6 signals are used as input signals, which are divided according to the tool wear stage.

[0009] Furthermore, the data corresponding to two of the three tools are used as the training set, and the data of the remaining tool is used as the test set. The training set uses the classified data, and the test set uses the original data. The data in the training set is divided into three parts: initial wear, normal wear, and excessive wear, corresponding to the three inner loop training cycles of meta-learning, and a support set and a query set are also defined. Equal amounts of samples are selected from the signal data of the test set for each of the three stages, which are used as the support set and input into the network model for fine-tuning. The remaining data is used as the query set.

[0010] Furthermore, all truth labels are set to the maximum wear value.

[0011] Furthermore, the formula for standardization is as follows:

[0012]

[0013]

[0014] in, Let be the amplitude of the i-th sampling point. u The average value of the amplitude at all sampling points. s Standard deviation, NLet i be the number of sampling points, i = 1 to N. for The calculation results after standardization.

[0015] Furthermore, the formula for normalizing the actual maximum value of tool wear is as follows:

[0016] in, Let j be the normalized value of the actual wear of the j-th sample. Let j be the actual wear value of the j-th sample. The minimum value of the current milling cutter's true value. This represents the maximum value of the current milling cutter's true value.

[0017] According to another aspect of the present invention, a tool wear condition monitoring model product based on meta-learning is provided, characterized in that it is trained using the training method described in any of the preceding claims.

[0018] According to another aspect of the present invention, a tool wear condition monitoring method based on meta-learning is provided, which uses the tool wear condition monitoring model product as described above to monitor tool wear condition.

[0019] According to another aspect of the present invention, a computer-readable storage medium is provided that stores the tool wear condition monitoring model product based on meta-learning as described above.

[0020] According to another aspect of the present invention, a tool wear condition monitoring system is provided, including a memory, a processor, and a computer program stored in the memory, wherein the computer program is the tool wear condition monitoring model product based on meta-learning as described above.

[0021] In summary, the technical solutions conceived in this invention, compared with the prior art, can achieve the following beneficial effects: This invention constructs a network model for predicting tool wear values. Addressing the limitations of traditional prediction models in terms of limited accuracy and low utilization of acquired signals, this paper proposes an improved model based on the fundamental algorithm network model, effectively enhancing both prediction accuracy and signal utilization. To address the issues of small training data volume and uneven data distribution, this invention introduces a meta-learning training method and integrates it with the improved algorithm network model, proposing a tool wear state monitoring network model under small sample conditions. This model effectively improves prediction accuracy, reduces the risk of overfitting, and significantly shortens the training time. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the principle framework of a preferred embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the variation trend of the maximum wear value of the milling cutter and the division of tool wear stages according to a preferred embodiment of the present invention; Figure 3 This is the original signal of the milling cutter in a preferred embodiment of the present invention; Figure 4 Is Figure 3 The signals after the tool infeed and retraction phases are removed from the basic structure; Figure 5 Is Figure 4 The signal segment after data truncation based on the data; Figure 6 Yes Figure 5 The signal segment after standardization; Figure 7 This is a schematic diagram of the basic framework of the algorithm network model of a preferred embodiment of the present invention; Figure 8 This is a schematic diagram of the algorithm network model training process according to a preferred embodiment of the present invention; Figure 9 This is the predicted wear value of the milling cutter according to a preferred embodiment of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0024] Figure 1 This is a flowchart of a tool wear state monitoring method based on meta-learning according to the present invention, which is specifically carried out according to the following steps: The composition of the PHM2010 public dataset was analyzed. This dataset used a total of six end mills, numbered C1, C2, C3, C4, C5, and C6, recording the entire lifecycle of each end mill under the same working conditions. Each end mill underwent a total of 315 machining operations. During the machining process, signals from seven channels were collected, including force signals in the X direction. Force signal in the Y direction Force signal in the Z direction Vibration signal in the X direction Vibration signal in the Y direction Vibration signal in the Z direction The root mean square value AE of the acoustic emission signal is also included. After each machining operation, the wear values ​​(flute1, flute2, and flute3) of the three cutting edges of the milling cutter are measured using a wear measuring instrument, and these wear values ​​are used as true value labels. Among them, the three milling cutters numbered C1, C4, and C6 have wear values. To more accurately reflect the wear degree of the milling cutters, this invention selects the force signal in the X direction of the three milling cutters numbered C1, C4, and C6. Force signal in the Y direction Force signal in the Z direction Vibration signal in the X direction Vibration signal in the Y direction Vibration signal in the Z direction A total of 6 signals are used as input signals.

[0025] The tool wear stages are divided. Six signals from the first step are selected, and the tool wear value VB reflects the tool wear condition, which is then divided into three stages: initial wear, normal wear, and excessive wear. For example... Figure 2 As shown, the wear process of the C1 milling cutter can be divided into three stages: initial wear, normal wear, and excessive wear, which conforms to the general law of tool wear evolution. The remaining two milling cutters, C4 and C6, are classified in the same way, and the truth labels of C1, C4, and C6 are all set to the maximum wear value. According to the tool wear value, a total of 945 samples from the three milling cutters are divided into three categories. The tool wear process is classified according to the slope of the wear curve. Generally, the slope of the wear curve is larger in the initial wear stage and the excessive wear stage, while in the normal wear stage, the slope of the wear curve is moderate and relatively stable. In a preferred embodiment, the specific classification criteria adopted by the present invention are as follows: wear <85μm is considered initial wear, 85-125μm is considered normal wear, and >125μm is considered excessive wear. After classification, the number of initial wear samples for C1 tools is 20, the number of normal wear samples is 183, and the number of excessive wear samples is 112; the number of initial wear samples for C4 tools is 79, the number of normal wear samples is 166, and the number of excessive wear samples is 70; the number of initial wear samples for C6 tools is 22, the number of normal wear samples is 147, and the number of excessive wear samples is 146, totaling 121 samples in the initial wear stage, 496 samples in the normal wear stage, and 328 samples in the excessive wear stage. The input signals corresponding to the classified tool wear samples are preprocessed. For example... Figure 3As shown, there are obvious signal changes at the beginning and end, and the signal change pattern does not conform to the steady-state cutting signal change pattern, indicating that the tool is in the approach or retraction phase. These signals interfere with the normal training of the model; therefore, it is necessary to remove the signals from the approach and retraction phases to retain the steady-state cutting signal. In a preferred embodiment, the present invention uses the method of deleting the first 2% and the last 2% of the signal. The result after removal is as follows: Figure 4 As shown, the processed signal changes are more stable and can more clearly reflect the wear of the tool during machining. Furthermore, since more than 200,000 signal points are collected from a single channel in a single cut, inputting all of these signal points into the algorithm network would result in a very long network training time. Therefore, in this embodiment, the signal is truncated to shorten the signal length and improve the model training rate. Preferably, the truncated input signal length is 4000. The result after the stage is as follows... Figure 5 As shown, the model training speed is effectively improved after the truncation operation. Meanwhile, to accelerate algorithm model convergence and avoid gradient explosion, this invention performs standardization on all truncated data. The standardization formula is shown below:

[0026]

[0027]

[0028] in, Let be the amplitude of the i-th sampling point. u The average value of the amplitude at all sampling points. s Standard deviation, N Let i be the number of sampling points, i = 1 to N. for The calculation results after standardization. The signal after standardization is as follows: Figure 6 As shown. The data size obtained at this time is 3×315×6×4000.

[0029] The actual wear value of each tool is normalized to a value between (0, 1). The formulas for normalizing the actual wear of the three tools are shown below:

[0030] in, Let j be the normalized value of the actual wear of the j-th sample. Let j be the actual wear value of the j-th sample. The minimum value of the current milling cutter's true value. This represents the maximum true value of the current milling cutter, j = 1~315.

[0031] A tool wear condition monitoring network model based on MSCNN-BiLSTM-Attention was constructed. This model mainly consists of a multi-scale convolutional neural network, a bidirectional long short-term memory network, and a multi-head attention mechanism. The model structure is as follows: Figure 7 As shown, the tool wear monitoring network model specifically comprises an input layer, a multi-scale convolutional layer, a bidirectional long short-term memory (BiLSTM) network layer, an attention mechanism layer, a fully connected layer, and an output layer. Specifically, the multi-scale convolutional layer extracts multi-scale local features from the input signal and performs feature fusion; the BiLSTM network layer extracts temporal features from the fused multi-scale local features; the attention mechanism layer employs a multi-head attention mechanism to further filter key features from the extracted temporal features; the fully connected layer converts the filtered key features into a one-dimensional vector; and finally, the output layer performs weighted calculations on the one-dimensional vector to obtain the predicted tool wear value and outputs it.

[0032] Preferably, in this embodiment, the input layer is the preprocessed signal data of each sample (i.e., all of them). The network model has a size of 3×315×6×4000; the multi-scale CNN convolutional layer has 3 channels, each performing 5 convolutions with kernel sizes of 4, 6, and 8 respectively; a max pooling layer is used to compress the feature dimension after each convolution calculation, with a kernel size of 2; and a batch normalization layer is added after each max pooling calculation; there are 3 BiLSTM layers; 4 multi-head attention heads; and 128 and 32 fully connected layer nodes respectively; the output layer outputs the final tool wear prediction value. To reduce the risk of overfitting during network training, the proposed network model incorporates a Dropout strategy in the BiLSTM layers, multi-head attention mechanism, and fully connected layers, preferably with a Dropout value of 0.3. The Leaky ReLU activation function is used, preferably with a slope of 0.1 in the negative region. Furthermore, the Adam optimizer is used as the model optimizer, and the mean squared error loss function is used for calculation. During model training, issues such as model instability and overfitting are addressed by modifying certain parameters of the network model, such as the number of convolutional kernels, padding, and stride. Furthermore, dynamic adjustments to parameters like training batches, learning rate, and iteration count improve training efficiency and prediction accuracy. By building this initial basic model, the goal of monitoring tool wear can be effectively achieved.

[0033] Introduce a model-independent meta-learning training strategy. For example... Figure 8The diagram shows the flowchart of the meta-learning training method. Its core is to divide a complete task into multiple sub-tasks, each of which can be defined as a different type of specific problem. Simultaneously, the original data needs to be divided into a support set and a query set. For example, in this invention, since the tool has three wear stages, and the number of samples for each wear stage is not consistent, in traditional meta-learning model training, three wear stages constitute a complete training sample. This makes it difficult to use traditional methods for model training in this invention. Therefore, this invention needs to introduce a model-independent meta-learning training strategy to address this obvious sample imbalance problem. In this invention, a complete task is divided into three sub-tasks, corresponding to the training samples of the three wear stages of the tool. For ease of demonstration, Figure 8 Only two examples of subtask models are given, but it should be understood that all subtask models are processed in the same way.

[0034] The support set is a small collection of samples used by the model to "learn" relevant content for the current subtask. By providing specific contextual information, it allows the model to quickly learn new concepts. The query set is a collection of test samples used to test the model's performance, reflecting its predictive ability after rapid learning. The data distribution of the support set and query set in each subtask can be the same or different. During the completion of each subtask, the meta-learning training method uses a shared base model (in this invention, the tool wear state monitoring network model built in step four) as a foundation, setting different parameters for different tasks to adjust and build the optimal model for completing the current task. After all subtasks are completed, the learning results are summarized, merged, and fed back to the base model, thereby updating the parameters and improving the performance of the overall model.

[0035] The goal of meta-learning training is to find a set of initial parameters that allow the model to achieve optimal performance for any new task with only a small number of gradient updates, thus meeting the task requirements. During meta-learning training, each update iteration is divided into inner and outer loops. In the inner loop, each subtask model is assigned independent parameters, which are initialized to the meta-network parameters at the start of the current inner loop. At the start of the inner loop, the support set for that subtask performs a gradient update on the model parameters, resulting in the updated parameters; subsequently, the query set for that subtask uses the updated parameters to calculate the loss, obtaining the gradient of the loss with respect to the current parameters. After all subtask inner loops are completed, all inner loop losses are summed, and this sum is used as the input to the outer loop. The formula for calculating the outer loop loss is shown below: LOSS 外 =

[0036] Among them, LOSS 外The outer loop loss value, LOSS k Let N be the loss for the k-th sub-loop, and N be the number of sub-loops. k =1~3.

[0037] In the outer loop, the initial parameters of the meta-network are updated using gradient descent based on the inner loop loss. The updated, independent parameters are then used as inputs for the next inner loop, thus completing one round of training. The meta-learning training process ends when the set number of loops is reached or the model accuracy meets the predetermined requirements. At this point, the model performance can be validated by inputting a test set.

[0038] The effectiveness of this invention will be illustrated through the following test experiments: The classified data is used as input to test the model's performance. Two of the three milling cutters (C1, C4, and C6) are used as the training set, and the remaining cutter is used as the test set. The two milling cutters used for training use the classified data, while the test cutter uses the original data. The training data is divided into three parts: initial wear, normal wear, and excessive wear. The training completes the three sub-tasks of meta-learning and divides the data into support and query sets. Simultaneously, 10 samples from each of the three stages in the test set signal data, totaling 30 samples, are used as the support set input to the network model for fine-tuning. The remaining data serves as the query set.

[0039] Based on the experimental data grouping above, this test experiment uses C4 and C6 as the training set and C1 as the test set.

[0040] like Figure 9 The figure shows the predicted maximum wear value curve for the C1 milling cutter in the test set. Analysis of the experimental results shows that the predicted wear value curve for the C1 milling cutter matches the actual value curve quite well. To further evaluate the effectiveness of the algorithm network model, this paper uses the root mean square error (RMSE) as the standard for evaluating the model's prediction accuracy. RMSE is a commonly used evaluation index in prediction regression problems, used to calculate the absolute value of the difference between the predicted and actual values. It can calculate the magnitude of the prediction error, thus intuitively reflecting the overall prediction accuracy of the network model within a complete machining cycle. The RMSE calculation formula is shown below:

[0041] in, N Number of cuts For the first i Predicted tool wear condition for the next pass For the first i The actual wear condition of the tool during the next pass. Substituting the 315 wear prediction values ​​from the algorithm network model into the formula, the calculated RMSE is 6.561.

[0042] The model not only accurately predicts normal wear samples, but also shows minimal deviation between predicted and actual values ​​for the initial and excessive wear stages of the C1 end mill, effectively ensuring prediction accuracy across all stages. Therefore, this meta-learning-based tool wear state monitoring method can accurately monitor tool wear conditions.

[0043] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A training method for a tool wear state monitoring model based on meta-learning, trained using an MSCNN-BiLSTM-Attention model, characterized in that... Includes the following steps: The existing tool wear data is used to collect tool signals and then divide them according to the tool wear stage: the tool wear value VB reflects the tool wear condition and is divided into three stages: initial wear, normal wear, and excessive wear; then, the signals are divided into initial wear stage signals, normal wear stage signals, and excessive wear stage signals according to the truth value label. The signals corresponding to the classified tool wear samples are preprocessed: based on the classified signals, invalid signals collected during the tool feed and retraction stages are removed, and each valid signal is truncated. Then, all data in the truncated valid signals are standardized, and the maximum value of the actual tool wear is normalized so that its value is between (0,1). A meta-learning training strategy is introduced: preprocessed signal data is divided into a support set and a query set, which are then input into the MSCNN-BiLSTM-Attention network model for simultaneous training of the inner and outer loops. The MSCNN-BiLSTM-Attention network model sequentially includes an input layer, a multi-scale convolutional layer, a BiLSTM layer, an attention mechanism layer, a fully connected layer, and an output layer. The input layer is used to input data from the support set. The multi-scale convolutional layer extracts multi-scale local features from the input signal and performs feature fusion. The BiLSTM layer extracts temporal features from the fused multi-scale local features. The attention mechanism layer further filters key features from the extracted temporal features; the fully connected layer converts the filtered key features into a one-dimensional vector; finally, the output layer performs weighted calculations on the one-dimensional vector to obtain the tool wear prediction value and outputs it. The inner loop consists of three loops, each corresponding to the training task of the support set for the three wear stages of the tool. In each inner loop, the support set first performs a gradient update on the model parameters, resulting in updated parameters. Then, based on the corresponding query set, the updated parameters are used to calculate the loss, and the gradient of the loss on the current model parameters is obtained. After all inner loops are completed, the losses of all inner loops are summed to obtain the outer loop loss value. This outer loop loss value is used as the input to the outer loop to perform gradient descent updates on the initial parameters of the MSCNN-BiLSTM-Attention model. The updated, independent parameters are then used as the input to the next inner loop, thus completing one round of training. The inner and outer loop processes are then repeated until a preset termination condition is met, resulting in the tool wear state monitoring model.

2. The training method for a tool wear state monitoring model based on meta-learning as described in claim 1, characterized in that, Select the force signals in the X direction from three existing tool wear datasets. Force signal in the Y direction Force signal in the Z direction Vibration signal in the X direction Vibration signal in the Y direction Vibration signal in the Z direction A total of 6 signals are used as input signals, which are divided according to the tool wear stage.

3. The training method for a tool wear state monitoring model based on meta-learning as described in claim 2, characterized in that, The data from two of the three tools are used as the training set, and the data from the remaining tool is used as the test set. The training set uses the classified data, and the test set uses the original data. The training set data is divided into three parts: initial wear, normal wear, and excessive wear. The training sets are then divided into three inner loops for meta-learning training, and support and query sets are also defined. Equal amounts of samples from the three stages of the signal data in the test set are selected as the support set and input into the network model for fine-tuning. The remaining data is used as the query set.

4. The training method for a tool wear state monitoring model based on meta-learning as described in claim 1, characterized in that, All truth labels are set to the maximum wear value.

5. The training method for a tool wear state monitoring model based on meta-learning as described in claim 1, characterized in that, The formula for standardization is as follows: in, Let be the amplitude of the i-th sampling point. u The average value of the amplitude at all sampling points. s Standard deviation N Let i be the number of sampling points, i = 1 to N. for The calculation results after standardization.

6. The training method for a tool wear state monitoring model based on meta-learning as described in claim 1, characterized in that, The formula for normalizing the actual maximum value of tool wear is as follows: in, Let j be the normalized value of the actual wear of the j-th sample. Let j be the actual wear value of the j-th sample. The minimum value of the current milling cutter's true value. This represents the maximum value of the current milling cutter's true value.

7. A tool wear condition monitoring model product based on meta-learning, characterized in that, It is obtained by training using the training method described in any one of claims 1 to 6.

8. A tool wear state monitoring method based on meta-learning, characterized in that, The tool wear condition monitoring model product according to claim 7 is used to monitor tool wear condition.

9. A computer-readable storage medium, characterized in that, It stores the tool wear condition monitoring model product based on meta-learning as described in claim 7.

10. A tool wear condition monitoring system, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The computer program is the tool wear condition monitoring model product based on meta-learning as described in claim 7.

Images