Machine tool remaining service life prediction method based on conflict-free multi-task learning

Abstract

The invention discloses a machine tool cutter remaining service life prediction method based on conflict-free multi-task learning, and the method comprises the steps: firstly, obtaining an average similarity curve through feature extraction, earth movement distance and correlation coefficient; calibrating the health state of the numerical control machine tool cutter based on the average similarity curve; secondly, setting a main task and an auxiliary task for predicting the residual service life of the cutter of the numerical control machine tool, and constructing a conflict-free multi-task learning model CMLM; secondly, dividing a data set, inputting the data set into the CMLM, and training the CMLM by using a conflict avoidance gradient descent algorithm; and finally, after the training of the CMLM is completed, inputting a cutter state monitoring signal of the numerical control machine tool into the trained CMLM to obtain a residual service life prediction result.

Description

Technical Field

[0001] This invention relates to the field of predicting the remaining service life of CNC machine tool tools, and more particularly to a method for predicting the remaining service life of machine tool tools based on conflict-free multi-task learning. Background Technology

[0002] In CNC machine tool machining, accurately predicting the remaining service life of cutting tools is crucial. This not only allows for the prediction of tool failure points, effectively preventing decreased machining accuracy and batch workpiece scrap, ensuring product quality and process stability, but also significantly reduces unplanned downtime, equipment damage, and safety accidents caused by sudden tool breakage, thereby directly lowering production and operating costs and maintenance risks. Simultaneously, this prediction supports the optimization of scientific tool replacement strategies, helps control resource waste caused by excessive tool stockpiling, and extends the effective service life of tools, forming the cornerstone of achieving efficient, safe, and low-cost intelligent manufacturing.

[0003] To achieve the above goals, predictive and health management technologies primarily rely on two core tasks: health status assessment and remaining useful life prediction. Health status assessment, through real-time monitoring of key parameters, promptly identifies equipment anomalies and early signs of failure, providing a basis for maintenance decisions and reducing unexpected failures and unplanned downtime. Remaining useful life prediction aims to quantify the time or work cycles required from the current state to failure, providing key inputs for predictive maintenance planning and spare parts strategy optimization, further contributing to equipment lifecycle management, maintenance cost control, and operational efficiency improvement. Currently, data-driven methods are the mainstream approach to solving these two tasks. However, modeling health status assessment and remaining useful life prediction separately may not only lead to the model failing to fully capture the dynamic correlation of changes in equipment health status and increasing the risk of prediction error accumulation, but also result in a complex overall model structure and reduced computational efficiency.

[0004] While conventional multi-task learning architectures achieve initial joint learning of health status assessment and remaining useful life prediction tasks by sharing underlying feature networks, significant conflicts and even contradictory directions (negative gradient interference) often arise during backpropagation due to the fundamental differences between the objectives of health status assessment (discrete classification task) and remaining useful life prediction (continuous regression task). Existing multi-task learning optimization strategies cannot effectively coordinate such gradient conflicts between tasks, leading to difficulties in model training convergence and decreased generalization performance. Furthermore, existing techniques fail to fully utilize the temporal degradation evolution characteristics of CNC machine tool vibration signals, limiting the early accuracy of remaining useful life prediction. Therefore, there is an urgent need to develop a multi-task optimization method that can dynamically avoid task gradient conflicts and deeply mine temporal degradation features to improve the performance of CNC machine tool remaining useful life prediction. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention discloses a method for predicting the remaining service life of machine tool tools based on conflict-free multi-task learning. It combines time and frequency domain features, Earth's movement distance, and correlation coefficients to determine and calibrate the health status of CNC machine tool tools. A dual-branch network is constructed with remaining service life prediction as the primary task and health status assessment as the secondary task. Furthermore, a conflict-avoiding gradient descent algorithm is used to resolve gradient conflicts between different tasks, thereby improving the performance of the neural network in predicting the remaining service life of CNC machine tool tools.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A method for predicting the remaining service life of machine tool cutting tools based on conflict-free multi-task learning is characterized by the following steps:

[0008] Step S1: Obtain the average similarity curve through feature extraction, Earth movement distance, and correlation coefficient;

[0009] Step S2: Calibrate the health status of CNC machine tool cutting tools based on the average similarity curve;

[0010] Step S3: Set the main task and auxiliary task for predicting the remaining service life of CNC machine tool tools, and construct a conflict-free multi-task learning model (CMLM).

[0011] Step S4: Divide the dataset and input it into the conflict-free multi-task learning model CMLM. Use the conflict-avoidance gradient descent algorithm to train the conflict-free multi-task learning model CMLM.

[0012] Step S5: After completing the training of the conflict-free multi-task learning model CMLM, input the CNC machine tool tool condition monitoring signal into the trained conflict-free multi-task learning model CMLM to obtain the remaining service life prediction result.

[0013] As a preferred technical solution of the present invention, step S1 is specifically as follows:

[0014] S11. The state monitoring signal measured by the accelerometer is converted into a frequency domain signal by fast Fourier transform, and the common trend features of each tool of the CNC machine tool are extracted using time and frequency domain feature extraction methods. The formula for the common trend features of each tool is as follows:

[0015] entropy: ;

[0016] Interquartile range: IQR = Q3 - Q1;

[0017] Maximum value: Max = max(x) i );

[0018] Minimum value: Min = min(x)i );

[0019] Mean: ;

[0020] Standard deviation: ;

[0021] Root mean square: ;

[0022] Skewness: ;

[0023] Kuroshi: ;

[0024] Sum of absolute values: ;

[0025] Waveform factor: ;

[0026] Peak factor: ;

[0027] Pulse factor: ;

[0028] Frequency domain centroid: ;

[0029] Mean square frequency: ;

[0030] Frequency variance: ;

[0031] Spectral entropy: ;

[0032] Among them, x i Let Q1 and Q3 represent the time-domain amplitude of the i-th sampling point, respectively, and let T represent the first and third quantiles. , and Let M represent the power spectral amplitude, actual frequency value, and normalized probability distribution of the power spectral amplitude of the k-th frequency component, respectively, and M represent the total number of frequency components.

[0033] S12. Combine the 17 common trend features obtained in step S11 into a state vector at each time step. Use the Earth's movement distance to measure the similarity of feature distributions between pairs of time steps, thus forming a feature distribution similarity. The matrix W is given by the following formula:

[0034] ;

[0035] Among them, P i and P jRepresents the state vector distribution at time i and time j. P represents i and P j All possible joint probability distributions γ Indicates from One of the samples, This represents the maximum expected distance in the sample.

[0036] S13. Based on the feature distribution similarity matrix W, calculate the data a in the i-th row and j-th column using the Pearson correlation coefficient. i and b j The correlation coefficients between them form a correlation coefficient matrix ρ:

[0037] ;

[0038] S14. Based on the correlation coefficient matrix ρ, calculate the average correlation coefficient between two adjacent time points by sliding along its main diagonal to obtain the average similarity curve ASC:

[0039] .

[0040] As a preferred technical solution of the present invention, step S2 is characterized by the following:

[0041] The moment when the minimum value on the average similarity curve (ASC) is located is marked as the critical point between normal and abnormal CNC machine tool health status. CNC machine tool health status before this moment is marked as "normal", and CNC machine tool health status after this moment is marked as "abnormal".

[0042] As a preferred technical solution of the present invention, step S3 is as follows:

[0043] S31. The main task is to predict the remaining service life of CNC machine tool tools to learn the long-term degradation characteristics of CNC machine tool tools; the secondary task is to assess the health status of CNC machine tool tools to learn the current health information of CNC machine tool tools and promote the learning of the main task.

[0044] S32. Based on the main task and the auxiliary task, construct a dual-branch network. The dual-branch network includes two branches, namely branch one and branch two. Branch one and branch two share a common feature extractor. Branch one consists of a remaining useful life prediction feature extractor and a predictor, which is used to complete the remaining useful life prediction task. Branch two consists of a health status assessment feature extractor and a classifier, which is used to complete the health status assessment task.

[0045] As a preferred technical solution of the present invention: in step S32...

[0046] Dual-branch network: Branch 1 and Branch 2 share a common feature extractor to learn common knowledge for prediction and classification tasks and reduce the number of parameters; Branch 1 and Branch 2 each have a feature extractor for remaining lifetime prediction and a feature extractor for health status assessment, which are used to learn task-specific long-term degradation features and local features, respectively.

[0047] Shared feature extractor: consists of four dilated causal convolutional layers, each accompanied by a batch normalization layer, an ELU activation function, a dropout layer and a pooling layer;

[0048] The remaining lifetime prediction feature extractor consists of a multi-head self-attention mechanism, two batch normalization layers, two dropout layers, two fully connected layers, and a SwiGLU activation function.

[0049] The health status assessment feature extractor consists of an average pooling layer, a convolutional layer, a group normalization layer, and a sigmoid activation function.

[0050] As a preferred technical solution of the present invention: in step S32...

[0051] The predictor in branch one and the classifier in branch two each consist of two fully connected layers, and each of their first fully connected layers is accompanied by an ELU activation function and a Dropout layer.

[0052] As a preferred technical solution of the present invention, step S4 is characterized by the following:

[0053] S41. Use the leave-one-out method to set up the training set and the test set, and use each CNC machine tool tool as the test set and the remaining CNC machine tool tools as the training set.

[0054] S42. Each training set data is a state monitoring signal measured by an accelerometer containing two labels. One label is the actual remaining service life, and the other is the calibrated health status of the CNC machine tool. Input the state monitoring signal measured by each accelerometer and its actual remaining service life into branch one, and input the state monitoring signal measured by each accelerometer and its calibrated health status of the CNC machine tool into branch two.

[0055] S43. Optimize the parameters of the conflict-free multi-task learning model CMLM using the conflict-avoiding gradient descent algorithm. For the gradient θ, the conflict-avoiding gradient descent algorithm aims to find a vector d to update θ.

[0056] As a preferred technical solution of the present invention: in step S43...

[0057] The conflict-avoidance gradient descent algorithm performs a first-order Taylor expansion of the loss function for the q-th task at θ:

[0058] ;

[0059] in, It is a loss L q The gradient at θ, where α represents the learning rate. It represents infinitesimal terms, and T is an operator that represents transpose.

[0060] Let R(θ, d) be used as a metric to measure gradient conflict between different tasks:

[0061] ;

[0062] Based on the analysis of R(θ, d), the vector d can be determined to update θ and avoid gradient conflicts by solving the following optimization problem:

[0063] ;

[0064] Where c=0.5 is a hyperparameter controlling the convergence speed, m is the dimension of d, and the optimization problem can be solved with average gradient... Find an optimal update vector d in the sphere centered at the center to minimize the conflict loss as measured by R(θ, d).

[0065] As a preferred technical solution of the present invention, step S5 is as follows:

[0066] S51. Determine whether the conflict-free multi-task learning model CMLM has been trained. If it has been trained, save the conflict-free multi-task learning model CMLM with the minimum training set loss. If it has not been trained, repeat step S43.

[0067] S52. Input the test set into the saved conflict-free multi-task learning model CMLM to obtain the remaining useful life prediction results.

[0068] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0069] 1. The health status calibration method of this technical solution is based on traditional time and frequency domain feature extraction and statistical methods, which ensures the reliability of the proposed method.

[0070] 2. The dual-branch network of this technical solution allows two tasks to share some layers while also having their own specific layers, which reduces the number of network parameters and can extract task-specific features.

[0071] 3. This technical solution uses a conflict-avoidance gradient descent algorithm to collaboratively optimize two tasks, and uses health status assessment as a secondary task to improve the accuracy of remaining lifetime prediction as the primary task. Attached Figure Description

[0072] Figure 1 This is a flowchart illustrating the present invention;

[0073] Figure 2 This is the average similarity curve of the present invention;

[0074] Figure 3 This is a graph showing the predicted remaining service life of CNC machine tool cutting tools according to the present invention. Detailed Implementation

[0075] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0076] like Figure 1 As shown, the machine tool tool remaining service life prediction method based on conflict-free multi-task learning proposed in this invention includes the following steps:

[0077] Step S1: Obtain the average similarity curve through feature extraction, Earth movement distance, and correlation coefficient;

[0078] Specifically as follows:

[0079] S11. The state monitoring signal measured by the accelerometer is converted into a frequency domain signal by fast Fourier transform, and the common trend features of each CNC machine tool are extracted using time and frequency domain feature extraction methods.

[0080] Specifically, the process involves acquiring status monitoring signals by attaching an accelerometer to the tool holder of the CNC machine tool being tested, and then using time- and frequency-domain feature extraction methods to extract the common trend features of each CNC machine tool. The formula for the common trend features of each tool is as follows:

[0081] entropy: ;

[0082] Interquartile range: IQR = Q3 - Q1;

[0083] Maximum value: Max = max(x) i );

[0084] Minimum value: Min = min(x) i );

[0085] Mean: ;

[0086] Standard deviation: ;

[0087] Root mean square: ;

[0088] Skewness: ;

[0089] Kuroshi: ;

[0090] Sum of absolute values: ;

[0091] Waveform factor: ;

[0092] Peak factor: ;

[0093] Pulse factor: ;

[0094] Frequency domain centroid: ;

[0095] Mean square frequency: ;

[0096] Frequency variance: ;

[0097] Spectral entropy: ;

[0098] Among them, x i Let Q1 and Q3 represent the time-domain amplitude of the i-th sampling point, respectively, and let T represent the first and third quantiles. , and Let M represent the power spectral amplitude, actual frequency value, and normalized probability distribution of the power spectral amplitude of the k-th frequency component, respectively, and M represent the total number of frequency components.

[0099] S12. Combine the 17 common trend features obtained in step S11 into a state vector at each time step. Use the Earth's movement distance to measure the similarity of feature distributions between pairs of time steps, thus forming a feature distribution similarity. The matrix W is given by the following formula:

[0100] ;

[0101] Among them, P i and P j Represents the state vector distribution at time i and time j. P represents i and P j All possible joint probability distributions γ Indicates from One of the samples, This represents the maximum expected distance in the sample.

[0102] S13. Based on the feature distribution similarity matrix W, calculate the data a in the i-th row and j-th column using the Pearson correlation coefficient. i and b jThe correlation coefficients between them form a correlation coefficient matrix ρ:

[0103] ;

[0104] S14. Based on the correlation coefficient matrix ρ, the average correlation coefficient between two adjacent time points is calculated by sliding along its main diagonal to obtain the average similarity curve ASC. The obtained average similarity curve is shown below. Figure 2 As shown:

[0105] .

[0106] Step S2: Calibrate the health status of CNC machine tool cutting tools based on the average similarity curve;

[0107] Specifically as follows:

[0108] The moment when the minimum value on the average similarity curve (ASC) is located is marked as the critical point between normal and abnormal CNC machine tool health status. CNC machine tool health status before this moment is marked as "normal", and CNC machine tool health status after this moment is marked as "abnormal".

[0109] Step S3: Set the main task and auxiliary task for predicting the remaining service life of CNC machine tool tools, and construct a conflict-free multi-task learning model (CMLM).

[0110] Specifically as follows:

[0111] S31. The main task is to predict the remaining service life of CNC machine tool tools to learn the long-term degradation characteristics of CNC machine tool tools; the secondary task is to assess the health status of CNC machine tool tools to learn the current health information of CNC machine tool tools and promote the learning of the main task.

[0112] S32. Based on the main task and the auxiliary task, construct a dual-branch network. The dual-branch network includes two branches, namely branch one and branch two. Branch one and branch two share a common feature extractor. Branch one consists of a remaining useful life prediction feature extractor and a predictor, which is used to complete the remaining useful life prediction task. Branch two consists of a health status assessment feature extractor and a classifier, which is used to complete the health status assessment task.

[0113] Dual-branch network: Branch 1 and Branch 2 share a common feature extractor to learn common knowledge for prediction and classification tasks and reduce the number of parameters; Branch 1 and Branch 2 each have a feature extractor for remaining lifetime prediction and a feature extractor for health status assessment, which are used to learn task-specific long-term degradation features and local features, respectively.

[0114] Shared feature extractor: consists of four dilated causal convolutional layers, each accompanied by a batch normalization layer, an ELU activation function, a dropout layer and a pooling layer;

[0115] The remaining lifetime prediction feature extractor consists of a multi-head self-attention mechanism, two batch normalization layers, two dropout layers, two fully connected layers, and a SwiGLU activation function.

[0116] The health status assessment feature extractor consists of an average pooling layer, a convolutional layer, a group normalization layer, and a sigmoid activation function.

[0117] The predictor in branch one and the classifier in branch two each consist of two fully connected layers, and each of their first fully connected layers is accompanied by an ELU activation function and a Dropout layer.

[0118] The parameters of each of the above modules are shown in Table 1.

[0119] Table 1

[0120]

[0121] Step S4: Divide the dataset and input it into the conflict-free multi-task learning model CMLM. Use the conflict-avoidance gradient descent algorithm to train the conflict-free multi-task learning model CMLM.

[0122] Specifically as follows:

[0123] S41. Use the leave-one-out method to set up the training set and the test set, and use each CNC machine tool tool as the test set and the remaining CNC machine tool tools as the training set.

[0124] S42. Each training set data is a state monitoring signal measured by an accelerometer containing two labels. One label is the actual remaining service life, and the other is the calibrated health status of the CNC machine tool. Input the state monitoring signal measured by each accelerometer and its actual remaining service life into branch one, and input the state monitoring signal measured by each accelerometer and its calibrated health status of the CNC machine tool into branch two.

[0125] S43. Set the batch size to 64. Use the conflict-avoiding gradient descent algorithm to optimize the parameters of the conflict-free multi-task learning model CMLM in order to reduce gradient direction conflicts between different tasks and avoid gradient scale differences between different tasks. For gradient θ, the conflict-avoiding gradient descent algorithm aims to find a vector d to update θ.

[0126] The conflict-avoidance gradient descent algorithm performs a first-order Taylor expansion of the loss function for the q-th task at θ:

[0127] ;

[0128] in, It is a loss L q The gradient at θ, where α represents the learning rate. It represents infinitesimal terms, and T is an operator that represents transpose.

[0129] Let R(θ, d) be used as a metric to measure gradient conflict between different tasks:

[0130] ;

[0131] Based on the analysis of R(θ, d), the vector d can be determined to update θ and avoid gradient conflicts by solving the following optimization problem:

[0132] ;

[0133] Where c=0.5 is a hyperparameter controlling the convergence speed, m is the dimension of d, and the optimization problem can be solved with average gradient... Find an optimal update vector d in the sphere centered at the center to minimize the conflict loss as measured by R(θ, d).

[0134] Step S5: After completing the training of the conflict-free multi-task learning model CMLM, input the CNC machine tool tool condition monitoring signal into the trained conflict-free multi-task learning model CMLM to obtain the remaining service life prediction result.

[0135] Specifically as follows:

[0136] S51. Determine whether the conflict-free multi-task learning model (CMLM) has been trained. If it has, save the CMLM with the minimum training set loss. If it has not been trained, repeat step S43 as follows:

[0137] Determine whether the training set loss does not decrease for 10 consecutive iterations or whether it has reached the maximum number of iterations of 500. If either of the two conditions is met, the model has been trained and the model with the minimum training set loss is saved. If neither of the two conditions is met, repeat step S43.

[0138] S52. Input the test set into the saved conflict-free multi-task learning model CMLM to obtain the remaining useful life prediction results, such as... Figure 3 As shown.

[0139] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any modifications or equivalent changes made based on the technical essence of the present invention shall still fall within the scope of protection claimed by the present invention.

Claims

1. A method for predicting the remaining service life of machine tool cutting tools based on conflict-free multi-task learning, characterized in that, Includes the following steps: Step S1: Obtain the average similarity curve through feature extraction, Earth movement distance, and correlation coefficient; Step S2: Calibrate the health status of CNC machine tool cutting tools based on the average similarity curve; Step S3: Set the main task and auxiliary task for predicting the remaining service life of CNC machine tool tools, and construct a conflict-free multi-task learning model (CMLM). Step S4: Divide the dataset and input it into the conflict-free multi-task learning model CMLM. Use the conflict-avoidance gradient descent algorithm to train the conflict-free multi-task learning model CMLM. Step S5: After completing the training of the conflict-free multi-task learning model CMLM, input the CNC machine tool tool condition monitoring signal into the trained conflict-free multi-task learning model CMLM to obtain the remaining service life prediction result.

2. The method for predicting the remaining service life of machine tool cutting tools based on conflict-free multi-task learning according to claim 1, characterized in that, Step S1 is as follows: S11. The state monitoring signal measured by the accelerometer is converted into a frequency domain signal by fast Fourier transform, and the common trend features of each tool of the CNC machine tool are extracted using time and frequency domain feature extraction methods. The formula for the common trend features of each tool is as follows: entropy: ; Interquartile range: IQR = Q3 - Q1; Maximum value: Max = max(x) i ); Minimum value: Min = min(x) i ); Mean: ; Standard deviation: ; Root mean square: ; Skewness: ; Kuroshi: ; Sum of absolute values: ; Waveform factor: ; Peak factor: ; Pulse factor: ; Frequency domain centroid: ; Mean square frequency: ; Frequency variance: ; Spectral entropy: ; Among them, x i Let Q1 and Q3 represent the time-domain amplitude of the i-th sampling point, respectively, and let T represent the first and third quantiles. , and Let M represent the power spectral amplitude, actual frequency value, and normalized probability distribution of the power spectral amplitude of the k-th frequency component, respectively, and M represent the total number of frequency components. S12. Combine the 17 common trend features obtained in step S11 into a state vector at each time step. Use the Earth's movement distance to measure the similarity of feature distributions between pairs of time steps, thus forming a feature distribution similarity. The matrix W is given by the following formula: ； Among them, P i and P j Represents the state vector distribution at time i and time j. P represents i and P j All possible joint probability distributions γ Indicates from One of the samples, This represents the maximum expected distance in the sample. S13. Based on the feature distribution similarity matrix W, calculate the data a in the i-th row and j-th column using the Pearson correlation coefficient. i and b j The correlation coefficients between them form a correlation coefficient matrix ρ: ； S14. Based on the correlation coefficient matrix ρ, calculate the average correlation coefficient between two adjacent time points by sliding along its main diagonal to obtain the average similarity curve ASC: 。 3. The method for predicting the remaining service life of machine tool cutting tools based on conflict-free multi-task learning according to claim 1 or 2, characterized in that, Step S2 is as follows: The moment when the minimum value on the average similarity curve (ASC) is located is marked as the critical point between normal and abnormal CNC machine tool health status. CNC machine tool health status before this moment is marked as "normal", and CNC machine tool health status after this moment is marked as "abnormal".

4. The method for predicting the remaining service life of machine tool cutting tools based on conflict-free multi-task learning according to claim 1, characterized in that, Step S3 is as follows: S31. The main task is to predict the remaining service life of CNC machine tool tools to learn the long-term degradation characteristics of CNC machine tool tools; the secondary task is to assess the health status of CNC machine tool tools to learn the current health information of CNC machine tool tools and promote the learning of the main task. S32. Based on the main task and the auxiliary task, construct a dual-branch network. The dual-branch network includes two branches, namely branch one and branch two. Branch one and branch two share a common feature extractor. Branch one consists of a remaining useful life prediction feature extractor and a predictor, which is used to complete the remaining useful life prediction task. Branch two consists of a health status assessment feature extractor and a classifier, which is used to complete the health status assessment task.

5. The method for predicting the remaining service life of machine tool cutting tools based on conflict-free multi-task learning according to claim 4, characterized in that: In step S32, Dual-branch network: Branch 1 and Branch 2 share a common feature extractor to learn common knowledge for prediction and classification tasks and reduce the number of parameters; Branch 1 and Branch 2 each have a feature extractor for remaining lifetime prediction and a feature extractor for health status assessment, which are used to learn task-specific long-term degradation features and local features, respectively. Shared feature extractor: consists of four dilated causal convolutional layers, each accompanied by a batch normalization layer, an ELU activation function, a dropout layer and a pooling layer; The remaining lifetime prediction feature extractor consists of a multi-head self-attention mechanism, two batch normalization layers, two dropout layers, two fully connected layers, and a SwiGLU activation function. The health status assessment feature extractor consists of an average pooling layer, a convolutional layer, a group normalization layer, and a sigmoid activation function.

6. The method for predicting the remaining service life of machine tool cutting tools based on conflict-free multi-task learning according to claim 4, characterized in that: In step S32, The predictor in branch one and the classifier in branch two each consist of two fully connected layers, and each of their first fully connected layers is accompanied by an ELU activation function and a Dropout layer.

7. The method for predicting the remaining service life of machine tool cutting tools based on conflict-free multi-task learning according to claim 1 or 4, characterized in that, Step S4 is as follows: S41. Use the leave-one-out method to set up the training set and the test set, and use each CNC machine tool tool as the test set and the remaining CNC machine tool tools as the training set. S42. Each training set data is a state monitoring signal measured by an accelerometer containing two labels. One label is the actual remaining service life, and the other is the calibrated health status of the CNC machine tool. Input the state monitoring signal measured by each accelerometer and its actual remaining service life into branch one, and input the state monitoring signal measured by each accelerometer and its calibrated health status of the CNC machine tool into branch two. S43. Optimize the parameters of the conflict-free multi-task learning model CMLM using the conflict-avoiding gradient descent algorithm. For the gradient θ, the conflict-avoiding gradient descent algorithm aims to find a vector d to update θ.

8. The method for predicting the remaining service life of machine tool cutting tools based on conflict-free multi-task learning according to claim 7, characterized in that: In step S43, The conflict-avoidance gradient descent algorithm performs a first-order Taylor expansion of the loss function for the q-th task at θ: ； in, It is a loss L q The gradient at θ, where α represents the learning rate. It represents infinitesimal terms, and T is an operator that represents transpose. Let R(θ, d) be used as a metric to measure gradient conflict between different tasks: ； Based on the analysis of R(θ, d), the vector d can be determined to update θ and avoid gradient conflicts by solving the following optimization problem: ； Where c=0.5 is a hyperparameter controlling the convergence speed, m is the dimension of d, and the optimization problem can be solved with average gradient... Find an optimal update vector d in the sphere centered at the center to minimize the conflict loss as measured by R(θ, d).

9. The method for predicting the remaining service life of machine tool cutting tools based on conflict-free multi-task learning according to claim 7, characterized in that, Step S5 is as follows: S51. Determine whether the conflict-free multi-task learning model CMLM has been trained. If it has been trained, save the conflict-free multi-task learning model CMLM with the minimum training set loss. If it has not been trained, repeat step S43. S52. Input the test set into the saved conflict-free multi-task learning model CMLM to obtain the remaining useful life prediction results.

Images