A method for predicting the residual life of a milling tool based on a deep convolutional neural network
By optimizing data preprocessing and feature selection, a deep convolutional neural network model without pooling layers was built, which solved the problems of insufficient accuracy and generalization ability in milling tool remaining life prediction, realized high-precision tool condition monitoring and remaining life prediction, and promoted the intelligent development of CNC machining.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HUAZHONG UNIV OF SCI & TECH
- Filing Date
- 2026-03-07
- Publication Date
- 2026-06-05
Smart Images
Figure CN122154465A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of machining tool condition monitoring technology, specifically to a method for predicting the remaining life of milling tools based on a deep convolutional neural network, which is applicable to real-time monitoring of tool wear and accurate prediction of remaining life during CNC milling. Background Technology
[0002] Milling is one of the core production methods in manufacturing. The cutting tool, as the "teeth" of the cutting process, directly affects the surface quality of the workpiece, machining efficiency, and production costs. Factors such as cutting heat and impact during machining can cause normal wear on the tool, such as rake face wear and flank face wear, or abnormal wear such as chipping and delamination. Statistics show that machine tool downtime due to tool wear accounts for as much as 20%. Traditional tool replacement strategies rely on periodic tool changes based on production experience, which can lead to insufficient tool utilization or tool exceeding its lifespan, resulting in decreased workpiece accuracy. Therefore, accurately predicting the remaining lifespan of milling tools is of great significance for intelligent manufacturing.
[0003] Currently, intelligent tool wear monitoring technology has gradually developed. Monitoring systems mainly include signal acquisition, feature extraction, and pattern recognition modules. Commonly monitored signals include cutting force, vibration, and current. Feature extraction covers time domain, frequency domain, and time-frequency domain analysis. Prediction models include traditional machine learning models such as artificial neural networks, hidden Markov models, and support vector machines, as well as deep learning models. However, existing technologies still have many shortcomings:
[0004] While foreign commercial monitoring systems are feature-rich, they cannot correlate monitoring signals with actual tool wear values and cannot reflect tool life status in real time; domestic monitoring models need improvement in accuracy, real-time performance, and versatility.
[0005] A single monitoring signal cannot fully characterize the tool wear pattern. After feature extraction, strong correlation features are not fully screened, and redundant features will reduce the model's prediction efficiency.
[0006] Traditional dataset partitioning is based on runtime, resulting in training set features being singular and model generalization ability being poor.
[0007] Traditional deep convolutional neural networks contain pooling layers, which can lead to the loss of some features during dimensionality reduction, thus reducing the model's learning ability. Traditional machine learning models have low prediction accuracy and high computational complexity, making them difficult to meet the requirements of intelligent manufacturing.
[0008] Predicting the remaining life of milling tools based on deep learning is a current research hotspot. Deep convolutional neural network models have powerful feature extraction and nonlinear fitting capabilities. Therefore, in order to address the above-mentioned technical shortcomings, optimizing the structure of deep convolutional neural network models, improving data preprocessing and feature selection methods, and rationally dividing the dataset, it is urgent to solve the technical problem in this field to construct a high-precision and highly generalizable method for predicting the remaining life of milling tools. Summary of the Invention
[0009] The purpose of this invention is to overcome the shortcomings of existing milling tool remaining life prediction methods and provide a milling tool remaining life prediction method based on deep convolutional neural networks. By optimizing data preprocessing, multi-domain feature filtering, dataset hierarchical partitioning, and deep convolutional neural network model structure, the prediction accuracy and generalization ability are improved, enabling accurate prediction of tool remaining life, making full use of tool life, and reducing machining downtime and production costs.
[0010] To achieve the above objectives, the present invention provides the following technical solution:
[0011] A method for predicting the remaining life of milling tools based on deep convolutional neural networks includes the following steps:
[0012] S1. Preprocessing of milling signals: Select the Z-axis cutting force signal, Y-axis cutting torque signal and spindle vibration signal in the milling process as the original monitoring signals. After removing the interference signals of entry and exit, the data is denoised using the 3sigma criterion. The signal is divided into several processing processes based on the time series. The tool wear data is completed by the polynomial fitting method to obtain the preprocessed dataset.
[0013] S2. Multi-domain feature extraction and screening: Time domain, frequency domain, and time-frequency domain features are extracted from the three preprocessed signals respectively. The correlation between each feature and the tool wear pattern is analyzed, and features strongly correlated with tool wear are screened to form a feature set.
[0014] S3. Dataset hierarchical partitioning: The feature set is hierarchically partitioned according to the processing time series by grouping 5 monitoring points. Each group is divided into a training subset and a test subset in a 4:1 ratio. All training subsets and test subsets are integrated to obtain the model training set and test set respectively.
[0015] S4. Building a Deep Convolutional Neural Network Model: Build a deep convolutional neural network model containing 3 convolutional layers and 2 fully connected layers. Remove pooling layers from the model to avoid feature loss. Set the kernel size, stride, and number of features to be extracted for the convolutional layers. Select ReLU as the activation function and Adam as the optimization function. Set the dropout parameter to 0.7 and use cross-entropy as the objective function.
[0016] S5. Model Training and Validation: Input the training set into the deep convolutional neural network model for training. Tune the parameters to minimize cross-entropy using the backpropagation algorithm. Validate the model using the test set in conjunction with RMSE and R-squared. 2 Adjusted_R 2 The Score metric is used to verify the model's prediction accuracy, and tool wear datasets under different working conditions are selected to verify the model's generalization ability.
[0017] S6. Milling tool remaining life prediction: After processing the milling signal to be predicted through steps S1-S2, input it into the trained deep convolutional neural network model, and output the predicted value of the remaining life of the milling tool.
[0018] Furthermore, in step S1, the specific operation of the 3sigma criterion denoising is as follows:
[0019] Calculate the mean μ and standard deviation σ of the signal data, and then... The 3σ value is considered normal data. 6σ was considered suspicious data. σ is considered abnormal data, and suspicious and abnormal data are removed; the original monitoring signal sampling frequency is 300Hz, and after preprocessing, the signal is divided into 400 processing steps, each containing 493 sampling points; the polynomial fitting function is: .
[0020] Furthermore, in step S2, the time-domain features include 10 dimensional features and 6 dimensionless features. The dimensional features are maximum value, maximum absolute value, minimum value, mean, absolute average value, root mean square value, root mean square amplitude, standard deviation, kurtosis value, and skewness value. The dimensionless features are margin index, waveform index, impulse index, peak index, kurtosis index, and skewness index. The frequency-domain features are centroid frequency, average frequency, root mean square frequency, and frequency variance. The time-frequency domain features are extracted into 8 sub-frequency band features through wavelet packet decomposition. The features strongly correlated with tool wear are the mean feature and waveform index feature of the Y-axis cutting torque, and the centroid frequency feature, sub-frequency band 3 feature, and sub-frequency band 7 feature of the spindle vibration.
[0021] Furthermore, in step S4, the parameters of the three convolutional layers are set as follows:
[0022] The first convolutional layer has a kernel size of 5×1, a kernel stride of 3, and extracts 16 types of features.
[0023] The second convolutional layer has a kernel size of 2×1, a kernel stride of 2, and extracts 32 features.
[0024] The third convolutional layer has a kernel size of 1×1, a kernel stride of 1, and extracts 64 features.
[0025] Before the convolution operation, the input feature vector is padded with the same padding method to ensure that the size of the vector remains unchanged before and after the convolution. The dimension of the model input feature vector is 400×8, where 400 represents 400 processing steps and 8 represents the category of the feature.
[0026] Furthermore, in step S5, the calculation formula for the model evaluation index is as follows:
[0027] ;
[0028] ;
[0029] ;
[0030] in, This represents the actual lifespan value corresponding to the i-th time point. This represents the predicted lifetime value corresponding to the i-th time point. To predict the mean lifetime, N is the number of samples in the dataset, and P is the number of features; the Score metric is used to evaluate the degree of over-prediction and under-prediction of the model, and the closer its value is to 0, the better the model's predictive performance.
[0031] Furthermore, in step S1, the original monitoring signal is derived from the orthogonal experimental conditions of the NUAA_Ideahouse tool wear dataset, the tool material is solid carbide, and the workpiece material is TC4.
[0032] Furthermore, the specific operation of wavelet packet decomposition is as follows: select wavelet packet basis functions, establish a wavelet packet tree by decomposing the signal subbands step by step, calculate the wavelet packet coefficients of each wavelet packet basis function in the signal, and select important wavelet packet coefficients as time-frequency domain features.
[0033] Furthermore, in step S5, the model's optimal prediction accuracy reached 80.65%, with a corresponding RMSE of 0.0469 and Adjusted_R. 2 The value was 0.8024 and the score was 0.2933. The selected different working condition datasets had different parameters such as feed rate fz, spindle speed n, and depth of cut ap compared with the original working condition dataset. The prediction accuracy of the model under different working conditions was 69.67% and 74.36%, respectively, with the accuracy variation within 14%.
[0034] Compared with the prior art, the present invention has the following advantages:
[0035] 1. High prediction accuracy: By using the 3sigma criterion for standard classification and noise reduction, and precise screening of multi-domain features, redundant and interfering features are eliminated. A deep convolutional neural network model without pooling layers is built to avoid feature loss. The model's optimal prediction accuracy reaches 80.65%, which is far higher than traditional machine learning models and can accurately reflect the remaining tool life.
[0036] 2. Strong generalization ability: The dataset is divided by hierarchical extraction, so that the training set contains all the features of the machining process. The prediction accuracy of the model under different working conditions is 69.67% and 74.36% respectively, with the accuracy variation within 14%. It is suitable for tool life prediction under different milling machining conditions.
[0037] 3. Precise feature representation: Selecting signals from multiple sensors such as cutting force, torque, and vibration, and combining time domain, frequency domain, and time-frequency domain feature extraction, five features strongly correlated with tool wear are selected to comprehensively and accurately represent the tool wear pattern, overcoming the limitations of representation by a single signal / feature;
[0038] 4. High practicality: This method is validated based on publicly available tool wear datasets. The optimized model parameters are clearly presented in a table. The preprocessing and feature extraction steps can be standardized, making it easy to port to actual CNC machining systems. This enables online monitoring of tool status and prediction of remaining tool life, making full use of tool life, reducing machine tool downtime and machining costs, and promoting the intelligent development of CNC machining. Attached Figure Description
[0039] Figure 1: Classification of milling cutter wear patterns;
[0040] Figure 2: Schematic diagram of milling process and flank wear;
[0041] Figure 3: Curve showing the variation of the tool flank wear value VB;
[0042] Figure 4: Comparison of original spindle vibration signal and signal extracted after stable cutting process;
[0043] Figure 5: Noise map of cutting force, torque, and spindle vibration signals before noise reduction;
[0044] Figure 6: Polynomial fitting plot of tool wear data;
[0045] Figure 7: Schematic diagram of wavelet decomposition and wavelet packet decomposition;
[0046] Figure 8: Structure diagram of a deep convolutional neural network model without pooling layers;
[0047] Figure 9: Schematic diagram of convolution operation;
[0048] Figure 10: Comparison of traditional and hierarchical partitioning of datasets. Detailed Implementation
[0049] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.
[0050] This invention provides a method for predicting the remaining life of milling tools based on deep convolutional neural networks, comprising the following steps:
[0051] (a) Preprocessing of milling machining signals
[0052] The NUAA_Ideahouse tool wear dataset was selected. This dataset contains single-factor experiments and orthogonal experimental datasets. The orthogonal experiment design has 3 factors and 3 levels (W1-W9). The specific parameters are shown in Table 1. In this invention, the Z-axis cutting force signal, Y-axis cutting torque signal, and spindle vibration signal under the orthogonal experiment W5 condition were selected as the original monitoring signals. The sampling frequency under this condition is 300Hz. The total data sample contains 198272 sampling points. The tool material is solid carbide and the workpiece material is TC4.
[0053] Table 1. Orthogonal Experiment Parameters
[0054]
[0055] First, interference signals from the cutting-in and cutting-out stages in the original signal are removed to extract the effective signal of the stable cutting process. Then, the 3sigma criterion is used to denoise the data. First, the mean μ and standard deviation σ of the signal data are calculated. Then, the data are divided into normal data, suspicious data and abnormal data according to the criteria in Table 2. Suspicious data and abnormal data are removed to ensure the validity of the signal data.
[0056] Table 2 3sigma Data Classification Table
[0057]
[0058] Next, based on the time series, the preprocessed stable signal was divided into 400 machining processes, each containing 493 sampling points. Finally, since the original tool wear data only had 8 data points, which could not correspond to the 400 machining processes, a polynomial fitting method was used to complete the tool wear data. The fitting function is as follows:
[0059] The function completes the signal preprocessing process by providing 400 tool wear values, resulting in a standardized dataset.
[0060] (II) Multi-domain feature extraction and screening
[0061] The preprocessed Z-axis cutting force, Y-axis cutting torque, and spindle vibration signals were subjected to multi-dimensional feature extraction in the time domain, frequency domain, and time-frequency domain to comprehensively mine the feature information related to tool wear in the signals.
[0062] Temporal features: 10 dimensional features and 6 dimensionless features were extracted. The dimensional features are maximum value, maximum absolute value, minimum value, mean, absolute mean, root mean square value, root mean square amplitude, standard deviation, kurtosis value, and skewness value. The dimensionless features are margin index, waveform index, impulse index, peak index, kurtosis index, and skewness index.
[0063] Frequency domain features: Extract four core features: centroid frequency, average frequency, root mean square frequency, and frequency variance, to reflect the frequency distribution pattern of the signal;
[0064] Time-frequency domain features: Eight sub-band features are extracted using wavelet packet decomposition. First, appropriate wavelet packet basis functions are selected, and a wavelet packet tree is established by decomposing the signal sub-bands step by step. The wavelet packet coefficients corresponding to each wavelet packet basis function are calculated, and finally, important wavelet packet coefficients are selected as time-frequency domain features.
[0065] After completing the multi-domain feature extraction, the correlation between each feature and the three stages of tool wear (initial wear, normal wear, and rapid wear) was analyzed. Redundant features with weak correlation to tool wear, large fluctuations, and many abrupt change points were eliminated. Finally, five strongly correlated features were selected: the mean value feature and waveform index feature of the Y-axis cutting torque, the center-of-gravity frequency feature, sub-band 3 feature, and sub-band 7 feature of the spindle vibration. These five features were used to form the feature set for model input, providing high-quality feature data for subsequent model training.
[0066] (III) Dataset Hierarchical Partitioning
[0067] To avoid the problem of limited training set features caused by traditional dataset partitioning methods based on runtime, this invention employs a hierarchical extraction method to partition the feature set. Specifically, the feature set is divided into groups of five monitoring points according to the processing time sequence. Within each group, it is further divided into a training subset and a test subset in a 4:1 ratio. Finally, all training subsets are integrated to obtain the model training set, and all test subsets are integrated to obtain the model test set. This partitioning method ensures that the training set contains all features of the entire milling process, improving the model's ability to identify and generalize unknown features.
[0068] (iv) Building a deep convolutional neural network model
[0069] A deep convolutional neural network model consisting of 3 convolutional layers and 2 fully connected layers was constructed. The core innovation is the removal of pooling layers, which avoids the loss of feature information during dimensionality reduction and improves the model's feature learning ability. Before the convolution operation, the input feature vector (dimension 400×8, where 400 represents 400 processing steps and 8 represents the feature category) is padded with zeros around the vector to ensure that the vector size remains unchanged before and after the convolution operation, allowing the feature information to be fully learned. The core parameters of the 3 convolutional layers were determined after multiple optimizations. The specific kernel size, kernel stride, and number of feature extraction types are shown in Table 3.
[0070] Table 3. Parameters of Convolutional Layers in Deep Convolutional Neural Network Models
[0071]
[0072] The model chooses ReLU as the activation function, and its expression is: This effectively avoids gradient explosion or gradient vanishing phenomena and improves the model convergence speed. Setting the dropout parameter to 0.7 means that only 70% of the neurons are activated to participate in training each time the neural network is trained, preventing the model from overfitting. Adam is selected as the optimization function, with cross-entropy as the objective function. The high-level features extracted by the convolutional layer are nonlinearly mapped through the fully connected layer, and finally the remaining tool life prediction value is output by the output layer.
[0073] (v) Model training and validation
[0074] The training set obtained by the hierarchical partitioning is input into the deep convolutional neural network model without pooling layers built above for training. The model parameters are continuously adjusted through the backpropagation algorithm until the cross-entropy reaches the minimum value. At this time, the optimal model after training is saved.
[0075] The prediction accuracy of the trained model was validated using a test set, employing the root mean square error (RMSE) and goodness-of-fit (R²) values. 2 ), Adjust the goodness-of-fit value (Adjusted_R) 2 The Score function is used as the model evaluation metric. The model is trained and optimized multiple times, and the evaluation metric values for each training iteration are shown in Table 4.
[0076] As shown in Table 4, the optimal evaluation index values for the model are: RMSE = 0.0469, R... 2 =0.8065, Adjusted_R 2=0.8024, Score=0.2933, meaning the model's optimal prediction accuracy reaches 80.65%, meeting the accuracy requirements for tool remaining life prediction.
[0077] Table 4 Evaluation metrics for deep convolutional neural network models
[0078]
[0079] To verify the generalization ability of the model, two milling condition datasets different from the original working condition (W5) were selected as generalization validation sets. The comparison working condition 1 was fz=0.045mm / r, n=1700r / min, ap=2.5mm, and the comparison working condition 2 was fz=0.055mm / r, n=1850r / min, ap=3mm. After preprocessing and feature filtering, the datasets of the two comparison working conditions were input into the trained model, and the model evaluation index values under different working conditions are shown in Table 5.
[0080] As shown in Table 5, the prediction accuracy of comparison condition 1 is 69.67%, and the prediction accuracy of comparison condition 2 is 74.36%. Compared with the original condition's 80.65%, the accuracy change is within 14%, which verifies that the deep convolutional neural network model built in this invention has good generalization ability and is suitable for different milling machining conditions.
[0081] Table 5. Comparison and evaluation index values of the model under different working conditions
[0082]
[0083] (vi) Prediction of remaining life of milling tools
[0084] The signal to be predicted, collected during the actual CNC milling process, is processed through the above-mentioned signal preprocessing, multi-domain feature extraction and filtering process, and then input into a trained and verified deep convolutional neural network model without pooling layers. The model will quickly output the predicted value of the remaining life of the milling tool, realizing accurate and real-time prediction of the remaining tool life.
[0085] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and the data in the tables above.
[0086] Example 1:
[0087] This embodiment uses the NUAA_Ideahouse tool wear dataset to perform a full-process verification of the milling tool remaining life prediction method of the present invention. The specific steps are as follows:
[0088] Signal preprocessing:
[0089] Select the Z-axis cutting force, Y-axis cutting torque, and spindle vibration signals for the W5 working condition in Table 1. After removing entry and exit interference signals, denoise the signals according to the 3 Sigma classification standard. Suspicious and abnormal data in 3σ; the signal is divided into 400 processing steps, each with 493 sampling points; 400 tool wear values are completed by using a polynomial fitting function to complete the preprocessing;
[0090] Multi-domain feature extraction and filtering:
[0091] Twenty-eight features were extracted from the three preprocessed signals: 20 in the time domain, 4 in the frequency domain, and 8 in the time-frequency domain. The correlation between each feature and the three stages of tool wear was analyzed. Five strongly correlated features were selected: the mean value of Y-axis cutting torque, waveform index, spindle vibration center of gravity frequency, sub-band 3, and sub-band 7. These features were then used to form a feature set.
[0092] Dataset hierarchical partitioning:
[0093] The feature set is stratified by grouping 5 monitoring points into a set. Each set is divided into a training subset and a test subset in a 4:1 ratio. After integration, the training set and test set are obtained, ensuring that the training set contains features of the entire processing process.
[0094] Building a deep convolutional neural network model:
[0095] Construct a deep convolutional neural network model without pooling layers. Set the parameters of the three convolutional layers according to Table 3, using the ReLU activation function, Adam optimization function, dropout=0.7, cross-entropy as the objective function, and equal padding before convolution.
[0096] Model training and validation:
[0097] The training set was used to train the model, and the parameters were back-tuned until the cross-entropy was minimized. The optimal metrics in Table 4 were verified using the test set: RMSE = 0.0469, R0 = 0.0469. 2 =0.8065, prediction accuracy 80.65%; Select two comparative working condition datasets in Table 5 to verify the generalization ability, and the prediction accuracy is 69.67% and 74.36% respectively. The accuracy variation range is ≤14%, and the model performance meets the standard.
[0098] Remaining life prediction:
[0099] After preprocessing and feature filtering, the milling signal to be predicted under W5 working condition is input into the trained deep convolutional neural network model. The model outputs the predicted value of the remaining tool life, and the error between the predicted value and the actual wear value is within an acceptable range, thus achieving accurate prediction.
[0100] Example 2
[0101] The difference between this embodiment and Embodiment 1 is that the W9 working condition in Table 1 is selected as the object to be predicted. The cutting force, torque, and vibration signals of the W9 working condition are denoised by the 3sigma criterion, and multi-domain features are extracted and filtered before being input into the deep convolutional neural network model trained in Embodiment 1. The model outputs the predicted value of the remaining life of the milling tool under the W9 working condition, with a prediction accuracy of 72.15%, further verifying the generalization ability of the method of the present invention to different orthogonal experimental working conditions.
[0102] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.
Claims
1. A method for predicting the remaining life of milling tools based on deep convolutional neural networks, characterized in that, Includes the following steps: S1. Preprocessing of milling signals: Select the Z-axis cutting force signal, Y-axis cutting torque signal and spindle vibration signal in the milling process as the original monitoring signals. After removing the interference signals of entry and exit, the data is denoised using the 3sigma criterion. The signal is divided into several processing processes based on the time series. The tool wear data is completed by the polynomial fitting method to obtain the preprocessed dataset. S2. Multi-domain feature extraction and screening: Time domain, frequency domain, and time-frequency domain features are extracted from the three preprocessed signals respectively. The correlation between each feature and the tool wear pattern is analyzed, and features strongly correlated with tool wear are screened to form a feature set. S3. Dataset hierarchical partitioning: The feature set is hierarchically partitioned according to the processing time series by grouping 5 monitoring points. Each group is divided into a training subset and a test subset in a 4:1 ratio. All training subsets and test subsets are integrated to obtain the model training set and test set respectively. S4. Building a Deep Convolutional Neural Network Model: Build a deep convolutional neural network model containing 3 convolutional layers and 2 fully connected layers. Remove pooling layers from the model to avoid feature loss. Set the kernel size, stride, and number of features to be extracted for the convolutional layers. Select ReLU as the activation function and Adam as the optimization function. Set the dropout parameter to 0.7 and use cross-entropy as the objective function. S5. Model Training and Validation: Input the training set into the deep convolutional neural network model for training. Tune the parameters to minimize cross-entropy using the backpropagation algorithm. Validate the model using the test set in conjunction with RMSE and R-squared. 2 Adjusted_R 2 The Score metric is used to verify the model's prediction accuracy, and tool wear datasets under different working conditions are selected to verify the model's generalization ability. S6. Milling tool remaining life prediction: After processing the milling signal to be predicted through steps S1-S2, input it into the trained deep convolutional neural network model, and output the predicted value of the remaining life of the milling tool.
2. The method for predicting the remaining life of milling tools based on a deep convolutional neural network according to claim 1, characterized in that, In step S1, the specific operation of the 3sigma criterion denoising is as follows: Calculate the mean μ and standard deviation σ of the signal data, and then... The 3σ value is considered normal data. 6σ was considered suspicious data. σ is considered abnormal data, and suspicious and abnormal data are removed; the original monitoring signal sampling frequency is 300Hz, and after preprocessing, the signal is divided into 400 processing steps, each containing 493 sampling points; the polynomial fitting function is: 。 3. The method for predicting the remaining life of milling tools based on a deep convolutional neural network according to claim 1, characterized in that, In step S2, the time-domain features include 10 dimensional features and 6 dimensionless features. The dimensional features are maximum value, maximum absolute value, minimum value, mean, absolute average value, root mean square value, root mean square amplitude, standard deviation, kurtosis value, and skewness value. The dimensionless features are margin index, waveform index, impulse index, peak value, kurtosis index, and skewness index. The frequency-domain features are centroid frequency, average frequency, root mean square frequency, and frequency variance. The time-frequency domain features are extracted into 8 sub-frequency band features through wavelet packet decomposition. The features strongly correlated with tool wear were the mean value and waveform index features of the Y-axis cutting torque, and the center-of-gravity frequency, sub-band 3, and sub-band 7 features of the spindle vibration.
4. The method for predicting the remaining life of milling tools based on a deep convolutional neural network according to claim 1, characterized in that, In step S4, the parameters of the three convolutional layers are set as follows: The first convolutional layer has a kernel size of 5×1, a kernel stride of 3, and extracts 16 types of features. The second convolutional layer has a kernel size of 2×1, a kernel stride of 2, and extracts 32 features. The third convolutional layer has a kernel size of 1×1, a kernel stride of 1, and extracts 64 features. Before the convolution operation, the input feature vector is padded with the same padding method to ensure that the size of the vector remains unchanged before and after the convolution. The dimension of the model input feature vector is 400×8, where 400 represents 400 processing steps and 8 represents the category of the feature.
5. The method for predicting the remaining life of milling tools based on a deep convolutional neural network according to claim 1, characterized in that, In step S5, the calculation formula for the model evaluation index is as follows: ； ； ； in, This represents the actual lifespan value corresponding to the i-th time point. This represents the predicted lifetime value corresponding to the i-th time point. To predict the mean lifetime, N is the number of samples in the dataset, and P is the number of features; the Score metric is used to evaluate the degree of over-prediction and under-prediction of the model, and the closer its value is to 0, the better the model's predictive performance.
6. The method for predicting the remaining life of milling tools based on a deep convolutional neural network according to claim 1, characterized in that, In step S1, the original monitoring signal comes from the orthogonal experimental conditions of the NUAA_Ideahouse tool wear dataset, the tool material is solid carbide, and the workpiece material is TC4.
7. The method for predicting the remaining life of milling tools based on a deep convolutional neural network according to claim 3, characterized in that, The specific operation of wavelet packet decomposition is as follows: select wavelet packet basis functions, establish a wavelet packet tree by decomposing the signal subbands step by step, calculate the wavelet packet coefficients of each wavelet packet basis function in the signal, and select important wavelet packet coefficients as time-frequency domain features.
8. The method for predicting the remaining life of milling tools based on a deep convolutional neural network according to claim 5, characterized in that, In step S5, the model's optimal prediction accuracy reached 80.65%, with a corresponding RMSE of 0.0469 and Adjusted_R. 2 The value was 0.8024 and the score was 0.2933. The selected different working condition datasets had different parameters such as feed rate fz, spindle speed n, and depth of cut ap compared with the original working condition dataset. The prediction accuracy of the model under different working conditions was 69.67% and 74.36%, respectively, with the accuracy variation within 14%.