Small sample cutter wear prediction method based on parallel multi-head residual memory network
By combining parallel multi-head residual memory networks with data preprocessing and feature extraction methods, the problem of accurate prediction of tool wear monitoring under small sample conditions is solved, and efficient tool wear monitoring under low cost conditions is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-30
- Publication Date
- 2026-03-31
AI Technical Summary
Existing technologies for tool wear monitoring require a large amount of data and computing power, making it difficult to train accurate models with small sample sizes, resulting in high costs and low efficiency.
A parallel multi-head residual memory network is employed, combined with discrete wavelet transform, empirical mode decomposition, and local weighted regression. The dataset is expanded through data augmentation and spline interpolation. Furthermore, the parallel residual network, multi-head self-attention mechanism, and stacked bidirectional long short-term memory network are used to adaptively extract multi-scale features, thereby achieving accurate prediction of tool wear.
It improves model training accuracy under small sample conditions, reduces the cost of data acquisition and model training, and enhances the accuracy and efficiency of tool wear monitoring.
Smart Images
Figure CN121766136A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of CNC machine tool wear monitoring and prediction, specifically involving a small-sample tool wear prediction method based on a parallel multi-head residual memory network. Background Technology
[0002] With the shift from traditional manufacturing to intelligent manufacturing, health monitoring systems are being deployed more and more frequently to ensure the reliability of machining quality. During the cutting process, as material is removed from the workpiece, the cutting tool is continuously worn due to the effects of cutting forces and temperature, which in turn affects dimensional errors and surface integrity. The condition of the cutting tool directly affects product quality and machining efficiency. In severe cases, it can even damage the machine tool. Previous studies have shown that 10-40% of machine tool downtime is caused by abnormal tool condition. Therefore, real-time monitoring of tool wear during machining is of great significance for improving production efficiency and quality, and saving costs.
[0003] In the era of big data, AI-driven data analysis methods have shown great potential in processing tool wear signals. Specifically, in the field of tool wear monitoring, the relationship between machining signals and tool condition is used to achieve real-time prediction of tool wear. Although traditional learning models have achieved good results in tool wear monitoring, they mainly extract surface-level feature information. Therefore, these models require a large amount of data and computing power to extract sufficient data features to meet task requirements. How to train a more accurate model with a small sample size remains a major challenge.
[0004] In view of this, the present invention proposes a small-sample tool wear prediction method based on a parallel multi-head residual memory network. Summary of the Invention
[0005] The purpose of this invention is to solve the problems of long time consumption and high cost in obtaining sample data in actual industrial processing, and to provide a small sample tool wear prediction method based on parallel multi-head residual memory network.
[0006] The prediction framework described in this method can significantly improve data quality and solve the problem of insufficient sample quantity. The model used achieves adaptive extraction of multi-scale features from input data and captures time-frequency domain features related to tool wear evolution, ultimately achieving accurate prediction of tool wear state.
[0007] To achieve the above functions, the technical solution of the present invention specifically includes the following technical steps:
[0008] Small-sample tool wear prediction methods based on parallel multi-head residual memory networks include:
[0009] 1) Collect CNC machine tool machining data:
[0010] The vibration signal data of the spindle front end and the AC current signal data of the spindle are collected during the machining process of CNC machine tool. The tool wear data is measured with a microscope and the above data are saved at equal time intervals as the initial data for model training.
[0011] 2) Noise reduction preprocessing:
[0012] Delete the invalid parts at the beginning and end of each spindle front-end vibration signal data and spindle AC current signal data;
[0013] Spindle front end vibration signal data: Empirical mode decomposition is used to decompose the spindle front end vibration signal data, discrete wavelet transform is used to process the high-frequency intrinsic mode functions, and then a soft threshold function is used to reconstruct the signal to obtain the noise-reduced spindle front end vibration signal data.
[0014] Main spindle AC current signal data: The discrete wavelet transform is directly used for processing and reconstructed using a soft thresholding function to obtain the noise-reduced main spindle AC current signal data;
[0015] 3) Further process the vibration signal data of the noise-reducing spindle front end, the AC current signal data of the noise-reducing spindle, and the tool wear data obtained in step 1):
[0016] a. For the vibration signal data of the front end of the noise reduction spindle and the AC current signal data of the noise reduction spindle, each cutting process is divided into multiple smaller cutting sequences to obtain enhanced input data;
[0017] For tool wear data, cubic spline interpolation is used for processing; the mathematical expression for cubic spline interpolation is as follows:
[0018] (1)
[0019] in, It is an interval cubic spline polynomial on, yes The function value at that point, These are the coefficients of the first derivative. These are the coefficients of the second derivative. These are the coefficients of the third derivative.
[0020] b. Perform feature value calculation on the enhanced input data to extract a total of 12 time-frequency domain features and 24 time-frequency domain feature values;
[0021] c. Process the 24 time-frequency domain feature values obtained in step b) using local weighted regression:
[0022] The mathematical expression for the locally weighted regression is as follows:
[0023] (2)
[0024] in, It is a prediction point The predicted value, These are the response values of the training data points. It is a weighting function that depends on the prediction point. and training data points The distance between them.
[0025] 4) Train the tool wear prediction model using a parallel multi-head residual memory neural network:
[0026] The tool wear prediction model consists of a parallel residual network, a multi-head self-attention mechanism, and a stacked bidirectional long short-term memory network.
[0027] a. 24 time-frequency domain features are extracted using conventional feature extraction methods through convolutional layers and max-pooling layers;
[0028] b. The extracted conventional features are fed into a parallel residual network, where each branch uses convolutional kernels of different sizes to extract multi-scale local features;
[0029] In its first three branches, a multi-head self-attention mechanism is added to replace the convolutional layer as the first weight layer of the residual network in order to extract local features;
[0030] The fourth branch consists of convolutional layers, which act as short connections in the parallel network and are responsible for global feature extraction.
[0031] c. Global and local features are input into a stacked bidirectional long short-term memory network for sequential feature extraction; finally, they are fed into a fully connected layer to complete the training process and output the training results.
[0032] 5) Using the method in step 1), collect CNC machine tool machining data and use the trained tool wear prediction model to perform real-time tool wear prediction.
[0033] The weighting function formula mentioned in step 3) is:
[0034] (3)
[0035] In the formula, for In With target observations The distance between them The maximum value of this distance; the selected neighboring observations In the diagram, the position with the largest distance has a weight of 0, while the position with the smallest distance has a weight of 0. Its own weight is 1;
[0036] Step 2) The vibration signal data of the front end of the spindle is first decomposed into 9 intrinsic mode functions and 1 residual component using empirical mode decomposition. Discrete wavelet transform is performed only on the high-frequency intrinsic mode functions. The high-frequency intrinsic mode functions are then decomposed into three-level wavelet based on the sym3 wavelet basis function and quantized using the soft thresholding method.
[0037] Step 2) The main shaft AC current signal data is quantized using a soft thresholding method after undergoing three-level wavelet decomposition based on the sym3 wavelet basis function.
[0038] The vibration signal data at the front end of the spindle mentioned in step 1) is collected by an accelerometer installed on the front end of the spindle;
[0039] The spindle AC current signal data mentioned in step 1) is collected by a current sensor installed on the spindle motor.
[0040] The time-frequency domain features mentioned in step 3) include: arithmetic mean, peak factor, peak-to-peak value, margin factor, root mean square, waveform factor, standard deviation, centroid frequency, skewness, root mean square frequency, skewness, and frequency standard deviation.
[0041] This invention provides a small-sample tool wear prediction method based on a parallel multi-head residual memory network (PMS) for CNC machine tool wear monitoring and prediction. First, methods such as discrete wavelet transform, empirical mode decomposition, and local weighted regression are used to improve data quality. Then, data augmentation and spline interpolation are used to expand the dataset to address the problem of insufficient sample size. Second, the PMS integrates a parallel residual network (PResNet), a multi-head self-attention (MHSA) mechanism, and a stacked bidirectional long short-term memory network (SBiLSTM), enabling adaptive extraction of multi-scale features from the data and capturing time-frequency domain features related to tool wear evolution. Finally, the fully connected layer outputs the predicted tool wear value. Experimental results verify the effectiveness of the method, showing that the system maintains high prediction accuracy even with a small number of training samples. This framework provides a valuable reference for tool wear monitoring in practical industrial environments.
[0042] In this invention, traditional short connections are further improved. Inside the parallel network, a multi-head self-attention mechanism is used instead of convolutional layers as the first weight layer of the parallel residual network. This gives the model a larger receptive field and enhances its global perception capability. Outside the parallel network, convolutional layers are used as short connections. This allows information to flow freely and preserves the enhanced details of the input data. The existence of internal and external short connections further optimizes the training process, enabling the network to approach the global optimum more quickly and reducing the difficulty of network training, thereby improving prediction accuracy.
[0043] Compared with existing technologies and methods, the present invention has the following advantages:
[0044] (1) In terms of data preprocessing, methods such as discrete wavelet transform, empirical mode decomposition, time-frequency domain feature extraction and local weighted regression are combined to improve data quality and thus improve the training accuracy of the model under small sample size.
[0045] (2) Use data augmentation and spline interpolation to expand the dataset, alleviate the problem of insufficient training samples, and reduce the cost of model training data in actual industrial production.
[0046] (3) In the model design, PMS combines a parallel residual network with a multi-head self-attention mechanism to adaptively extract multi-scale local features from the data. Then, conventional convolutional layers are applied to supplement global features. Stacked bidirectional long short-term memory networks are used to capture the sequential features related to tool wear evolution, which fully solves the problem of insufficient model accuracy under small sample input and has great reference value for practical production applications. Attached Figure Description
[0047] Figure 1 This is a flowchart illustrating the tool wear prediction method based on a parallel multi-head residual memory network according to the present invention.
[0048] Figure 2 This is a diagram of the parallel multi-head residual memory network structure.
[0049] Figure 3 A schematic diagram of the signal acquisition setup.
[0050] Figure 4 This refers to the noise reduction process for vibration signal data at the front end of the main spindle.
[0051] Figure 5 This is the noise reduction process for the main spindle AC current signal data.
[0052] Figure 6 A bar chart comparing the tool wear prediction results of the proposed parallel multi-head residual memory network and common neural networks.
[0053] Figure 7 This is a comparison chart of the predicted tool wear results and the actual results for sample 1.
[0054] Figure 8 This is a comparison chart of the predicted tool wear results and the actual results for sample 2. Detailed Implementation
[0055] Example 1:
[0056] Please see Figure 1-5 A small-sample tool wear prediction method based on parallel multi-head residual memory networks includes:
[0057] 1) Collect CNC machine tool machining data:
[0058] The vibration signal data of the spindle front end and the AC current signal data of the spindle are collected during the CNC machine tool machining process. The tool wear is measured with a microscope and the above data are saved at equal time intervals.
[0059] 2) Perform noise reduction preprocessing on the spindle front end vibration signal data and spindle AC current signal data obtained in step 1):
[0060] First, depending on the operating conditions, approximately 2.5% of the invalid portions at the beginning and end of each spindle front-end vibration signal data and spindle AC current signal data are removed. Then, noise reduction processing is performed. Empirical Mode Decomposition (EMD) is used to decompose the spindle front-end vibration signal data, and Discrete Wavelet Transform (DWT) is used to process the high-frequency intrinsic mode functions. The signal is then reconstructed using a soft thresholding function. The spindle AC current signal data is directly processed using Discrete Wavelet Transform and reconstructed using a soft thresholding function.
[0061] 3) Further process the vibration signal data of the noise-reducing spindle front end, the AC current signal data of the noise-reducing spindle, and the tool wear data obtained in step 2) and step 1):
[0062] a. For the vibration signal data and AC current signal data of the noise-reducing spindle front end, each cutting process is divided into multiple smaller cutting sequences to obtain enhanced input data. For the tool wear data, cubic spline interpolation is used to process the data, thereby increasing the data sample size. The mathematical expression for cubic spline interpolation is shown below:
[0063] (1)
[0064] in, It is an interval cubic spline polynomial on, yes The function value at that point, These are the coefficients of the first derivative. These are the coefficients of the second derivative. Third derivative coefficients.
[0065] b. Perform feature value calculation on the enhanced input data and extract a total of 12 types and 24 time-frequency domain feature values;
[0066] c. Process the 24 time-frequency domain feature values obtained in step b) using local weighted regression:
[0067] Local weighted regression is applied to 24 time-frequency domain features to eliminate periodic fluctuations and outliers. The mathematical expression for local weighted regression is shown below:
[0068] (2)
[0069] in, It is a prediction point The predicted value, These are the response values of the training data points. It is a weighting function that depends on the prediction point. and training data points The distance between them. In this study, a cubic function is chosen as the weighting function:
[0070] (3)
[0071] In the formula, for In With target observations The distance between them This is the maximum value of that distance. The selected neighboring observations. In the diagram, the position with the largest distance has a weight of 0, while the position with the smallest distance has a weight of 0. Its own weight is 1.
[0072] 4) Training a tool wear prediction model using a parallel multi-head residual memory network includes the following sub-steps:
[0073] a. Establish a tool wear prediction model, which consists of a parallel residual network, a multi-head self-attention mechanism, and a stacked bidirectional long short-term memory network;
[0074] b. First, 24 time-frequency domain features are extracted using conventional features through convolutional and max-pooling layers. Then, the extracted features are fed into a parallel residual network. Each branch uses convolutional kernels of different sizes to extract multi-scale local features. In the first three branches of the parallel residual network, a multi-head self-attention mechanism replaces the convolutional layers as the first weight layer to extract local features. The fourth branch consists of convolutional layers responsible for global feature extraction. After this, the global and local features are fed into a stacked bidirectional long short-term memory network for sequential feature extraction. Finally, the results are passed to a fully connected layer to complete the training process and output the training results.
[0075] Experimental example:
[0076] Please see Figure 1-8 .
[0077] I. Data Collection:
[0078] Experimental parameters: Data acquisition experiments were conducted on a high-speed CNC milling machine; the spindle speed was set to 826 r / min, and the feed rate was 1555 mm / min. Depending on the experimental conditions, the depth of cut was set to 1.5 mm and 0.75 mm, and the feed rate was set to 0.5 mm / r and 0.25 mm / r, respectively. Two types of materials were used in the experiment: cast iron and J45 stainless steel. In the experiment, a 70 mm face cutter with 6 inserts was used as the cutting tool to perform dry milling on a workpiece with dimensions of 483 mm x 178 mm x 51 mm.
[0079] Sensor setup: Tool wear was measured using a microscope. An accelerometer mounted on the spindle tip and a CTA 213 current sensor mounted on the spindle motor were used to capture the spindle tip vibration and spindle AC current signals, respectively. Experimental results were stored as a 1x167 MATLAB structure array.
[0080] II. Data Denoising:
[0081] Spindle front-end vibration signal data: First, empirical mode decomposition (EMD) is used to decompose the spindle front-end vibration signal data into 9 intrinsic mode functions (IMFs) and 1 residual component. To avoid signal distortion, discrete wavelet transform is performed only on the high-frequency IMFs. The high-frequency IMFs are then subjected to three-level wavelet decomposition based on the sym3 wavelet basis function, followed by soft thresholding quantization. Finally, the signal is reconstructed to obtain the denoised spindle front-end vibration signal data.
[0082] Spindle AC current signal: For the spindle AC current signal, only discrete wavelet transform is used for processing. After three-level wavelet decomposition based on sym3 wavelet basis function, the denoised spindle AC current signal is obtained by soft thresholding and reconstruction.
[0083] III. Data Processing:
[0084] 1) Data expansion
[0085] This experiment uses two working conditions: Sample 1 and Sample 2. For Sample 1, there are a total of 13 tool passes. Due to the relatively small number of tool passes, the sample data of one tool pass is divided into sample data of eight tool passes, and the data dimension is changed from (1, 4800, 24) to (8, 600, 24), thus obtaining 8 times the sample data. The number of tool passes in the sample data is expanded from 13 to 104. Accordingly, 7 tool wear values are inserted between every 2 tool wear values. For Sample 2, there are a total of 20 tool passes. The data dimension of each tool pass is (1, 5000, 24), which is changed to (5, 1000, 24) after data augmentation. The number of tool passes in the sample data is expanded from 20 to 100. Spline interpolation is performed on the tool wear values, and 4 tool wear values are inserted between every 2 tool wear values.
[0086] 2) Time-frequency domain feature extraction
[0087] Twelve time-frequency domain features, including mean square error and skewness, were extracted from the processed spindle front-end vibration signal data and spindle AC current signal data, respectively, resulting in a total of 24 time-frequency domain feature values, which were used as input data for the model.
[0088] Table 1: 24 Time-Frequency Domain Features
[0089]
[0090] 3) Locally weighted regression
[0091] By observing the extracted time-frequency domain feature values, it was found that the time-frequency domain feature values of the spindle AC current signal and the vibration signal data at the front end of the spindle exhibit significant periodic fluctuations. Furthermore, the time-frequency domain feature values of the vibration signal data at the front end of the spindle also show numerous abrupt changes, which significantly impacts the model's accuracy and convergence speed. Therefore, a local weighted regression was performed on the time-frequency domain feature values. By appropriately setting the parameters, data quality was improved and the model convergence speed was accelerated while preserving data characteristics.
[0092] IV. Network Setup:
[0093] A parallel multi-head residual memory network is established as a predictive model for tool wear.
[0094] First, 24 time-frequency domain features are extracted using conventional features through convolutional and max-pooling layers. Then, the extracted features are fed into a parallel residual network. The first three branches of the parallel residual network use convolutional kernels of different sizes (3×3, 5×5, 7×7) to extract multi-scale local features.
[0095] Secondly, a multi-head self-attention mechanism is added to the first three branches of the parallel residual network to replace the convolutional layers as the first weight layer for extracting local features. The fourth branch consists of convolutional layers and is responsible for global feature extraction. By combining local and global features, a more comprehensive representation of the data is achieved. In the multi-head self-attention mechanism, the input data is processed through multiple self-attention layers, with each input data corresponding to an output value. The data is refined and decomposed, different features are assigned different weights, and the final output is obtained through a linear layer transformation after concatenating the first and last features.
[0096] Finally, the branches of the parallel residual network are connected into a sequence to achieve multi-feature fusion. The fused features are then fed into a stacked bidirectional long short-term memory network for sequential feature extraction.
[0097] V. Comparison of Experimental Results
[0098] To verify the effectiveness and accuracy of the proposed model, we trained conventional machine learning models and mainstream deep learning models on the same dataset and compared their prediction performance with that of the proposed method.
[0099] SVR: Support Vector Regression.
[0100] ResNet: Residual Network.
[0101] ResNet–SBiLSTM: A hybrid model consisting of a residual network and a stacked bidirectional long short-term memory network.
[0102] PCNN–SBiLSTM: A hybrid model consisting of parallel convolutional networks and stacked bidirectional long short-term memory networks.
[0103] PRes–SBiLSTM: A hybrid model consisting of a parallel residual network and a stacked bidirectional long short-term memory network.
[0104] Parallel Multi-Head Residual Memory Network (PMS): The model proposed in this invention.
[0105] Table 2 compares the prediction results and evaluation metrics of each model in different experiments:
[0106]
[0107] In three experiments, 24 time-frequency domain feature values of the spindle AC current signal and the vibration signal of the spindle front end were used as inputs for each model. Tool wear is a continuously changing process, and there must be a relationship between the preceding and following time-series information. Therefore, the preceding and following time-series information of the data is very important for accurate prediction.
[0108] In processing the temporal information of the data, the parallel multi-head residual memory network model shows a significant performance advantage over SVR, ResNet, ResNet-SBiLSTM and PCNN-SBiLSTM models.
[0109] Parallel multi-head residual memory (MMRM) models can efficiently capture long-term dependencies and complex contextual information in time series data, a capability lacking in SVR models. Compared to ResNet and ResNet-SBiLSTM models, MMRM, by introducing a multi-head self-attention mechanism, can more flexibly identify key information in the sequence, thereby enhancing the model's feature learning ability. Furthermore, compared to PCNN-SBiLSTM models, MMRM exhibits superior performance when handling nonlinear relationships and dynamically changing time series data, and can more effectively adapt to the complexity of time series data.
[0110] Therefore, when modeling complex time series tasks, parallel multi-head residual memory networks typically offer higher prediction accuracy and superior performance. In experiments, comparisons with SVR, ResNet, ResNet-SBiLSTM, and PCNN-SBiLSTM models revealed that the proposed model outperforms these models in both RMSE and MAE, demonstrating its extremely high accuracy.
[0111] This invention presents a model that integrates a multi-head self-attention mechanism, a parallel residual network, and a stacked bidirectional long short-term memory network. It can handle multi-source data and leverages the powerful feature extraction, feature fusion, and abstract generalization capabilities of deep learning. This overcomes the limitations of traditional tool wear prediction methods, which require large datasets and suffer from high sensor costs. It is of great significance for ensuring workpiece quality, improving processing efficiency, and promoting automated and intelligent machining.
Claims
1. A small sample tool wear prediction method based on a parallel multi-head residual memory network, comprising: 1) Collecting numerical control machine tool processing data: Collect spindle front-end vibration signal data and spindle alternating current signal data in the processing of the numerical control machine tool, measure tool wear data with a microscope, and save the above data at equal time intervals as initial data for model training; 2) Noise reduction preprocessing: Delete the invalid part at the beginning and end of each spindle front-end vibration signal data and spindle alternating current signal data; Spindle front-end vibration signal data: use empirical mode decomposition to decompose the spindle front-end vibration signal data, use discrete wavelet transform to process the high-frequency intrinsic mode function, and then use a soft threshold function to reconstruct the signal to obtain the denoised spindle front-end vibration signal data; Spindle alternating current signal data: directly use discrete wavelet transform and a soft threshold function to reconstruct to obtain the denoised spindle alternating current signal data; 3) Further processing of the denoised spindle front-end vibration signal data, the denoised spindle alternating current signal data, and the tool wear data obtained in step 1): a. For the denoised spindle front-end vibration signal data and the denoised spindle alternating current signal data, each cutting process is divided into multiple smaller cutting sequences to obtain enhanced input data; For tool wear data, use cubic spline interpolation for processing; the mathematical expression of cubic spline interpolation is as follows: (1) wherein is a cubic spline polynomial on the interval is a function value at is a first derivative coefficient, is a second derivative coefficient, is a third derivative coefficient; b. Calculate the eigenvalues of the enhanced input data, extract the time-frequency domain features, and obtain the time-frequency domain eigenvalues; c. Use local weighted regression to process the time-frequency domain eigenvalues obtained in step b: The mathematical expression of the local weighted regression is as follows: (2) wherein is a prediction point , is a response value of a training data point, is a weight function that depends on the distance between the prediction point and the training data point ; 4) Training a tool wear prediction model by means of a parallel multi-head residual memory network: The tool wear prediction model is composed of a parallel residual network, a multi-head self-attention mechanism, and a stacked bidirectional long short-term memory network; a. The time-frequency domain eigenvalues are subjected to regular feature extraction through a convolution layer and a max-pooling layer; b. The extracted regular features are input into the parallel residual network, and different sizes of convolution kernels are used in each branch to extract multi-scale local features; The multi-head self-attention mechanism is added to the first three branches to replace the convolution layer as the first weight layer of the residual network to extract local features; The fourth branch is composed of a convolution layer, which is a short connection of the parallel network and is responsible for global feature extraction; c. The global features and local features are input into the stacked bidirectional long short-term memory network for forward and backward time sequence feature extraction; finally, they are input into a fully connected layer to complete the training process and output the training results; 5) Collecting numerical control machine tool processing data using the method of step 1), using the trained tool wear prediction model to perform real-time tool wear prediction.
2. The small sample tool wear prediction method based on a parallel multi-head residual memory network according to claim 1, characterized in that: The weight function formula in step 3) is: (3) wherein is the distance between the target observation and the selected neighboring observation , and is the maximum value of the distance. The selected neighboring observation with the largest distance has a weight of 0, while the position with the smallest distance, i.e. the closest observation , has a weight of 1.
3. The small sample tool wear prediction method based on a parallel multi-head residual memory network according to claim 2, characterized in that: Step 2) The spindle front-end vibration signal data is first decomposed into 9 intrinsic mode functions and 1 residual component using empirical mode decomposition, and only the high-frequency intrinsic mode functions are subjected to discrete wavelet transform; the high-frequency intrinsic mode functions are subjected to three-layer wavelet decomposition based on a sym3 wavelet basis function, and then subjected to quantization processing using a soft threshold method.
4. The small sample tool wear prediction method based on the parallel multi-head residual memory network according to claim 3, characterized in that: Step 2) The spindle AC current signal data is subjected to three-layer wavelet decomposition based on a sym3 wavelet basis function, and then subjected to quantization using a soft threshold method.
5. The small sample tool wear prediction method based on the parallel multi-head residual memory network according to claim 4, characterized in that: Step 1) The spindle front-end vibration signal data is collected by an accelerometer installed at the front end of the spindle; The spindle AC current signal data is collected by a current sensor installed on the spindle motor.
6. The small sample tool wear prediction method based on the parallel multi-head residual memory network according to claim 5, characterized in that: Step 3) The time-frequency domain features include: arithmetic mean, peak factor, peak-to-peak value, margin factor, root mean square, waveform factor, mean square deviation, barycenter frequency, skewness, root mean square frequency, kurtosis, and frequency standard deviation.