A Tool Wear Prediction Method Based on a Multi-Scale DenseNet-ResNet-GRU Model

Abstract

The present invention proposes a tool wear amount prediction method based on a multi-scale DenseNet-ResNet-GRU model, which avoids the problems of inaccurate judgment of wear amount based on subjective experience and the time-consuming and laborious extraction of signal features by humans in machine learning. The method includes the following steps: selecting cutting force signals to predict the tool wear amount; performing signal preprocessing, using the Grubbs criterion method to remove data that deviates far in the dataset, reducing the influence of outliers on the overall data; using wavelet threshold denoising and applying the progressive semi-soft threshold function to the cutting force signal denoising process; constructing the denoised signal into an input matrix; constructing a multi-scale DenseNet-ResNet-GRU model, using convolutional kernels of different scales to extract features, and using the different advantages of DenseNet and ResNet to extract different spatial features of sample features, and then extracting sample sequence features through GRU, constructing a fully connected layer and a regression layer to predict the tool wear amount.

Description

Technical Field

[0001] The present invention relates to the field of machining, and in particular to a method for predicting tool wear based on a multi-scale DenseNet-ResNet-GRU model. Background Art

[0002] A tool is one of the important components of a numerically controlled machine tool. Tool wear is inevitable during the cutting process, and its wear amount gradually increases over time and finally develops into severe wear and then fails.

[0003] Especially during the milling process, the milling cutter rotates at a high speed and makes discontinuous contact with the workpiece, which is very likely to cause tool wear.

[0004] Tool wear will cause the quality of the machined workpiece to decline. When the wear is severe, phenomena such as fracture and chipping may occur, which can damage the machine tool and even cause safety accidents.

[0005] Research shows that 20% of the downtime is caused by tool wear or tool breakage.

[0006] The traditional tool wear state is judged by workers based on experience, mainly by judging the tool wear degree through factors such as the color and state of the chips during the cutting process and the noise and vibration during the machining process.

[0007] Judging the wear amount based on personal experience alone is not accurate enough. If the tool is replaced before reaching the dulling standard, it may lead to insufficient utilization of the tool and increase the manufacturing cost; if the tool has reached the dulling standard but is not replaced in time, it will affect the workpiece quality and cause economic losses.

[0008] Therefore, on-line prediction of tool wear amount is of great significance in the machining process.

[0009] At present, the tool wear prediction methods can be divided into direct methods and indirect methods.

[0010] The direct methods mainly include optical measurement, ray measurement, resistance measurement, and computer image processing, etc.

[0011] However, the direct monitoring method requires repeated disassembly of the tool for off-line monitoring, which affects the machining efficiency. Therefore, it cannot be widely applied in the machining process.

[0012] The indirect monitoring method is to install various sensors on the machine tool or the workpiece to be machined, collect the sensor signals related to tool wear, and indirectly realize tool wear monitoring, which can monitor tool wear during the tool cutting process.

[0013] The indirect method constructs a non-linear mapping model between the tool wear amount and the monitoring signals during the cutting process through signal processing technology and machine learning algorithms or deep learning algorithms to monitor or predict the tool wear amount.

[0014] Deep learning can overcome the problem that machine learning requires manual extraction of tool wear characteristics, relies heavily on the expertise of experts in related fields, is time-consuming and labor-intensive, reduces the efficiency of tool wear prediction, and makes it difficult to control processing and manufacturing costs. Summary of the Invention

[0015] The purpose of the present invention is to address the deficiencies in the prior art and propose a tool wear prediction method based on a multi-scale DenseNet-ResNet-GRU model. The method mainly solves the problems that the direct method cannot make online predictions and the tool wear prediction based on machine learning cannot adaptively extract features and needs to extract features based on subjective experience, which is time-consuming and labor-intensive.

[0016] To achieve the above objectives, the technical solution of the present invention is as follows: The present invention proposes a tool wear prediction method based on a multi-scale DenseNet-ResNet-GRU model, which adopts the following steps:

[0017] S1: Obtain the three-dimensional cutting force signal data during the cutting process and record the flank wear width VB value after each pass.

[0018] S2: Data preprocessing is performed on the original force signal, including outlier processing in the data and denoising of the processed data; outlier data in the data are removed using the Grubbs criterion method, and linear interpolation of adjacent non-outlier values is used for filling.

[0019] S3: The cutting force signal is denoised using the wavelet threshold method and a progressive semi-soft threshold function. First, the input signal is subjected to a three-layer wavelet transform to obtain the third-layer approximation coefficient and each layer detail coefficient. The threshold is determined using the Birge-Massart strategy. The wavelet coefficients are then modified using a progressive semi-soft threshold function. Finally, the denoised signal is obtained through an inverse wavelet transform. The calculation formula using the progressive semi-soft threshold function is as follows:

[0020]

[0021] Where: ω j,k is the wavelet coefficient, λ is the set threshold, is the wavelet coefficient after threshold function processing.

[0022] S4: Normalize the denoised cutting force data and construct it into an input matrix; the normalization formula is as follows:

[0023] y=(xx min ) / (x max -x min )

[0024] Where: xmax is the maximum value in the data, x min is the minimum value in the feature data, x is the data to be normalized; construct a training set and a test set according to the normalized data; use the training set to train the tool wear prediction model, and use the test set to verify the trained model.

[0025] S5: Construct a multi-scale DenseNet-ResNet-GRU model;

[0026] S5.1: This model is mainly composed of DenseNet, ResNet, GRU and a fully connected layer regression layer, where DenseNet and ResNet are connected in parallel;

[0027] S5.2: DenseNet and ResNet each have two parallel channels. The two parallel channels respectively use convolutional kernels with scales of 3×1 and 5×1 to perform convolutions on the input matrix at different scales, and then are connected to the DenseNet module or ResNet module through the BN layer and the max pooling layer;

[0028] S5.3: Each DenseNet channel has two DenseNet modules, and each ResNet channel has two ResNet modules;

[0029] S5.4: The DenseNet module first sets a bottleneck layer, including a BN layer, a ReLU activation function, and a convolutional layer with a convolutional kernel of 1×1; then sets the BN layer again, and after being activated by the ReLU activation function, connects a convolutional layer with a convolutional kernel of 3×3; the last layer in the DenseNet module is connected (concat) with the input layer in the channel dimension; between the two DenseNet modules, they are connected through a Transition layer, and the Transition layer can reduce the size of the feature map; the Transition layer includes a convolutional layer with a convolutional kernel of 1×1 and an average pooling layer of 2×2;

[0030] S5.5: The first ResNet module in each ResNet channel includes two branches. The main branch is composed of two convolutional layers and two BN layers alternately stacked, using ReLU as the activation function, and the branch branch is the input tensor of the ResNet module; the main branch of the second ResNet module in each ResNet channel is composed of two convolutional layers and a BN layer alternately stacked, using ReLU as the activation function, and the branch branch is composed of a convolutional layer and a BN layer stacked; after the tensor input to the residual block passes through the main path and the branch path, the two new tensors are added, and finally ReLU is used as the activation function to form the output of a residual block;

[0031] S5.6: The global average pooling layer is adopted for the last layer of each channel, and the outputs of the four channels are subjected to feature fusion through the concatenation layer to obtain multi-scale features; the convolutional layer is used for feature extraction to obtain a feature matrix; the pooling layer is used to reduce the feature matrix;

[0032] S5.7: The deep GRU network is concatenated, and the deep GRU network is constructed to extract features of different time scales. The deep GRU network includes the first-layer GRU network and the second-layer GRU network;

[0033] S5.8: A fully connected layer is constructed after the deep GRU network, and a regression layer is constructed based on the output result of the fully connected layer;

[0034] S5.9: The Adam algorithm is used to automatically obtain the influence weights of these feature components on the tool wear amount, and then a model from the signal to the wear amount is established according to the method of supervised learning, and the prediction of the tool wear amount can be realized.

[0035] S6: In order to verify the effectiveness of the model proposed by the present invention, three performance indicators are used to evaluate the model, including the mean absolute error (MAE), the root mean square error (RMSE), and the mean absolute percentage error (MAPE); the calculation formulas of these three performance indicators are as follows:

[0036]

[0037]

[0038]

[0039] In the formula: N is the number of test data, X(t) is the actual value, is the predicted value; the smaller the values of the three indicators, the higher the prediction accuracy.

[0040] The beneficial effects of the present invention are as follows:

[0041] The present invention adopts the Grubbs criterion method to eliminate the data in the dataset that deviates far, which can reduce the influence of outliers on the overall data.

[0042] The progressive semi-soft threshold function is used to denoise the cutting force signal. This function is continuous in the wavelet domain, which can make up for the problems of poor continuity of the hard threshold function and loss of signal features caused by the constant deviation of the soft threshold function.

[0043] The present invention collects the cutting force signal to predict the tool wear amount of the numerical control machine tool, and normalizes the preprocessed cutting force signal, which can improve the accuracy of the prediction result.

[0044] The present invention can adaptively extract features by constructing a multi-scale DenseNet-ResNet-GRU model, without the need to extract signal features based on subjective experience. The preprocessed signal can be directly input into the model, which has universal applicability.

[0045] The present invention extracts features from the input matrix using convolution kernels of different scales, which can improve the robustness of the model and the representation ability of signal features.

[0046] Based on the feature connection of DenseNet in the channel (dimension) and the characteristics of ResNet to slow down the network degradation speed, the present invention utilizes the different advantages of the two networks to extract different spatial features of the samples; then extracts the sample sequence features through a deep GRU, constructs a fully connected layer based on the deep GRU model, and maps the features to the sample label space; finally, constructs a regression layer after the fully connected layer to predict the tool wear amount; this model has relatively superior performance, better training effects and higher prediction accuracy.

[0047] The present invention adds a batch normalization layer, also known as the BN (Batch Normalization) layer, to the model. After the input data passes through one or more layers of operations of the neural network, the distribution of the data may no longer follow a normal distribution. Adding a BN layer to the hidden layer can make the distribution of the data follow a normal distribution again; therefore, it can improve the robustness of the network model and has many advantages such as preventing overfitting and accelerating the model convergence speed. Description of the Drawings

[0048] Figure 1 is a flowchart of the method of the present invention;

[0049] Figure 2 is a comparison diagram of the progressive semi-soft threshold function and the soft and hard threshold functions;

[0050] Figure 3 is a schematic diagram of the multi-scale DenseNet-ResNet-GRU model;

[0051] Figure 4 is a structural diagram of the DenseNet module;

[0052] Figure 5 is a structural diagram of the ResNet module;

[0053] Figure 6 is a schematic diagram of the deep GRU network;

[0054] Figure 7 is a structural diagram inside a single GRU;

[0055] Figure 8 is a structural schematic diagram of the tool wear state prediction device in the embodiment;

[0056] Figure 9 is the original signal;

[0057] Figure 10 is the signal after outlier processing;

[0058] Figure 11 is the prediction diagram of tool wear amount by the CNN-GRU model;

[0059] Figure 12 is the prediction diagram of tool wear amount by the multi-scale DenseNet-ResNet-GRU model. Specific implementation manner

[0060] To describe and explain the technical content and effects of the present invention in detail, the following is described and explained in conjunction with the accompanying drawings and examples. It should be understood that the embodiments described herein are only used to illustrate and explain the present invention and are not used to limit the present invention.

[0061] As Figure 1 , in the present invention, a method for predicting tool wear amount based on a multi-scale DenseNet-ResNet-GRU model is provided. This method mainly predicts the tool wear amount by collecting the cutting force signal on the numerical control machine tool through a force sensor. The specific steps are as follows:

[0062] S1: Install a dynamometer on the fixture of the numerical control machine tool workbench, and use the tool to process the workpiece; use a data acquisition card to collect the original signal data of the X, Y, and Z three-direction cutting forces during the whole life cycle of the tool processing the workpiece; after each feed is completed, observe and record the flank wear width and VB value of the tool with an electron microscope as the tool wear amount.

[0063] S2: Perform data preprocessing on the original cutting force data, including outlier processing and signal denoising in the data. The outlier processing is as follows:

[0064] (1) For the signal X n (n = 1, 2, 3, ···, N), establish the Grubbs test statistic formula as follows:

[0065]

[0066] In the formula: x represents the sample mean, σ represents the sample standard deviation, and N represents the sample length;

[0067] (2) Set the significance level to α, and determine g0(N, α) according to the probability formula P{g n ≥ g0(N, α)} = α, where g0(N, α) represents the Grubbs critical value corresponding to the data length of N and the significance level of α. In the present invention, α = 0.05;

[0068] (3) If then eliminate x b , and fill it with linear interpolation of adjacent non-outliers.

[0069] S3: Denoise the cutting force signal X(k) using the wavelet threshold method. First, use the Mallat algorithm, and the formula is as follows:

[0070]

[0071] In the formula: i = 1, 2, 3, ···, m, j = 1, 2, 3, ···, J, m is the number of sampling points in the sampling sequence X(k), J is the number of wavelet transform layers, C j+1,m and D j+1,m are the approximation coefficient and the detail coefficient obtained by the wavelet transform of the j-th layer respectively; h * (k - 2m) and g * (k - 2m) are the low-pass filter coefficient and the high-pass filter coefficient respectively.

[0072] Perform three-layer wavelet transform on the signal through the db3 wavelet basis function to obtain the approximation coefficient and the detail coefficients of each layer; then use the progressive semi-soft threshold function, and the formula is as follows:

[0073]

[0074] In the formula: ω j,k is the wavelet coefficient, λ is the set threshold, is the wavelet coefficient after being processed by the threshold function. The comparison between the hard and soft threshold functions and the progressive semi-soft threshold function is as Figure 2 shown; correct the approximation coefficient and the detail coefficients of each layer, and finally obtain the denoised signal through the inverse wavelet transform.

[0075] The wavelet reconstruction formula is as follows:

[0076]

[0077] Introduce the infinite matrix, and the formula is as follows:

[0078]

[0079]

[0080] In the formula: H m,k = h * (k - 2m), G m,k = g * (k - 2m), then the above formula can be respectively expressed as:

[0081]

[0082] C j = H * C j+1 + G * D j+1 j = J, J - 1, ···, 1, 0

[0083] Where: H * and G * are the conjugate transpose matrices of H and G respectively; the threshold λ is determined by the Birge - Massart strategy.

[0084] S4: Normalize the pre - processed cutting force data and construct it into an input matrix; the normalization formula is as follows:

[0085] y = (x - x min ) / (x max - x min )

[0086] Where: x max is the maximum value in the data, x min is the minimum value in the characteristic data, and x is the data to be normalized; divide the normalized data set and the corresponding wear amount labels into a training set and a test set; use the constructed training set to train the tool wear amount prediction model, and use the test set to verify the trained model.

[0087] S5: Construct a multi - scale DenseNet - ResNet - GRU model, as Figure 3 shown;

[0088] S5.1: This model is mainly composed of DenseNet, ResNet, GRU, a fully - connected layer, and a regression layer. First, construct DenseNet and ResNet, and DenseNet and ResNet are connected in parallel;

[0089] S5.2: Each of DenseNet and ResNet has two parallel channels. The two parallel channels respectively use convolutional kernels with scales of 3×1 and 5×1 to perform convolutions on the input matrix at different scales, and then are connected to the DenseNet module or ResNet module through the BN layer and the max - pooling layer;

[0090] S5.3: Each DenseNet channel has two DenseNet modules, and each ResNet channel has two ResNet modules.

[0091] S5.4: Refer to Figure 4, the DenseNet module first sets up a bottleneck layer, which includes a BN layer, a ReLU activation function, and a convolutional layer with a kernel size of 1×1, which can play a role in dimensionality reduction and improving computational efficiency; then, a BN layer is set up again, and after activation by the ReLU activation function, it is connected to a convolutional layer with a kernel size of 3×3; each layer will be concatenated (concat) with all the previous layers in the channel dimension; between two DenseNet modules, they are connected through a Transition layer, and the Transition layer can reduce the size of the feature map; the Transition layer includes a convolutional layer with a kernel size of 1×1 and an average pooling layer of 2×2; the formula for the spatial feature extraction process of the DenseNet module is as follows:

[0092] x l =H l ([x0,x1,···,x l-1 )

[0093] In the formula: x l represents the extracted spatial feature, [x0,x1,···,x l-1 represents the concatenation of the output feature maps of layers 0 to l-1 in the channel dimension, and H l (·) represents a non-linear transformation function, which is a combined operation that may include a series of BN (Batch Normalization), ReLU, pooling, and convolutional operations; between two DenseNet modules, they are connected through a Transition layer, and the Transition layer can reduce the size of the feature map; the Transition layer includes a convolutional layer with a kernel size of 1×1 and an average pooling layer of 2×2.

[0094] S5.5: Refer to Figure 5 , the first ResNet module of each ResNet channel includes two branches. The main branch consists of two convolutional layers and two BN layers stacked alternately, using ReLU as the activation function, and the branch is the input of this ResNet module; the main branch of the second ResNet module of each ResNet channel consists of two convolutional layers and a BN layer stacked alternately, using ReLU as the activation function, and the branch consists of a convolutional layer and a BN layer stacked; after the tensors of the input residual block pass through the main path and the branch respectively, the two new tensors are added together, and finally ReLU is used as the activation function to form the output of a residual block; the formula for the spatial feature extraction process of the ResNet module is as follows:

[0095]

[0096] In the formula: x is the input of the residual block, y is the output of the residual block, W i and Ws is a convolution operation, and F(x, W i ) is a residual mapping;

[0097] S5.6: At the last layer of each channel, a global average pooling layer is adopted, and the outputs of the four channels are subjected to feature fusion through a concatenation layer to obtain multi-scale features; the convolutional layer is used for feature extraction to obtain a feature matrix; the pooling layer is used to reduce the size of the feature matrix;

[0098] S5.7: Refer to Figure 6 After the concatenation layer, a deep GRU network is cascaded to construct a deep GRU network to extract features of different time scales. The deep GRU network includes a first-layer GRU network and a second-layer GRU network, and the number of GRU neurons is 128; the internal structure of a single GRU is as Figure 7 shown, and its main parameters are as follows: The formula for the reset gate is as follows:

[0099] r t = σ(W r · [h t-1 , x t )

[0100] In the formula: σ is the sigmoid function, W r is the reset gate weight matrix, x t is the input at the t-th time step, that is, the t-th component of the entire input sample sequence, and h t-1 is the hidden information saved at the previous time step;

[0101] The formula for the update gate is as follows:

[0102] z t = σ(W z · [h t-1 , x t )

[0103] In the formula: W z is the update gate weight matrix, and other variables are the same as those of the reset gate;

[0104] The formula for the candidate hidden layer state is:

[0105]

[0106] In the formula: W is the candidate state weight matrix, r t *h t-1 is the element-wise multiplication of the reset gate r t and the hidden information h t-1 at the previous time step;

[0107] The formula for the final hidden state is:

[0108]

[0109] Where: h t is the current state is the candidate hidden state;

[0110] S5.8: Construct a fully connected layer after the deep GRU network, and construct a regression layer based on the output result of the fully connected layer.

[0111] S5.9: Use the Adam algorithm to automatically obtain the influence weights of these feature components on the tool wear amount; then establish a model from the signal to the wear amount according to the method of supervised learning, and the prediction of the tool wear amount can be realized.

[0112] S6: In order to verify the effectiveness of the model proposed by the present invention, three performance indicators are used to evaluate the model, including the mean absolute error (MAE), the root mean square error (RMSE), and the mean absolute percentage error (MAPE); the calculation formulas of these three performance indicators are as follows:

[0113]

[0114]

[0115]

[0116] Where: N is the number of test data, X(t) is the actual value, is the predicted value; the smaller the values of the three indicators, the higher the prediction accuracy.

[0117] Example: Predict the milling cutter wear, and verify the effectiveness and feasibility of the present invention by predicting the tool wear amount.

[0118] I. Adopt the dataset announced by the prediction competition of the tool health state of the high-speed milling machine tool of the American PHM Society (2010). This dataset is divided into 6 groups of experimental data of high-speed CNC milling machine cutters, numbered C1 to C6, and C1 is a milling cutter dataset containing the tool wear amount label; each group of experimental data contains 315 milling records, and each record collects the signal data of the tool under the milling operation through a force sensor, an acceleration sensor and an acoustic emission sensor; the present invention uses the signal collected by the force sensor in C1 and the corresponding tool wear amount label to verify the present invention; the specific installation positions are as Figure 8 shown, and the main equipment and processing parameters used in the experiment are shown in Table 1:

[0119] Table 1 Main equipment and processing parameters of the experiment

[0120]

[0121] II. Outlier handling: The Grubbs criterion method is used to remove the data that deviates far from the dataset, reducing the influence of outliers on the overall data. Taking the partial time-domain signal of Y_Force in the first milling process of the C1 dataset as an example, the data before and after processing is as Figure 9 follows. 10 as shown.

[0122] Wavelet threshold denoising: To test the denoising effect of the progressive semi-soft threshold function on the cutting force signal, taking the y-direction force of the first group of C1 as an example, the cutting force signal is processed by hard threshold and soft threshold wavelet denoising respectively. To visually verify the denoising effect, the signal-to-noise ratio (SNR) and root mean square error (RMSE) are introduced as evaluation indicators to measure the wavelet denoising effect. The definitions of SNR and RMSE are as follows:

[0123]

[0124]

[0125] where s i is the original signal, f i is the reconstructed signal, and N represents the signal length. The denoising effect is shown in Table 2:

[0126] Table 2 Denoising results of the hard threshold method, soft threshold method, and the threshold processing method used in step S3

[0127]

[0128] As can be seen from the table, the progressive semi-soft threshold function has the largest SNR and the smallest RMSE, and the denoising effect is the best.

[0129] III. According to each group of experimental data containing 315 milling records, one milling record is selected as the test set every two milling records, and the rest are used as the training set. The average value of the flank wear of the 3 cutting edges of the milling cutter is selected as the tool wear. The normalized data is input into the multi-scale DenseNet-ResNet-GRU model, and the CNN-GRU model is used for comparison. The prediction results of the CNN-GRU model are as Figure 11 shown, and the prediction results of the multi-scale DenseNet-ResNet-GRU model are as Figure 12 shown. The calculation results of the three indicators are shown in Table 3:

[0130] Table 3 Comparison of tool wear prediction results

[0131]

[0132] IV. The Grubbs criterion method is used to process signal outliers, and the progressive semi-soft threshold function is applied to the noise reduction process of the cutting force signal, which has a high signal-to-noise ratio and a low root mean square error, and can better reduce redundant signals and noise interference in the signal; and a good prediction effect on the tool wear amount is achieved based on the multi-scale DenseNet-ResNet-GRU model; compared with CNN-GRU, the error is smaller and the feasibility is higher.

[0133] The present invention can provide a certain theoretical basis and technical support for the prediction of tool wear amount in numerical control machine tools.

Claims

1. A tool wear amount prediction method based on a multi-scale DenseNet-ResNet-GRU model, the method comprising the following steps: S1: Obtain the original data of the three-directional forces during the tool machining process through the force sensor data, and construct the original force signal samples; S2: Perform data preprocessing on the original force signal, including the processing of outliers in the data and the denoising of the processed data; For the processing of outliers, the Grubbs criterion method is used to remove the data with large deviations in the data; S3: Perform wavelet threshold denoising processing on the cutting force signal after the outlier processing; S4: Perform normalization processing on the data, and the normalization formula is as follows: y=(x - x min ) / (x max - x min ) Where: x max is the maximum value in the data, x min is the minimum value in the feature data, and x is the data to be normalized; S5: Input the normalized cutting force data into the multi-scale DenseNet-ResNet-GRU model. In the multi-scale DenseNet-ResNet-GRU model, DenseNet and ResNet are connected in parallel. Convolutions of different scales are performed on the input data matrix to extract features, and different spatial features of the samples are extracted by using the different advantages of DenseNet and ResNet. Then, the sample sequence features are extracted by GRU, and a fully connected layer and a regression layer are constructed to map the features to the sample label space, complete the prediction of the tool wear amount, and obtain the tool wear amount; S6: Use the mean absolute error, root mean square error, and mean absolute percentage error to verify the effectiveness of the proposed model.

2. The tool wear prediction method based on the multi-scale DenseNet-ResNet-GRU model according to claim 1, characterized in that: The specific operation method of S1 is as follows. Based on the tool, establish a three-dimensional coordinate system, and collect the cutting force signal data in the x, y, and z directions of the real-time numerical control machine tool.

3. The tool wear prediction method based on the multi-scale DenseNet-ResNet-GRU model according to claim 1, characterized in that: The specific operation method of S3 is as follows. Perform three-layer wavelet transform on the input signal through the db3 wavelet basis function to obtain the third-layer approximation coefficient and the detail coefficients of each layer. The progressive semi-soft threshold function formula is as follows: where: ω j,k is the wavelet coefficient, λ is the set threshold value, is the wavelet coefficient after being processed by the threshold function; the wavelet coefficient ω is corrected to obtain the wavelet coefficient Then perform the inverse wavelet transform to finally obtain the denoised reconstructed signal.

4. The tool wear prediction method based on the multi-scale DenseNet-ResNet-GRU model according to claim 3, wherein: The threshold λ is the threshold determined by the Birge-Massart strategy.

5. The tool wear prediction method based on the multi-scale DenseNet-ResNet-GRU model according to claim 1, wherein: Both DenseNet and ResNet are provided with two parallel multi-scale channels, and convolution kernels of 3×1 and 5×1 scales are used to perform convolutions of different scales on the input signal.

Images