Tool Wear State Monitoring Method Based on Multi-Scale Spatio-Temporal Fusion Network Model

Abstract

The present invention provides a tool wear state monitoring method based on a multi-scale spatiotemporal fusion network model, relating to the technical field of tool state detection. The method first collects raw signals from the tool machining process to obtain a tool wear state dataset; then preprocesses the tool wear state dataset and divides it into a training set and a test set; constructs a multi-scale spatiotemporal fusion network model as a tool wear state prediction model; trains the tool wear state prediction model using the training dataset and tests it using the test set; then uses the trained tool wear state prediction model to predict the tool wear state to be monitored; finally, smoothes the predicted tool wear state to obtain a smoothed predicted tool wear state. Compared to other deep learning methods, this method improves network interpretability while significantly enhancing prediction accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of tool state detection, and in particular to a tool wear state monitoring method based on a multi-scale spatiotemporal fusion network model. Background Art

[0002] Milling plays a vital role in advanced modern manufacturing due to its high precision and flexibility. As an integral part of the milling process, the condition of the tool directly impacts the quality and precision of the finished product. Continuous tool wear causes changes in the contact between the cutting edge and the workpiece. Failure to promptly replace the tool will result in workpiece failure and equipment downtime, ultimately causing significant economic losses. Therefore, research on tool wear condition monitoring (TCM) methods can help improve the efficiency and quality of milled products and reduce production costs.

[0003] Current methods for TCM can be implemented through model-driven or data-driven approaches. Model-driven approaches require the establishment of a physical failure model based on domain knowledge, analysis of each component of the system, and the use of this model to predict wear values and current wear status. However, tool wear is a complex dynamic phenomenon, and it is difficult to capture its wear mechanism, and these faults are difficult to predict using simple mathematical models. In this case, data-driven can be used as a solution, which uses the signals in the processing steps (cutting force signals, vibration signals, acoustic emission signals, etc.) and the tool wear status as features and labels for model training, and obtains the mapping relationship between input and output through an internal learning algorithm.

[0004] In data-driven TCM models, feature extraction and predictive recognition algorithms are crucial, and the quality of the extracted features significantly impacts TCM model performance. After extracting the time, frequency, and time-frequency domain features of the original signal, feature selection techniques such as the Pearson correlation coefficient, monotonicity analysis, and principal component analysis are introduced to reduce dimensionality and learn more useful discriminative features. Finally, the resulting features are fed into machine learning algorithms such as support vector machines (SVMs), hidden Markov models (HMMs), Kalman filters, and artificial neural networks (ANNs) for training. These traditional machine learning-based models require manual extraction of complex features, which is time-consuming and inefficient. Furthermore, the extracted features lack representativeness and rely heavily on domain expertise, failing to meet the high-precision requirements of real-time monitoring.

[0005] In recent years, deep learning models such as convolutional neural networks (CNNs), recurrent neural networks (RNNs), stacked sparse autoencoders (SAEs), and deep belief networks (DBNs) have been widely used in tool wear condition monitoring due to their excellent adaptive feature extraction capabilities for complex data. A single deep learning network may suffer from incomplete feature extraction. Combining two or more deep learning networks can extract richer feature information. Alternatively, through reasonable network design, the number of network parameters can be minimized while ensuring performance, thereby reducing time consumption. Currently, the mainstream approach is to combine CNNs and RNNs in series to enhance spatial and temporal feature extraction capabilities for complex data analysis and processing. However, simply stacking or cascading network models has two drawbacks. First, using a single 1D-CNN or RNN can only extract isolated spatial or temporal features, failing to fully exploit the network's potential for feature extraction. Second, using CNNs and RNNs in series to extract features can cause interference between the extracted temporal and spatial features, resulting in performance losses and a lack of effective fusion of temporal and spatial interpretability.

[0006] Current research on tool wear methods is able to monitor tool wear without relying on physical models. This is due to the large amount of data collected by high-precision sensors and the excellent nonlinear fitting capabilities of data-driven models. Although the analysis of data-driven model algorithms is key to the entire monitoring process, practical issues such as source sensor data collection and the influence of noise are rarely considered in complex and dynamic manufacturing processes. Cutting operations are affected by many factors such as workpiece material, tool characteristics, and machining technology, and the continuous signals generated are very complex. The signals collected by each sensor are subject to noise or other interference, and the type of sensor installed on the machine tool is also limited by factors including the shape of the workpiece and the machining environment.

[0007] Therefore, in actual machining, it is necessary to ensure that the sensor accurately collects signals while also fully considering the difficulty and cost of sensor installation. Current tool wear monitoring methods focus on selecting the appropriate sensor and performing targeted data preprocessing to address its flaws. Furthermore, addressing the "black box" problem within deep learning networks, research on how to effectively design a spatiotemporal network fusion model that fully leverages expert knowledge is also a pressing issue in tool wear monitoring technology. Summary of the Invention

[0008] The technical problem to be solved by the present invention is to address the deficiencies of the above-mentioned existing technologies and provide a tool wear status monitoring method based on a multi-scale spatiotemporal fusion network model. This method fully considers the extraction of characteristic time and space, and combines the pre-data preprocessing and post-smoothing algorithm based on expert knowledge to solve the data defects collected by the sensor, so as to achieve high-precision monitoring and prediction of the tool wear status.

[0009] To solve the above technical problems, the technical solution adopted by the present invention is: a tool wear state monitoring method based on a multi-scale spatiotemporal fusion network model, comprising the following steps:

[0010] Step 1: Collect the original signal of the tool processing process to obtain the tool wear status data set;

[0011] Step 2: Preprocess the tool wear state dataset and divide it into a training set and a test set;

[0012] (1) Abnormal data processing: The median filtering method based on the adaptive sliding window is used to detect abnormal mutation data. After filtering, the abnormal mutation data is replaced with normal data, while other normal signal data are not affected;

[0013] (2) Data normalization;

[0014] Step 3: Construct a multi-scale spatiotemporal fusion network model as a tool wear state prediction model;

[0015] The multi-scale spatiotemporal fusion network model includes a multi-scale residual network module, a parallel attention mechanism-based spatial feature extraction module and a deep GRU temporal feature extraction module, and a feature fusion prediction module;

[0016] The pre-processed tool wear state data is used as the input of the multi-scale spatiotemporal fusion network model. The multi-scale residual network module performs shallow feature extraction. The extracted feature vectors are then spliced and input into the parallel spatial feature extraction module and the deep GRU temporal feature extraction module. The spatial feature extraction module assigns different weights to each feature to extract spatial features, while the temporal feature extraction module extracts features of the time series. Finally, the features extracted by the spatial feature extraction module and the temporal feature extraction module are fused and input into the feature fusion prediction module to predict tool wear.

[0017] The multi-scale residual network module is composed of three 1D-Conv-Residual residual networks of different scales in parallel. The concatenation operation is used to connect all the outputs of the three parallel 1D-Conv-Residual residual network structures as the input of the next layer to extract shallow features;

[0018] The 1D-Conv-Residual network is composed of two different CNNs; each CNN contains five layers, including two convolutional layers, one maximum pooling layer, and two batch normalization layers. The convolution kernel size of each CNN is different. The nonlinear identity mapping relationship between the input vector and the output vector is constructed by introducing shortcut connections.

[0019] The spatial feature extraction module based on the attention mechanism includes an attention module SE_Block, a Dropout layer and a Flatten layer connected in sequence;

[0020] The deep GRU temporal feature extraction module consists of two stacked GRU layers. In the two-layer GRU, the output of each hidden state in one layer is propagated in time and also serves as the input to the hidden state in the next layer. At time t, the fused features extracted by the multi-scale residual network module are input to the GRU in the first layer. The GRU in the first layer nonlinearly expresses, stores, and selects each feature through a gating mechanism. The hidden state vector output by each gating unit in the first layer continues to serve as the input to the second layer GRU, and the hidden state output by the second layer GRU is finally used as the final extracted feature vector.

[0021] The feature fusion prediction module consists of two layers of fully connected networks, with the number of neurons in each layer being 128 and 1 respectively; the activation function is set to ReLU;

[0022] Step 4: Use the training data set to train the tool wear state prediction model and use the test set to test it;

[0023] Three milling cutters with actual wear values were selected as experimental subjects. Three tool wear status datasets were obtained. The tool wear status prediction model was cross-validated using the leave-one-out method. Two of the datasets were used as training sets, and the remaining dataset was used for testing and verification. The Huber loss function was used as the loss function in the training process, and Nadam was used as the optimizer to update the model weights.

[0024] Step 5: Use the trained tool wear state prediction model to predict the tool wear state to be monitored;

[0025] Step 6: Smoothing the predicted tool wear state to obtain a smoothed predicted tool wear state;

[0026] The SG algorithm is used to smooth and correct the predicted tool wear state to obtain the smoothed predicted tool wear value, as shown in the following formula:

[0027]

[0028] in, is the smoothed predicted tool wear value, Predicted tool wear state, h i is the smoothing coefficient, H is the smoothing constant, which is obtained by fitting the polynomial using the least squares method, w is the window width in the SG algorithm, and k is the degree of the polynomial.

[0029] The beneficial effects of adopting the above technical solution are: the tool wear state monitoring method based on the multi-scale spatiotemporal fusion network model provided by the present invention is

[0030] (1) A tool wear state prediction model is proposed, which consists of data processing of expert knowledge and data-driven multi-scale spatiotemporal fusion network. It greatly reduces RMSE and MAE and improves the determination coefficient R 2 , which provides a new approach for the prediction of tool wear.

[0031] (2) Data processing based on expert knowledge reduces the impact of abnormal data collected by sensors on model prediction through pre-processing algorithms, thereby improving the robustness of the model. The predicted values are smoothed through post-SG processing algorithms, making the model's predicted values more consistent with the actual processing conditions.

[0032] (3) We designed a novel multi-scale spatiotemporal network structure, which uses a multi-scale residual structure to mine richer shallow features. We then extract spatial and temporal features separately. Finally, we fuse the features. Compared with other deep learning methods, this method improves the network interpretability and significantly enhances prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 A flowchart of a tool wear state monitoring method based on a multi-scale spatiotemporal fusion network model provided by an embodiment of the present invention;

[0034] Figure 2 A schematic diagram of a tool wear prediction process based on a multi-scale spatiotemporal fusion network model provided by an embodiment of the present invention;

[0035] Figure 3 A schematic diagram of the structure of a 1D-Conv-Residual network in a multi-scale residual module provided by an embodiment of the present invention;

[0036] Figure 4 A schematic diagram of the structure of a spatial feature extraction module based on the attention mechanism provided in an embodiment of the present invention;

[0037] Figure 5 Graph showing tool wear state prediction results according to an embodiment of the present invention, where (a) is the prediction result based on dataset C1, (b) is the prediction result based on dataset C4, and (c) is the prediction result based on dataset C6.

[0038] Figure 6 A loss comparison chart of the MSFNet model and other models under the prediction dataset C1 provided by an embodiment of the invention;

[0039] Figure 7 Graph showing tool wear state prediction results of the post-smoothing algorithm provided in an embodiment of the present invention, wherein (a) is the prediction result on dataset C1, (b) is the prediction result on dataset C4, and (c) is the prediction result on dataset C6. DETAILED DESCRIPTION

[0040] The following embodiments of the present invention are described in further detail with reference to the accompanying drawings and examples. The following examples are used to illustrate the present invention but are not intended to limit the scope of the present invention.

[0041] In this embodiment, a tool wear state monitoring method based on a multi-scale spatiotemporal fusion network model is used. Figure 1 、 2 As shown, the following steps are included:

[0042] Step 1: Collect the original signal of the tool processing process through experimental methods to obtain the tool wear status data set;

[0043] In this embodiment, considering the price and difficulty of collecting signals during the actual monitoring process, it is necessary to select suitable sensors to collect data. A large amount of data will significantly reduce the speed of model training. In view of this, this embodiment only uses a single type of signal for status monitoring. Commonly used signal acquisition sensors include cutting force sensors, acceleration sensors, and acoustic emission sensors. Cutting force sensors have size restrictions on the workpiece table and the acquisition cost is relatively high. Acceleration sensors are used to measure vibration signals and are easily affected by noise. Since the sampling frequency of acoustic emission sensors is very high, the signals are not easily interfered with, and their detection range is long, the response is fast, and the installation is easy. Although the acoustic emission sensor has high data collection requirements, the high-frequency signal it emits is very sensitive to tool wear. Therefore, this embodiment uses an acoustic emission sensor to reduce the cost of collecting data to increase the time for subsequent training and testing.

[0044] This example conducted a full life cycle experiment on six milling cutters (C1 to C6). The original acoustic emission signals from the CNC machine tool were collected using an acoustic emission sensor. Furthermore, after each cut, the actual flank wear of the cutter was measured using a microscope. 315 tool wear values were obtained for each cutter.

[0045] In this embodiment, the acquired acoustic emission (AE) root mean square value signal It is a one-dimensional sequence. Due to the high sampling frequency, for the acoustic emission signal channel, the sampling sequence is divided by 512 to obtain several segments. This embodiment uses the first 40 segments. Therefore, each original acoustic emission signal sequence is converted into a benchmark with a length of 40. Therefore, at each time step, the dimension is 3072. One-dimensional sequence data is obtained The input shape of the network during training is 630×40×3072, and the input shape of the network during testing is 315×40×3072.

[0046] Step 2: Preprocess the tool wear state dataset and divide it into a training set and a test set;

[0047] (1) Abnormal data processing: The cutting process is complex and subject to environmental interference, so some abnormal data will inevitably appear. The median filtering method based on adaptive sliding window is used to detect abnormal mutation data. After filtering, the abnormal mutation data is replaced with normal data, while other normal signal data is not affected.

[0048] Table 1 Median filtering method based on adaptive sliding window

[0049]

[0050]

[0051] Adaptive sliding window filtering adjusts the window length based on the network effect feedback to remove abnormal data. In this embodiment, the half window width k=4000 and the lower bound coefficient l δ =3, that is, if a sample differs from the median by more than three standard deviations, the sample is replaced by the median.

[0052] (2) Data normalization: In order to eliminate the influence of the unit and scale differences of the tool wear signal and treat each dimension of the feature equally, the original data needs to be normalized. Using data normalization before model training will significantly smooth the parameter optimization process and make it easier to correctly converge to the optimal solution. This embodiment selects maximum and minimum value normalization. The normalization process is shown as follows:

[0053]

[0054] Where i represents the sequence number of the value, is the normalized data, x i is the original data, x max and x min are the maximum and minimum values of the original input data.

[0055] Step 3: Construct a multi-scale spatiotemporal fusion network model (MSFNet) as a tool wear state prediction model;

[0056] The multi-scale spatiotemporal fusion network model includes a multi-scale residual network module, a parallel attention mechanism-based spatial feature extraction module and a deep GRU temporal feature extraction module, and a feature fusion prediction module;

[0057] The preprocessed tool wear status data is used as the input of the multi-scale spatiotemporal fusion network model. Shallow features are extracted through the multi-scale residual network module. The extracted feature vectors are then spliced and input into the parallel spatial feature extraction module and the deep GRU temporal feature extraction module. The spatial feature extraction module assigns different weights to each feature to extract spatial features, and the temporal feature extraction module extracts features of the time series. Finally, the features extracted by the spatial feature extraction module and the temporal feature extraction module are fused and input into the feature fusion prediction module to predict tool wear.

[0058] 1. The multi-scale residual network module is composed of three 1D-Conv-Residual residual networks of different scales in parallel. The concatenation operation is used to connect all the outputs of the three parallel 1D-Conv-Residual residual network structures as the input of the next layer to extract shallow features;

[0059] Inspired by the residual block in ResNet, a variant of CNN, a 1D-Conv-Residual residual network with shortcut connections is designed, such as Figure 3 As shown in the figure, each 1D-Conv-Residual residual network is composed of two different CNNs; each CNN contains five layers, including two convolutional layers, one max pooling layer, and two batch normalization layers, and the convolution kernel size of each CNN is different. By introducing shortcut connections, a nonlinear identity mapping relationship between the input vector and the output vector is constructed. Compared with simple CNNs, the 1D-Conv-Residual residual network based on the improved ResNet framework has a regularization effect and has a certain inhibitory effect on overfitting. At the same time, the parallel structure of three 1D-Conv-Residual residual networks adjusts the deep network structure to a parallel network, supporting backpropagation of multi-layer structures and better feature expression, optimizing the network structure and accelerating the convergence of the model. Finally, the concatenation operation is used to connect all the outputs of the three parallel 1D-Conv-Residual residual network structures as the input of the next layer. That is, every feature extracted by the three parallel 1D-Conv-Residual residual networks is reused for subsequent processing.

[0060] 2. The spatial feature extraction module includes an attention module SE_Block, a Dropout layer, and a Flatten layer connected in sequence; the process of extracting spatial features by the spatial feature extraction module is as follows: Figure 4 As shown, specifically:

[0061] After shallow feature extraction using the multi-scale residual network module, the tool wear signature is a one-dimensional signal sequence. However, not all features in this sequence are useful for tool wear monitoring. The attention mechanism can extract effective features from complex information, acquiring target information while adjusting weights to filter or de-emphasize redundant information. Therefore, applying the attention mechanism to the feature vector sequence extracted by the multi-scale residual network module can reveal important information for tool wear monitoring.

[0062] First, the attention module SE_Block is used to adaptively adjust the feature responses between channels through feature recalibration. To reduce the number of parameters and improve computational efficiency, the attention module SE_Block adopts a "dimensionality reduction followed by dimensionality increase" strategy. It uses two multi-layer perceptrons to learn the correlations between different channels. From the perspective of the channel domain, it assigns different weights to different feature locations to obtain more important feature information.

[0063] After the attention module SE_Block, to prevent overfitting caused by insufficient data, regularization or network size reduction are often considered. The Dropout layer allows the network to ignore randomly selected neurons during forward propagation and prevents the network from over-relying on certain features. Therefore, the Dropout layer is used to improve the network's generalization ability.

[0064] Finally, the Flatten layer is used to "flatten" the features, that is, to convert the extracted multi-dimensional features into one dimension. The output dimension is equal to the number of input channels, so the dimension of the obtained feature vector is consistent, which is crucial for the feature extraction of tool wear signals.

[0065] The present invention adds the attention module of the spatial feature extraction module behind the residual network of the multi-scale residual network module, which can learn more useful information based on shallow features.

[0066] 3. The temporal feature extraction module is composed of two stacked GRU layers;

[0067] The ability of neural networks can be improved by going deeper. Inspired by this idea, by stacking two layers of GRU, the network can more effectively learn higher-level representations at different time scales.

[0068] The single-layer GRU operation steps are as follows:

[0069] z t =σ(W z x t +U z h t-1 +b z )

[0070] r t =σ(W r x t +U r h t-1 +b r )

[0071]

[0072]

[0073] Among them, h t is the hidden state output at the current time step t, r t is the update gate, z t To reset the gate, is the candidate hidden state, x t is the data of the current time step t, σ is the sigmoid activation function, W z , U z , W r , U r , W h and U h are the shared weight matrices learned during training, b z ,b r ,b h are the learnable biases for the reset gate, update gate, and hidden state, respectively;

[0074] The input of a single-layer GRU is the hidden state h of the previous time step t-1 t-1 and the data x at the current time step t t , the output is the hidden state h t ; Output h t Depends on the previous hidden state h t-1 , update gate r t , reset gate z t and candidate hidden states

[0075] In a two-layer GRU, the output of each hidden state in one layer propagates in time and also serves as the input to the hidden state in the next layer. At time t, the fused features extracted by the multi-scale residual network module are input to the GRU in the first layer. The GRU in the first layer uses a gating mechanism to nonlinearly represent, store, and select each feature. The output hidden state vector of each gating unit in the first layer continues to serve as the input to the second layer GRU, and the hidden state output by the second layer GRU is finally used as the final extracted feature vector.

[0076] 4. The feature fusion prediction module consists of two layers of fully connected networks, with the number of neurons in each layer being 128 and 1 respectively; the activation function is set to ReLU;

[0077] The features extracted by the parallel spatial feature extraction module and the temporal feature extraction module are used as the input of the fully connected network to realize the mapping between the features and the tool wear value; the fully connected network performs weighted regression on the input features and updates the fully connected network parameters (weight matrix and bias vector) through back propagation to predict the tool wear value corresponding to the i-th acoustic emission signal sequence As shown in the following formula:

[0078]

[0079] Among them, A i-1 is the feature vector of the input fully connected network, w k and b k are the weight matrix and bias vector of the fully connected network respectively;

[0080] In the overall Multi-Scale Spatiotemporal Fusion Network (MSFNet) monitoring process, shallow features contain more details, so a multi-scale residual extraction module is used to extract shallow features at different scales. Simultaneously, parallel spatiotemporal feature fusion prevents interference between temporal and spatial features, improving the interpretability of the network model.

[0081] Step 4: Use the training data set to train the tool wear state prediction model and use the test set to test it;

[0082] The purpose of network model training is to use the training data set to optimize the network parameters so that it can achieve better performance on the training set, while making the network generalize to other data outside the training set as much as possible.

[0083] In this example, milling cutters C1, C4, and C6, each with actual wear values, were selected as experimental subjects. Three tool wear state datasets were obtained. The tool wear state prediction model was cross-validated using the leave-one-out method, as shown in Table 2. Two of the datasets were used as training sets, and the remaining dataset was used for testing and validation. The multi-scale spatiotemporal fusion network model was trained on the training set; the test set data was not used for training in any experiments.

[0084] Table 2 Training set and test set

[0085] Name Settings training set Test set C1 C4+C6 C1 C4 C1+C6 C4 C6 C1+C4 C6

[0086] Secondly, the loss function is the key to network model training. This embodiment uses the Huber loss function as the loss function in the training process to evaluate the gap between the predicted value and the true value; the Huber loss function can not only maintain the loss function with a continuous derivative, but also use the characteristic that the MSE gradient decreases with the error to obtain a more accurate minimum value, and also has better robustness to outliers. The good performance of the Huber loss function is due to the carefully trained hyperparameter δ, where the hyperparameter δ can be selected through cross-validation to select the best value, as shown in the following formula:

[0087]

[0088] Among them, y represents the true value, f(x) is the predicted value, and L δ (y,f(x)) is the Huber loss;

[0089] At the same time, Nadam was selected as the optimizer to update the model weights. The Nadam algorithm uses the second-order moment estimate of the gradient to achieve adaptive adjustment of the learning rate α. Combining the adaptive moment estimate (Adam) with the Nesterov accelerated gradient algorithm, Nadam can iterate and update faster than other optimizers, achieving optimal parameter updates. The Nadam update rule is described below.

[0090]

[0091]

[0092]

[0093] Among them, θ is the weight or threshold to be optimized, m t 、 are the first-order momentum deviation and the corrected first-order momentum deviation, V t 、 are the second-order momentum deviation and the modified second-order momentum deviation, t is the current iteration, β1, β2 and ε are the correction coefficients; gt is the gradient vector, and α is the learning rate. In the training test of this embodiment, the learning rate α is set to 1e-5, and β1 and β2 are set to 0.9 and 0.99.

[0094] In this embodiment, the tool wear state prediction results on data sets C1, C4 and C6 are as follows: Figure 5 shown.

[0095] Step 5: Use the trained tool wear state prediction model to predict the tool wear state to be monitored;

[0096] This embodiment uses the tool wear state prediction model to predict tool wear, and evaluates and compares it with five deep learning network models: CNN, ResNet, LSTM, MC-GRU, and ResC-GRU. The comparison results are shown in the figure. Figure 6 As shown. Calculate the root mean square error RMSE, mean absolute error MAE and determination coefficient R based on the predicted wear amount and the actual wear amount 2 , test the prediction accuracy of tool wear. The smaller the RMSE and MAE values are, the higher the prediction accuracy is. 2 The larger the value of , the higher the prediction accuracy.

[0097]

[0098]

[0099]

[0100] Among them, y i is the actual tool wear value, is the predicted tool wear value, is the average value of tool wear prediction.

[0101] Step 6: Use the SG algorithm to smooth the predicted tool wear state to obtain a smoothed predicted tool wear state;

[0102] Using SG algorithm to predict tool wear status Perform smoothing correction to obtain smoothed predicted tool wear values As shown in the following formula:

[0103]

[0104] Among them, h i is the smoothing coefficient, H is the smoothing constant, which is obtained by fitting the polynomial using the least squares method, w is the window width in the SG algorithm, and k is the degree of the polynomial.

[0105] During pre-processing, some abnormal tool status information was removed. However, in actual working conditions, irregular data points are still inevitably present in a large amount of data, resulting in large deviations between the predicted values and the actual values in the prediction curve, especially in the C6 dataset experiment. Therefore, combining pre-processing with post-processing smoothing correction can significantly improve the accuracy of tool wear prediction.

[0106] In this embodiment, on the data sets C1, C4 and C6, the tool wear state prediction results of the post-smoothing algorithm are as follows: Figure 7 As shown, compared Figure 7 and Figure 6 It can be seen from the prediction results that after smoothing correction, the tool wear prediction value and the amplitude of the abnormal points at adjacent moments are significantly reduced, the RMSE and MAE of the prediction results are reduced to varying degrees, and the fitting degree R 2 The closer it is to 1, the smoother the tool wear prediction curve.

[0107] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some or all of the technical features therein. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the claims of the present invention.

Claims

1. A tool wear condition monitoring method based on a multi-scale spatiotemporal fusion network model, characterized by: The following steps are involved: Step 1: Collect the original signal of the tool processing process to obtain the tool wear status data set; Step 2: Preprocess the tool wear state dataset and divide it into a training set and a test set; Step 3: Construct a multi-scale spatiotemporal fusion network model as a tool wear state prediction model; The multi-scale spatiotemporal fusion network model includes a multi-scale residual network module, a parallel attention mechanism-based spatial feature extraction module and a deep GRU temporal feature extraction module, and a feature fusion prediction module; The pre-processed tool wear state data is used as the input of the multi-scale spatiotemporal fusion network model. The multi-scale residual network module performs shallow feature extraction. The extracted feature vectors are then spliced and input into the parallel spatial feature extraction module and the deep GRU temporal feature extraction module. The spatial feature extraction module assigns different weights to each feature to extract spatial features, while the temporal feature extraction module extracts features of the time series. Finally, the features extracted by the spatial feature extraction module and the temporal feature extraction module are fused and input into the feature fusion prediction module to predict tool wear. The deep GRU temporal feature extraction module consists of two stacked GRU layers. In the two-layer GRU, the output of each hidden state in one layer is propagated in time and also serves as the input to the hidden state in the next layer. At time t, the fused features extracted by the multi-scale residual network module are input to the GRU in the first layer. The GRU in the first layer nonlinearly expresses, stores, and selects each feature through a gating mechanism. The hidden state vector output by each gating unit in the first layer continues to serve as the input to the second layer GRU, and the hidden state output by the second layer GRU is finally used as the final extracted feature vector. Step 4: Use the training data set to train the tool wear state prediction model and use the test set to test it; Step 5: Use the trained tool wear state prediction model to predict the tool wear state to be monitored; Step 6: Smoothing the predicted tool wear state to obtain a smoothed predicted tool wear state.

2. The tool wear state monitoring method based on the multi-scale spatiotemporal fusion network model according to claim 1 is characterized in that: The step 2 of preprocessing the tool wear state data set includes: (1) Abnormal data processing: The median filtering method based on the adaptive sliding window is used to detect abnormal mutation data. After filtering, the abnormal mutation data is replaced with normal data, while other normal signal data are not affected; (2) Data normalization.

3. The tool wear state monitoring method based on the multi-scale spatiotemporal fusion network model according to claim 1 is characterized in that: The multi-scale residual network module is composed of three 1D-Conv-Residual residual networks of different scales in parallel. The concatenation operation is used to connect all the outputs of the three parallel 1D-Conv-Residual residual network structures as the input of the next layer to extract shallow features; The 1D-Conv-Residual residual network is composed of two different CNNs; each CNN contains five layers, including two convolutional layers, one maximum pooling layer and two batch normalization layers, and the convolution kernel size of each CNN is different; and a nonlinear identity mapping relationship between the input vector and the output vector is constructed by introducing shortcut connections.

4. The tool wear state monitoring method based on a multi-scale spatiotemporal fusion network model according to claim 1 is characterized in that: The spatial feature extraction module based on the attention mechanism includes an attention module SE_Block, a Dropout layer and a Flatten layer connected in sequence.

5. The tool wear state monitoring method based on a multi-scale spatiotemporal fusion network model according to claim 1 is characterized in that: The feature fusion prediction module consists of two layers of fully connected networks, with the number of neurons in each layer being 128 and 1 respectively; the activation function is set to ReLU.

6. The tool wear state monitoring method based on a multi-scale spatiotemporal fusion network model according to claim 1 is characterized in that: In step 4, three milling cutters with actual wear values are selected as experimental objects, three tool wear status data sets are obtained, and the tool wear status prediction model is cross-validated using the leave-one-out method. Two of the data sets are used as training sets, and the remaining data set is used for testing and verification. The Huber loss function is used as the loss function in the training process, and Nadam is used as the optimizer to update the model weights.

7. The tool wear state monitoring method based on a multi-scale spatiotemporal fusion network model according to claim 1 is characterized in that: The step 6 uses the SG algorithm to perform smoothing correction processing on the predicted tool wear state to obtain a smoothed predicted tool wear value, as shown in the following formula: Among them, y i is the smoothed predicted tool wear value, y i Predicted tool wear state, h i is the smoothing coefficient, H is the smoothing constant, which is obtained by fitting the polynomial using the least squares method, w is the window width in the SG algorithm, and k is the degree of the polynomial.

Images