A tool wear monitoring method based on channel-time enhancement and global compensation network

Abstract

The application discloses a tool wear monitoring method based on channel-time enhancement and global compensation network, and steps include: 1) preprocessing multi-source sensor data set; 2) using tool wear monitoring model to carry out channel information interaction and time information enhancement processing on the preprocessed multi-source sensor data; 4) carrying out global time sequence feature mining on the channel information interaction and time information enhanced feature data; 5) carrying out local refinement on the global time sequence feature to form global-local time sequence feature; 6) carrying out global feature compensation on the global-local time sequence feature to enhance feature long-term dependence, and obtaining enhanced global-local time sequence feature; 7) decoding the compensated enhanced global-local time sequence feature, establishing the mapping between the feature and the tool wear value, and outputting the tool wear monitoring value. The application has simple operability and high monitoring precision, and can provide guidance for tool wear monitoring method selection.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of combining tool wear monitoring with artificial intelligence, specifically a tool wear monitoring method based on channel-time enhancement and global compensation networks. Background Technology

[0002] As the end effector of a CNC machine tool, the cutting tool experiences wear during machining due to mechanical friction with the workpiece, cutting heat, and cutting forces. Over time, the degree of tool wear gradually increases. If the tool wear exceeds the tool failure threshold and is not replaced in time, it will severely affect the workpiece's machining quality and the machine tool's machining efficiency, resulting in economic losses. Therefore, accurate and timely monitoring of tool wear is crucial for stable machine tool processing.

[0003] Currently, statistical, traditional machine learning, and physical model-based methods have all achieved ideal results in tool wear monitoring. However, statistical or traditional machine learning methods typically rely heavily on expert experience and are limited by their shallow model structures, resulting in poor nonlinear fitting capabilities and significant limitations in practical applications. Furthermore, physical model-based tool wear monitoring methods require modeling the mechanisms of the tool wear process, but the tool wear process is highly complex, and the working conditions of the tool during machining are highly variable, leading to lower accuracy for physical model-based methods. Deep learning methods, as typical data-driven methods, possess strong nonlinear fitting capabilities and can effectively establish the mapping relationship between tool wear and real-time sensor signals. However, in tool wear monitoring using multi-source sensor data, deep learning methods still face the following challenges:

[0004] (1) Convolutional Neural Networks (CNNs) have been proven to have powerful feature extraction capabilities in computer vision, but they suffer from insufficient global contextual understanding when dealing with temporal problems. In tool wear monitoring, this problem manifests as the difficulty in establishing the global long-term dependencies of the tool wear process, which greatly limits the application of CNNs in tool wear monitoring.

[0005] (2) The correlation between data collected by different sensors in a multi-source sensor dataset and the tool wear trend varies. Indiscriminately assigning the same weight to each sensor's data will result in feature redundancy, thus affecting the model's testing accuracy. Some existing feature enhancement methods use feature correlation to allocate sensor weights. Furthermore, feature compression is widely used to reduce model complexity and parameter count. However, the correlation between multi-source sensor data and tool wear during the tool wear process is dynamic; therefore, weight allocation should also be dynamic. Moreover, feature compression typically requires expert experience; insufficient feature compression makes it difficult to eliminate redundant features, while excessive compression leads to the loss of useful information.

[0006] (3) Data preprocessing methods are widely used in data time series problems, such as feature extraction and data denoising. However, overly complex data preprocessing processes often consume a lot of resources, which is difficult to meet the timeliness requirements of tool wear monitoring. In addition, feature extraction, feature screening, and denoising methods often require operators to have rich expert experience, which makes the monitoring model based on complex data preprocessing have great limitations. Summary of the Invention

[0007] The purpose of this invention is to provide a tool wear monitoring method based on channel-time enhancement and global compensation networks, comprising the following steps:

[0008] 1) During the tool operation process, vibration signals, cutting force signals and acoustic emission signals of the worktable and tool are collected to construct multi-source sensor data;

[0009] 2) Preprocess the multi-source sensor dataset;

[0010] 3) Perform channel information interaction and time information enhancement processing on the preprocessed multi-source sensor data to obtain channel-time enhancement;

[0011] 4) Perform global temporal feature mining on the channel-time enhancement features to obtain global temporal features;

[0012] 5) Refine the global temporal features locally to form global-local temporal features;

[0013] 6) Perform global feature compensation on global-local temporal features to enhance long-term feature dependencies and obtain enhanced global-local temporal features;

[0014] 7) Decode the compensated enhanced global-local temporal features, establish a mapping between the features and tool wear values, and output the tool wear monitoring values.

[0015] Furthermore, the vibration signal is acquired through a triaxial vibration sensor;

[0016] The cutting force signal is acquired through a triaxial force sensor;

[0017] The acoustic emission signal is acquired by an acoustic emission sensor.

[0018] Furthermore, step 2) involves preprocessing the multi-source sensor dataset, including:

[0019] 2.1) Truncate and remove the data from the multi-source sensor data, including the data that is cut in and cut out.

[0020] 2.2) The truncated multivariate data is sliced, and the root mean square value of each data segment is extracted to form the original dataset;

[0021] 2.3) The original dataset is normalized using Min-Max normalization so that the RMS values ​​obtained by different sensors are limited to [0,1], thus generating the input data for the tool wear monitoring model.

[0022] Furthermore, in step 2.1), the first and last 2.5% of the collected multi-source sensor data points are used as the entry and exit data, and are truncated and discarded.

[0023] Furthermore, in step 2.2), the root mean square value of each data segment is as follows:

[0024]

[0025] Where: RMS j Let X represent the root mean square value of the j-th data segment, L represent the length of a single segment, and X represent the root mean square value of the j-th data segment. i This represents the i-th data point.

[0026] In step 2.3), the normalized data is shown below:

[0027]

[0028] Where: RMS k,s and These are the raw data and the normalized data at the k-th RMS data point of the s-th sensor, respectively. min,s and RMS max,s This represents the minimum and maximum values ​​on the s-th sensor.

[0029] Furthermore, in step 3), the steps of using the tool wear monitoring model to perform channel information interaction and time information enhancement processing on the preprocessed multi-source sensor data include:

[0030] 3.1) Obtain the overall feature information v of each sensor data using global max pooling and global average pooling. (n) ,Right now:

[0031] v mi =GMP1D(x i ) = max(x i (3)

[0032]

[0033] v (n) =concat(v m ,v a (5)

[0034] In the formula: x i This represents the data from the i-th sensor; v mi and v ai This represents the overall information of the i-th channel data obtained by global max pooling and global average pooling along the data time direction; T represents the sample time window size, and concat represents vector concatenation; GMP1D(x i ), GAP1D(x i ) represent one-dimensional global max pooling and global average pooling operations, respectively;

[0035] 3.2) Utilizing fully connected layers to process the overall feature information v (n) Scaling is applied to achieve global channel information interaction, and an activation function is used to assign channel weights, resulting in the channel information interaction feature y. (n) ,Right now:

[0036] ac (n) =σ(w F2 (w F1 v (n) (6)

[0037]

[0038] In the formula: σ represents the sigmoid activation function, w F1 and w F2 These are the parameters of the two fully connected layers; b represents the bias; ac (n) This is scaled feature information;

[0039] 3.3) The data after the channel information exchange is marked using position coding, and the marked data is encoded using cosine and sine functions of different frequencies to obtain coding features;

[0040] 3.4) Along the channel direction, global max pooling and global average pooling are used to obtain the overall information t of the feature time direction.(n) ,Right now:

[0041] t mj =GMP(P j ) = max(P j (8)

[0042]

[0043] t (n) =concat(t m ,t a (10)

[0044] In the formula, t mj and t aj These are the global max pooling and global average pooling results for the input data at time point j, respectively; P j Let F represent the number of multi-source sensors, where F is the j-th encoded feature; GMP(P) j ), GAP (P j ) represent global max pooling and global average pooling operations, respectively;

[0045] 3.5) Utilizing one-dimensional convolution to process the overall information t (n) Local enhancement is performed, and temporal weights are assigned to obtain feature data enhanced by channel information interaction and temporal information. (n) ,Right now:

[0046] at (n) =σ(Conv1D(θ) Conv1D ,t (n) (11)

[0047]

[0048] In the formula: θ Conv1D P represents the convolution kernel parameters. (n) The channel information interaction characteristics after location encoding For element-wise multiplication; at (n) This represents local enhancement features. Conv1D represents a one-dimensional convolution operation.

[0049] Furthermore, in step 4), a deep convolutional neural network is used to perform global temporal feature mining on the feature data that has undergone channel information interaction and temporal information enhancement.

[0050] The process of feature extraction by a deep convolutional neural network is as follows:

[0051]

[0052] Mout (n) =MP(Ge2,ps) (17)

[0053] In the formula, BN represents batch normalization, GeLU represents the GeLU activation function, MP represents max pooling, and ps is the pooling kernel size. (n) These are the output features extracted by the deep convolutional neural network. BNout1, Ge1, BNout3, and Ge2 are intermediate feature parameters in the feature extraction process of the deep convolutional neural network.

[0054] Furthermore, in step 5), the global-local temporal features are as follows:

[0055] Conv=Conv1D(θ Conv1D ,GeLU(LN(Conv1D(θ Conv1D Mout (n) (18)

[0056] Cout (n) =add(Conv,Mout) (n) (19)

[0057] In the formula, Mout (n) For global temporal features, GeLU represents the GeLU activation function, LN represents layer normalization, and add represents element-wise addition; Cout (n) This represents global-local temporal features. Conv is an intermediate parameter.

[0058] Furthermore, in step 6), the step of performing global feature compensation on the global-local temporal features includes:

[0059] 6.1) Based on global temporal features, calculate the contribution of tool wear features to obtain the attention factor β of the hidden state vector. (n) ,Right now:

[0060] β (n) =δ(Cout) (n) ⊙GMP(Mout (n) (20)

[0061] In the formula, δ represents the softmax activation function, and ⊙ represents the element-wise dot product.

[0062] 6.2) Based on attention factor β (n) Memory enhancement processing is applied to global-local temporal features to obtain enhanced global-local temporal features. Right now:

[0063]

[0064] In the formula: This represents the global temporal characteristics after memory enhancement. This represents the enhanced global-local temporal features after global feature compensation.

[0065] Furthermore, in step 7), a fully connected regression subnetwork is used to establish a mapping between features and tool wear values;

[0066] The fully connected regression subnetwork includes three fully connected layers. The first two fully connected layers are followed by a ReLU activation function and Dropout regularization, and the third fully connected layer is followed by a Linear activation function. The activation function and Dropout regularization are used to enhance the nonlinear fitting ability of the subnetwork.

[0067] Furthermore, the tool wear monitoring values ​​are shown below:

[0068] Moni (n) =l(w F5 ReLU,(w F4 ReLU,(w F3 Fout (n) ))) (twenty three)

[0069] In the formula, Moni (n) The value represents the tool wear monitoring value; l represents the Linear activation function;

[0070] w F3 ,w F4 ,w F5 Fout represents the parameters of the three fully connected layers of the fully connected regression subnetwork. (n) The features are enhanced by Flattening to improve the global-local temporal features.

[0071] The technical effects of this invention are undeniable, and its beneficial effects are as follows:

[0072] (1) This invention proposes a tool wear monitoring method based on channel-time enhancement and global compensation networks. First, multi-source sensor signals are collected during the tool machining process using multi-source sensors. Then, the data is processed by in-cut-out rejection, RMS extraction of data slices, and Min-Max normalization as input data for the model. Next, the preprocessed data is input into a channel-time enhancement sub-network to achieve channel information interaction and enhance temporal information. The enhanced data is then input into a feature learning sub-network containing DCNN, residual blocks, and global compensation to mine global-local temporal features related to tool wear. Finally, a fully connected regression sub-network maps the global-local temporal features to tool wear monitoring values.

[0073] (2) This invention does not involve data denoising, extraction of various time-domain and frequency-domain features, or feature selection. Therefore, the use of this invention does not require the operator to have extensive expert experience in this field.

[0074] (3) The present invention uses a channel-time enhancement subnetwork to achieve data information interaction and timing enhancement in a lightweight structure, which can suppress useless information unrelated to tool wear and enhance useful information.

[0075] (4) This invention combines local refinement, memory enhancement and global compensation to effectively alleviate the problem of CNN's difficulty in establishing long-term dependencies in handling time-series problems, thereby improving the accuracy of tool wear monitoring. Attached Figure Description

[0076] Figure 1 This is a flowchart of the tool wear monitoring method based on channel-time enhancement and global compensation network of the present invention;

[0077] Figure 2 This is a structural diagram of the deep neural network monitoring model of the present invention;

[0078] Figure 3 This is a framework diagram of DCNN in the model structure of this invention;

[0079] Figure 4 This is a diagram of the multi-source sensor data acquisition device for tool wear in an embodiment of the present invention;

[0080] Figure 5 This is a preprocessed image of multi-source sensor data in an embodiment of the present invention;

[0081] Figure 6 This is a schematic diagram of the wear monitoring results of the centralized test set of milling tool wear data in an embodiment of the present invention; in the figure, (a) represents the tool wear monitoring results of test subset C1; (b) represents the tool wear monitoring results of test subset C4; and (c) represents the tool wear monitoring results of test subset C6. Detailed Implementation

[0082] The present invention will be further described below with reference to embodiments, but it should not be construed that the scope of the present invention is limited to the following embodiments. Various substitutions and modifications made based on ordinary technical knowledge and common practices in the art without departing from the above-described technical concept of the present invention should be included within the scope of protection of the present invention.

[0083] Example 1:

[0084] See Figures 1 to 6 A tool wear monitoring method based on channel-time enhancement and global compensation network includes the following steps:

[0085] 1) During the operation of tool 1, vibration signals, cutting force signals and acoustic emission signals of worktable 2 and tool 1 are collected to construct multi-source sensor data;

[0086] 2) Preprocess the multi-source sensor dataset;

[0087] 3) Perform channel information interaction and time information enhancement processing on the preprocessed multi-source sensor data to obtain channel-time enhancement features;

[0088] 4) Perform global temporal feature mining on the channel-time enhancement features to obtain global temporal features;

[0089] 5) Refine the global temporal features locally to form global-local temporal features;

[0090] 6) Perform global feature compensation on global-local temporal features to enhance long-term feature dependencies and obtain enhanced global-local temporal features;

[0091] 7) Decode the compensated enhanced global-local temporal features, establish a mapping between the features and tool wear values, and output the tool wear monitoring values.

[0092] The vibration signal was acquired by a triaxial vibration sensor;

[0093] The cutting force signal is acquired through a triaxial force sensor;

[0094] The acoustic emission signal is acquired by an acoustic emission sensor.

[0095] Step 2) includes the following steps for preprocessing the multi-source sensor dataset:

[0096] 2.1) Truncate and remove the data from the multi-source sensor data, including the data that is cut in and cut out.

[0097] 2.2) The truncated multivariate data is sliced, and the root mean square value of each data segment is extracted to form the original dataset;

[0098] 2.3) The original dataset is normalized using Min-Max normalization so that the RMS values ​​obtained by different sensors are limited to [0,1], thus generating the input data for the tool wear monitoring model.

[0099] In step 2.1), 2.5% of the collected sensor signals before and after are considered invalid and are discarded.

[0100] In step 2.2), the root mean square value of each data segment is shown below:

[0101]

[0102] Where: RMS j Let X represent the root mean square value of the j-th data segment, L represent the length of a single segment, and X represent the root mean square value of the j-th data segment. i This represents the i-th data point.

[0103] In step 2.3, the original dataset is normalized using Min-Max normalization to ensure that the RMS values ​​obtained from different sensors are limited to [0,1]. The calculation formula is shown below:

[0104]

[0105] Where: RMS k,s and These are the raw data and the normalized data at the k-th RMS data point of the s-th sensor, respectively. min,s and RMS max,s This represents the minimum and maximum values ​​on the s-th sensor.

[0106] Step 3), which involves using the tool wear monitoring model to perform channel information interaction and temporal information enhancement processing on the preprocessed multi-source sensor data, includes the following steps:

[0107] 3.1) Obtain the overall feature information v of each sensor data using global max pooling and global average pooling. (n) ,Right now:

[0108] v mi =GMP1D(x i ) = max(x i (3)

[0109]

[0110] v (n) =concat(v m ,v a (5)

[0111] In the formula: x i This represents the data from the i-th sensor; v mi and v ai This represents the overall information of the i-th channel data obtained by global max pooling and global average pooling along the data time direction; T represents the sample time window size, and concat represents vector concatenation; GMP1D(x i ), GAP1D(x i ) represent one-dimensional global max pooling and global average pooling operations, respectively;

[0112] 3.2) Utilizing fully connected layers to process the overall feature information v(n) Scaling is applied to achieve global channel information interaction, and an activation function is used to assign channel weights, resulting in the channel information interaction feature y. (n) ,Right now:

[0113] ac (n) =σ(w F2 (w F1 v (n) (6)

[0114]

[0115] In the formula: σ represents the sigmoid activation function, w F1 and w F2 These are the parameters of the two fully connected layers; b represents the bias; ac (n) This is scaled feature information;

[0116] 3.3) The data after the channel information exchange is marked using position coding, and the marked data is encoded using cosine and sine functions of different frequencies to obtain coding features;

[0117] 3.4) Along the channel direction, global max pooling and global average pooling are used to obtain the overall information t of the feature time direction. (n) ,Right now:

[0118] t mj =GMP(P j ) = max(P j (8)

[0119]

[0120] t (n) =concat(t m ,t a (10)

[0121] In the formula, t mj and t aj These are the global max pooling and global average pooling results for the input data at time point j, respectively; P j Let F represent the number of multi-source sensors, where F is the j-th encoded feature; GMP(P) j ), GAP (P j ) represent global max pooling and global average pooling operations, respectively;

[0122] 3.5) Utilizing one-dimensional convolution to process the overall information t (n) Local enhancement is performed, and temporal weights are assigned to obtain feature data enhanced by channel information interaction and temporal information. (n) ,Right now:

[0123] at (n) =σ(Conv1D(θ) Conv1D ,t (n) (11)

[0124]

[0125] In the formula: θ Conv1D P represents the convolution kernel parameters. (n) The channel information interaction characteristics after location encoding For element-wise multiplication; at (n) This represents local enhancement features. Conv1D represents a one-dimensional convolution operation.

[0126] In step 4), a deep convolutional neural network is used to perform global temporal feature mining on the feature data enhanced by channel information interaction and temporal information. The feature extraction by the deep convolutional neural network can be represented as:

[0127]

[0128] Mout (n) =MP(Ge2,ps) (17)

[0129] In the formula, BN represents batch normalization, GeLU represents the GeLU activation function, MP represents max pooling, and ps is the pooling kernel size. (n) These are the output features extracted by the deep convolutional neural network. BNout1, Ge1, BNout3, and Ge2 are intermediate feature parameters in the feature extraction process of the deep convolutional neural network.

[0130] In step 5), the global-local temporal features are shown below:

[0131] Conv=Conv1D(θ Conv1D ,GeLU(LN(Conv1D(θ Conv1D Mout (n) (18)

[0132] Cout (n) =add(Conv,Mout) (n) (19)

[0133] In the formula, Mout (n) For global temporal features, GeLU represents the GeLU activation function, LN represents layer normalization, and add represents element-wise addition; Cout (n) This represents global-local temporal features. Conv is an intermediate parameter.

[0134] Step 6) involves performing global feature compensation on the global-local temporal features, including:

[0135] 6.1) Based on global temporal features, calculate the contribution of tool wear features to obtain the attention factor β of the hidden state vector. (n) ,Right now:

[0136] β (n) =δ(Cout) (n) ⊙GMP(Mout (n) (20)

[0137] In the formula, δ represents the softmax activation function, and ⊙ represents the element-wise dot product.

[0138] 6.2) Based on attention factor β (n) Memory enhancement processing is applied to global-local temporal features to obtain enhanced global-local temporal features. Right now:

[0139]

[0140] In the formula: This represents the global temporal characteristics after memory enhancement. This represents the enhanced global-local temporal features after global feature compensation.

[0141] In step 7), a fully connected regression subnetwork is used to establish a mapping between features and tool wear values;

[0142] The fully connected regression subnetwork includes three fully connected layers. The first two fully connected layers are followed by a ReLU activation function and Dropout regularization, and the third fully connected layer is followed by a Linear activation function. The activation function and Dropout regularization are used to enhance the nonlinear fitting ability of the subnetwork.

[0143] The tool wear monitoring values ​​are shown below:

[0144] Moni (n) =l(w F5 ReLU,(w F4 ReLU,(w F3 Fout (n) ))) (twenty three)

[0145] In the formula, Moni (n) The value represents the tool wear monitoring value; l represents the Linear activation function;

[0146] w F3 ,w F4 ,w F5Fout represents the parameters of the three fully connected layers of the fully connected regression subnetwork. (n) The features are enhanced by Flattening to improve the global-local temporal features.

[0147] Example 2:

[0148] A tool wear monitoring method based on channel-time enhancement and global compensation networks includes the following steps:

[0149] 1) During the operation of tool 1, vibration signals, cutting force signals and acoustic emission signals of worktable 2 and tool 1 are collected to construct multi-source sensor data;

[0150] 2) Preprocess the multi-source sensor dataset;

[0151] 3) Perform channel information interaction and time information enhancement processing on the preprocessed multi-source sensor data to obtain channel-time enhancement features;

[0152] 4) Perform global temporal feature mining on the channel-time enhancement features to obtain global temporal features;

[0153] 5) Refine the global temporal features locally to form global-local temporal features;

[0154] 6) Perform global feature compensation on global-local temporal features to enhance long-term feature dependencies and obtain enhanced global-local temporal features;

[0155] 7) Decode the compensated enhanced global-local temporal features, establish a mapping between the features and tool wear values, and output the tool wear monitoring values.

[0156] Example 3:

[0157] A tool wear monitoring method based on channel-time enhancement and global compensation network, with the same technical content as in Embodiment 2, further wherein the vibration signal is acquired by a triaxial vibration sensor;

[0158] The cutting force signal is acquired through a triaxial force sensor;

[0159] The acoustic emission signal is acquired by an acoustic emission sensor.

[0160] Example 4:

[0161] A tool wear monitoring method based on channel-time enhancement and global compensation network, with the same technical content as any one of embodiments 2-3, further comprising the following steps in step 2): Preprocessing the multi-source sensor dataset.

[0162] 2.1) Truncate and remove the data from the multi-source sensor data, including the data that is cut in and cut out.

[0163] 2.2) The truncated multivariate data is sliced, and the root mean square value of each data segment is extracted to form the original dataset;

[0164] 2.3) The original dataset is normalized using Min-Max normalization so that the RMS values ​​obtained by different sensors are limited to [0,1], thus generating the input data for the tool wear monitoring model.

[0165] Example 5:

[0166] A tool wear monitoring method based on channel-time enhancement and global compensation network, with the same technical content as any one of Examples 2-4:

[0167] Furthermore, in step 2.1), 2.5% of the data points before and after the acquisition of multi-source sensor signals are used as the cut-in and cut-out data, which are then truncated and discarded.

[0168] Furthermore, in step 2.2), the root mean square value of each data segment is as follows:

[0169]

[0170] Where: RMS j Let X represent the root mean square value of the j-th data segment, L represent the length of a single segment, and X represent the root mean square value of the j-th data segment. i This represents the i-th data point.

[0171] Furthermore, in step 2.3), the original dataset is normalized using Min-Max normalization so that the RMS values ​​obtained by different sensors are limited to [0,1]. The calculation formula is shown below:

[0172]

[0173] Where: RMS k,s and These are the raw data and the normalized data at the k-th RMS data point of the s-th sensor, respectively. min,s and RMS max,s This represents the minimum and maximum values ​​on the s-th sensor.

[0174] Example 6:

[0175] A tool wear monitoring method based on channel-time enhancement and global compensation network, with the same technical content as any one of embodiments 2-5, further comprising the following steps in step 3), which involve using the tool wear monitoring model to perform channel information interaction and time information enhancement processing on the preprocessed multi-source sensor data:

[0176] 3.1) Obtain the overall feature information v of each sensor data using global max pooling and global average pooling. (n) ,Right now:

[0177] v mi =GMP1D(x i ) = max(x i (3)

[0178]

[0179] v (n) =concat(v m ,v a (5)

[0180] In the formula: x i This represents the data from the i-th sensor; v mi and v ai This represents the overall information of the i-th channel data obtained by global max pooling and global average pooling along the data time direction; T represents the sample time window size, and concat represents vector concatenation.

[0181] 3.2) Utilizing fully connected layers to process the overall feature information v (n) Scaling is applied to achieve global channel information interaction, and an activation function is used to assign channel weights, resulting in the channel information interaction feature y. (n) ,Right now:

[0182] ac (n) =σ(w F2 (w F1 v (n) (6)

[0183]

[0184] In the formula: σ represents the sigmoid activation function, w F1 and w F2 These are the parameters of the two fully connected layers; b represents the bias; ac (n) This is scaled feature information;

[0185] 3.3) The data after the channel information exchange is marked using position coding, and the marked data is encoded using cosine and sine functions of different frequencies to obtain coding features;

[0186] 3.4) Along the channel direction, global max pooling and global average pooling are used to obtain the overall information t of the feature time direction. (n) ,Right now:

[0187] tmj =GMP(P j ) = max(P j (8)

[0188]

[0189] t (n) =concat(t m ,t a (10)

[0190] In the formula, t mj and t aj These are the global max pooling and global average pooling results for the input data at time point j, respectively; P j Let F represent the number of multi-source sensors, where j is the j-th encoded feature.

[0191] 3.5) Utilizing one-dimensional convolution to process the overall information t (n) Local enhancement is performed, and temporal weights are assigned to obtain feature data enhanced by channel information interaction and temporal information. (n) ,Right now:

[0192] at (n) =σ(Conv1D(θ) Conv1D ,t (n) (11)

[0193]

[0194] In the formula: θ Conv1D P represents the convolution kernel parameters. (n) The channel information interaction characteristics after location encoding For element-wise multiplication; at (n) This is a local enhancement feature.

[0195] Example 7:

[0196] A tool wear monitoring method based on channel-time enhancement and global compensation network, with technical content identical to any one of embodiments 2-6, further comprising the following step 4): utilizing a deep convolutional neural network to perform global temporal feature mining on the feature data enhanced by channel information interaction and temporal information. The features extracted by the deep convolutional neural network can be expressed as:

[0197]

[0198] Mout (n) =MP(Ge2,ps) (17)

[0199] In the formula, BN represents batch normalization, GeLU represents the GeLU activation function, MP represents max pooling, and ps is the pooling kernel size. (n) These are the output features extracted by a deep convolutional neural network.

[0200] Example 8:

[0201] A tool wear monitoring method based on channel-time enhancement and global compensation network, with the same technical content as any one of embodiments 2-7, further wherein, in step 5), the global-local temporal characteristics are as follows:

[0202] Conv=Conv1D(θ Conv1D ,GeLU(LN(Conv1D(θ Conv1D Mout (n) (18)

[0203] Cout (n) =add(Conv,Mout) (n) (19)

[0204] In the formula, Mout (n) For global temporal features, GeLU represents the GeLU activation function, LN represents layer normalization, and add represents element-wise addition; Cout (n) It represents global-local temporal features.

[0205] Example 9:

[0206] A tool wear monitoring method based on channel-time enhancement and global compensation network, with the same technical content as any one of embodiments 2-8, further comprising, in step 6), the step of performing global feature compensation on global-local temporal features includes:

[0207] 6.1) Based on global temporal features, calculate the contribution of tool wear features to obtain the attention factor β of the hidden state vector. (n) ,Right now:

[0208] β (n) =δ(Cout) (n) ⊙GMP(Mout (n) (20)

[0209] In the formula, δ represents the softmax activation function, and ⊙ represents the element-wise dot product.

[0210] 6.2) Based on attention factor β (n) Memory enhancement processing is applied to global-local temporal features to obtain enhanced global-local temporal features. Right now:

[0211]

[0212] In the formula: This represents the global temporal characteristics after memory enhancement. This represents the enhanced global-local temporal features after global feature compensation.

[0213] Example 10:

[0214] A tool wear monitoring method based on channel-time enhancement and global compensation network, with the same technical content as any one of embodiments 2-9, further wherein, in step 7), a fully connected regression subnetwork is used to establish a mapping between features and tool wear values;

[0215] The fully connected regression subnetwork includes three fully connected layers. The first two fully connected layers are followed by a ReLU activation function and Dropout regularization, and the third fully connected layer is followed by a Linear activation function. The activation function and Dropout regularization are used to enhance the nonlinear fitting ability of the subnetwork.

[0216] Example 11:

[0217] A tool wear monitoring method based on channel-time enhancement and global compensation network, with the same technical content as any one of embodiments 2-10, and further, the tool wear monitoring values ​​are as follows:

[0218] Moni (n) =l(w F5 ReLU,(w F4 ReLU,(w F3 Fout (n) ))) (twenty three)

[0219] In the formula, Moni (n) The value represents the tool wear monitoring value; l represents the Linear activation function; w F3 ,w F4 ,w F5 Fout represents the parameters of the three fully connected layers of the fully connected regression subnetwork. (n) The features are enhanced by Flattening to improve the global-local temporal features.

[0220] Example 12:

[0221] A tool wear monitoring method based on channel-time enhancement and global compensation network is described below:

[0222] Figure 1 The diagram shows the flow chart of a tool wear monitoring method based on channel-time enhancement and global compensation network according to the present invention.

[0223] After acquiring multi-source sensor signals during the tool machining process, the raw sensor data is preprocessed. Then, a channel-time enhancement subnetwork is used to perform global information interaction and temporal information enhancement on the preprocessed input data to suppress invalid information and enhance effective information. Next, the enhanced features are input into the feature learning subnetwork for tool wear feature extraction: (1) DCNN is used to mine global temporal features related to tool wear; (2) residual blocks are used to locally refine the global temporal features mined by DCNN to obtain global-local temporal features; (3) memory enhancement and global feature compensation are used to enhance the global long-term dependence of global-local temporal features. Finally, the extracted features are mapped to tool wear values ​​through a fully connected regression subnetwork.

[0224] Specifically, the steps include the following:

[0225] Step 1: A triaxial vibration sensor (accelerometer 5), a triaxial force sensor (cutting force sensor 6), and an acoustic emission sensor 7 are used to collect the vibration signals, cutting force signals, and acoustic emission signals of the worktable 2 and the tool 1 during the operation of the tool 1. These signals constitute multi-source sensor signals and are transmitted to the data acquisition system 8. The worktable 2 is located on the machine tool 3.

[0226] Step 2: The data acquisition system 8 performs data preprocessing on the acquired signals to ensure that the data from each sensor have the same order of magnitude;

[0227] Step 3: Enhance the preprocessed data with channel information and temporal information using the trained tool wear monitoring model;

[0228] Step 4: Perform global temporal feature mining on multi-source sensor data with enhanced channel information interaction and temporal information;

[0229] Step 5: Refine the mined global temporal features locally to form global-local temporal features;

[0230] Step 6: Perform global feature compensation on the global-local temporal features to enhance the long-term dependency of the features and obtain enhanced global-local temporal features;

[0231] Step 7: Decode the compensated enhanced global-local temporal features, establish the mapping between the features and the tool wear value, and output the tool wear monitoring value.

[0232] The multi-source sensor data preprocessing described in step 2 specifically includes:

[0233] The first and last 2.5% of the data points in the multi-source sensor data are truncated and removed as the cut-out portion;

[0234] The truncated multivariate data was sliced, and the root mean square (RMS) value of each data segment was extracted to form the original dataset; the RMS value of each data segment is shown below:

[0235]

[0236] Where: RMS j Let X represent the root mean square value of the j-th data segment, L represent the length of a single segment, and X represent the root mean square value of the j-th data segment. i This represents the i-th data point.

[0237] The original dataset was normalized using Min-Max normalization to ensure that the RMS values ​​obtained from different sensors are limited to the range [0,1]. The calculation formula is shown below:

[0238]

[0239] Where: RMS k,s and These are the raw data and the normalized data at the k-th RMS data point of the s-th sensor, respectively. min,s and RMS max,s This represents the minimum and maximum values ​​on the s-th sensor.

[0240] The channel information interaction and time information enhancement mentioned in step 3 are specifically implemented as follows:

[0241] (1) Channel information interaction

[0242] Global max pooling and global average pooling are used to obtain the overall data information of each sensor data;

[0243] v mi =GMP1D(x i ) = max(x i (3)

[0244]

[0245] v (n) =concat(v m ,v a (5)

[0246] In the formula: x i This represents the data from the i-th sensor, i.e., the data from the i-th channel after data preprocessing; v mi and v ai This represents the overall information of the i-th channel data obtained by global max pooling and global average pooling along the data time direction. T represents the sample time window size, and concat represents vector concatenation.

[0247] Based on the overall information obtained after vector concatenation, two fully connected layers are used to compress and expand the overall information to achieve global channel information interaction, and a sigmoid activation function is used to allocate channel weights, thereby obtaining channel information interaction features:

[0248] ac (n) =σ(w F2 (w F1 v (n) (6)

[0249]

[0250] In the formula: σ represents the sigmoid activation function, w F1 and w F2 ...

[0251] (2) Enhanced time information

[0252] Position coding is introduced to mark the data after the channel information is exchanged, and cosine and sine functions of different frequencies are used to encode the marked sensor data.

[0253] Global max pooling and global average pooling are used along the channel direction to obtain the overall information of the features along the time direction.

[0254] t mj =GMP(P j ) = max(P j (8)

[0255]

[0256] t (n) =concat(t m ,t a (10)

[0257] In the formula, t mj and t aj These are the global max pooling and global average pooling results for the data at time point j, respectively; P j Let F represent the number of multi-source sensors, where j is the j-th encoded data.

[0258] Based on the overall temporal information, one-dimensional convolution is used to locally enhance the overall temporal information, and temporal weights are assigned to obtain temporally enhanced features:

[0259] at (n) =σ(Conv1D(θ) Conv1D ,t (n)(11)

[0260]

[0261] In the formula: θ Conv1D P represents the convolution kernel parameters. (n) The channel information interaction characteristics after location encoding Multiplying elements together.

[0262] Step 4 describes using a deep convolutional neural network (DCNN) for feature mining to obtain global temporal features. The feature extraction process using a deep convolutional neural network can be represented as:

[0263]

[0264]

[0265] Mout (n) =MP(Ge2,ps) (17)

[0266] In the formula, BN represents batch normalization, GeLU represents the GeLU activation function, MP represents max pooling, and ps is the pooling kernel size. (n) These are the output features extracted by a deep convolutional neural network.

[0267] The global-local temporal features described in step 5 are obtained by locally refining the global temporal features using a residual block with a kernel of 1. The specific process is as follows:

[0268] Conv=Conv1D(θ Conv1D ,GeLU(LN(Conv1D(θ Conv1D Mout (n) (17)

[0269] Cout (n) =add(Conv,Mout) (n) (18)

[0270] In the formula, Mout (n) For global temporal features mined by DCNN, GeLU represents the GeLU activation function, LN is the layer normalization process, and add means element-wise addition.

[0271] The enhanced global-local temporal features mentioned in step 6 are obtained through memory enhancement and global compensation mechanisms, and the specific process is as follows:

[0272] The attention factor of the hidden state vector is calculated based on the contribution of global temporal features mined by DCNN to tool wear features.

[0273] β (n) =δ(Cout) (n) ⊙GMP(Mout (n) (19)

[0274] In the formula, δ represents the softmax activation function, and ⊙ represents the element-wise dot product.

[0275] Based on an attention factor, memory enhancement processing is applied to global-local temporal features to strengthen their long-term global dependencies. Furthermore, to compensate for information loss caused by local feature refinement, global temporal features acquired by DCNN are used as compensation and merged with the memory-enhanced features to form enhanced global-local temporal features.

[0276]

[0277] The step 7, which establishes the mapping relationship between features and tool wear values, relies on a fully connected regression subnetwork. This subnetwork consists of three fully connected layers, with a ReLU activation function added after the first two fully connected layers and a Linear activation function added after the third fully connected layer. The activation functions in the fully connected regression subnetwork are used to enhance the nonlinear fitting ability of the subnetwork, and the combination of ReLU and Linear activation functions avoids monitoring model overfitting.

[0278] In addition, Dropout regularization is introduced after the first two fully connected layers to further prevent the model from overfitting;

[0279] The fully connected regression subnetwork maps the extracted enhanced global-local temporal features into tool wear values ​​(Moni) after one tool pass. (n) This enables tool wear monitoring;

[0280] The specific calculation process for the mapping is as follows:

[0281] Moni (n) =l(w F5 ReLU,(w F4 ReLU,(w F3 Fout (n) ))) (twenty two)

[0282] In the formula, Moni (n) The monitored tool wear value; l represents the Linear activation function; w F3 ,w F4 ,w F5 These represent the parameters of the three fully connected layers, Fout (n) The result is obtained after Flattening the local temporal features.

[0283] Example 13:

[0284] The verification of a tool wear monitoring method based on channel-time enhancement and global compensation network is as follows:

[0285] In this embodiment, the tool wear dataset used is the PHM2010 competition dataset. The experimental setup and parameters are detailed below. Figure 4 See Table 1. Based on the cutting operations of the tool machining process, this dataset contains three tool life cutting subsets that include tool wear values. Each subset consists of 315 tool passes. For each pass, vibration signals and cutting force signals in the x, y, and z axes were collected, and acoustic emission signals during the tool pass were also collected using an acoustic emission sensor. Based on subset completeness, the dataset was divided into a training set and a test set using a three-fold cross-partitioning method as shown in Table 2. Specifically, two subsets were used as the training set each time, and the remaining subset was used as the test set.

[0286] Table 1

[0287]

[0288] Table 2

[0289]

[0290]

[0291] After acquiring multi-source sensor tool machining data, such as Figure 1 As shown, the implementation of this invention should also include the following steps:

[0292] Step 1: Data Preprocessing

[0293] Sub-step (1), such as Figure 5 As shown, the multi-source sensor data acquired by the tool during each machining process includes vibration signals and cutting force signals in the x, y, and z axes, as well as an acoustic emission signal, totaling seven sensor data points as raw data. Input and output filtering is performed on the data acquired during each machining process; in this embodiment, the first 2.5% and the last 2.5% of the signal are extracted as the entry and exit points, respectively.

[0294] Sub-step (2), such as Figure 5 As shown, the extracted data is divided into N segments, where N is 30 in this embodiment. The RMS value of each segment is calculated to form the original dataset. The RMS calculation formula is as follows:

[0295]

[0296] Where: RMS j Let X represent the RMS of the j-th segment, L represent the length of a single segment, and X represent the length of the segment.i This represents the i-th data point.

[0297] Sub-step (3) uses Min-Max normalization to normalize the original dataset, so that the RMS values ​​obtained by different sensors are limited to [0,1]. The calculation formula is shown below:

[0298]

[0299] Where: RMS k,s and These are the raw data and the normalized data at the k-th RMS data point of the s-th sensor, respectively. min,s and RMS max,s This represents the minimum and maximum values ​​on the s-th sensor.

[0300] Step 2: Channel information interaction and time information enhancement

[0301] In this step, through Figure 2 The channel-time enhancement subnetwork shown in the figure includes channel information exchange and time information enhancement. The specific process is as follows:

[0302] (1) Channel information interaction

[0303] Use global max pooling and global average pooling to obtain overall information about each sensor's data;

[0304] v mi =GMP1D(x i ) = max(x i (3)

[0305]

[0306] v (n) =concat(v m ,v a (5)

[0307] In the formula: x i This represents the data from the i-th sensor, i.e., the data from the i-th channel after data preprocessing; v mi and v ai This represents the overall information of the i-th channel data obtained by global max pooling and global average pooling along the data time direction. T represents the sample time window size, and concat represents vector concatenation.

[0308] Based on the overall information obtained after vector concatenation, two fully connected layers are used to compress and expand the overall information to achieve global channel information interaction, and a sigmoid activation function is used to allocate channel weights, thereby obtaining channel information interaction features:

[0309] ac (n) =σ(w F2 (w F1 v (n) (6)

[0310]

[0311] In the formula: σ represents the sigmoid activation function, w F1 and w F2 represents the parameters of fully connected layer 1 and fully connected layer 2, and b represents the deviation.

[0312] (2) Enhanced time information

[0313] Position coding is introduced to mark the data after the channel information is exchanged, and cosine and sine functions of different frequencies are used to encode the marked sensor data.

[0314] Global max pooling and global average pooling are used along the channel direction to obtain overall information along the time direction.

[0315] t mj =GMP(P j ) = max(P j (8)

[0316]

[0317] t (n) =concat(t m ,t a (10)

[0318] In the formula, t mj and t aj These are the global max pooling and global average pooling results for the multi-source sensor features at time point j, respectively; P j Let F be the j-th encoded feature, and let F represent the number of multi-source sensors.

[0319] Based on the overall temporal information, one-dimensional convolution is used to locally enhance the overall temporal information, and temporal weights are assigned to obtain time-enhanced data:

[0320] at (n) =σ(Conv1D(θ) Conv1D ,t (n) (11)

[0321]

[0322] In the formula: θ Conv1D P represents the convolution kernel parameters. (n) The channel information interaction characteristics after location encoding Multiplying elements together.

[0323] Step 3: Global Temporal Feature Mining

[0324] In this step, a structure such as Figure 3 The DCNN shown extracts global temporal features related to tool wear information from channel-time augmented data mining in step 2. The features extracted by the deep convolutional neural network can be represented as:

[0325]

[0326] Mout (n) =MP(Ge2,ps) (17)

[0327] In the formula, BN represents batch normalization, GeLU represents the GeLU activation function, MP represents max pooling, and ps is the pooling kernel size. (n) These are the output features extracted by a deep convolutional neural network.

[0328] Step 4: Local refinement of global temporal features

[0329] In this step, the global temporal features obtained in step 3 are locally refined using a residual block with a convolutional kernel of 1, thereby obtaining global-local temporal features. The process can be represented as follows:

[0330] Conv=Conv1D(θ Conv1D ,GeLU(LN(Conv1D(θ Conv1D Mout (n) (18)

[0331] Cout (n) =add(Conv,Mout) (n) (19)

[0332] In the formula, Mout (n) For global temporal features mined by DCNN, GeLU represents the GeLU activation function, LN is the layer normalization process, and add means element-wise addition.

[0333] Step 5: Global Feature Compensation

[0334] In this step, the global long-term dependency of the global-local features obtained in step 4 is enhanced through memory enhancement and global feature compensation mechanisms. The specific process is as follows:

[0335] The attention factor of the hidden state vector is calculated based on the contribution of global features mined by DCNN to tool wear features.

[0336] β (n) =δ(Cout) (n) ⊙GMP(Mout (n) (20)

[0337] In the formula, δ represents the softmax activation function, and ⊙ represents the element-wise dot product.

[0338] Based on an attention factor, memory enhancement processing is applied to global-local features to improve their temporal dependence. Furthermore, to compensate for information loss caused by the refinement of local features, global features obtained from DCNN are used as compensation and merged with the memory-enhanced features to form enhanced global-local features.

[0339]

[0340] Step 6: Feature Decoding and Wear Monitoring Value Mapping

[0341] In this step, the following is adopted: Figure 2 The fully connected regression subnetwork in the diagram maps the enhanced global-local temporal features obtained in step 5 to tool wear values. This subnetwork consists of three fully connected layers, with ReLU activation functions added after the first two fully connected layers and Linear activation functions added after the third fully connected layer. The activation functions in the fully connected regression subnetwork are used to enhance the nonlinear fitting ability of the subnetwork, and the combination of ReLU and Linear activation functions avoids monitoring model overfitting.

[0342] In addition, Dropout regularization is introduced after the first two fully connected layers to further prevent the model from overfitting;

[0343] The fully connected regression subnetwork maps the extracted enhanced global-local temporal features into tool wear values ​​(Moni) after one tool pass. (n) This enables tool wear monitoring;

[0344] The specific calculation process for the mapping is as follows:

[0345] Moni (n) =l(w F5 ReLU,(w F4 ReLU,(w F3 Fout (n)))) (twenty three)

[0346] In the formula, Moni (n) The monitored tool wear value; l represents the Linear activation function; w F3 ,w F4 ,w F5 These represent the parameters of the three fully connected layers, Fout (n) To enhance local features, the features are processed using Flatten.

[0347] In this embodiment, the root mean square error (RMSE) and mean absolute error (MAE) are used to evaluate the tool wear monitoring performance. The expressions for RMSE and MAE are as follows:

[0348]

[0349] In the formula: n represents the total number of samples in the test set, i represents the i-th test sample, and tw i and Moni i These represent the actual tool wear value for the i-th pass and the monitored tool wear value mapped from the model, respectively.

[0350] The tool wear monitoring method based on channel-time enhancement and global compensation network described above has RMSE and MAE results on three test subsets, as shown in Table 3:

[0351] Table 3

[0352]

[0353] To more intuitively represent the error between the actual and monitored values ​​of tool wear, Figure 6 The figure shows the fitting results of the monitoring values ​​of three test subsets. The dashed line in the figure represents the actual tool wear value measured by microscope 4 (the actual tool wear value measured by microscope 4 is transmitted to the data acquisition unit 9 for display), and the solid line represents the tool wear monitoring value given by the tool wear monitoring method based on channel-time enhancement and global compensation network described in this invention.

[0354] Through Table 3 and Figure 6 As can be seen, the tool wear monitoring method based on channel-time enhancement and global compensation network proposed in this invention can effectively monitor tool wear values, thereby providing users with appropriate decision-making basis for tool replacement before tool failure, avoiding problems such as low processing efficiency and poor workpiece quality caused by tool failure and failure to replace in time, thus reducing unnecessary economic losses.

Claims

1. A tool wear monitoring method based on channel-time enhancement and global compensation networks, characterized in that, Includes the following steps: Step 1) During the operation of the tool (1), the vibration signal, cutting force signal and acoustic emission signal of the worktable (2) and the tool (1) are collected to construct multi-source sensor data; Step 2) Preprocess the multi-source sensor dataset; Step 3) Use the tool wear monitoring model to perform channel information interaction and time information enhancement processing on the preprocessed multi-source sensor data to obtain feature data after channel information interaction and time information enhancement; Step 4) Use a deep convolutional neural network to perform global temporal feature mining on the feature data enhanced by channel information interaction and temporal information to obtain global temporal features; Step 5) Refine the global temporal features locally to form global-local temporal features; Step 6) Perform global feature compensation on the global-local temporal features to enhance the long-term dependency of the features and obtain enhanced global-local temporal features; Step 7) Use a fully connected regression subnetwork to decode the compensated enhanced global-local temporal features, establish a mapping between the features and tool wear values, and output tool wear monitoring values.

2. The tool wear monitoring method based on channel-time enhancement and global compensation network according to claim 1, characterized in that, The vibration signal was acquired by a triaxial vibration sensor; The cutting force signal is acquired through a triaxial force sensor; The acoustic emission signal is acquired by an acoustic emission sensor.

3. The tool wear monitoring method based on channel-time enhancement and global compensation network according to claim 1, characterized in that, Step 2) includes the following preprocessing steps for the multi-source sensor dataset: Step 2.1) Truncate and remove the data from the multi-source sensor data, including the data that was cut in and cut out. Step 2.2) Slice the truncated multivariate data and extract the root mean square value of each data segment to form the original dataset; Step 2.3) Normalize the original dataset using Min-Max normalization so that the RMS values ​​obtained by different sensors are limited to [0, 1], and generate the input data for the tool wear monitoring model.

4. The tool wear monitoring method based on channel-time enhancement and global compensation network according to claim 3, characterized in that: In step 2.1), the first and last 2.5% of the collected multi-source sensor data points are used as the entry and exit data, and are truncated and discarded. In step 2.2), the root mean square value of each data segment is shown below: （1） In the formula: Let L represent the root mean square value of the j-th data segment, and L represent the length of a single segment. This represents the i-th data point; In step 2.3), the normalized data is shown below: （2） In the formula: and These are the raw data and the normalized data at the k-th RMS data point of the s-th sensor, respectively. and This represents the minimum and maximum values ​​on the s-th sensor.

5. The tool wear monitoring method based on channel-time enhancement and global compensation network according to claim 1, characterized in that, Step 3), which involves using the tool wear monitoring model to perform channel information interaction and temporal information enhancement processing on the preprocessed multi-source sensor data, includes the following steps: Step 3.1) Obtain the overall feature information of each sensor data using global max pooling and global average pooling. ,Right now: （3） （4） （5） In the formula: This represents the data from the i-th sensor; and This represents the overall information of the i-th channel data obtained by global max pooling and global average pooling along the data time direction; T represents the sample time window size, and concat represents vector concatenation. , These represent one-dimensional global max pooling and global average pooling operations, respectively. Step 3.2) Utilize fully connected layers to process overall feature information Scaling is applied to achieve global channel information interaction, and an activation function is used to assign channel weights to obtain channel information interaction features. ,Right now: （6） （7） In the formula: This represents the sigmoid activation function. and These are the parameters for the two fully connected layers, respectively. Indicates deviation; This is scaled feature information; For sensor data; Step 3.3) Use position coding to mark the data after the channel information exchange, and use cosine and sine functions of different frequencies to encode the marked data to obtain the coding features; Step 3.4) Along the channel direction, use global max pooling and global average pooling to obtain the overall information of the feature time direction. ,Right now: （8） （9） （10） In the formula, and These are the global max pooling and global average pooling results for the input data at the j-th time point, respectively. Let F represent the number of multi-source sensors, where j is the j-th encoded feature. , These represent global max pooling and global average pooling operations, respectively. Step 3.5) Utilize one-dimensional convolution to process the overall information Local enhancement is performed, and temporal weights are assigned to obtain feature data enhanced by channel information interaction and temporal information. ,Right now: （11） （12） In the formula: Represents the convolution kernel parameters. The channel information interaction characteristics after location encoding Multiplication of elements; This is a local enhancement feature; This represents a one-dimensional convolution operation.

6. The tool wear monitoring method based on channel-time enhancement and global compensation network according to claim 1, characterized in that, In step 4), a deep convolutional neural network is used to perform global temporal feature mining on the feature data that has been enhanced by channel information interaction and temporal information. The process of feature extraction by a deep convolutional neural network is as follows: （13） （14） （15） （16） （17） In the formula, BN represents batch normalization, GeLU represents the GeLU activation function, MP represents max pooling, and ps is the pooling kernel size; Output features extracted by a deep convolutional neural network; , , , Intermediate feature parameters in the feature extraction process of deep convolutional neural networks.

7. The tool wear monitoring method based on channel-time enhancement and global compensation network according to claim 1, characterized in that, In step 5), the global-local temporal features are shown below: （18） （19） In the formula, LN represents layer normalization, and add represents element addition; It represents global-local temporal features; This is an intermediate parameter.

8. The tool wear monitoring method based on channel-time enhancement and global compensation network according to claim 1, characterized in that, Step 6) involves performing global feature compensation on the global-local temporal features, including: Step 6.1) Based on global temporal features, calculate the contribution of tool wear features to obtain the attention factor of the hidden state vector. ,Right now: （20） In the formula, This represents the softmax activation function. Dot product of elements; Step 6.2) Based on attention factor Memory enhancement processing is applied to global-local temporal features to obtain enhanced global-local temporal features. ,Right now: （21） （22） In the formula: This represents the global temporal characteristics after memory enhancement. This represents the enhanced global-local temporal features after global feature compensation.

9. The tool wear monitoring method based on channel-time enhancement and global compensation network according to claim 1, characterized in that, In step 7), a fully connected regression subnetwork is used to establish a mapping between features and tool wear values; The fully connected regression subnetwork includes three fully connected layers. The first two fully connected layers are followed by a ReLU activation function and Dropout regularization, and the third fully connected layer is followed by a Linear activation function. The activation function and Dropout regularization are used to enhance the nonlinear fitting ability of the subnetwork.

10. The tool wear monitoring method based on channel-time enhancement and global compensation network according to claim 1, characterized in that, The tool wear monitoring values ​​are shown below: （23） In the formula, This is the value monitored for tool wear. This represents the Linear activation function; These represent the parameters of the three fully connected layers of the fully connected regression subnetwork. The features are enhanced by Flattening to improve the global-local temporal features.

Images