Cutter wear monitoring method based on physical guidance dynamic gating chart convolution framework

Abstract

The invention discloses a cutter wear monitoring method based on a physical guidance dynamic gating chart convolution framework, and the method comprises the steps: 1, collecting a cutting force signal, a bending moment signal and a vibration signal of a cutter and a workpiece in the operation process of the cutter; 2, performing data preprocessing on the acquired multi-source sensor signal sequence; 3, calculating a Pearson correlation coefficient between each feature in the original feature set and the tool wear; 4, screening the first Top-k features to construct a sensitive feature set integrating force, vibration and bending moment; 5, extracting cutter wear information in the sensitive feature set by adopting a dynamic gated graph convolutional network, and establishing long-term dependency between features; and 6, performing output decoding on the features after the long-term dependency is established, establishing a mapping relation between the tool wear value and the input features, and outputting a tool wear monitoring value. According to the method, deep fusion of the graph structure and the time sequence features is realized, and the accuracy and stability of tool wear prediction are effectively enhanced.

Description

TECHNICAL FIELD

[0001] The application relates to the field of the combination of artificial intelligence and mechanical engineering, and particularly relates to a tool wear monitoring method based on a physically guided dynamic gated graph convolution framework. BACKGROUND

[0002] With the rapid development of intelligent manufacturing technology, as a key component in the metal cutting process, the tool directly affects the machining precision, surface quality and production efficiency. Especially in the manufacturing scene with increasing requirements for high precision, high efficiency and high reliability, tool wear has become an important factor restricting the stability of the numerical control machining process and the consistency of parts. Therefore, how to accurately identify and monitor the tool wear state has become one of the core technical problems to be solved in the field of intelligent manufacturing.

[0003] At present, tool wear monitoring technology can be mainly divided into two categories: mechanism modeling-based method and data-driven method. The former relies on physical models such as cutting force, heat conduction and wear mechanism, and derives the wear evolution law through the establishment of mathematical expressions, and has certain interpretability and universality, but it is difficult to adapt to the dynamic changes under complex actual working conditions; the latter collects multi-source sensing signals (such as vibration, cutting force, acoustic emission, etc.), and models the tool state by means of machine learning and deep learning, and has strong adaptability and accuracy, but the model training is seriously dependent on data quality, and is easily disturbed by noise, and lacks physical interpretability.

[0004] In recent years, GCN has shown great potential in processing non-Euclidean structure data, and some research has tried to apply it to the tool wear monitoring task to mine the association structure between different sensor signals. However, the existing graph convolution method mostly uses static graph structure, which cannot dynamically perceive the complex relationship between time-varying machining conditions and wear state. In addition, the existing method generally ignores the physical prior knowledge in the machining process, such as the monotonicity of tool wear, the stable evolution trend and the modulation effect of cutting parameters on the wear rate, resulting in problems such as insufficient generalization ability and large prediction fluctuations of the model.

[0005] Therefore, there is an urgent need for a tool wear monitoring method that combines physical knowledge and graph learning ability and has a dynamic structure perception mechanism to improve the stability, accuracy and industrial adaptability of the model. SUMMARY

[0006] The purpose of the application is to provide a tool wear monitoring method based on a physically guided dynamic gated graph convolution framework, and the steps include:

[0007] Step 1: During the tool operation, the cutting force signal, bending moment signal and vibration signal of the tool and workpiece are collected to form a multi-source sensor signal sequence;

[0008] Step 2: Perform data preprocessing on the acquired multi-source sensor signal sequences to obtain the original feature set;

[0009] Step 3: Calculate the Pearson correlation coefficient between each feature in the original feature set and tool wear;

[0010] Step 4: Based on the Pearson correlation coefficient, select the top-k features to construct a sensitive feature set integrating force, vibration, and bending moment;

[0011] Step 5: Use a dynamic gated graph convolutional network to extract tool wear information from the sensitive feature set and establish long-term dependencies between features;

[0012] Step 6: Decode the output of the features after establishing long-term dependency, establish the mapping relationship between tool wear value and input features, and output tool wear monitoring value.

[0013] Furthermore, the cutting force signal and bending moment signal of the tool are acquired by a triaxial force sensor and a bending moment sensor integrated on the tool;

[0014] The vibration signal of the workpiece is collected by a vibration sensor attached to the workpiece.

[0015] Furthermore, step 2, the data preprocessing steps for the acquired multi-source sensor signal sequences, include:

[0016] 2.1) The collected sensor signals are truncated and discarded;

[0017] 2.2) Extract time-domain, frequency-domain, and time-frequency-domain features from the truncated and discarded sensor signals;

[0018] 2.3) Normalize the extracted time-domain, frequency-domain, and time-frequency-domain features to obtain the original feature set;

[0019] The normalization process is as follows:

[0020]

[0021] In the formula: f i s This represents the i-th feature of the s-th sensor. and These are the corresponding minimum and maximum values. This represents the normalized features.

[0022] Furthermore, minimum value and maximum value It comes from the training sample set.

[0023] Furthermore, the Pearson correlation coefficient r between each feature and tool wear is shown below:

[0024]

[0025] In the formula: |·| represents the absolute value operation, which is used to eliminate the influence of the positive or negative correlation on the analysis. Let y represent the j-th feature point of the ith feature of the s-th sensor. j This represents the tool wear value corresponding to the j-th cut. and These represent the average value of the i-th feature of the s-th sensor and the average value of tool wear, respectively, and N represents the total number of cutting operations.

[0026] Furthermore, step 5, which involves using a dynamic gated graph convolutional network to extract tool wear information from the sensitive feature set, includes:

[0027] 5.1) Construct the dynamic trajectory matrix; where the trajectory matrix corresponding to the l-th gated convolution is shown below:

[0028]

[0029] In the formula: δ(·) represents the softmax activation function; A (l) This represents the trajectory matrix before normalization;

[0030] 5.2) Introduce an adaptive threshold strategy to establish connections or disconnections between features, i.e.:

[0031]

[0032] A (l) =ReLU(S (l) -τ (l) (5)

[0033]

[0034] In the formula: S (l) Let represent the cosine similarity matrix of the l-th layer. This represents the standardized input of the l-th layer. for The transpose matrix; ReLU(·) denotes the ReLU activation function, τ (l) The adaptive threshold for the l-th layer. μ(S) (l) ) and α(S (l) Let be the mean and deviation of the similarity matrices. When l is 1, For the selection of sensitive feature set x (n)The standardized processing results. Let be the cosine similarity parameter corresponding to the j-th neuron in the l-th layer.

[0035] 5.3) Utilize gated graph convolutional networks to extract tool wear information, thereby establishing long-term dependencies;

[0036] The gated graph convolutional network is shown below:

[0037]

[0038] In the formula: Gr i (l) and These are the update gate, reset gate, and candidate state output for the i-th time step of the l-th gated graph convolutional layer, respectively. and These are the weight matrices for the update gate, reset gate, and candidate states of the l-th gated graph convolutional layer, respectively. and This is the corresponding bias. and This represents the current time step input and the previous hidden layer state vector of the l-th gated graph convolutional layer; This represents the hidden state vector at the i-th time step of the l-th gated graph convolutional layer; This represents the weights of the update gate, reset gate, and candidate state; This represents the candidate state vector of the hidden layer.

[0039] Furthermore, when the gated graph convolutional network is a multi-layer dynamic gated graph convolutional network, the input time series x (n) The hidden layer state vectors are processed into a graph structure, i.e.:

[0040]

[0041] In the formula: DGGCN(·) represents a dynamically gated graph convolutional layer; when l is 1, H (1) That is, the sensitive feature set x to be selected. (n) . H represents the hidden state vector at the i-th time step of the (l+1)-th gated graph convolutional layer; (l+1) This represents the current time step input of the (l+1)th gated graph convolutional layer.

[0042] Furthermore, the gated graph convolutional network is trained on a historical dataset; the historical dataset includes a set of historical sensitive features integrating force, vibration, and bending moment, as well as corresponding tool wear information.

[0043] The loss function for gated graph convolutional networks during training is shown below:

[0044] Loss = λ1·Loss d +λ2·(Loss p +Loss k (15)

[0045] In the formula: λ1 and λ2 represent the data-driven loss, respectively. d Weights of physical guidance loss; Loss p Loss k This represents the loss of wear monotonicity and the loss of physical consistency.

[0046] Furthermore, data-driven loss d As shown below:

[0047]

[0048] In the formula: ||·||2 represents the L2 norm, y and These are the actual tool wear value and the tool wear value estimated by the model, respectively.

[0049] The soft constraint on wear monotonicity is shown below:

[0050]

[0051] In the formula: Indicates tool wear rate; Loss k This indicates the wear monotonicity loss;

[0052] The physical consistency constraints are as follows:

[0053]

[0054] f(t) = a·ln(b·t+1) + c·t 3 (19)

[0055] In the formula: f represents the physical model of tool wear; a, b, and c are adjustable empirical parameters; t is time. p This represents the loss of physical consistency.

[0056] Furthermore, in step 6, the mapping relationship between tool wear values ​​and input features is constructed through two fully connected layers;

[0057] The mapping relationship between tool wear values ​​and input features is shown below:

[0058]

[0059] In the formula, The value represents the tool wear monitored by the model; l(·) represents the Linear activation function; w F1and w F2 H represents the trainable parameters of the two fully connected layers respectively. (L) This represents the output of a dynamic gated graph convolutional network with a total of L layers.

[0060] The technical effects of this invention are undeniable. The beneficial effects of this invention are:

[0061] (1) Utilize physical guidance mechanisms to enhance the model's ability to capture wear evolution trends and solve the problem of lack of physical interpretation in traditional data-driven methods;

[0062] (2) Introduce a dynamic gated graph convolution structure to enable the model to adaptively model the relationship between features in multiple domains and improve the robustness of monitoring under complex working conditions;

[0063] (3) It achieves deep integration of graph structure and time series features, effectively enhancing the accuracy and stability of tool wear prediction. Attached Figure Description

[0064] Figure 1 This is a flowchart of the tool wear monitoring method based on a physically guided dynamic gated graph convolutional network according to the present invention;

[0065] Figure 2 This is a graph structure feature construction graph of the present invention;

[0066] Figure 3 This is the baseline model diagram of the convolutional neural network structure of the present invention;

[0067] Figure 4 This is the composite loss function of physical constraints used for model training in this invention;

[0068] Figure 5 This is a diagram of the multi-source sensor data acquisition device for tool wear in an embodiment of the present invention;

[0069] Figure 6 This is a preprocessed image of multi-sensor data in an embodiment of the present invention;

[0070] Figure 7 This is the multi-domain feature Pearson correlation coefficient matrix in this embodiment of the invention;

[0071] Figure 8 This is a schematic diagram of the wear monitoring results obtained from the test set in the milling tool wear experimental data set of this embodiment of the invention. (a) shows the monitoring results of test subset D1; (b) shows the monitoring results of test subset D2; and (c) shows the monitoring results of test subset D3. Detailed Implementation

[0072] 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.

[0073] Example 1:

[0074] See Figures 1 to 8 A method for monitoring tool wear based on a physically guided dynamic gating graph convolution framework, comprising the following steps:

[0075] Step 1: During the tool operation, the cutting force signal, bending moment signal and vibration signal of the tool and workpiece are collected to form a multi-source sensor signal sequence;

[0076] Step 2: Perform data preprocessing on the acquired multi-source sensor signal sequences to obtain the original feature set;

[0077] Step 3: Calculate the Pearson correlation coefficient between each feature in the original feature set and tool wear;

[0078] Step 4: Based on the Pearson correlation coefficient, select the top-k features to construct a sensitive feature set integrating force, vibration, and bending moment;

[0079] Step 5: Use a dynamic gated graph convolutional network to extract tool wear information from the sensitive feature set and establish long-term dependencies between features;

[0080] Step 6: Decode the output of the features after establishing long-term dependency, establish the mapping relationship between tool wear value and input features, and output tool wear monitoring value.

[0081] The cutting force and bending moment signals of the tool are acquired by a triaxial force sensor and a bending moment sensor integrated on the tool.

[0082] The vibration signal of the workpiece is collected by a vibration sensor attached to the workpiece.

[0083] Step 2, the data preprocessing steps for the acquired multi-source sensor signal sequences include:

[0084] 2.1) The collected sensor signals are truncated and discarded;

[0085] 2.2) Extract time-domain, frequency-domain, and time-frequency-domain features from the truncated and discarded sensor signals;

[0086] 2.3) Normalize the extracted time-domain, frequency-domain, and time-frequency-domain features to obtain the original feature set;

[0087] The normalization process is as follows:

[0088]

[0089] In the formula: f i s This represents the i-th feature of the s-th sensor. and These are the corresponding minimum and maximum values. This represents the normalized features.

[0090] Minimum value and maximum value It comes from the training sample set.

[0091] The Pearson correlation coefficient r between each feature and tool wear is shown below:

[0092]

[0093] In the formula: |·| represents the absolute value operation, which is used to eliminate the influence of the positive or negative correlation on the analysis. Let y represent the j-th feature point of the ith feature of the s-th sensor. j This represents the tool wear value corresponding to the j-th cut. and These represent the average value of the i-th feature of the s-th sensor and the average value of tool wear, respectively, and N represents the total number of cutting operations.

[0094] Step 5, which involves using a dynamic gated graph convolutional network to extract tool wear information from the sensitive feature set, includes:

[0095] 5.1) Construct the dynamic trajectory matrix; where the trajectory matrix corresponding to the l-th gated convolution is shown below:

[0096]

[0097] In the formula: δ(·) represents the softmax activation function (used for normalization); A (l) This represents the trajectory matrix before normalization;

[0098] 5.2) Introduce an adaptive threshold strategy to establish connections or disconnections between features, i.e.:

[0099]

[0100] A (l) =ReLU(S (l) -τ (l) (5)

[0101]

[0102] In the formula: S (l) Let represent the cosine similarity matrix of the l-th layer. This represents the standardized input of the l-th layer. for The transpose matrix; ReLU(·) denotes the ReLU activation function, τ (l) The adaptive threshold for the l-th layer. μ(S) (l) ) and α(S (l) Let be the mean and deviation of the similarity matrices. When l is 1, For the selection of sensitive feature set x (n) The standardized processing results. Let be the cosine similarity parameter corresponding to the j-th neuron in the l-th layer.

[0103] 5.3) Utilize gated graph convolutional networks to extract tool wear information, thereby establishing long-term dependencies;

[0104] The gated graph convolutional network is shown below:

[0105]

[0106]

[0107] In the formula: Gr i (l) and These are the update gate, reset gate, and candidate state output for the i-th time step of the l-th gated graph convolutional layer, respectively. and These are the weight matrices for the update gate, reset gate, and candidate states of the l-th gated graph convolutional layer, respectively. and This is the corresponding bias. and This represents the current time step input and the previous hidden layer state vector of the l-th gated graph convolutional layer; This represents the hidden state vector at the i-th time step of the l-th gated graph convolutional layer; The weights (obtained through training) represent the update gate, reset gate, and candidate state. This represents the candidate state vector of the hidden layer.

[0108] When the gated graph convolutional network is a multi-layer dynamic gated graph convolutional network, the input time series x (n) The hidden layer state vectors are processed into a graph structure, i.e.:

[0109]

[0110] In the formula: DGGCN(·) represents a dynamically gated graph convolutional layer; when l is 1, H (1) That is, the sensitive feature set x to be selected. (n) . H represents the hidden state vector at the i-th time step of the (l+1)-th gated graph convolutional layer; (l+1) This represents the current time step input of the (l+1)th gated graph convolutional layer.

[0111] The gated graph convolutional network is trained on a historical dataset; the historical dataset includes a set of historical sensitive features integrating force, vibration and bending moment, as well as corresponding tool wear information.

[0112] The loss function for gated graph convolutional networks during training is shown below:

[0113] Loss = λ1·Loss d +λ2·(Loss p +Loss k (15)

[0114] In the formula: λ1 and λ2 represent the data-driven loss, respectively. d Weights of physical guidance loss; Loss p Loss k This represents the loss of wear monotonicity and the loss of physical consistency.

[0115] Data-driven loss d As shown below:

[0116]

[0117] In the formula: ||·||2 represents the L2 norm, y and These are the actual tool wear value and the tool wear value estimated by the model, respectively.

[0118] The soft constraint on wear monotonicity is shown below:

[0119]

[0120] In the formula: Indicates tool wear rate; Loss k This indicates the wear monotonicity loss;

[0121] The physical consistency constraints are as follows:

[0122]

[0123] f(t) = a·ln(b·t+1) + c·t 3 (19)

[0124] In the formula: f represents the physical model of tool wear; a, b, and c are adjustable empirical parameters (obtained through training on a training set); t is time. Loss p This represents the loss of physical consistency.

[0125] In step 6, the mapping relationship between tool wear values ​​and input features is constructed through two fully connected layers;

[0126] The mapping relationship between tool wear values ​​and input features is shown below:

[0127]

[0128] In the formula, The value represents the tool wear monitored by the model; l(·) represents the Linear activation function; w F1 and w F2 H represents the trainable weight parameters of the two fully connected layers, respectively. (L) This represents the output of a dynamic gated graph convolutional network with a total of L layers.

[0129] Example 2:

[0130] A tool wear monitoring method based on a physically guided dynamic gating graph convolution framework, comprising the following steps:

[0131] Step 1: During the tool operation, the cutting force signal, bending moment signal and vibration signal of the tool and workpiece are collected to form a multi-source sensor signal sequence;

[0132] Step 2: Perform data preprocessing on the acquired multi-source sensor signal sequences to obtain the original feature set;

[0133] Step 3: Calculate the Pearson correlation coefficient between each feature in the original feature set and tool wear;

[0134] Step 4: Based on the Pearson correlation coefficient, select the top-k features to construct a sensitive feature set integrating force, vibration, and bending moment;

[0135] Step 5: Use a dynamic gated graph convolutional network to extract tool wear information from the sensitive feature set and establish long-term dependencies between features;

[0136] Step 6: Decode the output of the features after establishing long-term dependency, establish the mapping relationship between tool wear value and input features, and output tool wear monitoring value.

[0137] Example 3:

[0138] A tool wear monitoring method based on a physical-guided dynamic gating graph convolution framework, with the same technical content as in Embodiment 2, further wherein the cutting force signal and bending moment signal of the tool are acquired by a triaxial force sensor and a bending moment sensor integrated on the tool;

[0139] The vibration signal of the workpiece is collected by a vibration sensor attached to the workpiece.

[0140] Example 4:

[0141] A tool wear monitoring method based on a physically guided dynamic gating graph convolution framework, with the same technical content as any one of embodiments 2-3, further comprising the following steps in step 2: data preprocessing of the acquired multi-source sensor signal sequence:

[0142] 2.1) The collected sensor signals are truncated and discarded;

[0143] 2.2) Extract time-domain, frequency-domain, and time-frequency-domain features from the truncated and discarded sensor signals;

[0144] 2.3) Normalize the extracted time-domain, frequency-domain, and time-frequency-domain features to obtain the original feature set;

[0145] The normalization process is as follows:

[0146]

[0147] In the formula: f i s This represents the i-th feature of the s-th sensor. and These are the corresponding minimum and maximum values.

[0148] Example 5:

[0149] A tool wear monitoring method based on a physically guided dynamic gating graph convolution framework, with the same technical content as any one of embodiments 2-4, further comprising the minimum value. and maximum value It comes from the training sample set.

[0150] Example 6:

[0151] A tool wear monitoring method based on a physically guided dynamic gating graph convolution framework, with the same technical content as any one of embodiments 2-5, further wherein the Pearson correlation coefficient between each feature and tool wear is shown below:

[0152]

[0153] In the formula: |·| represents the absolute value operation, which is used to eliminate the influence of the positive or negative correlation on the analysis. Let y represent the j-th feature point of the ith feature of the s-th sensor. j This represents the tool wear value corresponding to the j-th cut. and These represent the average value of the i-th feature of the s-th sensor and the average value of tool wear, respectively, and N represents the total number of cutting operations.

[0154] Example 7:

[0155] A tool wear monitoring method based on a physically guided dynamic gating graph convolutional framework, with the same technical content as any one of embodiments 2-6, further comprising the following steps in step 5: extracting tool wear information from the sensitive feature set using a dynamic gating graph convolutional network:

[0156] 5.1) Construct the dynamic trajectory matrix; where the trajectory matrix corresponding to the l-th gated convolution is shown below:

[0157]

[0158] In the formula: δ(·) represents the softmax activation function;

[0159] 5.2) Introduce an adaptive threshold strategy to establish connections or disconnections between features, i.e.:

[0160]

[0161] A (l) =ReLU(S (l) -τ (l) (5)

[0162]

[0163] In the formula: S (l) Let represent the cosine similarity matrix of the l-th layer. This represents the standardized input of the l-th layer. Let τ be its transpose matrix. ReLU(·) denotes the ReLU activation function, where τ is the transpose matrix. (l) The adaptive threshold for the l-th layer. μ(S) (l) ) and α(S (l) Let be the mean and deviation of the similarity matrices. When l is 1, That is, the sensitive feature set x to be selected. (n) The standardized processing results.

[0164] 5.3) Utilize gated graph convolutional networks to extract tool wear information, thereby establishing long-term dependencies;

[0165] The gated graph convolutional network is shown below:

[0166]

[0167] In the formula: Gr i (l) and These are the update gate, reset gate, and candidate state output for the i-th time step of the l-th gated graph convolutional layer, respectively. and These are the weight matrices for the update gate, reset gate, and candidate states of the l-th gated graph convolutional layer, respectively. and This is the corresponding bias. and This represents the current time step input and the state vector of the previous hidden layer of the l-th gated graph convolutional layer. This represents the hidden layer state vector at the i-th time step of the l-th gated graph convolutional layer.

[0168] Example 8:

[0169] A tool wear monitoring method based on a physically guided dynamic gated graph convolutional framework, with the same technical content as any one of embodiments 2-7, further comprising the following steps: when the gated graph convolutional network is a multi-layer dynamic gated graph convolutional network, the input time series x... (n) The hidden layer state vectors are processed into a graph structure, i.e.:

[0170]

[0171] In the formula: DGGCN(·) represents a dynamically gated graph convolutional layer; when l is 1, H (1) That is, the sensitive feature set x to be selected. (n) .

[0172] Example 9:

[0173] A tool wear monitoring method based on a physically guided dynamic gated graph convolutional framework, with the same technical content as any one of embodiments 2-8, further wherein the gated graph convolutional network is trained from a historical dataset; the historical dataset includes a set of historical sensitive features integrating force, vibration and bending moment, as well as corresponding tool wear related information;

[0174] The loss function for gated graph convolutional networks during training is shown below:

[0175] Loss = λ1·Loss d +λ2·(Loss p +Loss k (15)

[0176] In the formula: λ1 and λ2 represent the data-driven loss, respectively. d Weights of physical guidance loss; Loss pLoss k This represents the loss of wear monotonicity and the loss of physical consistency.

[0177] Example 10:

[0178] A tool wear monitoring method based on a physically guided dynamic gating graph convolution framework, with the same technical content as any one of embodiments 2-9, further comprising data-driven loss. d As shown below:

[0179]

[0180] In the formula: ||·||2 represents the L2 norm, y and These are the actual tool wear value and the tool wear value estimated by the model, respectively.

[0181] The soft constraint on wear monotonicity is shown below:

[0182]

[0183] In the formula: Indicates the tool wear rate;

[0184] The physical consistency constraints are as follows:

[0185]

[0186] f(t) = a·ln(b·t+1) + c·t 3 (19)

[0187] In the formula: f represents the physical model of tool wear; a, b and c are adjustable empirical parameters.

[0188] Example 11:

[0189] A tool wear monitoring method based on a physically guided dynamic gating graph convolutional framework, with the same technical content as any one of embodiments 2-10, further wherein, in step 6, the mapping relationship between the tool wear value and the input feature is constructed through two fully connected layers;

[0190] The mapping relationship between tool wear values ​​and input features is shown below:

[0191]

[0192] In the formula, The value represents the tool wear monitored by the model; l(·) represents the Linear activation function; w F1 and w F2 H represents the trainable parameters of the two fully connected layers respectively. (L) This represents the output of a dynamic gated graph convolutional network with a total of L layers.

[0193] Example 12:

[0194] A tool wear monitoring method based on a physically guided dynamic gating graph convolution framework is described below:

[0195] Figure 1 The implementation flow of the tool wear monitoring method based on a physically guided dynamic gated graph convolutional network proposed in this invention is illustrated. First, multi-source sensor signals generated by the tool during machining are collected, including but not limited to typical physical quantities such as cutting force, vibration, and bending moment. Then, preprocessing operations are performed on the raw signal data, including in-cut and out-cut rejection, multi-domain feature extraction, and sensitive feature screening.

[0196] After data preprocessing, the selected sensitive feature set is input into the designed dynamic gated graph convolutional subnetwork. This subnetwork has a time-varying connection control mechanism, which can adjust the weights of information flow transmission in the graph according to physical changes, thereby modeling the dynamic dependencies between key nodes. Tool wear monotonicity and data physical consistency constraints are embedded during graph convolution to improve the physical consistency of feature representation and prediction stability.

[0197] The representation features extracted by the dynamic gated graph convolution module are input into the fully connected layer to complete the mapping and prediction of tool wear values. The output is the estimated tool wear value under the current machining state, which can be used for subsequent remaining tool life prediction or machining process optimization control.

[0198] This method achieves an integrated monitoring mechanism that combines "physical knowledge guidance + dynamic graph structure modeling + gating mechanism adjustment," which improves the accuracy, stability, and interpretability of tool wear prediction while ensuring model flexibility.

[0199] Specifically, the steps include the following:

[0200] Step 1: Use an intelligent cutting tool with integrated triaxial force sensor and bending moment sensor and a vibration sensor attached to the workpiece to collect the cutting force signal, bending moment signal and vibration signal of the cutting tool (1) and the workpiece (2) during the operation of the cutting tool (1), and form a multi-source sensor signal;

[0201] Step 2: Perform data preprocessing on the collected multi-source sensor signals to obtain the original feature set;

[0202] Step 3: Perform correlation analysis on the extracted original multi-domain features, and evaluate the richness of tool wear information contained in the features by analyzing the Pearson correlation coefficient between each feature and tool wear;

[0203] Step 4: Sort the correlation analysis results by numerical value and select the top-k features to form a sensitive feature set integrating force, vibration and bending moment;

[0204] Step 5: Use a gated dynamic graph convolutional network to extract tool wear-related information from the sensitive feature set and establish long-term dependencies between features;

[0205] Step 6: Decode the output of the features after establishing long-term dependency, establish the mapping relationship between tool wear value and input features, and output tool wear monitoring value.

[0206] Step 2, specifically the preprocessing of multi-source sensor signal data, includes:

[0207] 1) Cut-in and cut-out rejection

[0208] The collected sensor signals are truncated and discarded.

[0209] 2) Multi-domain feature extraction

[0210] For the truncated and discarded sensor signals, feature extraction is performed in the time domain, frequency domain, and time-frequency domain. Specifically, 12 time-domain features, such as mean and maximum value, and 12 frequency-domain features, such as spectral mean and spectral variance, are extracted. In addition, a three-level wavelet decomposition strategy is used to calculate energy to extract 8 time-frequency domain features, resulting in 32 features extracted for each sensor signal.

[0211] 3) Feature normalization processing

[0212] The extracted time-domain, frequency-domain, and time-frequency-domain features are normalized to form the original feature set. This ensures that different sensors and different features are within the same dimensional range.

[0213]

[0214] In the formula: f i s This represents the i-th feature of the s-th sensor. and These are the corresponding minimum and maximum values. It should be noted that the test samples are unknown during the actual processing; therefore, the maximum and minimum values ​​used for normalization of the test set are derived from the training set.

[0215] The correlation analysis described in step 3 can be expressed as:

[0216]

[0217] In the formula: |·| represents the absolute value operation, which is used to eliminate the influence of the positive or negative correlation on the analysis. Let y represent the j-th feature point of the ith feature of the s-th sensor. j This represents the tool wear value corresponding to the j-th cut. and These represent the average value of the i-th feature of the s-th sensor and the average value of tool wear, respectively, and N represents the total number of cutting operations.

[0218] The sensitive feature set in step 4 is constructed as follows:

[0219] Features are sorted by sensor category and feature index based on relevance, and the top-k features are selected to construct a feature set for model training and testing. It should be noted that complete test samples are unavailable during actual machining, and the tool wear values ​​corresponding to the test samples are also unavailable. Therefore, the index used for feature selection in the test set is based on the training set and is consistent with the training set index.

[0220] In step 5, a physically guided gated dynamic graph convolutional network is used to extract tool wear-related information from the sensitive feature set. The specific process is as follows:

[0221] 1) Construction of dynamic trajectory matrix:

[0222] For the l-th gated convolution, the corresponding trajectory matrix is ​​constructed as follows:

[0223]

[0224] In the formula: δ(·) represents the softmax activation function. Subsequently, an adaptive thresholding strategy is introduced to establish or disconnect connections between features, which can be expressed as:

[0225]

[0226] A (l) =ReLU(S (l) -τ (l) (5)

[0227]

[0228] In the formula: S (l) Let represent the cosine similarity matrix of the l-th layer. This represents the standardized input of the l-th layer. Let τ be its transpose matrix. ReLU(·) denotes the ReLU activation function, where τ is the transpose matrix. (l) The adaptive threshold for the l-th layer. μ(S) (l) ) and α(S (l) Let be the mean and deviation of the similarity matrices. When l is 1, That is, the sensitive feature set x to be selected. (n) The standardized processing results.

[0229] 2) Tool wear information extraction and long-term dependency establishment based on gated graph convolutional networks:

[0230] To obtain information related to tool wear from the selected sensitive feature set and establish long-term dependencies between features, gated graph convolution is used to learn features from the sensitive feature set. This process can be represented as follows:

[0231]

[0232] In the formula: Gr i (l) and These are the update gate, reset gate, and candidate state output for the i-th time step of the l-th gated graph convolutional layer, respectively. and These are the weight matrices for the update gate, reset gate, and candidate state of the l-th dynamically gated convolutional layer, respectively. and This is the corresponding bias. and This represents the current time step input and the state vector of the previous hidden layer of the l-th gated graph convolutional layer. This represents the hidden layer state vector at the i-th time step of the l-th gated graph convolutional layer.

[0233] For a multi-layer dynamic gated graph convolutional network, the input time series x (n) It can be processed as a hidden layer state vector with a graph structure:

[0234]

[0235] In the formula: DGGCN(·) represents a dynamically gated graph convolutional layer; when l is 1, H (1) That is, the sensitive feature set x to be selected. (n) .

[0236] 3) Construction of the composite loss function:

[0237] To achieve robust and physically consistent tool wear prediction, this invention constructs a composite loss function that integrates data-driven approaches and physical knowledge constraints to jointly optimize model performance and physical consistency. This loss function consists of three parts, jointly modeled through a weighted summation method, and its overall form is shown below:

[0238] Loss = λ1·Loss d +λ2·(Loss p +Loss k (15)

[0239] In the formula, λ1 and λ2 represent the weights of the data-driven loss and the physical guidance loss, respectively. To balance the dominant roles of each loss term during training, this invention designs an adaptive half-decay weighting strategy. The total number of training rounds is set to 5000. In the first 2500 rounds, the weights λ1 and λ2 of the data-driven loss and the physical guidance loss are both set to 1 to accelerate error convergence in the initial stage. After the 2500th round, if the condition λ1·Loss is satisfied... d ≥2·λ2·(Loss p +Loss k If λ1 is halved every 100 rounds, then λ1 is updated every 100 rounds until the inequality no longer holds. To maintain the symmetry of the weight adjustment mechanism, this invention also sets the same trigger condition for λ2: that is, when λ2·(Loss) = 1 / 2, the weight adjustment condition is 1 / 2. p +Loss k )≥2·λ1·Loss d At the same time, the half-life update strategy is also executed. Loss d Loss p and Loss k These represent the data-driven loss term, the wear monotonicity soft constraint, and the physical consistency constraint, respectively.

[0240] Each constraint term in the composite loss function is as follows:

[0241] Data-driven loss term: Mean squared error (MSE) is used as the basic supervised learning objective to minimize the difference between the model output value and the actual tool wear label.

[0242]

[0243] In the formula: ||·||2 represents the L2 norm, y and These represent the actual tool wear value and the model-estimated tool wear value, respectively. This term is mainly used to guide the model in making accurate regression predictions and has strong convergence and numerical stability.

[0244] Wear monotonicity constraint: In order to introduce the physical prior knowledge that "tool wear accumulates gradually over time and is irreversible", a soft constraint of the non-negativity of the first-order time derivative is introduced into the loss function.

[0245]

[0246] In the formula: This represents the tool wear rate. Specifically, if the model estimate shows a decreasing trend over time, it is constrained by penalizing the negative gradient term, thereby causing the estimated curve to maintain a monotonically increasing trend.

[0247] Physical consistency constraint: To improve the physical interpretability of the model predictions, the first derivative of wear calculated by a reference physical model is introduced as a reference curve, and its residual is compared with that of the first derivative of the data-driven prediction. By minimizing the difference between the two, the model output is further guided to maintain dynamic consistency with the physical prior.

[0248]

[0249] In the formula: f represents the physical model of tool wear. This loss term is used to evaluate the model estimate and the corresponding physical model value, so that the estimated value and the physical model output are consistent.

[0250] 4) Physical Model Construction

[0251] To more accurately describe the wear evolution of cutting tools during machining, this invention introduces an empirical physical model of tool wear with adjustable parameters.

[0252] f(t) = a·ln(b·t+1) + c·t 3 (19)

[0253] In the formula: a, b, and c are adjustable empirical parameters that can be obtained by fitting tool wear values.

[0254] The mapping relationship between tool wear values ​​and input features established in step 6 is constructed through two fully connected layers, which can be specifically described as follows:

[0255]

[0256] In the formula, The value represents the tool wear monitored by the model; l(·) represents the Linear activation function; w F1 and w F2 H represents the trainable parameters of the two fully connected layers respectively. (L) This represents the output of a dynamic gated graph convolutional network with a total of L layers.

[0257] Example 13:

[0258] The verification of a tool wear monitoring method based on a physically guided dynamic gated graph convolutional network is as follows:

[0259] In this embodiment, the tool wear dataset used comes from the milling tool life machining experiment at Shandong University. The experimental setup and key machining parameters are as follows: Figure 4As shown in Table 1, this dataset contains complete machining records for three full-life milling tools, each undergoing 100 pass operations. During each pass, the system simultaneously acquired multimodal physical signals, including vibration signals in the x, y, and z directions, cutting force signals, and bending moment signals, comprehensively recording the dynamic evolution information of the tool during the machining process.

[0260] To ensure the fairness of data partitioning and the robustness of model evaluation, this invention employs a three-fold cross-validation strategy to rationally partition the dataset. Specifically, in each round of partitioning, the full lifespan data of two tools are selected as the training set, and the data of the remaining tool is used as the test set. Training and validation are performed alternately in this manner, ensuring that each subset can participate in model learning and testing in different rounds, thereby improving the model's generalization ability and reliability.

[0261] Table 1

[0262]

[0263] Table 2

[0264]

[0265] After acquiring multi-source sensor signals during the milling process, such as Figure 1 As shown, a specific implementation of the present invention should also include the following steps:

[0266] Step 1: Data Preprocessing

[0267] Sub-step (1), such as Figure 5 As shown, during each cutting operation, the system synchronously acquires seven-dimensional multi-channel raw signals, specifically including triaxial vibration signals, triaxial cutting force signals, and single-channel bending moment signals. To ensure the high stability and representativeness of the extracted features, the raw signals are truncated during the preprocessing stage, systematically removing 2.5% of the non-steady-state segments at the beginning and end of the signal to avoid transient interference that may be introduced during the tool entry and exit stages, thereby improving the accuracy and robustness of subsequent modeling.

[0268] Sub-step (2), such as Figure 6 As shown, the data after input and output are subjected to multi-domain feature extraction, specifically including 12 time-domain features such as mean and maximum value, and 12 frequency-domain features such as spectral mean and spectral variance, as detailed in Table 3. Furthermore, a three-level wavelet decomposition method is used to extract energy and obtain 8 time-frequency domain features. That is, 32 features are acquired for each sensor signal.

[0269] Fea = [Fea] (1) ,Fea (2) ,...,Fea (s) ,...,Fea(S) (1)

[0270]

[0271] In the formula: S represents the total number of sensor signals, which is 7 in this use case. (s) Let represent the feature vector extracted from the s-th sensor signal.

[0272] Table 3

[0273]

[0274]

[0275] Sub-step (3) involves normalizing the collected multi-domain features. The corresponding formula can be expressed as:

[0276]

[0277] In the formula: f i s This represents the i-th feature of the s-th sensor. and These are the corresponding minimum and maximum values. It should be noted that the test samples are unknown during the actual processing; therefore, the maximum and minimum values ​​used for normalization of the test set are derived from the training set.

[0278] Sub-step (4), such as Figure 7 As shown, Pearson correlation analysis was performed on the normalized multi-domain feature set, and the top-k features with the highest correlation coefficients were selected to construct a sensitive feature set. The correlation analysis formula can be expressed as:

[0279]

[0280] In the formula: |·| represents the absolute value operation, which is used to eliminate the influence of the positive or negative correlation on the analysis. Let y represent the j-th feature point of the ith feature of the s-th sensor. j This represents the tool wear value corresponding to the j-th cut. and These represent the average value of the i-th feature of the s-th sensor and the average value of tool wear, respectively, and N represents the total number of cutting operations.

[0281] Step 2: Model Training

[0282] A physically guided gated dynamic graph convolutional network is used to extract tool wear-related information from a set of sensitive features. The specific process is as follows:

[0283] 1) Construction of dynamic trajectory matrix:

[0284] For the l-th gated convolution, the corresponding trajectory matrix is ​​constructed as follows:

[0285]

[0286] In the formula: δ(·) represents the softmax activation function. Subsequently, an adaptive thresholding strategy is introduced to establish or disconnect connections between features, which can be expressed as:

[0287]

[0288] A (l) =ReLU(S (l) -τ (l) (7)

[0289]

[0290]

[0291] In the formula: S (l) Let represent the cosine similarity matrix of the l-th layer. This represents the standardized input of the l-th layer. Let τ be its transpose matrix. ReLU(·) denotes the ReLU activation function, where τ is the transpose matrix. (l) The adaptive threshold for the l-th layer. μ(S) (l) ) and α(S (l) Let be the mean and deviation of the similarity matrices. When l is 1, That is, the sensitive feature set x to be selected. (n) The standardized processing results.

[0292] 2) Tool wear information extraction and long-term dependency establishment based on gated graph convolutional networks:

[0293] To obtain information related to tool wear from the selected sensitive feature set and establish long-term dependencies between features, gated graph convolution is used to learn features from the sensitive feature set. This process can be represented as follows:

[0294]

[0295] In the formula: Gr i (l) and These are the update gate, reset gate, and candidate state output for the i-th time step of the l-th gated graph convolutional layer, respectively. and These are the weight matrices for the update gate, reset gate, and candidate states of the l-th gated graph convolutional layer, respectively. and This is the corresponding bias. and This represents the current time step input and the state vector of the previous hidden layer of the l-th gated graph convolutional layer. This represents the hidden layer state vector at the i-th time step of the l-th gated graph convolutional layer.

[0296] For a multi-layer dynamic gated graph convolutional network, the input time series x (n) It can be processed as a hidden layer state vector with a graph structure:

[0297]

[0298] In the formula: DGGCN(·) represents a dynamically gated graph convolutional layer; when l is 1, H (1) That is, the sensitive feature set x to be selected. (n) .

[0299] 3) Construction of the composite loss function:

[0300] To achieve robust and physically consistent tool wear prediction, this invention constructs a composite loss function that integrates data-driven approaches and physical knowledge constraints to jointly optimize model performance and physical consistency. This loss function consists of three parts, jointly modeled through a weighted summation method, and its overall form is shown below:

[0301] Loss = λ1·Loss d +λ2·(Loss p +Loss k (16)

[0302] In the formula, λ1 and λ2 represent the weights of the data-driven loss and the physical guidance loss, respectively. To balance the dominant roles of each loss term during training, this invention designs an adaptive half-decay weighting strategy. The total number of training rounds is set to 5000. In the first 2500 rounds, the weights λ1 and λ2 of the data-driven loss and the physical guidance loss are both set to 1 to accelerate error convergence in the initial stage. After the 2500th round, if the condition λ1·Loss is satisfied... d ≥2·λ2·(Loss p +Loss k If λ1 is halved every 100 rounds, then λ1 is updated every 100 rounds until the inequality no longer holds. To maintain the symmetry of the weight adjustment mechanism, this invention also sets the same trigger condition for λ2: that is, when λ2·(Loss) = 1 / 2, the weight adjustment condition is 1 / 2. p +Loss k )≥2·λ1·Loss d At the same time, the half-life update strategy is also executed. Loss d Loss p and Loss k These represent the data-driven loss term, the wear monotonicity soft constraint, and the physical consistency constraint, respectively.

[0303] Each constraint term in the composite loss function is as follows:

[0304] Data-driven loss term: Mean squared error (MSE) is used as the basic supervised learning objective to minimize the difference between the model output value and the actual tool wear label.

[0305]

[0306] In the formula: ||·||2 represents the L2 norm, y and These represent the actual tool wear value and the model-estimated tool wear value, respectively. This term is mainly used to guide the model in making accurate regression predictions and has strong convergence and numerical stability.

[0307] Wear monotonicity constraint: In order to introduce the physical prior knowledge that "tool wear accumulates gradually over time and is irreversible", a soft constraint of non-negativity of the first derivative is introduced into the loss function.

[0308]

[0309] In the formula: This represents the tool wear rate. Specifically, if the model estimate shows a decreasing trend over time, it is constrained by penalizing the negative gradient term, thereby causing the estimated curve to maintain a monotonically increasing trend.

[0310] Physical consistency constraint: To improve the physical interpretability of the model predictions, the first derivative of wear calculated by a reference physical model is introduced as a reference curve, and its residual is compared with that of the first derivative of the data-driven prediction. By minimizing the difference between the two, the model output is further guided to maintain dynamic consistency with the physical prior.

[0311]

[0312] In the formula: f represents the physical model of tool wear. This loss term is used to evaluate the model estimate and the corresponding physical model value, so that the estimated value and the physical model output are consistent.

[0313] 4) Physical Model Construction

[0314] To more accurately describe the wear evolution of cutting tools during machining, this invention introduces an empirical physical model of tool wear with adjustable parameters.

[0315] f(t) = a·ln(b·t+1) + c·t 3 (20)

[0316] In the formula: a, b, and c are adjustable empirical parameters that can be obtained by fitting tool wear values.

[0317] Step 3: Establishing a feature-tool wear mapping mechanism

[0318] In this step, the mapping relationship between tool wear values ​​and input features is established through two fully connected layers, which can be specifically described as follows:

[0319]

[0320] In the formula, The value represents the tool wear monitored by the model; l(·) represents the Linear activation function; w F1 and w F2 H represents the trainable parameters of the two fully connected layers respectively. (L) This represents the output of a dynamic gated graph convolutional network with a total of L layers. In this embodiment, L is 2.

[0321] In this embodiment, the root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²) are used. 2 The model's performance was evaluated using RMSE, MAE, and R... 2 The formulas are as follows:

[0322]

[0323] In the formula: C represents the total number of passes throughout the tool's lifespan, and y and These represent the actual tool wear value and the tool wear value estimated by the model, respectively.

[0324] The tool wear monitoring method based on a physically guided dynamic gated graph convolutional network described above achieved excellent RMSE, MAE, and RMS performance on three test subsets. 2 As shown in Table 4:

[0325] Table 4

[0326]

[0327] To more intuitively reflect the deviation between the predicted tool wear and the actual wear, Figure 8 The figure shows the wear monitoring fitting curves of the method of the present invention on three test subsets. In the figure, the dashed line represents the actual tool wear value obtained by microscopic image measurement, and the solid line represents the wear result estimated by the tool wear monitoring method based on physical guided dynamic gated graph convolutional network proposed in this invention.

[0328] Combination Figure 8As can be observed from the results in Table 4, the tool wear monitoring model constructed in this invention achieves high prediction accuracy on multiple test subsets. This method can detect the degree of tool wear in a timely and accurate manner, providing a basis for users to make reasonable replacement decisions before tool failure, effectively avoiding the decline in processing efficiency and deterioration of workpiece quality caused by excessive wear, thereby reducing economic losses caused by equipment malfunctions or defective products.

Claims

1. A method for monitoring tool wear based on a physically guided dynamic gating graph convolution framework, characterized in that the steps include... include: Step 1: During the tool operation, the cutting force signal, bending moment signal and vibration signal of the tool and workpiece are collected to form a multi-source sensor signal sequence; Step 2: Perform data preprocessing on the acquired multi-source sensor signal sequences to obtain the original feature set; Step 3: Calculate the Pearson correlation coefficient between each feature in the original feature set and tool wear; Step 4: Based on the Pearson correlation coefficient, select the top-k features to construct a sensitive feature set integrating force, vibration, and bending moment; Step 5: Use a dynamic gated graph convolutional network to extract tool wear information from the sensitive feature set and establish long-term dependencies between features; Step 6: Decode the output of the features after establishing long-term dependency, establish the mapping relationship between tool wear value and input features, and output tool wear monitoring value.

2. The tool wear monitoring method based on a physically guided dynamic gating graph convolutional network according to claim 1, characterized in that: The cutting force and bending moment signals of the tool are acquired by a triaxial force sensor and a bending moment sensor integrated on the tool. The vibration signal of the workpiece is collected by a vibration sensor attached to the workpiece.

3. The tool wear monitoring method based on a physically guided dynamic gating graph convolutional network according to claim 1, characterized in that, Step 2, the data preprocessing steps for the acquired multi-source sensor signal sequences include: 2.1) The collected sensor signals are truncated and discarded; 2.2) Extract time-domain, frequency-domain, and time-frequency-domain features from the truncated and discarded sensor signals; 2.3) Normalize the extracted time-domain, frequency-domain, and time-frequency-domain features to obtain the original feature set; The normalization process is as follows: In the formula: This represents the i-th feature of the s-th sensor. and These are the corresponding minimum and maximum values. This represents the normalized features.

4. The tool wear monitoring method based on a physically guided dynamic gated graph convolutional network according to claim 3, characterized in that, Minimum value and maximum value It comes from the training sample set.

5. The tool wear monitoring method based on a physically guided dynamic gating graph convolutional network according to claim 1, characterized in that, The Pearson correlation coefficient r between each feature and tool wear is shown below: In the formula: |·| represents the absolute value operation, which is used to eliminate the influence of the positive or negative correlation on the analysis. Let y represent the j-th feature point of the ith feature of the s-th sensor. j This represents the tool wear value corresponding to the j-th cut. and These represent the average value of the i-th feature of the s-th sensor and the average value of tool wear, respectively, and N represents the total number of cutting operations.

6. The tool wear monitoring method based on a physically guided dynamic gated graph convolutional network according to claim 1, characterized in that, Step 5, which involves using a dynamic gated graph convolutional network to extract tool wear information from the sensitive feature set, includes: 5.1) Construct a dynamic trajectory matrix; where the trajectory matrix corresponds to the l-th gated convolution layer. As shown below: In the formula: δ(·) represents the softmax activation function; A (l) This represents the trajectory matrix before normalization; 5.2) Introduce an adaptive threshold strategy to establish connections or disconnections between features, i.e.: AM (l) JReLU(S (l) -τ (l) ) (5) In the formula: S (l) Let represent the cosine similarity matrix of the l-th layer. This represents the standardized input of the l-th layer. for The transpose matrix; ReLU(·) denotes the ReLU activation function, τ (l) The adaptive threshold for the l-th layer. μ(S) (l) ) and α(S (l) Let be the mean and deviation of the similarity matrices. When l is 1, For the selection of sensitive feature set x (n) The standardized processing results. Let be the cosine similarity parameter corresponding to the j-th neuron in the l-th layer. 5.3) Utilize gated graph convolutional networks to extract tool wear information, thereby establishing long-term dependencies; The gated graph convolutional network is shown below: In the formula: Gr i (l) and These are the update gate, reset gate, and candidate state output for the i-th time step of the l-th gated graph convolutional layer, respectively. and These are the weight matrices for the update gate, reset gate, and candidate states of the l-th gated graph convolutional layer, respectively. and This is the corresponding bias. and This represents the current time step input and the previous hidden layer state vector of the l-th gated graph convolutional layer; This represents the hidden layer state vector at the i-th time step of the l-th gated graph convolutional layer; This represents the weights of the update gate, reset gate, and candidate state; This represents the candidate state vector of the hidden layer.

7. The tool wear monitoring method based on a physically guided dynamic gated graph convolutional network according to claim 6, characterized in that, When the gated graph convolutional network is a multi-layer dynamic gated graph convolutional network, the input time series x (n) The hidden layer state vectors are processed into a graph structure, i.e.: In the formula: DGGCN(·) represents a dynamically gated graph convolutional layer; when l is 1, H (1) That is, the sensitive feature set x to be selected. (n) . H represents the hidden state vector at the i-th time step of the (l+1)-th gated graph convolutional layer; (l+1) This represents the current time step input of the (l+1)th gated graph convolutional layer.

8. The tool wear monitoring method based on a physically guided dynamic gated graph convolutional network according to claim 6, characterized in that, The gated graph convolutional network is trained on a historical dataset; the historical dataset includes a set of historical sensitive features integrating force, vibration and bending moment, as well as corresponding tool wear information. The loss function for gated graph convolutional networks during training is shown below: Loss=λ1·Loss d +λ2·(Loss p +Loss k ) (15) In the formula: λ1 and λ2 represent the data-driven loss, respectively. d Weights of physical guidance loss; Loss p Loss k This represents the loss of wear monotonicity and the loss of physical consistency.

9. The tool wear monitoring method based on a physically guided dynamic gated graph convolutional network according to claim 8, characterized in that, Data-driven loss d As shown below: In the formula: ||·||2 represents the L2 norm, y and These are the actual tool wear value and the tool wear value estimated by the model, respectively. The soft constraint on wear monotonicity is shown below: In the formula: Indicates tool wear rate; Loss k This indicates the wear monotonicity loss; The physical consistency constraints are as follows: f(t)=a·ln(b·t+1)+c·t 3 (19) In the formula: f represents the physical model of tool wear; a, b, and c are adjustable empirical parameters; t is time. p This represents the loss of physical consistency.

10. The tool wear monitoring method based on a physically guided dynamic gated graph convolutional network according to claim 1, characterized in that: In step 6, the mapping relationship between tool wear values ​​and input features is constructed through two fully connected layers; The mapping relationship between tool wear values ​​and input features is shown below: In the formula, The value represents the tool wear monitored by the model; l(·) represents the Linear activation function; w F1 and w F2 H represents the trainable parameters of the two fully connected layers respectively. (L) This represents the output of a dynamic gated graph convolutional network with a total of L layers.

Images