A multi-step prediction method for tunnel boring machine cutter wear

Abstract

The application discloses a multi-step prediction method for tunnel boring machine cutter wear, which comprises the following steps: constructing a complementary topological graph by combining an equal-scale and variable-scale double division strategy with a multi-feature voting mechanism; synchronously completing wear state classification and wear amount regression double-task learning by relying on a double-branch graph attention network and a cross-attention gate network; and eliminating gradient conflicts by using a gradient projection coordination optimization algorithm to improve small sample training stability, fusing a three-stage segmented wear mechanism model of a cutter, and constructing a rolling incremental multi-step prediction strategy to inhibit time series error accumulation, so that long-time series high-precision prediction of cutter wear can be realized, tunnel boring machine cutter condition-based replacement and predictive maintenance can be supported, and machining operation and maintenance costs can be effectively reduced.

Description

Technical Field

[0001] This invention relates to the field of tool wear condition monitoring technology, and in particular to a multi-step prediction method for cutter wear in tunnel boring machines. Background Technology

[0002] The cutter head is a core actuator of a tunnel boring machine (TBM). During the tunneling process, it is continuously subjected to mechanical loads and cutting heat, making it highly susceptible to wear. Cutter head wear directly affects tunneling speed and efficiency, and can even cause unplanned downtime of the TBM. Therefore, achieving accurate long-term prediction of cutter head wear is a core prerequisite for predictive maintenance of TBMs.

[0003] Existing tunneling machine wear prediction technologies have several shortcomings: traditional solutions directly feed Euclidean sensor data into the model without exploring the topological relationships between sensors, resulting in insufficient feature utilization; a single network simultaneously undertakes two different tasks of wear classification and wear regression, and joint training is prone to gradient cancellation and amplitude imbalance, making it difficult for the model to converge in small sample scenarios; moreover, most only support single-step short-term prediction, lacking constraints on cutting wear mechanisms, and long-term extrapolation errors continue to accumulate. At the same time, the single signal segmentation method cannot take into account both stable and time-varying cutting conditions, making it difficult to meet the needs of tool life cycle monitoring under limited industrial samples. Summary of the Invention

[0004] Technical Objective: This invention addresses the shortcomings of existing technologies, such as poor generalization with small samples, gradient conflicts in multiple tasks, accumulation of long-term prediction errors, and insufficient mining of multi-source sensor topological features. It provides a multi-step prediction method for cutter wear in tunnel boring machines (TBMs). By combining dual-scale subsample partitioning with multi-feature voting graph construction, parallel multi-head graph attention network, bidirectional cross-attention gating, PCO gradient collaborative optimization, and three-stage segmented wear rolling increment prediction, it simultaneously completes wear state classification and quantitative regression of wear amount, suppressing long-term prediction errors, improving the accuracy of multi-step cutter wear prediction with limited samples, and enabling condition-based cutter replacement and intelligent operation and maintenance for TBMs.

[0005] Technical solution: To achieve the above technical objectives, the present invention adopts the following technical solution.

[0006] A multi-step prediction method for cutter wear in tunnel boring machines includes the following steps: Step S1: Divide the acquired multi-channel sensor signals into sub-samples; Step S2: Extract the temporal features of sensor signal sub-samples to achieve data dimensionality reduction, and use correlation analysis to screen temporal features that are highly correlated with tool wear; Step S3: Transform the Euclidean structure data into a topological graph using a multi-feature voting method; Step S4: Extract two sets of global expert features from the input topology graph using a parallel multi-head graph attention network model. , ; Step S5: Utilize a dual cross-attention gating unit (CAG) comprising CAG1 and CAG2, where CAG1 is... For query, As a key value, CAG2 uses For query, Using key-value pairs, the features are concatenated after bidirectional interaction to generate classification-specific features. Returning to unique characteristics ; Step S6: In the wear state classification task, the KAN network backend is paired with... Activate and output the reference time. The corresponding tool wear stage; in the wear regression task, the KAN network is used to directly output the continuous wear values ​​at the reference time. ; Step S7: Gradient Co-optimization training module, with built-in gradient projection-coordinated joint optimization algorithm PCO, receives the task gradients corresponding to classification loss and regression loss, first projects to eliminate gradient direction conflicts, then adaptively scales and balances gradient magnitudes, and outputs all trainable parameters of the fused coordinated gradient update network. Step S8: The segmented wear mechanism reasoning module is connected to the multi-task feature fusion network, with a built-in three-stage tool wear model. Based on the wear state at the reference time, it matches the corresponding function and calculates the time-series wear increment. The rolling increment multi-step prediction output module combines the reference wear data and the wear increment to complete the rolling window K-step deduction and output the multi-step tool wear prediction results.

[0007] Further, step S1 includes the following steps: Step S11: Normalize the multi-channel sensor signals; Step S12: Select the sensor signal of the first channel. Set the number of subsamples to N, and the baseline length of the subsamples to N. ; Step S13: The nth subsample of the first sensor signal at the same scale The nth variable-scale subsample of the first sensor signal is ; Step S14: Repeat steps S12 and S13 to obtain a set of subsamples of the same scale for the M channel sensor signals. This yields a set of identically scaled subsamples of M channel sensor signals. .

[0008] Further, step S2 includes the following steps: Step S21: Extract the time-domain statistical features of the sensor signal sub-samples. The calculation formula is as follows: in, These represent the sets of statistical eigenvectors at the same scale and at varying scales, respectively. Indicates statistical feature extraction; Step S22: Calculate the correlation between time-domain statistical characteristics and tool wear values; Step S23: Filter for correlation values ​​greater than the threshold. The statistical characteristics are used to obtain new sets of statistical feature vectors of the same scale and variable scale. , ; Furthermore, step S3 includes the following steps: S31. Calculate the absolute difference between any two nodes in each feature dimension, and select the nodes with the smallest single-dimensional difference. Nodes are granted voting rights; S32. Accumulate the vote counts for all feature dimensions to obtain the total vote value among nodes; S33. Provide two adjacency determination rules: select the one with the highest total number of votes. Each node can be designated as a nearest neighbor, or a voting threshold can be set. If the number of votes exceeds a threshold, an edge is established. S34. Symmetricize the directed adjacency matrix to obtain the undirected graph adjacency matrix; S35, respectively for , Perform multi-feature voting to construct a graph, generating a graph of the same scale with complementary topological structures. Variable scale diagram ; Furthermore, in step S4, the parallel multi-head graph attention network model contains two parallel feature extraction branches. Each parallel branch is composed of multiple layers of multi-head graph attention networks stacked in series. Each layer of multi-head graph attention network synchronously mines the feature associations and topological dependencies of nodes in the graph through multiple sets of independent attention heads. The two parallel branches output the deep feature representations of the corresponding topological graphs, which are then pooled to obtain two sets of global expert features with complementary information. , ; Furthermore, step S5 includes the following steps: S51, Adaptive weight allocation using dual cross-attention gating units. , ; S52. After two-way interaction, specific features for classification are generated by splicing. Returning to unique characteristics Further, step S6 includes the following steps: S61. Input the classification-specific features output by the dual-cross attention gating unit into the KAN classification sub-network, configure the Softmax activation function at the back end of the KAN network, normalize the probability mapping of the output results, and output the baseline time. The probability distribution of the tool in three stages of wear: initial wear, stable wear, and severe wear is used, and the category corresponding to the maximum probability is selected as the tool wear stage at the reference time. S62. Input the regression-specific features output by the cross-attention gating unit into the Bi-LSTM time-series extraction module. The module uses forward and backward bidirectional LSTM to mine the temporal dependencies of wear features. After fusing the bidirectional hidden states to obtain a complete temporal representation, it is fed into the KAN regression sub-network. The KAN network fits the nonlinear mapping relationship of wear based on learnable B-spline basis functions, without additional activation layer constraints, and directly outputs the baseline time. Corresponding continuous flank wear value .

[0009] Furthermore, step S7 includes the following steps: S71. Gradient Conflict Detection and Orthogonal Projection Conflict Resolution: Input Wear Regression Gradient Wear classification gradient Calculate the gradient dot product if If a gradient conflict is detected, the two types of gradients are projected onto the direction orthogonal to the other gradient to obtain the projected gradients: If there is no conflict, then , Used for numerical stabilization; S72, Gradient Magnitude Calculation and Adaptive Cross-Scaling: Solve for the L2 norm magnitudes of the projected wear regression gradient and wear classification gradient respectively. Construct cross-scaling coefficients based on the magnitudes, and inversely scale the two types of projected gradients, weakening gradients with larger amplitudes and enhancing gradients with smaller amplitudes, thus balancing the difference in gradient magnitude between the two tasks. Used for numerical stabilization; S73, Weighted Coordination Correction Gradient: S74, Gradient Fusion and Network Parameter Update: The fused total gradient is obtained by adding the two coordinated gradients. By utilizing backpropagation of fused gradients, all trainable parameters of the network are updated uniformly, completing a multi-task collaborative optimization iteration.

[0010] Further, step S8 includes the following steps: S81. Reading reference parameters and matching wear segmentation function: Receiving reference time. Wear condition Compared with the reference wear amount ,in accordance with Matching the corresponding branch of the three-stage segmented wear model, the segmented model expression is as follows: S82. Calculation baseline and theoretical wear amount at the time to be predicted: The wear state is fixed within a single-window prediction period. Substitute the reference time into each. Predictable time The theoretical wear amount is obtained. , ; S83. Calculate the wear increment using the theoretical wear difference. ; S84. Scrolling window multi-step wear prediction: Obtaining the predicted value by superimposing the baseline wear value with the wear increment. It completes a single-window K-step time series simulation, and the window slides forward to output continuous wear prediction results in a loop.

[0011] Compared with the prior art, the present invention has the following beneficial effects: 1. Significantly improved prediction accuracy and feature mining capabilities with small sample sizes. This invention designs a dual-path subsample partitioning strategy with equal and variable scales, and constructs two sets of topological complementary graphs with a multi-feature voting mechanism. It extracts heterogeneous sensor complementary features by relying on a parallel multi-head graph attention network, and realizes the directional interaction of classification and regression task features through bidirectional cross-attention gating. This fully explores the intrinsic correlation information between the two types of tasks, overcomes the problem of insufficient feature utilization caused by traditional single-scale modeling and direct input of Euclidean data, significantly improves the model's generalization ability in industrial labeled sample-scarce scenarios, and effectively improves the accuracy of tool wear state recognition and quantitative prediction of wear amount.

[0012] 2. Multi-task joint training exhibits strong stability and faster convergence speed. We innovatively propose a gradient projection-coordinated joint optimization (PCO) algorithm. First, we eliminate the problem of gradient inverse cancellation between wear classification and wear regression tasks by using gradient orthogonal projection. Then, we adaptively scale and balance the gradient magnitudes of the two classes based on the gradient norm 2, avoiding the single task dominating parameter updates. This solves the problem that the traditional fixed-weight loss weighting method cannot adapt to gradient fluctuations in small samples, greatly alleviates model training oscillations, accelerates network convergence, and improves the stability of multi-task collaborative learning.

[0013] 3. By incorporating wear mechanism constraints, the accumulation of long-term prediction errors is effectively suppressed. By deeply coupling the data-driven multi-task network with the three-stage segmented wear mechanism of the tool, the reference wear stage and reference wear amount output by the model in real time are used as the initial boundary for time series inference. The time series wear increment is calculated by matching the corresponding mechanism function according to the wear rate change. A rolling window incremental multi-step prediction mechanism is adopted to make up for the shortcoming of the pure data-driven model that the error continues to amplify as the prediction step size increases, so as to achieve long-term high-precision continuous prediction of the tool's entire life cycle. Attached Figure Description

[0014] Figure 1 This is a diagram of the overall architecture for multi-step prediction of cutter wear in tunnel boring machines according to the present invention. The components in the attached diagram are labeled as follows: 1. Numerical control airport; 2. Sensor data acquisition; 3. Variable-scale subsample partitioning module; 4. Same-scale subsample partitioning module; 5. Variable-scale graph; 6. Same-scale graph; 7. Parallel multi-head graph attention network model; 8. Double cross-attention gating unit; 9. Regressor; 10. Classifier; 11. Baseline wear value; 12. Baseline wear state; 13. PCO gradient collaborative optimization module; 14. Segmented wear mechanism reasoning module; 15. Output multi-step wear value. Detailed Implementation

[0015] This embodiment provides a multi-step prediction method for hob wear in tunnel boring machines. The processing equipment is a vertical three-axis CNC milling machine, the cutting tool is a ball-end carbide end mill, and the workpiece is a nickel-based high-temperature alloy. Sensors collect six channels of timing signals, including X / Y / Z axis cutting force and spindle vibration, at a sampling frequency of 50kHz. The complete execution steps are as follows: S1, Multi-channel signal dual-scale subsample partitioning S11. Normalize the original vibration and cutting force signals of the 6 channels to 0-1. S12, Total length of single-channel signal Set the total number of subsamples Reference subsample length ; S13. Divide the samples into equal-scale subsamples. Variable-scale subsamples ; S14. Traverse all 6 channels to generate complete equal-scale and variable-scale subsample sets. , .

[0016] S2. Temporal Feature Extraction and Correlation Screening S21. Extract 32-dimensional time-domain statistical features such as mean, root mean square, kurtosis, and peak value to obtain the original feature set. , ; S22. Calculate the wear of each feature and the actual tool flank face. Pearson correlation coefficient; S23. Set the correlation threshold σ=0.9, retain 13 highly correlated features, and obtain the filtered feature set. , .

[0017] S3. Constructing a complementary topology graph using multi-feature voting. S31. Minimizes differences in selecting a single feature dimension. Each node votes; S32. Accumulate all 13-dimensional feature votes to obtain the total number of votes for each node; S33. Set vote threshold If the number of votes exceeds the threshold, a node connection is established; S34. Symmetricalizing the adjacency matrix yields an undirected graph; S35, respectively based on , Generate isoscale maps Variable scale diagram .

[0018] S4. Parallel multi-head graph attention network extracts global features and constructs two completely parallel multi-head GAT branches, which are then input to... , Each branch is stacked with 3 layers of multi-head graph attention network, and graph pooling is used to filter key nodes after every two layers; after pooling, the aggregated features yield two sets of complementary global features. , .

[0019] S5, bidirectional cross-attention gating feature fusion construction , Two sets of attention units are concatenated after bidirectional weighted interaction to generate classification-specific features. Returning to unique characteristics .

[0020] S6, Dual-task KAN network outputs reference wear information S61, The data is fed into the KAN classification network, and the end... Output the three types of wear probabilities, and take the maximum value to obtain the reference time. Wear condition (Stable wear); S62, The input is processed by a Bi-LSTM to extract the time-series dependencies, which are then fed into a KAN regression network to output the baseline wear value. .

[0021] S7, PCO gradient co-optimization iterative training to calculate classification cross-entropy gradient Regression MSE gradient If there is a conflict in the gradient dot product determination, perform orthogonal projection to eliminate the canceling components; construct scaling coefficients to balance the magnitude based on the gradient 2 norm; weightedly fuse the two gradients, backpropagate to update all network parameters, and iterate for 100 rounds until the loss converges.

[0022] S8, Multi-step prediction of rolling increment in segmented wear mechanism S81, according to Matching stable wear piecewise function ; S82. Calculate the theoretical wear at the reference time and the time to be predicted, respectively; S83, Calculate the wear increment for each time series. ; S84. Set the single-window prediction step number K=15 and the single cutting step length τ=4s; the reference wear amount is superimposed with the increment to obtain the continuous wear prediction value for the next 56 seconds; the window slides forward once every 60 seconds to continuously output the wear prediction results; the operation and maintenance terminal pushes tool replacement warnings in advance based on the predicted wear threshold.

Claims

1. A multi-step prediction method for cutter wear in tunnel boring machines, characterized in that, Includes the following steps: Step S1: Divide the acquired multi-channel sensor signals into sub-samples; Step S2: Extract the temporal features of sensor signal sub-samples to achieve data dimensionality reduction, and use correlation analysis to screen temporal features that are highly correlated with tool wear; Step S3: Transform the Euclidean structure data into a topological graph using a multi-feature voting method; Step S4: Extract two sets of global expert features from the input topology graph using a parallel multi-head graph attention network model. , ; Step S5: Utilize a dual cross-attention gating unit (CAG) comprising CAG1 and CAG2, where CAG1 is... For query, As a key value, CAG2 uses For query, Using key-value pairs, the features are concatenated after bidirectional interaction to generate classification-specific features. Returning to unique characteristics ; Step S6: In the wear state classification task, the KAN network backend is paired with... Activate and output the reference time. The corresponding tool wear stage; in the wear regression task, the KAN network is used to directly output the continuous wear values ​​at the reference time. ; Step S7: Gradient Co-optimization training module, with built-in gradient projection-coordinated joint optimization algorithm PCO, receives the task gradients corresponding to classification loss and regression loss, first projects to eliminate gradient direction conflicts, then adaptively scales and balances gradient magnitudes, and outputs all trainable parameters of the fused coordinated gradient update network. Step S8: The segmented wear mechanism reasoning module is connected to the multi-task feature fusion network, with a built-in three-stage tool wear model. Based on the wear state at the reference time, it matches the corresponding function and calculates the time-series wear increment. The rolling increment multi-step prediction output module combines the reference wear data and the wear increment to complete the rolling window K-step deduction and output the multi-step tool wear prediction results.

2. The multi-step prediction method for cutter wear of tunnel boring machines according to claim 1, characterized in that: Step S1 includes the following steps: Step S11: Normalize the multi-channel sensor signals; Step S12: Select the sensor signal of the first channel. Set the number of subsamples to N, and the baseline length of the subsamples to N. ; Step S13: The nth subsample of the first sensor signal at the same scale The nth variable-scale subsample of the first sensor signal is ; Step S14: Repeat steps S12 and S13 to obtain a set of subsamples of the same scale for the M channel sensor signals. This yields a set of identically scaled subsamples of M channel sensor signals. .

3. The multi-step prediction method for cutter wear of tunnel boring machines according to claim 1, characterized in that, Step S2 includes the following steps: Step S21: Extract the time-domain statistical features of the sensor signal sub-samples. The calculation formula is as follows: in, These represent the sets of statistical eigenvectors at the same scale and at varying scales, respectively. Indicates statistical feature extraction; Step S22: Calculate the correlation between time-domain statistical characteristics and tool wear values; Step S23: Filter for correlation values ​​greater than the threshold. The statistical characteristics are used to obtain new sets of statistical feature vectors of the same scale and variable scale. , .

4. The multi-step prediction method for cutter wear of tunnel boring machines according to claim 1, characterized in that, Step S3 includes the following steps: S31. Calculate the absolute difference between any two nodes in each feature dimension, and select the nodes with the smallest single-dimensional difference. Nodes are granted voting rights; S32. Accumulate the vote counts for all feature dimensions to obtain the total vote value among nodes; S33. Provide two adjacency determination rules: select the one with the highest total number of votes. Each node can be designated as a nearest neighbor, or a voting threshold can be set. If the number of votes exceeds a threshold, an edge is established. S34. Symmetricize the directed adjacency matrix to obtain the undirected graph adjacency matrix; S35, respectively for , Perform multi-feature voting to construct a graph, generating a graph of the same scale with complementary topological structures. Variable scale diagram .

5. The multi-step prediction method for cutter wear of tunnel boring machines according to claim 1, characterized in that, In step S4, the parallel multi-head graph attention network model contains two parallel feature extraction branches. Each parallel branch is composed of multiple layers of multi-head graph attention networks stacked in series. Each layer of multi-head graph attention network synchronously mines the feature associations and topological dependencies of nodes in the graph through multiple independent attention heads. The two parallel branches output the deep feature representations of the corresponding topological graphs, which are then pooled to obtain two sets of global expert features with complementary information. , .

6. The multi-step prediction method for cutter wear of tunnel boring machines according to claim 1, characterized in that, Step S5 includes the following steps: S51, Adaptive weight allocation using dual cross-attention gating units. , ; S52. After two-way interaction, specific features for classification are generated by splicing. Returning to unique characteristics .

7. A multi-step prediction method for cutter wear of tunnel boring machines according to claim 1, characterized in that, Step S6 includes the following steps: S61. Input the classification-specific features output by the dual-cross attention gating unit into the KAN classification sub-network, configure the Softmax activation function at the back end of the KAN network, normalize the probability mapping of the output results, and output the baseline time. The probability distribution of the tool in three stages of wear: initial wear, stable wear, and severe wear is used, and the category corresponding to the maximum probability is selected as the tool wear stage at the reference time. S62. The regression-specific features output by the dual-cross attention gating unit are input into the Bi-LSTM time-series extraction module. The forward and backward bidirectional LSTMs are used to mine the temporal dependencies of wear features. After fusing the bidirectional hidden states to obtain a complete temporal representation, it is fed into the KAN regression sub-network. The KAN network fits the nonlinear mapping relationship of wear based on learnable B-spline basis functions, without additional activation layer constraints, and directly outputs the baseline time. Corresponding continuous flank wear value .

8. A multi-step prediction method for cutter wear of tunnel boring machines according to claim 1, characterized in that, Step S7 includes the following steps: S71. Gradient Conflict Detection and Orthogonal Projection Conflict Resolution: Input Wear Regression Gradient Wear classification gradient Calculate the gradient dot product if If a gradient conflict is determined, the two types of gradients are projected onto the direction orthogonal to the other gradient to obtain the projected gradients: ； If there is no conflict, then , Used for numerical stability; S72, Gradient Magnitude Calculation and Adaptive Cross-Scaling: Solve for the L2 norm magnitudes of the projected wear regression gradient and wear classification gradient respectively. Construct cross-scaling coefficients based on the magnitudes and inversely scale the two types of projected gradients, weakening gradients with larger amplitudes and enhancing gradients with smaller amplitudes to balance the difference in gradient magnitudes between the two tasks. ； Used for numerical stability; S73, Weighted Coordination Correction Gradient: ； S74, Gradient Fusion and Network Parameter Update: The fused total gradient is obtained by adding the two coordinated gradients. ； By utilizing backpropagation of fused gradients, all trainable parameters of the network are updated uniformly, completing a multi-task collaborative optimization iteration.

9. A multi-step prediction method for cutter wear of tunnel boring machines according to claim 1, characterized in that, Step S8 includes the following steps: S81. Reading reference parameters and matching wear segmentation function: Receiving reference time. Wear condition , and the reference wear amount ,in accordance with Matching the corresponding branch of the three-stage segmented wear model, the segmented model expression is as follows: ； S82. Calculation baseline and theoretical wear amount at the time to be predicted: The wear state is fixed within a single-window prediction period. Substitute the reference time into each. Predictable time The theoretical wear amount is obtained. , ; S83. Calculate the wear increment using the theoretical wear difference. ; S84. Scrolling window multi-step wear prediction: Obtaining the predicted value by superimposing the baseline wear value with the wear increment. It completes a single-window K-step time series simulation, and the window slides forward to output continuous wear prediction results in a loop.

Images