KAN-Transform data mechanism dual-drive tool wear prediction method based on adaptive fusion

Abstract

The invention discloses a KAN-Transform data mechanism dual-drive tool wear prediction method based on adaptive fusion. The method comprises the following steps: respectively constructing a network architecture module based on a KAN and a network architecture module based on a Transform; performing standardization processing on the obtained cutter signal features, performing feature screening in combination with Pearson correlation, and dividing the signal features into a training set and a test set; respectively inputting the training set into a KAN network architecture module and a Transform network architecture module to obtain corresponding output characteristics; through a self-adaptive fusion layer, carrying out self-adaptive fusion on the output characteristics of the KAN network and the output characteristics of the Transform network; and constructing a loss function combining a tool wear mechanism and data loss, and predicting a tool wear value through the test set. According to the method, the prediction precision is improved, and meanwhile, the physical consistency and generalization ability are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of deep learning tool wear prediction, and particularly relates to a KAN-Transformer data mechanism dual-driven tool wear prediction method based on adaptive fusion. BACKGROUND

[0002] Accurate tool wear prediction is an important part of intelligent manufacturing systems. In numerical control milling operations, real-time tool condition monitoring plays a crucial role in ensuring product quality and reducing unplanned downtime. Early tool wear assessment mainly relies on the experience of operators. This often leads to premature replacement, wasting tool life, or late replacement, resulting in failure and damage to workpieces.

[0003] With the increasing popularity of sensor data and the success of deep learning, various data-driven models, such as convolutional neural networks, recurrent neural networks, and Transformer-based network architectures, have shown good performance in tool wear prediction tasks. These methods are good at learning complex mappings from sensor signals to wear states, but have the disadvantages of poor generalization ability, low interpretability, and high data dependency under domain conversion. In addition, current methods usually treat the system as a complete black box, ignoring the underlying physical degradation mechanisms. Existing methods cannot simultaneously model complex nonlinear interactions and time series while maintaining physical consistency.

[0004] Therefore, how to accurately predict tool wear is a technical problem that needs to be solved in the field. SUMMARY

[0005] To solve the above technical problems, the present application provides the following technical solutions: The present application provides a KAN-Transformer data mechanism dual-driven tool wear prediction method based on adaptive fusion, comprising: respectively constructing a KAN network architecture module and a Transformer network architecture module; standardizing the obtained tool signal features and performing feature selection based on Pearson correlation, and then dividing the signal features into a training set and a test set; inputting the training set into the KAN network architecture module and the Transformer network architecture module to obtain the output features of the KAN network and the output features of the Transformer network, respectively; through an adaptive fusion layer, adaptively fusing the output features of the KAN network and the output features of the Transformer network; A loss function combining tool wear mechanism and data loss is constructed, and tool wear value is predicted using the test set.

[0006] In one possible implementation, the KAN network architecture module includes a linear layer whose output is connected to the input of a nonlinear RBF layer, whose output is connected to the input of a ReLU activation function layer, and whose output is connected to the input of an adaptive fusion layer.

[0007] In one possible implementation, the Transformer network architecture module includes a feature encoding layer and a temporal encoding layer. The outputs of both the feature encoding layer and the temporal encoding layer are connected to the input of the fused feature layer. The output of the fused feature layer is connected to the input of the multi-head attention layer. The output of the multi-head attention layer is connected to the input of the layer normalization layer. The output of the layer normalization layer is connected to the input of the feedforward network. The output of the feedforward network is connected to the input of the adaptive fusion layer.

[0008] In one possible implementation, the standardization of the acquired tool signal features and the feature filtering based on Pearson correlation include: The acquired tool signal features are normalized, and the actual tool wear values ​​are also normalized. Map the normalized data to a preset range proportionally; The correlation coefficient between the acquired tool signal features and the output wear value is calculated using the Pearson correlation coefficient, and signal features with a correlation coefficient greater than a preset value are selected.

[0009] In one possible implementation, the training set is input into the KAN network architecture module to obtain the output features of the KAN network, including: The selected features are input into the KAN network architecture module for linear and nonlinear RBF spline transformations, calculated using the following formulas: Where W is the weight of the linear layer, b is the bias of the linear layer, x is the input feature, and c i k are the learnable coefficients. i This represents the position of the i-th node, and exp represents the exponential function. The calculated linear transformation and nonlinear RBF spline transformation are fused to obtain the output features of the KAN network. The calculation formula is as follows: in, For hyperparameters, The value of the linear transformation. represents the value of the nonlinear RBF spline transform.

[0010] In one possible implementation, the training set is input into the Transformer network architecture module to obtain the output features of the Transformer network, including: The selected features are input into the Transformer network architecture module. The input features are mapped into feature vectors through the feature encoding layer, and the time input is mapped into time feature vectors through the time encoding layer. The feature vector and the time feature vector are stacked according to their dimensions to construct a sequence, thus obtaining the fused feature sequence. The fused feature sequence is linearly mapped to obtain the query, key, and value, respectively, using the following formulas: , , in, For the input fused feature sequence, , , The weight matrix is ​​a learnable weight matrix; The multi-head attention mechanism uses a multi-head attention layer to embed the input and obtain the query, key, and value through linear transformations. The calculation formula is as follows: in, , Indicates the first Scaling attention score matrix for each attention head. This represents the transpose of the key matrix of the i-th head. , This represents the total embedding dimension of the model. The number of heads representing multi-head attention. This represents the dimension of each attention head. Let i be the query matrix for the i-th head. For the first The value matrix of each head; The output features of the Transformer network are obtained after layer normalization and a feedforward network.

[0011] In one possible implementation, the formula for calculating the output features of the Transformer network after layer normalization and feedforward network is as follows: Where LN represents layer normalization and FFN represents feedforward network.

[0012] In one possible implementation, the formula for adaptively fusing the output features of the KAN network and the output features of the Transformer network is as follows: in, and This represents the learnable parameters.

[0013] In one possible implementation, the formula for calculating the loss function that combines tool wear mechanism and data loss is as follows: in, , in, N This represents the total number of data collections. Indicates the first i The actual value of tool wear during the first data acquisition. Indicates the first i Wear values ​​predicted by the model in the second data collection. Represents the normalized derivative term. Let be the objective function. This represents the i-th sampling time point. The time-varying characteristics of wear rate, the exponential term Reflecting the nonlinear characteristics of the wear process, e, f, g 1 , g 2 These are the learnable physical parameters. This is a hyperparameter.

[0014] Compared with the prior art, the beneficial effects of this application are as follows: This application leverages the combination of KAN and Transformer to enhance nonlinear feature extraction and temporal modeling capabilities. By introducing wear rate formula constraints, it avoids prediction results that violate physical laws. At the same time, it integrates physical knowledge to ensure the model maintains stable performance during extrapolation of operating conditions. Cosine learning rate, dynamic physical weights, and early stopping mechanism ensure smooth convergence during the training process.

[0015] This application is applicable to tool condition monitoring in actual CNC machining processes, which is beneficial to improving the level of intelligence in the manufacturing process. Attached Figure Description

[0016] Figure 1 A flowchart illustrating a dual-drive tool wear prediction method based on adaptive fusion KAN-Transformer data mechanism provided in this application embodiment; Figure 2 This is a schematic diagram of the overall network structure provided in the embodiments of this application; Figure 3 The following are diagrams showing the prediction results of different tool wear in the embodiments of this application; Figure 4 The image shows the tool verification prediction results provided in the embodiments of this application. Detailed Implementation

[0017] The present solution will now be described in conjunction with the accompanying drawings and specific embodiments.

[0018] Figure 1 A flowchart illustrating a dual-drive tool wear prediction method based on adaptive fusion of KAN-Transformer data mechanism, provided in this application embodiment, is shown below. Figure 1 This application provides a method for predicting dual-drive tool wear based on adaptive fusion of KAN-Transformer data mechanism, including: S101, respectively construct a module based on the KAN network architecture and a module based on the Transformer network architecture.

[0019] See Figure 2 In this embodiment, the KAN network architecture module integrates a neural network based on linear transformation and radial basis function (RBF) spline approximation. It includes a linear layer, the output of which is connected to the input of a nonlinear RBF layer. The output of the nonlinear RBF layer is connected to the input of a ReLU activation function layer, and the output of the ReLU activation function layer is connected to the input of an adaptive fusion layer. The Transformer network architecture module includes a feature encoding layer, a temporal encoding layer, a fused feature layer, a multi-head attention layer, a layer normalization layer, and a feedforward network. The outputs of the feature encoding layer and the temporal encoding layer are both connected to the input of the fused feature layer. The output of the fused feature layer is connected to the input of the multi-head attention layer. The output of the multi-head attention layer is connected to the input of the layer normalization layer. The output of the layer normalization layer is connected to the input of the feedforward network, and the output of the feedforward network is connected to the input of the adaptive fusion layer. The KAN path uses an efficient RBF-based spline function approximation to extract the fine-grained nonlinear relationship between cutting parameters and wear behavior. In contrast, the Transformer path combines physical awareness with temporal encoding and a lightweight, efficient attention mechanism to simulate long-range temporal dependencies. This embodiment uses five learnable univariate cubic RBF-based spline functions. The number of multi-head attention mechanisms is set to four, and the dropout rate is set to 0.15.

[0020] S102, after standardizing the acquired tool signal features and performing feature filtering based on Pearson correlation, the signal features are divided into training and testing sets.

[0021] In this embodiment, the acquired tool signal features are normalized, and the actual tool wear values ​​are also normalized. The normalized data are mapped to a preset interval of [0.1] proportionally. The signal features are then divided, with 80% of the labeled data serving as the training set and 20% as the test set. The Pearson correlation coefficient is used to calculate the correlation coefficient between the acquired tool signal features and the output wear values, and signal features with a correlation coefficient greater than a preset value of 0.85 are selected.

[0022] S103, input the training set into the KAN network architecture module and the Transformer network architecture module respectively to obtain the output features of the corresponding KAN network and the output features of the Transformer network.

[0023] In this embodiment, the training set is input into the KAN network architecture module to obtain the output features of the KAN network, including: The selected features are input into the KAN network architecture module for linear and nonlinear RBF spline transformations, calculated using the following formulas: Where W is the weight of the linear layer, b is the bias of the linear layer, x is the input feature, and c i k are the learnable coefficients. i Let represent the position of the i-th node, and exp represent the exponential function.

[0024] The calculated linear transformation and nonlinear RBF spline transformation are fused to obtain the output features of the KAN network. The calculation formula is as follows: in, For hyperparameters, The value of the linear transformation. represents the value of the nonlinear RBF spline transform.

[0025] The training set is input into the Transformer network architecture module to obtain the output features of the Transformer network, including: The selected features are input into the Transformer network architecture module. The input features are mapped into feature vectors through the feature encoding layer, and the time input is mapped into time feature vectors through the time encoding layer. The feature vectors and time feature vectors are stacked according to their dimensions to construct a sequence, resulting in a fused feature sequence. The fused feature sequence is linearly mapped to obtain the query, key, and value, respectively, using the following formulas: , , in, For the input fused feature sequence, , , The weight matrix is ​​a learnable weight matrix; The multi-head attention mechanism uses a multi-head attention layer to embed the input and obtain the query, key, and value through linear transformations. The calculation formula is as follows: in, , Indicates the first Scaling attention score matrix for each attention head This represents the transpose of the key matrix of the i-th head. , This represents the total embedding dimension of the model. The number of heads representing multi-head attention. This represents the dimension of each attention head. Let i be the query matrix for the i-th head. For the first The value matrix of each head is concatenated with all attention head outputs, and then linearly mapped to output the result. , This represents matrix concatenation, where h is the number of attention heads. This is the output projection matrix.

[0026] Finally, after layer normalization and a feedforward network, the output features of the Transformer network are obtained, calculated using the following formula: Where LN represents layer normalization and FFN represents feedforward network.

[0027] S104 uses an adaptive fusion layer to adaptively fuse the output features of the KAN network and the Transformer network.

[0028] In this embodiment, an adaptive cross-fusion layer with learnable weights is introduced to dynamically integrate representations from two paths, enabling the model to adjust its dependence on spatial or temporal information based on the wear stage. The specific calculation formula for adaptive fusion is as follows: in, and This represents the learnable parameters.

[0029] S105: Construct a loss function that combines tool wear mechanism and data loss, and predict tool wear value using a test set.

[0030] In this embodiment, to ensure physical rationality, differentiable physical residuals with learnable coefficients are incorporated into the loss function, thereby enabling the soft execution of physical constraints without compromising flexibility. The formula for calculating the loss function that constructs the data loss component is as follows: in, N This represents the total number of data collections. Indicates the first i The actual value of tool wear during the first data acquisition. Indicates the first i The wear value predicted by the model in the second data collection.

[0031] The formula for calculating the loss function of the mechanism loss component is as follows: in, , in, Represents the normalized derivative term. Let be the objective function. This represents the i-th sampling time point. The time-varying characteristics of wear rate, the exponential term Reflecting the nonlinear characteristics of the wear process, e, f, g 1 , g 2 These are the learnable physical parameters.

[0032] Therefore, the total loss function can be expressed as: in, This is a hyperparameter used to control the weight balance between data error and physical residual.

[0033] During training, the Adam optimizer was used with an initial learning rate of 1e-3, a warm-up period of 100, and decaying cosine annealing for the learning rate. The batch size was set to 64. The best model from the training process was saved, and the final network outputs tool wear values ​​on the test set through linear layers. To ensure the reliability of the results, the average of three runs was taken. The best model from the training process was directly loaded, and tool wear prediction was performed on the partitioned test set. Figure 3 ,4 The comparison of prediction results of the proposed method on the test set is shown. From Figure 3 It is evident that the tool wear curve predicted by the proposed method closely matches the actual tool wear curve, and the overall distribution of the prediction error is also very small. Figure 4 This verifies the generalization ability of the proposed method. It can be seen that even when predicting on tools completely different from the training set, it still exhibits good prediction performance, demonstrating the excellent generalization ability of the proposed method. Experimental results show that this method improves both prediction accuracy and physical consistency and generalization ability. The proposed method provides a promising direction for integrating domain knowledge into neural architectures for tool wear prediction.

[0034] In this embodiment, "multiple" refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent the existence of A alone, the simultaneous existence of A and B, or the existence of B alone. A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0035] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0036] The above description is merely a specific embodiment of this application. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the protection scope of this application. The protection scope of this application should be determined by the protection scope of the claims.

Claims

1. A dual-drive tool wear prediction method based on adaptive fusion of KAN-Transformer data mechanism, characterized in that, include: Construct modules based on the KAN network architecture and modules based on the Transformer network architecture, respectively. The acquired tool signal features were standardized and then filtered using Pearson correlation. The signal features were then divided into training and testing sets. The training set is input into the KAN network architecture module and the Transformer network architecture module respectively to obtain the output features of the corresponding KAN network and the output features of the Transformer network; The output features of the KAN network and the output features of the Transformer network are adaptively fused through an adaptive fusion layer. A loss function combining tool wear mechanism and data loss is constructed, and tool wear value is predicted using the test set.

2. The method for predicting dual-drive tool wear based on adaptive fusion KAN-Transformer data mechanism according to claim 1, characterized in that, The KAN network architecture module includes a linear layer, the output of which is connected to the input of a nonlinear RBF layer, the output of which is connected to the input of a ReLU activation function layer, and the output of which is connected to the input of an adaptive fusion layer.

3. The method for predicting dual-drive tool wear based on adaptive fusion of KAN-Transformer data mechanism according to claim 1, characterized in that, The Transformer network architecture module includes a feature encoding layer and a temporal encoding layer. The outputs of the feature encoding layer and the temporal encoding layer are both connected to the input of the fused feature layer. The output of the fused feature layer is connected to the input of the multi-head attention layer. The output of the multi-head attention layer is connected to the input of the layer normalization layer. The output of the layer normalization layer is connected to the input of the feedforward network. The output of the feedforward network is connected to the input of the adaptive fusion layer.

4. The method for predicting dual-drive tool wear based on adaptive fusion KAN-Transformer data mechanism according to claim 1, characterized in that, The process of standardizing the acquired tool signal features and combining them with Pearson correlation for feature filtering includes: The acquired tool signal features are normalized, and the actual tool wear values ​​are also normalized. Map the normalized data to a preset range proportionally; The correlation coefficient between the acquired tool signal features and the output wear value is calculated using the Pearson correlation coefficient, and signal features with a correlation coefficient greater than a preset value are selected.

5. The method for predicting dual-drive tool wear based on adaptive fusion KAN-Transformer data mechanism according to claim 1, characterized in that, The training set is input into the KAN network architecture module to obtain the output features of the KAN network, including: The selected features are input into the KAN network architecture module for linear and nonlinear RBF spline transformations, calculated using the following formulas: Where W is the weight of the linear layer, b is the bias of the linear layer, x is the input feature, and c i k are the learnable coefficients. i This represents the position of the i-th node, and exp represents the exponential function. The calculated linear transformation and nonlinear RBF spline transformation are fused to obtain the output features of the KAN network. The calculation formula is as follows: in, For hyperparameters, The value of the linear transformation. represents the value of the nonlinear RBF spline transform.

6. The method for predicting dual-drive tool wear based on adaptive fusion of KAN-Transformer data mechanism according to claim 1, characterized in that, The training set is input into the Transformer network architecture module to obtain the output features of the Transformer network, including: The selected features are input into the Transformer network architecture module. The input features are mapped into feature vectors through the feature encoding layer, and the time input is mapped into time feature vectors through the time encoding layer. The feature vector and the time feature vector are stacked according to their dimensions to construct a sequence, thus obtaining the fused feature sequence. The fused feature sequence is linearly mapped to obtain the query, key, and value, respectively, using the following formulas: ， , in, For the input fused feature sequence, , , The weight matrix is ​​a learnable weight matrix; The multi-head attention mechanism uses a multi-head attention layer to embed the input and obtain the query, key, and value through linear transformations. The calculation formula is as follows: in, , Indicates the first Scaling attention score matrix for each attention head This represents the transpose of the key matrix of the i-th head. , This represents the total embedding dimension of the model. The number of heads representing multi-head attention. This represents the dimension of each attention head. Let i be the query matrix for the i-th head. For the first The value matrix of each head; The output features of the Transformer network are obtained after layer normalization and a feedforward network.

7. The method for predicting dual-drive tool wear based on adaptive fusion of KAN-Transformer data mechanism according to claim 6, characterized in that, The formula for calculating the output features of the Transformer network after layer normalization and feedforward network is as follows: Where LN represents layer normalization and FFN represents feedforward network.

8. The method for predicting dual-drive tool wear based on adaptive fusion KAN-Transformer data mechanism according to claim 1, characterized in that, The formula for adaptively fusing the output features of the KAN network and the output features of the Transformer network is as follows: in, and This represents the learnable parameters.

9. The method for predicting dual-drive tool wear based on adaptive fusion KAN-Transformer data mechanism according to claim 1, characterized in that, The formula for calculating the loss function that combines tool wear mechanism and data loss is as follows: in, , in, N This represents the total number of data collections. Indicates the first i The actual value of tool wear during the first data acquisition. Indicates the first i Wear values ​​predicted by the model in the second data collection. This represents the normalized derivative term. Let be the objective function. This represents the i-th sampling time point. The time-varying characteristics of wear rate, the exponential term Reflecting the nonlinear characteristics of the wear process, e, f, g 1 , g 2 These are the learnable physical parameters. This is a hyperparameter.

Images