A Smart Tool Condition Monitoring Method Based on Multi-Source Domain Transfer Learning

Abstract

This invention discloses an intelligent tool state monitoring method based on multi-source domain transfer learning. First, multiple source domain models are pre-trained using tool state samples from multiple source domains. Then, the tool state from the target domain is input into each source domain model to obtain information entropy and classification error, thereby obtaining the transferability between each source domain and the target domain. Next, tool state samples from multiple source and target domains are simultaneously input into a deep extreme value learning machine model with a softmax classifier to obtain the feature distance loss between each source and target domain and the classification loss of each source domain. Weighted loss backpropagation is used to train and update the deep extreme value learning machine classification model. Finally, the classification model is tested using tool state test samples from the target domain, thereby achieving the monitoring of the tool state in the target domain. The greatest advantage of this method is its ability to effectively solve the problem of insufficient tool state labels and improve the accuracy of tool state monitoring.

Description

Technical Field

[0001] This invention belongs to the field of machining condition monitoring technology, specifically relating to an intelligent tool condition monitoring method based on multi-source domain transfer learning. Background Technology

[0002] Tool wear is a common problem in metal cutting. Machining dulls the tool's cutting edge, increasing friction between the tool and workpiece, and also increasing power consumption. Failure to promptly assess tool wear can negatively impact machining quality and efficiency.

[0003] Thanks to the development of deep learning technology, using deep learning networks for indirect tool condition monitoring has become a very effective method. Li et al.'s paper, "Deep Hole Boring Tool Condition Monitoring Based on Long Short-Term Memory Networks," proposed a method for monitoring boring tool condition based on deep long short-term memory networks. The results showed that the deep long short-term memory network model has good prediction accuracy for tool condition. However, these methods require a large number of labels to train the deep learning network model (Zhou et al. Deep learning fault diagnosis method based on globaloptimizationGAN for unbalanced data). In actual factory operations, it is difficult to obtain sufficient labels for every machining condition because tool wear needs to be measured by specialized instruments, which is both time-consuming and labor-intensive.

[0004] Transfer learning offers a good solution to the labeling problem. It can apply knowledge rules learned in a given source domain to similar new tasks, i.e., the target domain, and monitor tool states in the target domain through a small amount of unsupervised training in the target domain. However, it is often difficult to select a suitable source domain to adapt to the target domain. More practically, tool state sample data from the source domains can be collected from multiple different machining conditions, making multi-source domain adaptation more valuable for performance improvement. Therefore, we consider applying multi-source domain adaptive transfer learning methods to tool state monitoring. However, treating all source domains equally is unreasonable; if one source domain has a large domain difference from the target domain, it will be detrimental to domain adaptation. Therefore, it is necessary to find better methods to fully utilize multiple source domains. Summary of the Invention

[0005] To address the aforementioned problems, this invention provides an intelligent tool state monitoring method based on multi-source domain transfer learning. Multiple source domain models are pre-trained using tool state samples from multiple source domains. Then, tool state samples from the target domain are input into each source domain model to obtain information entropy and classification error, thereby obtaining the transferability between each source domain and the target domain. A method is proposed that weights the feature distance between the source and target domains and the classification loss on the source domain based on transferability, to give greater focus to source domains with high transferability.

[0006] The technical solution of the present invention:

[0007] A method for intelligent tool state monitoring based on multi-source domain transfer learning is proposed. First, multiple source domain models are pre-trained using tool state samples from multiple source domains. Then, the tool state from the target domain is input into each source domain model to obtain information entropy and classification error, thus obtaining the transferability between each source and target domain. Next, tool state samples from multiple source and target domains are simultaneously input into a deep extremum learning machine model with a softmax classifier to obtain the feature distance loss between each source and target domain and the classification loss of the source domain. A method based on the weighted feature distance loss between the domains and the classification loss of the source domain is proposed to focus more on source domains with higher transferability. The weighted loss is used for backpropagation to train and update the deep extremum learning machine classification model. Finally, the classification model is tested using tool state test samples from the target domain, thereby achieving the monitoring of the tool state in the target domain. The specific steps are as follows:

[0008] The first step is to collect tool condition samples under different machining conditions.

[0009] A cutting experiment with varying cutting parameters was designed, where different cutting parameters represent different machining conditions. A data acquisition system was used to collect vibration signals from the tool holder, with a three-dimensional (XYZ) accelerometer fixed to the spindle. The vibration signal captured under each machining condition was used as a tool state sample for each domain.

[0010] The second step is to calculate the portability between each source and target domain.

[0011] To mitigate the adverse effects of different domains, we consider domain transferability, which measures the domain similarity between each source and target domain. Here, we introduce information entropy and classification error to calculate the transferability between each source and target domain. First, we pre-train multiple deep learning models using tool state samples from multiple source domains as source domain models, obtaining the decision boundaries learned by each source domain model. Then, we input tool state samples from the target domain into the source domain models. When the domain similarity between the target and source domains is high, it will achieve lower information entropy and lower classification error.

[0012] Given tool state samples from K source domains A target domain tool state sample Tool T The information entropy and classification error of the tool state samples in the target domain are given in equations (1) and (2), respectively.

[0013]

[0014] Cl k =err / m (2)

[0015] Where k is the k-th source domain, m is the number of samples in the target domain, c is the dimension of the softmax function output of the source domain model, and p j H is the probability value of the j-th dimension output by the softmax function, and err is the number of misclassified samples in the target domain. k and Cl k Since they are not on the same order of magnitude, they need to be normalized, as shown in equation (3).

[0016]

[0017] Therefore, the portability of each source and target domain is defined as shown in equation (4). k The higher the value, the greater the portability.

[0018]

[0019] The third step is to calculate the weighted loss and classification loss of the classification model based on transferability.

[0020] By combining the obtained domain transferability, a weighted feature distance loss and a classification loss are proposed. The feature distance loss is defined as the loss between the source domain tool state sample and the classification loss. S and target domain tool state sample Tool T The distance between the second-order statistics (covariance) of the features, i.e., the coral loss, is shown in equations (5)-(7).

[0021]

[0022]

[0023]

[0024] Among them, F S and F T These are the tool state samples from the source domain. S and target domain tool state sample Tool T Extracted features, C S and C TThese are Tools S and Tools T The characteristic covariance matrix, n S and n T These are Tools S and Tools T The number of nodes in the feature extraction layer is l. Let Frobenius norm represent the square matrix, and 1 is a column vector where all elements are equal to 1.

[0025] By combining the obtained domain portability w k The coral loss and classification loss are weighted and defined as shown in equations (8) and (9).

[0026]

[0027]

[0028] in, It is the coral loss of the k-th source and target domains. L is the classification loss for the k-th source domain. Coral It is the total coral loss of the deep extreme learning machine model, L Class This is the total classification loss of the deep extreme learning machine model. In multi-source domain adaptive transfer learning, the source domain with high transferability is given a larger weight, so as not to destroy the transfer effect of the positive source domain, while reducing the negative impact of the low transferability domain.

[0029] The loss function and parameter update process of the model are shown in (10)-(12), respectively.

[0030] L = L Class +αL Coral (10)

[0031]

[0032]

[0033] Where α is the weighting coefficient balancing the classification loss and the coral loss, η is the learning rate, and θ is the weighting coefficient. i θ is the parameter of the i-th feature extraction layer. out These are the parameters of the output layer.

[0034] This model can extract tool state sample features from different domains better and faster, and transfer the learned multi-source domain knowledge to the target domain to achieve tool state monitoring.

[0035] The fourth step is to verify the effectiveness of multi-source domain transfer learning.

[0036] The tool condition test samples in the target domain are input into the trained deep extreme learning machine model for classification, verifying the transfer effect of multi-source domain transfer learning in tool condition monitoring.

[0037] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0038] 1. The multi-source domain transfer learning method proposed in this invention calculates the transferability between the target domain and each source domain by introducing information entropy and classification error, thereby performing multi-source domain adaptive transfer learning based on transferability, avoiding negative transfer damage between different domains, and effectively improving the transfer effect of multi-source domains.

[0039] 2. This invention is used in the field of tool condition monitoring, and can effectively solve the problem of insufficient tool condition labels and improve the accuracy of tool condition monitoring. Attached Figure Description

[0040] Figure 1 This is a flowchart of a tool condition intelligent monitoring method based on multi-source domain transfer learning.

[0041] Figure 2 This is a schematic diagram showing the sensor installation location.

[0042] Figure 3 This is a structural diagram of the multi-source domain transfer learning model used in this invention.

[0043] Figure 4 The vibration signal is in the Y direction; a) Time domain under normal tool conditions; b) Time domain under abnormal tool conditions; c) Frequency domain under normal tool conditions; d) Frequency domain under abnormal tool conditions.

[0044] Figure 5 (a) Transfer weights obtained in target domain C5, (b) Transfer weights obtained in target domain C6, (c) Classification results of tool condition test samples in target domain C5, (d) Classification results of tool condition test samples in target domain C6.

[0045] Figure 2 In the middle: 1. Spindle; 2. Three-dimensional accelerometer; 3. Tool holder; 4. Tool; 5. Titanium alloy workpiece. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of this invention clearer, the accompanying drawings are provided. Figure 1 Taking the milling of titanium alloy on a three-axis vertical machining center as an example, the embodiments of the present invention will be described in detail.

[0047] This experiment used a 10mm diameter rapid-feed tool holder and inserts to mill the surface of a titanium alloy workpiece. Six sets of cutting experiments with varying cutting parameters were designed, each representing a different machining condition, as shown in Table 1. A data acquisition system was used to collect vibration signals from the tool holder, with a three-dimensional (XYZ) accelerometer fixed to the spindle. The vibration signal captured under each machining condition was used as a tool state sample for each domain. The number of tool state samples collected under each machining condition is shown in Table 1.

[0048] Table 1 Experimental cutting parameters

[0049]

[0050] Tool condition was inspected using a Keyence VHX-600E ultra-deep microscope (Japan) after milling approximately 500mm. According to national standard GB / T 16460-2016, the width of the side wear band, VB = 0.3mm, is the threshold for tool failure. Tool condition was categorized as normal (VB < 0.3mm) and abnormal (VB > 0.3mm). 5000 data points were extracted from the vibration signal in each direction for each sample, and the vibration signals from all three directions were fused together. We used frequency domain data as input to the multi-source domain transfer learning model and tool condition as the sample label. Frequency domain data is as follows: Figure 4 As shown, when the tool is machining a workpiece under abnormal conditions, the amplitude increases significantly, which will seriously affect the surface quality of the workpiece. Therefore, it is necessary to monitor the tool condition and replace the tool in a timely manner.

[0051] We selected C1, C2, C3, and C4 as source domains, and C5 and C6 as target domains. The number of nodes in the deep extreme learning machine network was set to 7503-600-80, and the activation function was the sigmoid function. The weight coefficient α was set to 1e5, with the criterion that the gradient generated by the coral loss is close to the gradient generated by the classification error. Under the conditions of C1, C2, C3, and C4, multiple source domain models were pre-trained using tool state samples. Then, tool condition samples from the target domain were input into each source domain model to obtain the transferability weights for each source and target domain. The transferability weight results are as follows: Figure 5 As shown in (a).

[0052] To verify the effectiveness of the proposed method, we simultaneously input tool state samples from the four source domains and training samples of tool state from the target domain into a deep extreme learning machine model for multi-source domain transfer learning. After training, we input test samples of tool state from the target domain into the trained model for testing, and the results are as follows. Figure 5The confusion matrix is ​​shown in (b). It can be seen that multi-source domain transfer learning incorporates more information and exhibits higher classification accuracy. The model trained on multi-source domains C1, C2, C3, and C4 is applicable to target domains C5 and C6, improving the model's generalization performance.

Claims

1. A method for intelligent monitoring of tool condition based on multi-source domain transfer learning, characterized in that, First, multiple source domain models are pre-trained using tool state samples from multiple source domains. Then, the tool state of the target domain is input into each source domain model to obtain information entropy and classification error, thereby obtaining the transferability between each source domain and the target domain. Then, tool state samples from multiple source and target domains are simultaneously input into a deep extreme value learning machine (DEM) classification model with a softmax classifier to obtain the feature distance loss between each source and target domain and the classification loss of the source domain. Weighted loss backpropagation is used to train and update the DEM classification model. Finally, the DEM classification model is tested with tool state test samples from the target domain, thereby achieving monitoring of the tool state in the target domain. The specific steps are as follows: The first step is to collect tool condition samples under different machining conditions. Design a cutting experiment with varying cutting parameters, where different cutting parameters represent different machining conditions; a data acquisition system is used to collect vibration signals from the tool holder, with a triaxial accelerometer fixed on the spindle; the vibration signal under each machining condition is extracted as a tool state sample for each source domain; The second step is to calculate the portability between each source and target domain. Information entropy and classification error are introduced to calculate the transferability between each source domain and target domain. First, multiple deep learning models are pre-trained using tool state samples from multiple source domains as source domain models to obtain the decision boundary learned by each source domain model. Then, tool state samples from the target domain are input into the source domain model. When the domain similarity between the target domain and the source domain is high, smaller information entropy and lower classification error are obtained. Given tool state samples from K source domains A target domain tool state sample Tool T The information entropy and classification error of the tool state samples in the target domain are given in equations (1) and (2), respectively. Cl k =err / m (2) Where k is the k-th source domain, m is the number of samples in the target domain, c is the dimension of the softmax function output of the source domain model, and p j is the probability value of the j-th dimension output by the softmax function, and err is the number of misclassified samples in the target domain; H k and Cl k Since they are not on the same order of magnitude, normalization is required, as shown in equation (3): Therefore, the portability of each source and target domain is defined as shown in equation (4); w k The larger the value, the higher the portability; The third step is to calculate the weighted loss and classification loss of the classification model based on transferability. By combining the obtained domain transferability, a weighted feature distance loss and a classification loss are proposed; the feature distance loss is defined as the source domain tool state sample. S and target domain tool state sample Tool T The second-order statistics of the features, namely the distance between covariances, i.e., the coral loss, are shown in equations (5)-(7): Among them, F S and F T These are the tool state samples from the source domain. S and target domain tool state sample Tool T Extracted features, C S and C T These are Tools S and Tools T The characteristic covariance matrix, n S and n T These are Tools S and Tools T The number of nodes, where l is the number of nodes in the feature extraction layer; represents the Frobenius norm of a square matrix, and 1 is a column vector where all elements are equal to 1; By combining the obtained domain portability w k The coral loss and classification loss are weighted and defined as shown in equations (8) and (9): in, It is the coral loss of the k-th source and target domains. L is the classification loss for the k-th source domain. Coral It is the total coral loss of the deep extremum learning machine classification model, L Class It is the total classification loss of the deep extreme value learning machine classification model; The loss function and parameter update process of the deep extremum learning machine classification model are shown in (10)-(12) respectively: L=L Class +αL Coral (10) Where α is the weighting coefficient balancing the classification loss and the coral loss, η is the learning rate, and θ is the weighting coefficient. i θ is the parameter of the i-th feature extraction layer. out These are the parameters of the output layer; The fourth step is to verify the effectiveness of multi-source domain transfer learning. The tool condition test samples in the target domain are input into a trained deep extremum learning machine classification model for classification, thus verifying the transfer effect of multi-source domain transfer learning in tool condition monitoring.

Images