A tool multi-task cooperative monitoring method based on a graph gated recurrent network

Abstract

The application discloses a tool multi-task cooperative monitoring method based on a graph gating recurrent network, and belongs to the technical field of fusing abnormality detection and residual useful life prediction technology. In view of the problem that existing tool multi-task monitoring methods cannot use the same intelligent model to realize multiple tasks, leading to increased costs of equipment development and deployment, the tool vibration signal and the cutting force signal are collected, and the collected signals are formed into a data set; a time and space feature is extracted by constructing a space-time graph data set; a graph gating recurrent network and a cooperative regularization function are constructed, an optimization algorithm is used to adjust the parameters of the network, so that the cooperative loss function reaches the minimum, and thus the training of the network is completed; and the test data set is input into the trained graph gating recurrent network, so that the abnormality detection and RUL prediction results are obtained. The application realizes the two tasks of abnormality detection and RUL prediction under the premise of training and deploying only one model, effectively reduces the cost of task completion, and improves the intelligence of the method.

Description

Technical Field

[0001] This invention belongs to the field of integrated anomaly detection and remaining service life prediction technology, specifically involving a tool multi-task collaborative monitoring method based on graph-gated cyclic networks. Background Technology

[0002] Cutting tools are critical cutting components of CNC machine tools, and they must operate safely and reliably during their service life. Prognostics and Health Management (PHM) is considered a fundamental technology for improving operational safety and the effectiveness of equipment maintenance. Accurate anomaly detection and Remaining Useful Life (RUL) prediction are two key tasks in the field of PHM, crucial for helping managers make timely decisions on selective maintenance or parts replacement. Therefore, there is a strong need to develop multi-task tool monitoring methods. However, practice shows that existing multi-task tool monitoring methods, limited by network structure, cannot use a single model to implement multiple tasks, leading to a significant waste of equipment resources. Based on this, it is necessary to invent a new multi-task monitoring method to address the problem that existing methods cannot use the same intelligent model to implement multiple tasks, resulting in increased equipment development and deployment costs. Summary of the Invention

[0003] To address the problem that existing tool multi-task monitoring methods cannot use the same intelligent model to implement multiple tasks, leading to increased equipment development and deployment costs, this invention provides a tool multi-task collaborative monitoring method based on graph-gated cyclic networks.

[0004] To achieve the above objectives, the present invention employs the following technical solutions:

[0005] A multi-task collaborative monitoring method for cutting tools based on graph-gated loop networks includes the following steps:

[0006] Step 1: Collect tool vibration signals and cutting force signals using an accelerometer and a triaxial force gauge, respectively. Compile the collected signals into a dataset, which includes a training dataset and a test dataset. The resulting dataset can be represented as:

[0007]

[0008]

[0009] y m ∈{1,2}

[0010] In the formula: M represents the number of data samples; x m y m and RULm Let represent the m-th sample in the dataset, and the corresponding ground label for the anomaly detection task and the ground label for the RUL prediction task, respectively. This represents a T×6 dimensional time series, where 6 represents the number of sensors; X t This represents a dataset with time-related characteristics.

[0011] The dataset contains data samples with different wear levels. Data samples with wear levels less than 0.16 mm are defined as normal samples and assigned the label 0; data samples with wear levels greater than or equal to 0.16 mm are defined as abnormal samples and assigned the label 1; based on this, the wear threshold for RUL prediction labels is set to 0.16 mm.

[0012] Step 2: Construct a spatiotemporal graph dataset based on the original dataset to extract temporal and spatial features from the data; the specific steps for constructing the spatiotemporal graph dataset are as follows:

[0013] First, find the neighboring sensors of each sensor, treating each sensor as a node. Randomly select a sensor and calculate its Euclidean distance to the other sensors, expressed as:

[0014]

[0015] In the formula: Let θ represent the distance between node i and node j; cos(θ) represents the cosine similarity between the feature vectors of the two nodes; using the Z-score normalization strategy, the square of the feature vector norm equals 1, i.e., |θ|. Therefore, the Euclidean distance between nodes simplifies to:

[0016]

[0017] As can be seen from the above formula, the distance between two nodes is determined by the value of cos(θ); the cosine similarity between each node is calculated to find the neighbors of the node, which is expressed as:

[0018]

[0019] In the formula: Let A be the neighboring nodes of node i, and obtain the adjacency matrix A of node i. This represents the adjacency matrix between node i and node j.

[0020] After the above process, a dataset is formed, represented as follows:

[0021] G = (V, E, X) t A)

[0022] In the formula: G represents the spatiotemporal graph dataset; |V|=N represents the vertex set; E represents the edge set; X t A represents the node feature matrix; A represents the adjacency matrix.

[0023] Step 3: Construct a graph-gated recurrent network and a collaborative regularization function, and input the training dataset into the graph-gated recurrent network. Use optimization algorithms to adjust the parameters of the entire network to minimize the collaborative loss function, thereby completing the network training. The graph-gated recurrent network consists of four layers: an input layer, a graph-gated recurrent unit layer, a fully connected layer, and an output layer.

[0024] The input layer receives spatiotemporal graph data samples from the training dataset, which are represented as follows:

[0025] G = (V, E, X) t A)

[0026] The graph gated loop unit layer consists of gated loop units and graph convolution operations. The gated loop unit is represented as follows:

[0027]

[0028]

[0029]

[0030]

[0031] In the formula: r t m Indicates the reset gate; u t m Indicates an update to the door; Indicates candidate activation; h t m The m-th sample in the hidden layer represents the feature output at time t; σ represents the sigmoid activation function; w r w u w h v r v u v h Both represent weight parameters; b r b u b h All represent bias parameters; tanh represents the tanh activation function; ⊙ represents the element-wise multiplication operator; This represents the m-th sample in a dataset with time characteristics.

[0032] The gated recurrent unit after adding graph convolution operation is represented as:

[0033]

[0034]

[0035]

[0036]

[0037] In the formula: This represents the normalized adjacency matrix after adding self-connections; express The degree diagonal matrix; R t U t , and H t H represents the outputs of the reset gate, update gate, candidate activation, and hidden layer at time t, respectively; t-1 W represents the output of the hidden layer at time t-1; σ represents the sigmoid activation function; W r W u W h V r V u V h Both represent weight parameters; W g1 W g2 W g3 V g1 V g2 V g3 represents the weight parameter; tanh represents the tanh activation function; ⊙ represents the element-wise multiplication operator.

[0038] The graph-gated recurrent unit layer performs time-series-based feature extraction on data samples in the spatiotemporal graph dataset, namely:

[0039]

[0040] H t This represents the output of the hidden layer at time t; This represents the feature output by the m-th sample in the hidden layer at time t;

[0041] The fully connected layer reduces the feature dimension of the data samples in the spatiotemporal graph dataset, as shown below:

[0042]

[0043]

[0044] Where: σ r denoted by ; w1 and w2 represent the weight parameters; b1 and b2 represent the bias parameters; These represent the outputs of the anomaly detection task and the remaining lifetime prediction task after being fully connected, respectively.

[0045] The output layer performs anomaly detection and RUL prediction on the data in the spatiotemporal graph dataset, as shown below:

[0046]

[0047]

[0048] In the formula: Indicates anomaly detection results; preRUL m Represents the RUL prediction results; θ = [θ1, θ2] and w m b represents the weight parameter; m Indicates the bias parameter; This represents the probability that the m-th sample is labeled 1 and belongs to the abnormal sample category; T represents the matrix transpose operator.

[0049] The collaborative loss function in step three is expressed as follows:

[0050]

[0051]

[0052] L = 0.4·L1 + 0.6·L2

[0053] In the formula, L1 and L2 represent the losses of anomaly detection and RUL prediction, respectively; L represents the total loss of the entire network. and y m Let preRUL represent the predicted label and the true label for the m-th sample anomaly detection task, respectively. m and realRUL m Let represent the predicted label and the true label of the m-th sample in the RUL prediction task, respectively.

[0054] Step 4: Input the test dataset into the trained graph-gated recurrent network to obtain the anomaly detection and RUL prediction results.

[0055] Compared with the prior art, the present invention has the following advantages:

[0056] This invention presents a multi-task collaborative tool monitoring method based on graph-gated cyclic networks. Based on a novel principle, it achieves both anomaly detection and RUL prediction by training and deploying only one model. This effectively reduces the cost of task completion and enhances the method's intelligence. Attached Figure Description

[0057] Figure 1 This is a flowchart illustrating the present invention;

[0058] Figure 2 This is a schematic diagram of the gated loop unit of the present invention;

[0059] Figure 3 This is a schematic diagram illustrating the principle of the present invention;

[0060] Figure 4 This is a comparison chart of the results of this invention and advanced deep learning methods based on anomaly detection tasks;

[0061] Figure 5 This is a comparison chart of the results of this invention and advanced deep learning methods based on the RUL task. Detailed Implementation

[0062] A multi-task collaborative monitoring method for cutting tools based on graph-gated loop networks includes the following steps:

[0063] Step 1: Collect tool vibration signals and cutting force signals using an accelerometer and a triaxial force gauge, respectively. Compile the collected signals into a dataset, which includes a training dataset and a test dataset. The resulting dataset can be represented as:

[0064]

[0065]

[0066] y m ∈{1,2}

[0067] In the formula: M represents the number of data samples; x m y m and RUL m Let represent the m-th data sample in the dataset, and the corresponding ground label for the anomaly detection task and the ground label for the RUL prediction task, respectively. This represents a T×6 dimensional time series, where 6 represents the number of sensors; X t This represents a dataset with time-related characteristics.

[0068] The dataset contains data samples with different wear levels. Data samples with wear levels less than 0.16 mm are defined as normal samples and assigned the label 0; data samples with wear levels greater than or equal to 0.16 mm are defined as abnormal samples and assigned the label 1; based on this, the wear threshold for RUL prediction labels is set to 0.16 mm.

[0069] Step 2: Construct a spatiotemporal graph dataset based on the original dataset to extract temporal and spatial features from the data; the specific steps for constructing the spatiotemporal graph dataset are as follows:

[0070] First, find the neighboring sensors of each sensor, treating each sensor as a node. Randomly select a sensor and calculate its Euclidean distance to the other sensors, expressed as:

[0071]

[0072] In the formula: Let θ represent the distance between node i and node j; cos(θ) represents the cosine similarity between the feature vectors of the two nodes; using the Z-score normalization strategy, the square of the feature vector norm equals 1, i.e. Therefore, the Euclidean distance between nodes simplifies to:

[0073]

[0074] As can be seen from the above formula, the distance between two nodes is determined by the value of cos(θ); the cosine similarity between each node is calculated to find the neighbors of the node, which is expressed as:

[0075]

[0076] In the formula: Let A be the neighboring nodes of node i, and obtain the adjacency matrix A of node i. This represents the adjacency matrix between node i and node j.

[0077] After the above process, a dataset is formed, represented as follows:

[0078] G = (V, E, X) t A)

[0079] In the formula: G represents the spatiotemporal graph dataset; |V|=N represents the vertex set; E represents the edge set; X t A represents the node feature matrix; A represents the adjacency matrix.

[0080] Step 3: Construct a graph-gated recurrent network and a collaborative regularization function, and input the training dataset into the graph-gated recurrent network. Use optimization algorithms to adjust the parameters of the entire network to minimize the collaborative loss function, thereby completing the network training. The graph-gated recurrent network consists of four layers: an input layer, a graph-gated recurrent unit layer, a fully connected layer, and an output layer.

[0081] The input layer receives spatiotemporal graph data samples from the training dataset, which are represented as follows:

[0082] G = (V, E, X) t A)

[0083] The graph gated loop unit layer consists of gated loop units and graph convolution operations. The gated loop unit is represented as follows:

[0084]

[0085]

[0086]

[0087]

[0088] In the formula: r t m Indicates the reset gate; Indicates an update to the door; Indicates candidate activation; The m-th sample in the hidden layer represents the feature output at time t; σ represents the sigmoid activation function; w r w u w h v r v u v h Both represent weight parameters; b r b u b h All represent bias parameters; tanh represents the tanh activation function; ⊙ represents the element-wise multiplication operator; This represents the m-th sample in a dataset with time characteristics.

[0089] The gated recurrent unit after adding graph convolution operation is represented as:

[0090]

[0091]

[0092]

[0093]

[0094] In the formula: This represents the normalized adjacency matrix after adding self-connections; express The degree diagonal matrix; R t U t , and H t H represents the outputs of the reset gate, update gate, candidate activation, and hidden layer at time t, respectively; t-1 W represents the output of the hidden layer at time t-1; σ represents the sigmoid activation function; W r W u W h V r V u Vh Both represent weight parameters; W g1 W g2 W g3 V g1 V g2 V g3 represents the weight parameter; tanh represents the tanh activation function; ⊙ represents the element-wise multiplication operator.

[0095] The graph-gated recurrent unit layer performs time-series-based feature extraction on data samples in the spatiotemporal graph dataset, namely:

[0096]

[0097] Where: H t This represents the output of the hidden layer at time t; This represents the feature output by the m-th sample in the hidden layer at time t;

[0098] The fully linked layer reduces the feature dimension of the data samples in the spatiotemporal graph dataset, and is represented as follows:

[0099]

[0100]

[0101] Where: σ r denoted by ; w1 and w2 represent the weight parameters; b1 and b2 represent the bias parameters; These represent the outputs of the anomaly detection task and the remaining lifetime prediction task after being fully connected, respectively.

[0102] The output layer performs anomaly detection and RUL prediction on the data in the spatiotemporal graph dataset, as shown below:

[0103]

[0104]

[0105] In the formula: Indicates anomaly detection results; preRUL m Represents the RUL prediction results; θ = [θ1, θ2] and w m b represents the weight parameter; m Indicates the bias parameter; This represents the probability that the m-th sample is labeled 1 and belongs to the abnormal sample category; T represents the matrix transpose operator.

[0106] The collaborative loss function in step three is expressed as follows:

[0107]

[0108]

[0109] L = 0.4·L1 + 0.6·L2

[0110] In the formula, L1 and L2 represent the losses of anomaly detection and RUL prediction, respectively; L represents the total loss of the entire network. and y m Let preRUL represent the predicted label and the true label for the m-th sample anomaly detection task, respectively. m and realRUL m Let represent the predicted label and the true label of the m-th sample in the RUL prediction task, respectively.

[0111] Step 4: Input the test dataset into the trained graph-gated recurrent network to obtain the anomaly detection and RUL prediction results.

[0112] Table 1 is a schematic diagram of the experimental data partitioning based on the publicly available PHM2010 dataset of this invention;

[0113] Table 2 is a comparison of the performance metrics of this invention and advanced deep learning methods for anomaly detection tasks.

[0114] Table 3 is a comparison table of the performance indicators of the present invention and advanced deep learning methods based on the RUL prediction task;

[0115] As shown in Table 2, the method of this invention achieves the highest accuracy and F1-score compared to other advanced deep learning methods, indicating that the model has good global classification quality and anomaly detection performance; Appendix Figure 4 The results comparison chart based on the anomaly detection task is shown below. The horizontal axis represents the predicted label, the vertical axis represents the true label, coordinate 0 represents anomaly samples, and coordinate 1 represents normal samples. It can be seen that the method of this invention has a high ability to distinguish anomaly samples. Table 2 and Appendix Figure 4 The results matched, demonstrating the effectiveness of the present invention for anomaly detection tasks.

[0116] As shown in Table 3, compared with other advanced deep learning methods, Benming's method achieves the lowest root mean square error and SF-score, indicating a better fit between the model's predicted labels and the true labels; Appendix Figure 5 The comparison chart of results based on the RUL task is shown below. The horizontal axis represents the tool cutting time, and the vertical axis represents the RUL label. It can be seen that this invention consistently produces the desired RUL prediction results with minimal error fluctuation. Table 3 and Appendix Figure 5 The results matched, demonstrating the effectiveness of the present invention for RUL prediction tasks.

[0117] Table 1

[0118]

[0119] Table 2

[0120]

[0121] Table 3

[0122]

[0123] Contents not described in detail in this specification are prior art known to those skilled in the art. Although illustrative specific embodiments of the invention have been described above to facilitate understanding by those skilled in the art, it should be understood that the invention is not limited to the scope of the specific embodiments. Various modifications are readily apparent to those skilled in the art as long as they fall within the spirit and scope of the invention as defined and determined by the appended claims, and all inventions utilizing the concept of this invention are protected.

Claims

1. A method for multi-task collaborative monitoring of cutting tools based on graph-gated cyclic networks, characterized in that, The method includes: Step 1: Collect tool vibration signals using an accelerometer and cutting force signals using a triaxial force gauge. Compile the collected signals into a dataset, which includes a training dataset and a test dataset. Step 2: Construct a spatiotemporal graph dataset based on the original dataset to extract temporal and spatial features from the data; Step 3: Construct a graph-gated recurrent network and a collaborative regularization function, and input the training dataset into the graph-gated recurrent network. Use optimization algorithms to adjust the parameters of the entire network to minimize the collaborative loss function, thereby completing the network training. Step 4: Input the test dataset into the trained graph-gated recurrent network to obtain the anomaly detection and RUL prediction results; The dataset in step one is represented as follows: ， ， ， In the formula: M represents the number of data samples; , and Let represent the m-th sample in the dataset, and the corresponding ground label for the anomaly detection task and the ground label for the RUL prediction task, respectively. express A 3D time series, where 6 represents the number of sensors; Represents a dataset with temporal characteristics; The dataset contains data samples with different wear levels. Data samples with wear levels less than 0.16 mm are defined as normal samples and assigned the label 0; data samples with wear levels greater than or equal to 0.16 mm are defined as abnormal samples and assigned the label 1; based on this, the wear threshold for RUL prediction labels is set to 0.16 mm. The specific steps for constructing the spatiotemporal graph dataset in step two are as follows: First, find the neighboring sensors of each sensor, treating each sensor as a node. Randomly select a sensor and calculate its Euclidean distance to the other sensors, expressed as: ， In the formula: This represents the distance between node i and node j; This represents the cosine similarity between the feature vectors of two nodes; using the Z-score normalization strategy, the square of the feature vector norm equals 1, i.e. Therefore, the Euclidean distance between nodes simplifies to: ， As can be seen from the above formula, the distance between two nodes is... The value of determines the cosine similarity between each node; the neighbors of each node are found, represented as: ， In the formula: It is a node The adjacent nodes are obtained as nodes. The adjacency matrix A; This represents the adjacency matrix between node i and node j; After the above process, a dataset is formed, represented as follows: ， In the formula: G represents the spatiotemporal graph dataset; E represents the vertex set; E represents the edge set; A represents the node feature matrix; A represents the adjacency matrix. The graph-gated recurrent network in step three consists of four layers: an input layer, a graph-gated recurrent unit layer, a fully connected layer, and an output layer. The input layer receives spatiotemporal graph data samples from the training dataset, which are represented as follows: ， The graph gated loop unit layer consists of gated loop units and graph convolution operations. The gated loop unit is represented as follows: ， ， ， ， In the formula: Indicates the reset gate; Indicates an update to the door; Indicates candidate activation; This represents the feature output by the m-th sample in the hidden layer at time t; This represents the sigmoid activation function; , , , , , Both represent weight parameters; , , Both represent bias parameters; tanh represents the tanh activation function; This represents the element-wise multiplication operator; Represents the first in a dataset with time characteristics One sample; The gated recurrent unit after adding graph convolution operation is represented as: ， ， ， ， In the formula: This represents the normalized adjacency matrix after adding self-connections; express The degree diagonal matrix; , , and These represent the outputs of the reset gate, update gate, candidate activation, and hidden layer at time t, respectively. Indicates hidden layers in Output at any moment; This represents the sigmoid activation function; , , , , , Both represent weight parameters; W g1 W g2 W g3 V g1 V g2 V g3 Represents the weight parameters; tanh represents the tanh activation function; This represents the element-wise multiplication operator; The graph-gated recurrent unit layer performs time-series-based feature extraction on data samples in the spatiotemporal graph dataset, namely: ， In the formula: This represents the output of the hidden layer at time t; This represents the feature output by the m-th sample in the hidden layer at time t; The fully linked layer reduces the feature dimension of the data samples in the spatiotemporal graph dataset, and is represented as follows: ， ， In the formula: Represents the ReLU activation function; , Indicates the weighting parameter; , Indicates the bias parameter; , These represent the outputs of the anomaly detection task and the remaining lifetime prediction task after being fully connected, respectively. The output layer performs anomaly detection and RUL prediction on the data in the spatiotemporal graph dataset, as shown below: ， ， In the formula: Indicates abnormal detection results; This indicates the RUL prediction result; and Indicates the weighting parameter; Indicates the bias parameter; Indicates the first A sample is labeled 1, representing the probability that it belongs to an outlier sample; T represents the matrix transpose operator.

2. The tool multi-task collaborative monitoring method based on graph-gated loop networks according to claim 1, characterized in that, The collaborative loss function in step three is expressed as follows: ， ， ， In the formula, and These represent the losses from anomaly detection and RUL prediction, respectively. This represents the total loss of the entire network; and They represent the first Predicted and true labels for an anomaly detection task for a single sample; and They represent the first The predicted and true labels for a sample RUL prediction task.

Images