Cross-stitch network-based multi-task high coordination monitoring method for cutting tools
By using feature extraction and fusion methods based on cross-stitch networks, the problem of low multi-task collaboration in tool monitoring was solved, achieving high-precision anomaly detection and prediction, and improving the overall performance of tool monitoring.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- TAIYUAN UNIVERSITY OF SCIENCE AND TECHNOLOGY
- Filing Date
- 2024-06-03
- Publication Date
- 2026-07-24
AI Technical Summary
Existing tool monitoring methods cannot effectively utilize the correlation between tasks, resulting in low multi-task coordination and difficulty in achieving high-precision anomaly detection and remaining service life prediction.
We employ a cross-stitch network for feature extraction and fusion. We extract features for anomaly detection and prediction tasks using wavelet neural networks and gated recurrent networks, respectively. We also utilize channel, spatial, and temporal co-attention modules for feature sharing and combine an automatic weight adjustment strategy to optimize the loss function, thereby achieving deep fusion of multiple tasks.
It significantly improves the coordination of multi-task tool monitoring, enhances the accuracy of anomaly detection and prediction, and has good stability and robustness.
Smart Images
Figure CN118699874B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of multi-task monitoring technology for cutting tools, and more specifically to a highly collaborative multi-task monitoring method for cutting tools based on cross-stitch networks. Background Technology
[0002] Cutting tools are critical cutting components in workpiece machining and are widely used in various fields of modern industry. During machine operation, continuous wear leads to tool degradation and damage. To ensure the safe use of cutting tools in actual production, anomaly detection and remaining useful life prediction are crucial. Tool monitoring is receiving increasing attention. However, existing methods can only meet the needs of single-task or multi-task learning based on hard parameter sharing, without further considering or researching methods to explore soft parameter sharing by utilizing the correlation between tasks, thus leading to low task synergy. Based on the current research status, it is necessary to optimize and improve existing tool monitoring methods to solve the problem of low synergy in existing multi-task tool monitoring methods.
[0003] It should be noted that the above content falls within the inventor's technical knowledge and does not necessarily constitute prior art. Summary of the Invention
[0004] The purpose of this invention is to solve the problems existing in the prior art and provide a highly collaborative monitoring method for multi-task tools based on cross-stitch networks. It innovatively introduces and optimizes cross-stitch networks and applies them to multi-task tool monitoring, which significantly improves the collaborative nature of multi-task tools, achieves higher classification and prediction accuracy, and has good stability.
[0005] The present invention achieves the above objectives by adopting the following technical solutions:
[0006] A multi-task, highly collaborative monitoring method for cutting tools based on cross-stitch networks includes the following steps:
[0007] S1. Collect tool vibration signals and construct a dataset from the collected vibration signals. The dataset includes a training dataset and a test dataset, as shown below:
[0008] (1)
[0009] (2)
[0010] In the formula, Represents a dataset, Indicates the number of data samples. , , They represent the first and second elements in the dataset, respectively. Each data sample and its corresponding ground truth label for the anomaly detection task Predict the true labels for the task;
[0011] The dataset contains data samples with different degrees of wear. Data samples with a wear level less than 0.16 mm are designated as normal samples and assigned label 0; data samples with a wear level greater than or equal to 0.16 mm are designated as abnormal samples and assigned label 1. The wear threshold for the predicted tag is set to 0.16 mm;
[0012] S2. To ensure the uniqueness of features for each task, different branches are constructed for feature extraction. The anomaly detection task uses a wavelet neural network for feature extraction. The prediction task uses a gated recurrent network for feature extraction;
[0013] S3. Construct a cross-stitch network, which includes an input layer, a channel co-attention module, a spatial co-attention module, a temporal co-attention module, a split-feature fusion module, and an output layer; input the features extracted by the wavelet neural network and the gated recurrent network into the cross-stitch network to perform anomaly detection tasks and... Deep fusion of prediction task features; the input layer is used to receive anomaly detection and... The prediction task output features are obtained by: the channel co-attention module generating attention matrices for each task by calculating the similarity between features from different channels; the spatial co-attention module generating a shared attention matrix by calculating the similarity between different feature spaces; and the temporal co-attention module generating a shared attention matrix by calculating the similarity between different feature times. A split-feature fusion module receives the output features from the channel, spatial, and temporal co-attention modules and performs deep fusion to obtain deep fused features. The output layer receives the deep fused features and performs anomaly detection and... Predict the output of the task;
[0014] S4. Construct a regularized loss function, input the training dataset, and use an automatic weight adjustment strategy to adaptively optimize the loss weights for different tasks, so that the regularized loss function reaches its minimum, and complete the network training.
[0015] S5. Input the test dataset into the trained network model to obtain the anomaly detection task and... Predict the outcome of the task.
[0016] In step S2, the anomaly detection task uses a wavelet neural network for feature extraction, as shown below:
[0017] (3)
[0018] (4)
[0019] (5)
[0020] In the formula: and They represent the hidden layers respectively. The input and output of each node; This represents the weights from the input layer to the hidden layer; Represents the Morlet wavelet function; and These represent the translation factor and the scaling factor, respectively.
[0021] The obtained anomaly detection task output features are represented as follows:
[0022] (6)
[0023] In the formula, This represents the Sigmoid activation function of the output layer.
[0024] In step S2, the The prediction task uses a gated recurrent network for feature extraction, represented as follows:
[0025] (7)
[0026] (8)
[0027] (9)
[0028] (10)
[0029] In the formula, Indicates the reset gate. Indicates an update to the door. Indicates candidate activation, Indicates hidden layers in Characteristics of output at any given moment; This represents the Sigmoid activation function. , , , , , All represent weight parameters. , , Both represent bias parameters. express Activation function This represents the element-wise multiplication transport symbol;
[0030] In step S3, the input layer receives the anomaly detection task from step S2 and The output features of the prediction task are represented as follows:
[0031] (11)
[0032] (12)
[0033] In the formula, , , These represent the number of channels, height, and width, respectively.
[0034] In step S3, the channel co-attention module generates the attention matrix for each task by calculating the similarity between features of different channels, which is represented as follows:
[0035] (13)
[0036] (14)
[0037] (15)
[0038] (16)
[0039] (17)
[0040] In the formula, and They represent and Along the channel axis The reconstructed features after expansion, among which ; This indicates the anomaly detection task number. Channels and Prediction Task 1 Attention weights between channels; and These represent attention weight normalization; and These represent the generated anomaly detection tasks and Attention score for the prediction task;
[0041] The output feature representation of the channel common attention module can be obtained through the above process as follows:
[0042] (18)
[0043] (19).
[0044] In step S3, the spatial co-attention module generates a shared attention matrix by calculating the similarity between different feature spaces, which is represented as:
[0045] (20)
[0046] (twenty one)
[0047] (twenty two)
[0048] In the formula, and These represent average pooling and max pooling operations, respectively. This represents the common attention matrix obtained after applying the standard convolutional transformation;
[0049] After the above process, the output feature representation of the spatial common attention module can be obtained as follows:
[0050] (twenty three)
[0051] (twenty four).
[0052] In step S3, the temporal co-attention module generates a shared attention matrix by calculating the similarity between different feature times, which is represented as follows:
[0053] (25)
[0054] (26)
[0055] (27)
[0056] In the formula, and These represent the features generated after compressing the channel dimensions. Represents the temporal co-attention matrix. ;
[0057] The output features of the temporal co-attention module can be represented as follows after the above process:
[0058] (28)
[0059] (29).
[0060] In step S3, the split-feature fusion module receives the output features of the channel co-attention module, the spatial co-attention module, and the temporal co-attention module, which are represented as follows:
[0061] (30)
[0062] (31)
[0063] (32)
[0064] (33)
[0065] (34)
[0066] (35)
[0067] In the formula, , , , , , This enables features to complement and guide each other.
[0068] The deep fusion features can be obtained through the above process, and are represented as follows:
[0069] (36)
[0070] (37).
[0071] In step S3, the output layer receives deep fusion features and performs anomaly detection and... The output of the prediction task is represented as follows:
[0072] (38)
[0073] (39)
[0074] In the formula, Indicates abnormal detection results; express Prediction results; and Indicates the weighting parameter; Indicates the bias parameter; Indicates the first The probability that a sample is labeled 1 is that it belongs to an abnormal sample; The transpose operator for a matrix.
[0075] In step S4, the regularization loss function includes the anomaly detection task loss. and Predicting task losses It is represented as:
[0076] (40)
[0077] (41)
[0078] (42)
[0079] In the formula, This represents the regularization loss function; and This represents the adaptive loss weight for the corresponding task.
[0080] The present invention, by employing the above method, can bring the following beneficial effects:
[0081] Based on a novel theory, a cross-stitch network model suitable for multi-task monitoring of cutting tools is constructed. Under the premise of deep feature extraction and split feature fusion, two tasks, anomaly detection and RUL prediction, are realized, overcoming the defects of hard parameter sharing methods and effectively improving the collaboration between multiple tasks. Attached Figure Description
[0082] Figure 1 This is a flowchart of the multi-task high-cooperative monitoring method for cutting tools based on cross-stitch networks according to the present invention;
[0083] Figure 2 This is a schematic diagram of the network model of the present invention;
[0084] Figure 3 This is a schematic diagram of the channel co-attention module, spatial co-attention module, and temporal co-attention module of the present invention;
[0085] Figure 4 This is a schematic diagram of the data acquisition principle in Verification 1 of the present invention;
[0086] Figure 5 This diagram illustrates the comparison of adaptive loss changes for different tasks in verification 1 of this invention.
[0087] Figure 6 This is a diagram illustrating the change process of adaptive loss weights for different tasks in verification 1 of this invention;
[0088] Figure 7 This invention verifies the effectiveness of different network modules using ablation experiments in Verification 1.
[0089] Figure 8 A comparative chart of experimental results using five advanced methods to verify the present invention;
[0090] Figure 9 This invention provides a confusion matrix diagram for verifying the anomaly detection task.
[0091] Figure 10 For verification of the present invention 1 Fitting curves for the prediction task;
[0092] Figure 11 This invention verifies the experimental system in section 2;
[0093] Figure 12 This is a confusion matrix diagram for the anomaly detection task in Verification 2 of this invention;
[0094] Figure 13 Verification 2 of this invention Fitting curves for the prediction task; Detailed Implementation
[0095] To more clearly illustrate the overall concept of the present invention, a detailed description will be provided below with reference to the accompanying drawings and examples.
[0096] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.
[0097] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.
[0098] like Figure 1-3 As shown, the multi-task high-cooperative monitoring method for cutting tools based on cross-stitch networks includes the following steps:
[0099] S1. Collect tool vibration signals and construct a dataset from the collected vibration signals. The dataset includes a training dataset and a test dataset, as shown below:
[0100] (1)
[0101] (2)
[0102] In the formula, Represents a dataset, Indicates the number of data samples. , , They represent the first and second elements in the dataset, respectively. Each data sample and its corresponding ground truth label for the anomaly detection task Predict the true labels for the task;
[0103] The dataset contains data samples with different degrees of wear. Data samples with a wear level less than 0.16 mm are designated as normal samples and assigned label 0; data samples with a wear level greater than or equal to 0.16 mm are designated as abnormal samples and assigned label 1. The wear threshold for the predicted tag is set to 0.16 mm;
[0104] S2. To ensure the uniqueness of features for each task, different branches are constructed for feature extraction. The anomaly detection task uses a wavelet neural network for feature extraction. The prediction task uses a gated recurrent network for feature extraction;
[0105] The anomaly detection task uses a wavelet neural network for feature extraction, represented as follows:
[0106] (3)
[0107] (4)
[0108] (5)
[0109] In the formula: and They represent the hidden layers respectively. The input and output of each node; This represents the weights from the input layer to the hidden layer; Represents the Morlet wavelet function; and These represent the translation factor and the scaling factor, respectively.
[0110] The obtained anomaly detection task output features are represented as follows:
[0111] (6)
[0112] In the formula, This represents the Sigmoid activation function of the output layer;
[0113] The The prediction task uses a gated recurrent network for feature extraction, represented as follows:
[0114] (7)
[0115] (8)
[0116] (9)
[0117] (10)
[0118] In the formula, Indicates the reset gate. Indicates an update to the door. Indicates candidate activation, Indicates hidden layers in The characteristics of the output at any time, This represents the Sigmoid activation function. , , , , , All represent weight parameters. , , Both represent bias parameters. express Activation function Represents the element-wise multiplication transport symbol;
[0119] S3. Construct a cross-stitch network, which includes an input layer, a channel co-attention module, a spatial co-attention module, a temporal co-attention module, a split-feature fusion module, and an output layer; input the features extracted by the wavelet neural network and the gated recurrent network into the cross-stitch network to perform anomaly detection tasks and... Deep fusion of prediction task features; the input layer is used to receive anomaly detection and... The prediction task output features are obtained by: the channel co-attention module generating attention matrices for each task by calculating the similarity between features from different channels; the spatial co-attention module generating a shared attention matrix by calculating the similarity between different feature spaces; and the temporal co-attention module generating a shared attention matrix by calculating the similarity between different feature times. A split-feature fusion module receives the output features from the channel, spatial, and temporal co-attention modules and performs deep fusion to obtain deep fused features. The output layer receives the deep fused features and performs anomaly detection and... Predict the output of the task;
[0120] Specifically, the input layer receives the anomaly detection task in step S2 and The output features of the prediction task are represented as follows:
[0121] (11)
[0122] (12)
[0123] In the formula, , , These represent the number of channels, height, and width, respectively.
[0124] The channel co-attention module generates the attention matrix for each task by calculating the similarity between features of different channels, which is represented as follows:
[0125] (13)
[0126] (14)
[0127] (15)
[0128] (16)
[0129] (17)
[0130] In the formula, and They represent and Along the channel axis The reconstructed features after expansion, among which ; This indicates the anomaly detection task number. Channels and Prediction Task 1 Attention weights between channels; and These represent attention weight normalization; and These represent the generated anomaly detection tasks and... Attention score for the prediction task;
[0131] The output feature representation of the channel common attention module can be obtained through the above process as follows:
[0132] (18)
[0133] (19)
[0134] The spatial co-attention module generates a shared attention matrix by calculating the similarity between different feature spaces, which is represented as follows:
[0135] (20)
[0136] (twenty one)
[0137] (twenty two)
[0138] In the formula, and These represent average pooling and max pooling operations, respectively. This represents the common attention matrix obtained using the standard convolutional transformation;
[0139] After the above process, the output feature representation of the spatial common attention module can be obtained as follows:
[0140] (twenty three)
[0141] (twenty four)
[0142] The temporal co-attention module generates a shared attention matrix by calculating the similarity between different feature times, which is represented as follows:
[0143] (25)
[0144] (26)
[0145] (27)
[0146] In the formula, and These represent the features generated after compressing the channel dimensions. Represents the temporal co-attention matrix;
[0147] The output features of the temporal co-attention module can be represented as follows after the above process:
[0148] (28)
[0149] (29)
[0150] The split-feature fusion module receives the output features from the channel co-attention module, the spatial co-attention module, and the temporal co-attention module, which are represented as follows:
[0151] (30)
[0152] (31)
[0153] (32)
[0154] (33)
[0155] (34)
[0156] (35)
[0157] In the formula, , , , , , This enables features to complement and guide each other.
[0158] The deep fusion features can be obtained through the above process, and are represented as follows:
[0159] (36)
[0160] (37)
[0161] In step S3, the output layer receives deep fusion features and performs anomaly detection and... The output of the prediction task is represented as follows:
[0162] (38)
[0163] (39)
[0164] In the formula, Indicates abnormal detection results; express Prediction results; and Indicates the weighting parameter; Indicates the bias parameter; Indicates the first The probability that a sample is labeled 1 is that it belongs to an abnormal sample; The transpose operator for a matrix;
[0165] S4. Construct a regularized loss function, input the training dataset, and use an automatic weight adjustment strategy to adaptively optimize the loss weights for different tasks, so that the regularized loss function reaches its minimum, and complete the network training.
[0166] Specifically, the regularization loss function includes the anomaly detection task loss. and Predicting task losses It is represented as:
[0167] (40)
[0168] (41)
[0169] (42)
[0170] In the formula, This represents the regularization loss function; and This represents the adaptive loss weights for the corresponding task;
[0171] S5. Input the test dataset into the trained network model to obtain the anomaly detection task and... Predict the outcome of the task.
[0172] To fully demonstrate the feasibility and significant technical effects of the method of the present invention, the following verifications were conducted:
[0173] Verification 1
[0174] (1) Using the PHM2010 dataset
[0175] PHM2010 was collected during operation of a high-speed CNC cutting machine (Roders Tech RFM760) with a spindle speed of up to 42,000 rpm. The data acquisition process is as follows: Figure 4 As shown, six tools were individually used to cut stainless steel workpieces, and the lateral wear of the tools was recorded as labels. Therefore, the tool wear dataset contains six distinct subsets, C1, C2, ..., C6. Each subset records 315 samples, corresponding to 315 wear values of the tool. Each sample includes data from... , and Vibration data were collected from three directions. Since subsets C2, C3, and C5 did not provide actual tool wear values, three subsets with real tool wear values, namely C1, C4, and C6, were selected for the experimental study. The resulting multi-task dataset is shown in Table 1.
[0176] Table 1 - PHM2010 Dataset
[0177]
[0178] (2) Network hyperparameter analysis
[0179] To demonstrate the effectiveness of the automatic weight adjustment strategy for task balancing, we take the C1 subset of the dataset as an example and compare its loss with that of a weighted experiment. A weighted experiment involves assigning equal weights to two tasks for which the strategy was not used. Its visualization results are as follows Figure 5 As shown, applying the automatic weight adjustment strategy during training reduces the losses of both tasks at a similar rate, which helps the model better balance and train multiple tasks. Furthermore, using 30 as a threshold, we observe the convergence speed of both tasks. Without the automatic weight adjustment strategy, the loss of the anomaly detection task is 1 and... The loss for the prediction task 2 requires 90 and 50 training iterations to converge to 30. However, using this strategy, the training loss for both tasks can converge to 30 in about 28 iterations.
[0180] Figure 6 The adaptive loss weight changes for the two tasks are shown. and These represent the anomaly detection task and... The weights for the prediction task. In the early stages, both weights are kept at 1. As training progresses, the weights begin to change, adaptively balancing between the two to minimize the loss.
[0181] (3) The necessity of combining different modules
[0182] To verify the effectiveness of the channel co-attention module, spatial co-attention module, temporal co-attention module, and split-feature fusion module for this invention, four variants were derived and discussed based on three subsets of the PHM2010 dataset. The ablation experiment results are as follows: Figure 7 As shown. Ablation 1, 2, 3, and 4 represent the removal of co-attention modules, spatial co-attention modules, temporal co-attention modules, and split-feature fusion modules from the network of this invention, respectively.
[0183] Ablation method 3, which lacks the ability to extract temporal features from vibration signals, performs worse than ablation method 2, which lacks the ability to extract spatial features. Ablation method 4, which removes the split-feature fusion module, cannot achieve shared representation of the generated feature vectors, significantly reducing model performance. The proposed method, which enables highly fused feature representation, performs well on the three datasets, demonstrating the effectiveness of combining the modules.
[0184] (4) Visualization of experimental results
[0185] To verify the advantages of the network proposed in this invention, it was compared with five deep learning models: Wavelet Neural Network (WNN), Gated Recurrent Network (GRU), Long Short-Term Memory Network (LSTM), Multi-Task Convolutional Neural Network (DCNN), and Joint Loss Convolutional Neural Network (JLCNN). The performance comparison is shown in Table 2. Higher accuracy and F1 score indicate better anomaly detection performance; lower mean squared error and SF score indicate better model performance. Better prediction performance is achieved. It can be observed that, compared to other methods, the method of this invention achieves the best results across all evaluation metrics on the three datasets, verifying the superiority of this invention for multi-tasking tool applications.
[0186] Table 2 - Performance Comparison of the Invention with Five Networks
[0187]
[0188] To more intuitively compare the performance of different methods, a BOX diagram based on the C1 subset dataset was plotted, as shown below. Figure 8 As shown in the BOX diagram, the orange solid line represents the median, the diamond represents the mean, and the circle represents outliers. The smaller the box length, the better the robustness of the model. It can be seen that, compared with other methods, the method proposed in this invention not only achieves higher classification and prediction accuracy but also exhibits outstanding stability.
[0189] Figure 9 and Figure 10 Anomaly detection and analysis of the PHM2010 dataset throughout its entire lifecycle The prediction results were visualized. Figure 9 In the diagram, the horizontal axis represents the predicted label, the vertical axis represents the true label, coordinate 0 represents an anomaly sample, and coordinate 1 represents a normal sample. The model performs worse on the C1 dataset than on the C4 and C6 datasets, which is due to the smaller number of training samples on the C1 dataset. However, overall, the method of this invention achieves satisfactory results in distinguishing anomaly samples. Figure 10 In the diagram, the horizontal axis represents the tool cutting time, and the vertical axis represents the RUL label. The green shaded area represents the range of the corresponding confidence interval, indicating a 20% tolerance for prediction bias. The green dashed line represents the upper and lower bounds of the confidence interval. As training progresses, the model shows good fitting between the predicted and true labels in later stages, with minimal error fluctuation, demonstrating its ability to perform well. Prediction task.
[0190] Verification 2
[0191] (1) Tool wear dataset
[0192] The data acquisition experiment was conducted at the Mechanical Engineering Basic Experiment Center of Xi'an Jiaotong University. A JT-GL8-V five-axis vertical CNC machine tool manufactured by Jiatai CNC Technology Co., Ltd. was used. Vibration signals were acquired using an IFPE piezoelectric accelerometer. The experimental platform was as follows: Figure 11 As shown. The sensor collects data during the cutting process at a sampling frequency of 5000Hz. , and Vibration signals in three directions were recorded from the moment the tool was changed until the tool was severely worn and could no longer work, resulting in a dataset of approximately 1050000×6 matrices. After processing, 900 samples were obtained, with 80% selected as the training set and 20% as the test set.
[0193] (2) Visualization of experimental results
[0194] Table 3 shows the experimental comparison results between this embodiment and other methods. As can be seen from the table, the proposed method achieved satisfactory index scores, further validating the effectiveness of this invention in anomaly detection and... Effectiveness in prediction.
[0195] Table 3 - Performance Comparison of the Invention with Five Networks
[0196]
[0197] Furthermore, in order to observe the performance of this embodiment in anomaly detection and... The performance of the prediction tasks was visualized. Figure 12 The confusion matrix shown demonstrates that the method of this invention can effectively distinguish between normal and abnormal samples, proving its effective discriminative ability. Figure 13 In The visualization of the prediction task results shows that in the initial stage, the data tracking performance was poor, and the error fluctuations were large. However, as time went on, the model's generated results gradually approached the ideal.
[0198] The above experiments fully demonstrate the feasibility of the multi-task high-cooperative monitoring method for cutting tools based on cross-stitch networks, and realize anomaly detection and... This method predicts two monitoring tasks, overcomes the shortcomings of hard parameter sharing methods, effectively improves the collaboration between multiple tasks, achieves higher classification and prediction accuracy, and has good stability.
[0199] The above specific embodiments should not be construed as limiting the scope of protection of the present invention. For those skilled in the art, any alternative improvements or modifications made to the embodiments of the present invention shall fall within the scope of protection of the present invention.
[0200] Any aspects of this invention not described in detail are well-known to those skilled in the art.
Claims
1. A multi-task, highly collaborative monitoring method for cutting tools based on cross-stitch networks, characterized in that, Includes the following steps: S1. Collect tool vibration signals and construct a dataset from the collected vibration signals. The dataset includes a training dataset and a test dataset, as shown below: （1） （2） In the formula, Represents a dataset, Indicates the number of data samples. , , They represent the first and second elements in the dataset, respectively. Each data sample and its corresponding ground truth label for the anomaly detection task Predict the true labels for the task; The dataset contains data samples with different degrees of wear. Data samples with a wear level less than 0.16 mm are designated as normal samples and assigned label 0; data samples with a wear level greater than or equal to 0.16 mm are designated as abnormal samples and assigned label 1. The wear threshold for the predicted tag is set to 0.16 mm; S2. To ensure the uniqueness of features for each task, different branches are constructed for feature extraction. The anomaly detection task uses a wavelet neural network for feature extraction. The prediction task uses a gated recurrent network for feature extraction; S3. Construct a cross-stitch network, which includes an input layer, a channel co-attention module, a spatial co-attention module, a temporal co-attention module, a split-feature fusion module, and an output layer; input the features extracted by the wavelet neural network and the gated recurrent network into the cross-stitch network to perform anomaly detection tasks and... Deep fusion of prediction task features; The input layer is used to receive anomaly detection and... The channel co-attention module generates the attention matrix for each task by calculating the similarity between features of different channels, thereby obtaining the output features of the channel co-attention module. The spatial co-attention module generates a shared attention matrix by calculating the similarity between different feature spaces, thereby obtaining the output features of the spatial co-attention module; The temporal co-attention module generates a shared attention matrix by calculating the similarity between different feature times, thereby obtaining the output features of the temporal co-attention module; The split-feature fusion module receives the output features from the channel co-attention module, spatial co-attention module, and temporal co-attention module and performs deep fusion to obtain deep fused features; the output layer receives the deep fused features and performs anomaly detection and... Predict the output of the task; S4. Construct a regularized loss function, input the training dataset, and use an automatic weight adjustment strategy to adaptively optimize the loss weights for different tasks, so that the regularized loss function reaches its minimum, and complete the network training. S5. Input the test dataset into the trained network model to obtain the anomaly detection task and... Predict the outcome of the task.
2. The method for high-cooperative multi-task monitoring of cutting tools based on cross-stitch networks according to claim 1, characterized in that, In step S2, the anomaly detection task uses a wavelet neural network for feature extraction, as shown below: （3） （4） （5） In the formula: and They represent the hidden layers respectively. The input and output of each node; This represents the weights from the input layer to the hidden layer; Represents the Morlet wavelet function; and These represent the translation factor and the scaling factor, respectively. The obtained anomaly detection task output features are represented as follows: （6） In the formula, This represents the Sigmoid activation function of the output layer.
3. The method for high-coordination multi-task monitoring of cutting tools based on cross-stitch networks according to claim 2, characterized in that, In step S2, the The prediction task uses a gated recurrent network for feature extraction, represented as follows: （7） （8） （9） （10） In the formula, Indicates the reset gate. Indicates an update to the door. Indicates candidate activation, Indicates hidden layers in Characteristics of output at any given moment; This represents the Sigmoid activation function. , , , , , All represent weight parameters. , , Both represent bias parameters. express Activation function This represents the element-wise multiplication transport symbol.
4. The method for high-coordination multi-task monitoring of cutting tools based on cross-stitch networks according to claim 3, characterized in that, In step S3, the input layer receives the anomaly detection task from step S2 and The output features of the prediction task are represented as follows: （11） （12） In the formula, , , These represent the number of channels, height, and width, respectively.
5. The method for high-coordination multi-task monitoring of cutting tools based on cross-stitch networks according to claim 4, characterized in that, In step S3, the channel co-attention module generates the attention matrix for each task by calculating the similarity between features of different channels, which is represented as follows: （13） （14） （15） （16） （17） In the formula, and They represent and Along the channel axis The reconstructed features after expansion, among which ; This indicates the anomaly detection task number. Channels and Prediction Task 1 Attention weights between channels; and These represent attention weight normalization; and These represent the generated anomaly detection tasks and Attention score for the prediction task; The output feature representation of the channel common attention module can be obtained through the above process as follows: （18） （19）。 6. The method for high-coordination multi-task monitoring of cutting tools based on cross-stitch networks according to claim 5, characterized in that, In step S3, the spatial co-attention module generates a shared attention matrix by calculating the similarity between different feature spaces, which is represented as: （20） （21） （22） In the formula, and These represent average pooling and max pooling operations, respectively. This represents the common attention matrix obtained using the standard convolutional transformation; After the above process, the output feature representation of the spatial common attention module can be obtained as follows: （23） （24）。 7. The method for high-cooperative multi-task monitoring of cutting tools based on cross-stitch networks according to claim 6, characterized in that, In step S3, the temporal co-attention module generates a shared attention matrix by calculating the similarity between different feature times, which is represented as follows: （25） （26） （27） In the formula, and These represent the features generated after compressing the channel dimensions. Represents the temporal co-attention matrix. ; The output features of the temporal co-attention module can be represented as follows after the above process: （28） （29）。 8. The method for high-cooperative multi-task monitoring of cutting tools based on cross-stitch networks according to claim 7, characterized in that, In step S3, the split-feature fusion module receives the output features of the channel co-attention module, the spatial co-attention module, and the temporal co-attention module, which are represented as follows: （30） （31） （32） （33） （34） （35） In the formula, , , , , , This enables features to complement and guide each other. The deep fusion features can be obtained through the above process, and are represented as follows: （36） （37）。 9. The method for high-coordination multi-task monitoring of cutting tools based on cross-stitch networks according to claim 8, characterized in that, In step S3, the output layer receives deep fusion features and performs anomaly detection and... The output of the prediction task is represented as follows: （38） （39） In the formula, Indicates abnormal detection results; express Prediction results; and Indicates the weighting parameter; Indicates the bias parameter; Indicates the first The probability that a sample is labeled 1 is that it belongs to an abnormal sample; The transpose operator for a matrix.
10. The method for high-coordination multi-task monitoring of cutting tools based on cross-stitch networks according to claim 9, characterized in that, In step S4, the regularization loss function includes the anomaly detection task loss. and Predicting task losses It is represented as: （40） （41） （42） In the formula, This represents the regularization loss function; and This represents the adaptive loss weight for the corresponding task.