A method for detecting tool wear in milling based on a double self-attention mechanism

Abstract

The present application belongs to the technical field of mechanical cutting tool wear detection, and specifically discloses a milling tool wear detection method based on a double self-attention mechanism, comprising the following steps: sensor arrangement and experimental parameter setting, signal acquisition sample selection, model training, milling tool wear detection, and milling tool wear detection. The detection method effectively extracts features and reduces data dimensions of various sensor signals in the tool machining process, better utilizes the information extraction capability of deep learning technology to achieve higher precision tool wear prediction, and reduces the complexity of the model through more sophisticated model optimization design, so that the model requires less hardware adaptation, making it possible to truly realize tool wear detection, thereby effectively improving system reliability and reducing the risk of accidents, thus ensuring the safe operation of the machine tool. Not only does it overcome the limitations of traditional detection methods, but it also brings more efficient monitoring means to the milling industry.

Description

Technical Field

[0001] The present invention belongs to the technical field of mechanical cutting tool wear detection, and in particular relates to a milling tool wear detection method based on a dual self-attention mechanism. Background Art

[0002] With the rapid development of modern manufacturing, the requirements for tool performance and life are increasing. Monitoring and evaluating tool wear has become a critical step in improving machining quality and efficiency. Existing tool wear detection methods are primarily categorized into two main types. One involves real-time image analysis of the tool flank using machine vision technology to determine tool wear values. However, this method is often affected by factors such as cutting fluid and chips, making it incompatible with complex machining environments. Another method uses machine learning to analyze various sensor signals during tool machining to predict tool wear. This type of method is further divided into two types: one that utilizes traditional machine learning algorithms and the other that utilizes deep learning algorithms.

[0003] Traditional machine learning algorithms, unable to learn complex feature representations, typically rely on feature engineering, requiring manual data selection and feature extraction. This results in the model's performance being largely dependent on the quality and relevance of the input features. Our research also found that the correlation between the statistical indicators of sensor signals extracted during machining, as extracted through feature engineering, and tool wear changes with changes in tool type, cutting conditions, and machining materials. In other words, feature engineering developed by analyzing sensor data obtained under specific tools, machining conditions, and machining materials is generally only applicable to those specific working conditions. The development of deep learning, which enables automatic feature extraction from data, has made this problem a potential solution. Since a single tool cutting operation can potentially capture over 200,000 data points, effective feature extraction and data dimensionality reduction remain challenging. However, the existing methods of using deep learning models to detect tool wear usually divide the data into N segments, take some statistical data indicators of each segment, such as maximum value, average value, variance, etc., and then reorganize the data. On the one hand, this method has the limitations of traditional machine learning methods. During the data reconstruction process, a lot of information related to tool wear will be lost, reducing the accuracy of the prediction. On the other hand, these statistical indicators are not necessarily applicable to all tool wear data. It may happen that the selected statistical indicators are only applicable to the current tool. In addition, this method involves a large amount of data storage calls, which is not easy to achieve online monitoring.

[0004] Therefore, it is very necessary to develop a new method for detecting tool wear in milling. Summary of the Invention

[0005] The purpose of the present invention is to overcome the defects of the existing technology and provide a milling tool wear detection method based on a dual self-attention mechanism. Through higher-dimensional feature mapping and effective time series modeling, the accuracy and practicality of tool wear prediction are improved, the limitations of traditional detection methods are overcome, and at the same time, more efficient monitoring means are brought to the milling industry, promoting the further development of intelligent manufacturing; the complexity of the model is reduced through model optimization design, so that the model's requirements for compatible hardware are reduced, the system reliability is effectively improved, and the risk of accidents is reduced, thereby ensuring the safe operation of the machine tool.

[0006] In order to achieve the above objectives, the technical solution of the present invention is: a milling tool wear detection method based on a dual self-attention mechanism, comprising the following steps:

[0007] (1) Sensor arrangement and experimental parameter setting: After determining the range of cutting parameters, choose to conduct a single cutting parameter or multiple cutting parameter milling experiment; when conducting a single cutting parameter milling experiment, use bolts to install the dynamometer on the workbench, the workpiece is connected to the dynamometer by bolts, the vibration sensor is installed on the spindle, the acoustic emission sensor is placed near the workpiece, and the total power of the machine tool is collected using a power meter; after the sensor arrangement is completed, cutting is carried out under the same cutting parameters and processing conditions to obtain the data of each sensor during the processing; when conducting a multiple cutting parameter milling experiment, the workpiece is clamped by a vise; the arrangement of the acceleration sensor and acoustic emission sensor is the same as that of the single cutting parameter milling experiment;

[0008] (2) Selection of signal acquisition samples: Use an electron microscope to measure and record the tool wear after a single machining operation, and take the maximum wear value of the back tool face as the wear value; correspond the obtained single machining data to the tool wear value one by one according to the number of machining operations; each tool wear value corresponds to multiple tool passes; when extracting vibration signal samples and acoustic emission signals, first select the tool pass process signal corresponding to each tool wear value, and then select the last 4s of the stable sensor signal of each signal in the single tool pass process as the signal acquisition sample. Then, each tool wear value corresponds to one sample, and each sample contains vibration signals in the XY and Z directions, cutting force signals in the XY and Z directions, and acoustic emission signals;

[0009] (3) Model training: The one-to-one correspondence between the stable sensor signal and the tool wear value in step (2) is input into the tool wear prediction algorithm based on dual attention for training;

[0010] (4) Milling tool wear detection: The milling processing signal to be detected is input into the trained model to predict the wear of the milling tool and provide the corresponding evaluation results;

[0011] (5) Milling tool wear detection: The milling processing signal to be detected is input into the trained model to predict the wear of the milling tool and give the corresponding evaluation results.

[0012] In a preferred embodiment of the present invention, in step (1), vibration, acoustic emission, and cutting force data are collected during multiple identical machining processes, and the original input data is represented as XS, and the single machining data is represented as XSn, where n=1, 2, ..., n, and n represents the number of machining times.

[0013] In a preferred embodiment of the present invention, the specific method of training the model in step (3) comprises the following steps:

[0014] ① Data folding: The data is folded in the time series dimension using the number of points collected during one rotation of the tool as a period, and then folded in the channel dimension. After folding the data, the number of neurons required for data mapping within the period can be effectively reduced, effectively reducing the complexity of the model while performing high-dimensional feature mapping.

[0015] ② Intra-cycle data mapping: By building a intra-cycle data mapping layer based on MLP, this mapping layer takes into account the sequential nature of the data and comprehensively considers all data points within a cycle, mapping the data to a higher dimension to avoid the impact of errors in various sensor data at different times.

[0016] ③ Channel dimension expansion of data: When performing feature mapping within a cycle, since the design goal is to perform high-dimensional mapping within the cycle using data points within the cycle, the data is also folded along the channel dimension. When performing feature extraction within the cycle, it is necessary to fuse the features of signals from different channels, so the data needs to be expanded along the channel dimension;

[0017] ④ Feature extraction within a cycle: After expanding the data along the channel dimension, it is input into the cross-attention module for feature extraction within the cycle, establishing the connection between different time steps within the cycle. In the cross-attention module, the time step dimension of the matrix Q is set to 1, thereby effectively reducing the dimensionality of the data;

[0018] ⑤ Inter-cycle time series modeling: The data that has undergone intra-cycle feature extraction is input into the self-attention mechanism. Since the data has undergone intra-cycle feature extraction, the time step dimension of the data within a tool rotation cycle becomes 1. The self-attention mechanism then establishes the connection between different time steps within the cycle, thereby performing inter-cycle time series modeling on the data.

[0019] ⑥ Data aggregation: The data after multiple attention mechanisms are input into the data aggregation layer based on MLP. GELU is set as the activation function between each layer to finally obtain the predicted value;

[0020] ⑦ Calculate the parameters of each layer: The predicted value obtained above and the actual wear value are calculated through the cross entropy loss function, and then the size of each parameter in the above-mentioned layers is calculated through the back propagation algorithm.

[0021] In a preferred embodiment of the present invention, the milling tool wear detection method also includes, in order to effectively extract and reduce the data, after the data passes through the intra-cycle data mapping layer, expanding the data along the channel dimension, and inputting the data into the cross-attention mechanism of the first layer in sequence according to the cycle order to perform intra-cycle feature extraction, and then performing time series modeling between cycles through self-attention calculation between cycles.

[0022] Compared with the prior art, the present invention has the following beneficial effects:

[0023] 1. The self-attention mechanism used in this invention can take into account the relationship between each data point when processing data, fully exploit its temporal characteristics, and achieve effective fusion of multi-channel data;

[0024] 2. The tool wear detection method of the present invention is based on a dual self-attention mechanism, which improves the accuracy and practicality of tool wear prediction through higher-dimensional feature mapping and effective time series modeling;

[0025] 3. The tool wear detection method based on the dual self-attention mechanism of the present invention can not only overcome the limitations of traditional detection methods, but also bring more efficient monitoring means to the milling industry and promote the further development of intelligent manufacturing. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 is a flow chart of the tool wear detection method of the present invention;

[0027] Figure 2 Schematic diagram of the tool wear prediction model of the present invention;

[0028] Figure 3 This is a schematic diagram of the workpiece specifications and processing direction of the present invention;

[0029] Figure 4 This is a diagram showing the verification results of the model performance of the present invention. DETAILED DESCRIPTION

[0030] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is described in more detail below with reference to the accompanying drawings and specific embodiments, but the protection scope of the present invention is not limited to these embodiments.

[0031] A milling tool wear detection method based on a dual self-attention mechanism includes the following steps:

[0032] (1) Sensor arrangement and experimental parameter setting: Determine the cutting parameter range and conduct milling experiments with single or multiple cutting parameters. After the sensor arrangement is completed, perform cutting under the same cutting parameters and machining conditions to obtain data from each sensor during the machining process.

[0033] (2) Signal acquisition sample selection: The tool wear after a single machining process is measured and recorded using an electron microscope. The wear value is the maximum wear value of the flank face. The stable sensor signal of the last 4 seconds of a single machining process is extracted as the signal acquisition sample, including vibration signal, cutting force signal and acoustic emission signal. The obtained single machining data is matched with the tool wear value according to the number of machining times. Each tool wear value corresponds to a single cutting process.

[0034] (3) Model training: The one-to-one correspondence between the stable sensor signal and the tool wear value in step (2) is input into the tool wear prediction algorithm based on dual attention for training;

[0035] (4) Milling tool wear detection: The milling processing signal to be detected is input into the trained model to predict the wear of the milling tool and provide the corresponding evaluation results;

[0036] (5) Milling tool wear detection: The milling processing signal to be detected is input into the trained model to predict the wear of the milling tool and give the corresponding evaluation results.

[0037] In the step (1), vibration, acoustic emission and cutting force data of the same machining process are collected multiple times, and the original input data is expressed as XS, and the single machining data is expressed as XSn, n=1, 2, ..., n, where n represents the number of machining times.

[0038] When performing a single cutting parameter milling experiment in step (1), a dynamometer is mounted on a workbench using bolts, a workpiece is connected to the dynamometer via bolts, a vibration sensor is mounted on a spindle, and an acoustic emission sensor is placed near the workpiece.

[0039] When the multi-cutting parameter milling experiment is performed in step (1), the workpiece is clamped by a vise; the arrangement of the acceleration sensor and the acoustic emission sensor is the same as that of the single cutting parameter milling experiment.

[0040] The vibration signal sample and the acoustic emission signal in the step (2) are extracted. Firstly, the 1-time tool path whole process signal corresponding to each tool wear value is selected. Secondly, the stable signal of the last 4s in the single tool path whole process signal is selected as the sample. Each tool wear value corresponds to one sample, and each sample contains the XY and Z three-direction vibration signals, the XY and Z three-direction cutting force signals and the acoustic emission signal.

[0041] In the step (2), the power signal sample is extracted. Each tool wear value corresponds to the extraction of the 1-time tool path whole process power signal. The stable signal of the last 4s in the single tool path whole process power signal is extracted as the power signal sample. Each sample contains the XY and Z three-direction vibration signals, the XY and Z three-direction cutting force signals and the acoustic emission signal.

[0042] The specific method of the model training in the step (3) includes the following steps.

[0043] ①Data folding: the data collected in one rotation of the tool is folded in the time sequence dimension, and then the data is folded in the channel dimension. After the data is folded, the number of neurons needed for data mapping in the cycle can be effectively reduced, and the high-dimensional feature mapping is performed while the complexity of the model is effectively reduced.

[0044] ②In-cycle data mapping: a mapping layer of in-cycle data is established based on MLP. Through this mapping layer, the order of the data is considered, all data points in one rotation cycle are considered comprehensively, and the data is mapped to a higher dimension to avoid the influence of errors of various sensors at different times.

[0045] ③Channel dimension unfolding of data: when performing in-cycle feature extraction, the design purpose is to map the data points in the cycle to a high dimension in the cycle, so the data is also folded along the channel dimension. When performing in-cycle feature extraction, the signals of different channels need to be fused, so the data needs to be unfolded along the channel dimension.

[0046] ④In-cycle feature extraction: after the data is unfolded along the channel dimension, it is input into the cross-attention module for in-cycle feature extraction, and the relationship between different time steps in the cycle is established. In the cross-attention module, the time step dimension of the matrix Q is set to 1, so that the data is effectively reduced.

[0047] ⑤Inter-cycle time sequence modeling: the data extracted in the cycle is input into the self-attention mechanism. Since the data has been subjected to in-cycle feature extraction, the time step dimension of the data in one rotation cycle of the tool becomes 1. Then, the self-attention mechanism is used to establish the relationship between different time steps in the cycle, so as to model the inter-cycle time sequence.

[0048] ⑥ Data aggregation: The data after multiple attention mechanisms are input into the data aggregation layer based on MLP. GELU is set as the activation function between each layer to finally obtain the predicted value;

[0049] ⑦ Calculate the parameters of each layer: The predicted value obtained above is compared with the actual wear value through the MSE loss function, and then the size of each parameter in the above-mentioned layers is calculated through the back propagation algorithm.

[0050] The milling tool wear detection method also includes, in order to effectively extract and reduce the data, after the data passes through the intra-cycle data mapping layer, expanding the data along the channel dimension, and inputting the data into the cross-attention mechanism of the first layer in sequence according to the cycle order to perform intra-cycle feature extraction, and then performing time series modeling between cycles through self-attention calculation between cycles.

[0051] Example 1

[0052] The milling tool wear detection method based on the dual self-attention mechanism is used to perform actual detection of tool wear. The specific steps are as follows:

[0053] (1) Sensor arrangement and experimental parameter setting: The experiment was conducted under a single cutting parameter. The dynamometer was installed on the workbench using bolts. The workpiece was connected to the dynamometer by bolts. The vibration sensor was installed on the spindle, and the acoustic emission sensor was placed near the workpiece. After the sensor arrangement was completed, cutting was performed under the same cutting parameters and machining conditions. The data of each sensor during the machining process were obtained. The cutting condition settings are shown in Table 1. The models of the tool and cutter body are shown in Table 2 and Table 3, respectively.

[0054] The cutting process is as follows:

[0055] ①Workpiece specifications such as Figure 3 As shown, an indexable ball end milling cutter of the above specifications is used to process once along the illustrated processing direction, and the three-dimensional cutting force, vibration signal and acoustic emission signal during the processing are recorded.

[0056] ② After every 5000mm of machining, record the tool flank wear value.

[0057] ③The total processing length of each knife is 60 meters (or stop if the blade breaks in the middle and mark it).

[0058] ④ Test three identical knives in the same way under the same processing conditions.

[0059] (2) Selection of signal collection samples: the wear of the single processed tool is measured and recorded by an electron microscope, and the maximum wear value of the tool face is taken as the wear value. The processing data obtained by measuring the wear value each time is taken as the last 4s data, and the last 4s processing data measured each time is corresponding to the tool wear value. Each tool wear value corresponds to 3 axis vibration signals, 3 cutting force signals and acoustic emission signals, and the sampling frequency of the sensor signal is 30kHZ, so there are 120000 points in a single sample, and there are 8 columns of data in total;

[0060] (3) Model training: the processed tool sensor signal is corresponding to the tool wear value, the data set is divided by adopting three-fold cross-validation, and the tool wear prediction algorithm based on double attention is input for training. The loss function in the training process takes the MSE function, the optimizer selects Adam, in the input model hyperparameters, the batch size is set to 16, the period number is set to 600, the single period size is 200, the cross-attention mechanism (the first heavy attention) multi-head number is 2, the model dimension is 7, the multi-head number in the self-attention mechanism (the second heavy attention) is set to 256, the dropout rate is set to 0.1, the multi-head number is set to 4, the feedforward network dimension is set to 512, the layer number is set to 2, and the activation function is Gelu.

[0061] Specifically, the specific method of the model training in the step (3) in the application example of the application includes the following steps:

[0062] ① Data folding: the data collected in the tool rotation process is folded in the time sequence dimension, and then the data is folded in the channel dimension, and the data dimension after folding is changed from (16, 120000, 7) to (112, 600, 200);

[0063] ② Data mapping in the period: the mapping layer of the data in the period is established based on the MLP, the number of neurons of the mapping layer is 60, and the dropout rate is set to 0.1. After passing through the mapping layer, the data dimension is (112, 600, 200);

[0064] ③ Channel dimension expansion of data: in order to better fuse the channel information, the data is expanded along the channel dimension, and the data dimension is changed to (16, 7, 600, 200);

[0065] ④ Feature extraction in the period: after the data is expanded along the channel dimension, the data is input into the cross-attention module in a period cycle for feature extraction in the period, the cross-attention mechanism (the first heavy attention) multi-head number is 2, the model dimension is 512, and the dropout rate is 0.1. After extraction, the data dimension is changed to (16, 7, 600);

[0066] ⑤ Time series modeling between cycles: swap the positions of the second and third dimensions of the data extracted from the features within the cycle and input them into the self-attention mechanism. The number of multi-heads in the self-attention mechanism (second attention) is set to 256, the drop rate is set to 0.1, the number of multi-heads is set to 4, the feedforward network dimension is set to 512, the number of layers is set to 2, and the activation function is Gelu. After time series modeling, the data dimension is (16, 600, 256);

[0067] ⑥ Data aggregation: The data after the multiple attention mechanisms is expanded to a dimension of (16, 153600) and input into the data aggregation layer based on MLP. GELU is set as the activation function between each layer. The number of neurons in each layer is 12000, 600, and 1 respectively, and the predicted value is finally obtained.

[0068] ⑦ Calculate the parameters of each layer: The predicted value obtained above is compared with the actual wear value through the MSE loss function, and then the size of each parameter in the above-mentioned layers is calculated through the back propagation algorithm.

[0069] (4) Milling tool wear detection: The milling processing signal to be detected is input into the trained model to predict the wear of the milling tool and provide the corresponding evaluation results;

[0070] Since the data set was subjected to three-fold cross validation, three different knives were used as test sets to verify the model performance. The final verification results are as follows: Figure 4 As shown in the figure, after calculation, the average absolute error between the predicted value and the true value is 12.86μm, and the root mean square error is 18.65μm, which shows good performance.

[0071] Table 1 Cutting parameters in Example 1

[0072]

[0073] Table 2 Tool models and specifications in Example 1

[0074] factory Material code / blade model Specification Golden Heron QTD3007V-S QTD3007V-S

[0075] Table 3 Model and size of the blade in Example 1

[0076] factory Cutter body model Diameter D(mm) Hitachi ABPF30S32W 30

[0077] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A milling tool wear detection method based on a dual self-attention mechanism, characterized in that: The following steps are involved: (1) Sensor arrangement and experimental parameter setting: Determine the cutting parameter range and conduct milling experiments with single cutting parameters or multiple cutting parameters. After the sensor arrangement is completed, a single cutting parameter is cut under the same cutting parameters and processing conditions to obtain the data of each sensor during the processing. The multiple cutting parameter milling experiment is cut under variable cutting parameters. (2) Signal acquisition sample selection: The tool wear after a single machining process is measured and recorded using an electron microscope, and the wear value is the maximum wear value of the flank face; the stable sensor signal of the last 4 seconds of a single machining process is extracted as the signal acquisition sample, including vibration signal, cutting force signal and acoustic emission signal. The obtained single machining data is matched with the tool wear value according to the number of machining times, and each tool wear value corresponds to a single cutting process; (3) Model training: The one-to-one correspondence between the stable sensor signal and the tool wear value in step (2) is input into the tool wear prediction algorithm based on dual attention for training; the training method includes the following steps: ① Data folding: The data is folded in the time series dimension using the number of points collected during one rotation of the tool as a period, and then folded in the channel dimension. After folding the data, the number of neurons required for data mapping within the period can be effectively reduced, effectively reducing the complexity of the model while performing high-dimensional feature mapping. ② Intra-cycle data mapping: By building a intra-cycle data mapping layer based on MLP, this mapping layer takes into account the sequential nature of the data and comprehensively considers all data points within a cycle, mapping the data to a higher dimension to avoid the impact of errors in various sensor data at different times. ③ Channel dimension expansion of data: When performing feature mapping within a cycle, since the design goal is to perform high-dimensional mapping within the cycle using data points within the cycle, the data is also folded along the channel dimension. When performing feature extraction within the cycle, it is necessary to fuse the features of signals from different channels, so the data needs to be expanded along the channel dimension; ④ Feature extraction within a cycle: After expanding the data along the channel dimension, it is input into the cross-attention module for feature extraction within the cycle, establishing the connection between different time steps within the cycle. In the cross-attention module, the time step dimension of the matrix Q is set to 1, thereby effectively reducing the dimensionality of the data; ⑤ Inter-cycle time series modeling: The data that has undergone intra-cycle feature extraction is input into the self-attention mechanism. Since the data has undergone intra-cycle feature extraction, the time step dimension of the data within a tool rotation cycle becomes 1. The self-attention mechanism then establishes the connection between different time steps within the cycle, thereby performing inter-cycle time series modeling on the data. ⑥ Data aggregation: The data after multiple attention mechanisms are input into the data aggregation layer based on MLP. GELU is set as the activation function between each layer to finally obtain the predicted value; ⑦Calculate the parameters of each layer: The predicted value and the actual wear value are calculated through the cross entropy loss function, and then the size of each parameter in the above-mentioned layers is calculated through the back propagation algorithm; (4) Milling tool wear detection: The milling processing signal to be detected is input into the trained model to predict the wear of the milling tool and provide the corresponding evaluation results; (5) Milling tool wear detection: The milling processing signal to be detected is input into the trained model to predict the wear of the milling tool and give the corresponding evaluation results.

2. The milling tool wear detection method based on the dual self-attention mechanism according to claim 1, characterized in that: In the step (1), vibration, acoustic emission, and cutting force data are collected during multiple identical machining processes, and the original input data is represented as XS, and the single machining data is represented as XSn, where n=1, 2, ..., n, and n represents the number of machining times.

3. The milling tool wear detection method based on the dual self-attention mechanism according to claim 1, characterized in that: When performing the single cutting parameter milling experiment in step (1), the dynamometer is mounted on the workbench using bolts, the workpiece is connected to the dynamometer via bolts, the vibration sensor is mounted on the spindle, and the acoustic emission sensor is placed near the workpiece.

4. The milling tool wear detection method based on the dual self-attention mechanism according to claim 1, characterized in that: When the multi-cutting parameter milling experiment is performed in step (1), the workpiece is clamped by a vise, and the arrangement of the acceleration sensor and the acoustic emission sensor is the same as that of the single cutting parameter milling experiment.

5. The milling tool wear detection method based on the dual self-attention mechanism according to claim 1, characterized in that: When extracting the vibration signal samples and acoustic emission signals in step (2), firstly, the whole tool path process signal corresponding to each tool wear value is selected, including the cutting force, vibration and acoustic emission signals in the three directions of XY and Z.

6. The milling tool wear detection method based on the dual self-attention mechanism according to claim 1, characterized in that: When extracting the power signal sample in step (2): each tool wear value also corresponds to the stable signal of the last 4 seconds of the collected wear signal.

7. The milling tool wear detection method based on the dual self-attention mechanism according to claim 1, characterized in that: The milling tool wear detection method also includes expanding the data along the channel dimension after the data passes through the intra-cycle data mapping layer, and inputting the data into the cross-attention mechanism of the first layer in sequence according to the cycle order to perform intra-cycle feature extraction, and then performing time series modeling between cycles by calculating the self-attention between cycles.

Images