Vibration cutting tool wear state monitoring method suitable for biological soft tissue

Abstract

The invention belongs to the technical field of combination of biomedical engineering and intelligent monitoring, and discloses a vibration cutting tool wear state monitoring method suitable for biological soft tissue, which comprises the following steps: S1, executing vibration cutting on the biological soft tissue, and obtaining a three-channel cutting force original signal; s2, inputting the three-channel cutting force original signal into a one-dimensional residual network, and expanding the three-channel cutting force original signal into a multi-feature channel; s3, a channel attention mechanism is applied, and first enhancement is achieved for some feature channels with wear sensitivity; s4, continuing to apply the time attention mechanism, and focusing for the key time interval; s5, regression is carried out to obtain a corresponding cutter flank abrasion loss prediction value; and S6, the real-time state of the cutter is judged. Compared with the prior art, the method has the advantages that the monitoring precision and robustness in a soft tissue cutting application scene can be remarkably improved, and the high quality of subsequent microscopic optical section tomography combined with cutting, the continuity of three-dimensional modeling and the data reliability are correspondingly and effectively guaranteed.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of biomedical engineering combined with intelligent monitoring, and more particularly relates to a vibration cutting tool wear state monitoring method suitable for biological soft tissue. BACKGROUND

[0002] In the field of traditional mechanical machining, the tool state monitoring method mainly relies on cutting force, vibration, acoustic emission, spindle power and temperature signals, and realizes vibration cutting tool wear recognition through threshold judgment, time domain and frequency domain feature analysis, time-frequency transformation or modeling based on statistical quantities. In recent years, with the development of machine learning and deep learning, researchers have proposed support vector machines, random forests, hidden Markov models, and convolutional neural networks, recurrent neural networks and other methods in order to improve the accuracy and robustness of monitoring under complex working conditions. These methods have been relatively mature in the field of machining of rigid materials such as metal cutting and composite materials, and can achieve real-time monitoring and working condition adaptation to some extent.

[0003] However, further research shows that when the above existing technologies are directly applied to the vibration cutting process of biological soft tissue, there are significant adaptability problems, mainly reflected in the following aspects: First, in terms of cutting mode, the cutting force amplitude generated in the vibration cutting process of biological soft tissue is much lower than that of metal materials, usually only in the order of millinewton, which belongs to the category of weak signals; this signal is easily disturbed by environmental noise and system disturbance, resulting in a significant decrease in signal-to-noise ratio. Therefore, the traditional monitoring method relying on threshold judgment or narrowband spectral indicators cannot achieve reliable recognition. In addition, unlike metal cutting (such as turning, milling, grinding), the dominant cutting force direction of biological soft tissue is not stably concentrated in the tangential or normal direction of the cutting force, but shows significant differences due to the anisotropy and viscoelasticity of the tissue. In the vibration cutting mode, the dominant force often exhibits alternating coupling along the tool feed direction and the vibration direction, and dynamically changes with the type and water content of the tissue. This is fundamentally different from the mechanical mode in the metal cutting process, which is dominated by tangential force or normal force and has more regularity.

[0004] Second, in terms of material properties, biological soft tissue belongs to viscoelastic material, whose Young's modulus is usually two orders of magnitude lower than that of metal, and the mechanical constitutive relation shows significant nonlinearity; in the initial deformation stage, it may exhibit approximately linear behavior, but as the deformation increases, the nonlinearity becomes more and more obvious, accompanied by complex characteristics such as anisotropy and non-uniformity; in this case, the traditional frequency domain analysis method based on linear assumption or the time-frequency analysis method with fixed parameters cannot effectively capture the dynamic response characteristics of this type of material, resulting in insufficient stability of feature extraction.

[0005] Finally, most of the prior art relies on manual feature engineering, with high feature redundancy and poor generalization ability when the sample and working condition change, often requiring re-adjustment of parameters, limiting the practical application feasibility; among them, the traditional machine learning method has limited representation ability for the dynamic evolution process of tool wear; and the existing deep learning model is designed for high signal-to-noise ratio and high-energy cutting signal, and when directly migrated to the biological soft tissue vibration cutting scene, it is easy to have problems such as overfitting to noise or underfitting to signal. At the same time, the existing method generally lacks explicit modeling of the complementary and redundant relationship between multiple channel signals, and cannot distinguish the importance of features of different sensing channels, and lacks sensitive detection mechanism and explainable analysis means for early wear state. SUMMARY

[0006] In view of one or more of the above defects or needs of the prior art, the present application provides a vibration cutting tool wear state monitoring method suitable for biological soft tissue, wherein the entire monitoring process flow including signal acquisition-feature extraction-channel attention mechanism optimization-time attention mechanism optimization-tool wear state determination and the like is redesigned by fully combining the actual characteristics and specific needs of vibration cutting of biological soft tissue, and some specific operations of key steps are improved accordingly. Compared with the prior art, the monitoring accuracy and robustness in the soft tissue cutting application scene can be significantly improved, and the quality of subsequent microscopic optical sectioning imaging, the continuity of three-dimensional modeling and the reliability of data are effectively guaranteed.

[0007] To achieve the above-mentioned purpose, according to the present application, a vibration cutting tool wear state monitoring method suitable for biological soft tissue is provided, characterized in that the method comprises the following steps: S1, performing vibration cutting on biological soft tissue according to a preset length, and establishing an independent sample for each cutting segment, and acquiring three-channel cutting force signals including the cutting tool feed direction Fx, the vibration direction Fy and the normal direction Fz in real time; S2, inputting the acquired three-channel cutting force signals into a one-dimensional residual network, and expanding the corresponding output number from three to tens to hundreds of multi-feature channels; S3, applying a channel attention mechanism to the multi-feature channels output by step S2, in which process, the characteristic channels related to the characteristics of biological soft tissue material, the cutting tool feed direction, the vibration direction and the normal force are enhanced, and the processed feature sequence is output; S4, continuing to apply a time attention mechanism to the processed feature sequence of step S3, in which process, the key time intervals of tool entering, stable cutting and tool leaving are focused on for each cutting segment to enhance the time sequence features sensitive to wear; S5, regression is performed on the feature sequence processed in step S4 to obtain a predicted value of the tool flank wear amount corresponding to each cutting segment; S6, based on the predicted value obtained in step S5, the tool state is determined in combination with a preset determination threshold value, and when the predicted value reaches or exceeds the threshold value, the tool is prompted to be replaced.

[0008] As a further preferred embodiment of the application, after step S3, the processed feature sequence is preferably input into a bidirectional long short-term memory network for processing, and then the time attention mechanism is continued to be applied on the basis of the output thereof; in this process, short-term fluctuations and gradual trajectories of tool wear from light to heavy over time can be captured.

[0009] As a further preferred embodiment of the application, in step S1, the three-channel cutting force original signals are collected by a force sensor, and the sensor sampling frequency is preferably set in the range of hundreds of hertz to thousands of hertz to adapt to different accuracy and cost requirements; the three-channel cutting force original signals are preferably subjected to zero offset, filtering and light smoothing processing for suppressing sensor drift and background noise.

[0010] As a further preferred embodiment of the application, in step S1, each cutting segment is preferably configured with a uniform length under the premise of maintaining the integrity of the cutting period, and its timing division can use zero-crossing detection of a smoothed signal, hysteresis threshold or envelope detection to enhance boundary stability.

[0011] As a further preferred embodiment of the application, in step S3, the channel attention mechanism preferably realizes adaptive weighting through feature compression and gate re-scaling of each feature channel, and dynamically enhances the characteristic channels related to biological soft tissue material characteristics, cutting tool feed direction, vibration direction and normal force.

[0012] As a further preferred embodiment of the application, in step S4, the time attention mechanism preferably uses a learnable similarity measure to construct weights, and then normalizes and weights the importance of each time of the feature sequence, thereby focusing on the key time interval; in addition, a visualized time weight distribution is generated for result tracing and explainability analysis.

[0013] As a further preferred embodiment of the present application, in step S6, the determination threshold can be preset according to the predicted value of different application scenarios; when the application scenario is single neuron morphology reconstruction, blood vessel reconstruction, synaptic connection tracking, subcellular structure observation or other similar scenarios, the determination threshold is set to a first threshold at a lower level to ensure slice continuity and imaging accuracy; when the application scenario is brain region partitioning, macroscopic three-dimensional modeling or other similar scenarios, the determination threshold is set to a second threshold at a higher level and allows greater cutting redundancy to improve tool utilization.

[0014] As a further preferred embodiment of the present application, in step S6, the tool real-time state is preferably divided into normal, initial wear, moderate wear, severe wear and tool collapse.

[0015] As a further preferred embodiment of the present application, preferably in step S1, a sample library is established, which takes three-channel cutting force original signals as input, takes tool relief wear as label, and is divided into training set, validation set and test set and kept from crossing.

[0016] As a further preferred embodiment of the present application, the sample library is trained by steps S2-S4 based on an end-to-end deep learning manner, so as to obtain the predicted value of tool relief wear, and the predicted value is used to trigger the early warning of tool replacement in advance.

[0017] Overall, compared with the prior art, the above technical solutions conceived by the present application mainly have the following technical advantages: 1. The present application fully combines the actual characteristics and specific needs of vibration cutting processing of biological soft tissue, and redesigns the whole monitoring process including signal acquisition-feature extraction-channel attention mechanism optimization-time attention mechanism optimization-tool wear state determination, etc. The one-dimensional residual network can significantly expand the number of feature channels, thereby improving the expression ability of weak and non-stationary features in soft tissue cutting. By applying channel attention mechanism to multi-channel features, some key information related to biological tissue material characteristics, cutting feed direction, vibration direction and normal force can be further highlighted. By continuing to apply time attention mechanism, the key time intervals such as tool entry, stable cutting and tool withdrawal in the cutting process can be further focused, the interference of redundant intervals can be weakened, and the identification sensitivity of sudden degradation and gradual wear can be improved. 2、The application further designs the step of inputting the characteristic sequence enhanced by the channel attention into the bidirectional long short-term memory network, so that short-term fluctuations and long-term evolution can be captured at the same time, the gradual change trajectory of tool wear from light to heavy over time can be mastered, and guidance information can be provided for the subsequent optimization step; in addition, the specific operations in the steps of signal acquisition, channel attention mechanism optimization and time attention mechanism optimization are optimized, so that the action mechanism of each step can be better played, and the required operation effect can be ensured. 3、The application further proposes to construct a soft tissue cutting force signal sample library in view of the characteristics of low amplitude, easy noise interference and high non-stationarity of biological soft tissue vibration cutting signals, wherein three-direction cutting forces are used as core inputs, and tool flank wear is used as a label, so as to form a modeling data basis highly matched with the soft tissue experimental scene, so that the tool wear state monitoring can be performed, and continuous prediction and intelligent warning of tool wear can be realized in the soft tissue cutting process. 4、The tool wear state monitoring method of the application is convenient to operate as a whole, and compared with the prior art, the monitoring accuracy and robustness in the soft tissue cutting application scene can be significantly improved, so that the high quality of subsequent microscopic optical section imaging, the continuity of three-dimensional modeling and the data reliability are effectively guaranteed, and therefore the application is especially suitable for vibration cutting application occasions of biological soft tissues for various medical engineering. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 is a basic flowchart of the vibration cutting tool wear state monitoring method suitable for biological soft tissues according to the application; Figure 2 is a flowchart for more specifically showing the collection and preprocessing of three-channel cutting force signals according to one preferred embodiment of the application; Figure 3 is a schematic diagram for exemplarily showing the principle and structure of the channel attention mechanism; Figure 4 is a schematic diagram for exemplarily showing the principle and structure of the time attention mechanism. DETAILED DESCRIPTION

[0019] In order to make the purpose, technical scheme and advantages of the embodiments of the application clearer, the technical scheme in the embodiments of the application will be described clearly and completely with reference to the drawings in the embodiments of the application. It should be understood that the specific embodiments described herein are only used to explain the application, and are not used to limit the application. In addition, the technical features involved in each embodiment of the application described below can be combined with each other as long as they do not conflict with each other.

[0020] As known to those skilled in the art, the micro-optical sectioning tomography technology combined with cutting is a powerful tool for three-dimensional imaging of intact organs with single-cell resolution. In this technology, a vibration cutting system overcomes the imaging depth limitation of optical microscopes due to tissue scattering by performing high-precision, continuous thin layer cutting on biological tissue samples, thereby realizing non-destructive, continuous three-dimensional reconstruction of large volume samples. The cutting quality directly determines the clarity and alignment accuracy of subsequent optical imaging. Only high-quality cutting slices with smooth surfaces and uniform thickness can ensure truly single-cell level accurate imaging. However, during continuous cutting, tool wear is inevitable. Tool wear will cause rough cutting surface, scratches or tissue tearing, which not only reduces the quality of single slice, but also causes thickness inconsistency and image misalignment between consecutive slices, seriously damaging the continuity and accuracy of three-dimensional reconstruction data. Therefore, in order to realize stable and reliable high-resolution imaging, the wear state of the cutting tool must be monitored in real time and accurately, and the tool must be replaced in time before the performance degradation affects the imaging quality, to ensure the continuity of the imaging process and the reliability of the data.

[0021] As analyzed in the "BACKGROUND" section, various cutting tool wear state monitoring methods in the prior art often have the problems of weak signal amplitude, strong non-stationarity, high dependence on artificial features, etc., resulting in low wear recognition accuracy, poor robustness, and inability to realize real-time monitoring. Therefore, the present application fully combines the actual characteristics and specific needs of vibration cutting of biological soft tissue and redesigns the entire monitoring process including signal acquisition-feature extraction-channel attention mechanism optimization-time attention mechanism optimization-tool wear state determination and other main steps.

[0022] Figure 1 is the basic flowchart of the vibration cutting tool wear state monitoring method for biological soft tissue according to the present application, and the present application will be explained more specifically with reference to Figure 1 .

[0023] Step one, perform vibration cutting on biological soft tissue according to the preset length, and establish independent samples for each cutting segment, and real-time acquire three-channel cutting force signals including cutting tool feed direction Fx, vibration direction Fy and normal direction Fz.

[0024] More specifically, referring to Figure 2In the process of vibration cutting soft tissue, a high-sensitivity three-component force sensor can be used to obtain multi-channel cutting force original signals in the feed direction Fx, the vibration direction Fy and the normal direction Fz. Then, the original signals continuously collected are subjected to zero offset, filtering and light smoothing processing for suppressing sensor drift and background noise. In addition, the continuous process can be time-sequentially divided based on the zero-crossing detection of the smoothed signals, and each fixed-length cutting segment is standardized to establish an independent sample. The sampling frequency and length are uniformly configured to ensure input consistency.

[0025] In step two, the obtained three-channel cutting force signals are input into a one-dimensional residual network, and the corresponding output quantity is expanded from three to dozens to hundreds of multi-feature channels.

[0026] More specifically, the three-channel cutting force signals can be constructed into a time sequence tensor and input into a one-dimensional residual network to extract local time sequence patterns and multi-scale trend information through multi-scale convolution and residual connection, thereby improving the expression ability of weak and non-stationary features in soft tissue cutting. The basic principles and structural forms of the one-dimensional residual network are well known in the art, and will not be described here. In this step, the convolution operation maps the original three channels to a higher-dimensional feature space, and the number of output feature channels is expanded from 3 to dozens to hundreds, providing rich feature expression for subsequent time sequence modeling.

[0027] In step three, the multi-feature channels output by the one-dimensional residual network are subjected to a channel attention mechanism. In this process, the characteristic channels related to the characteristics of biological soft tissue materials, the feed direction of the cutting tool, the vibration direction and the normal force are dynamically enhanced, and the processed feature sequence is output.

[0028] Figure 3 The principle and structure of the channel attention mechanism are exemplarily shown. The basic principles and component structures of the channel attention mechanism are well known to those skilled in the art, and will not be described here. More specifically, in this step, the channel attention mechanism can achieve adaptive weighting through feature compression and gate re-labeling of each feature channel, so as to adjust the contribution of different feature channels in different working conditions and stages, and suppress the interference of redundant feature channels. This mechanism can highlight some key information in a high feature space, thereby balancing physical meaning and automatic feature selection. In the present application, these key information are designed to maintain some characteristic channels related to the characteristics of biological soft tissue materials, the feed direction of the cutting tool, the vibration direction and the normal force.

[0029] In step four, the processed feature sequence is subjected to a time attention mechanism. In this process, the key time intervals of tool engagement, stable cutting and tool withdrawal are focused on for each cutting segment to enhance the time sequence features sensitive to wear.

[0030] Figure 4 The principle and structure of the time attention mechanism are demonstrated. The basic principle and structure of the time attention mechanism are also known to those skilled in the art, and thus will not be described here. More specifically, in this step, the time attention mechanism constructs weights with a learnable similarity metric, and then normalizes and weights the importance of each time of the feature sequence, thereby achieving focusing on the key time intervals. In the present application, these key time intervals are designed to correspond to the stages of entering the cutting, stable cutting, and withdrawing the cutting of each cutting segment, which can further weaken the interference of redundant intervals on the regression results and improve the identification sensitivity to sudden degradation and gradual wear. The mechanism can also output a time weight distribution diagram to assist in result tracing and explainability analysis.

[0031] According to a preferred embodiment of the present application, the feature sequence enhanced by the channel attention can be input into a bidirectional long short-term memory network, which can simultaneously capture short-term fluctuations and long-term evolution, and grasp the gradual trajectory of tool wear from light to heavy over time, which helps to provide guidance information for the subsequent optimization step.

[0032] Step five, the feature sequence processed by the time attention mechanism is subjected to regression to obtain the predicted value of the tool flank wear of each cutting segment. More specifically, the global feature sequence processed by the time attention mechanism is input into a fully connected layer to perform regression.

[0033] Step six, based on the predicted value of the tool flank wear obtained in step S5, and in combination with a predetermined decision threshold, the tool state is determined. When the predicted value reaches or exceeds the threshold, the tool is prompted to be replaced.

[0034] More specifically, in step S6, the decision threshold can be preset according to the predicted value in different application scenarios; when the application scenario is single neuron morphological reconstruction, blood vessel reconstruction, synaptic connection tracking, subcellular structure observation or other similar scenarios, the decision threshold is set to a lower level of first threshold to ensure slice continuity and imaging accuracy; when the application scenario is brain region partitioning, macro three-dimensional modeling or other similar scenarios, the decision threshold is set to a higher level of second threshold, and greater cutting redundancy is allowed to improve tool utilization.

[0035] In addition, the real-time state of the tool can be divided into normal, initial wear, moderate wear, severe wear, and tool collapse, and different states can be managed with different preset thresholds.

[0036] According to another preferred embodiment of the present application, for the above process, a sample library taking the three-channel cutting force original signal as input and the tool relief wear as label can also be established, and the sample library is divided into training set, validation set and test set and kept from crossing each other.

[0037] On this basis, the above sample library can be trained through the above steps of feature extraction, channel attention mechanism optimization and time attention mechanism optimization based on the end-to-end deep learning.

[0038] Finally, the trained model is applied to the online signal monitoring of soft tissue vibration cutting, and real-time wear prediction and early warning can be realized. When the predicted tool relief wear reaches the corresponding preset threshold, the system triggers the tool replacement prompt in time, avoiding the roughness of the cutting surface or the uneven thickness caused by tool failure, and ensuring the high quality of the microscopic optical section imaging and the continuity and accuracy of the three-dimensional reconstruction.

[0039] A specific example is given below to more clearly and specifically explain the present application.

[0040] In the specific example, the biological soft tissue is mouse brain tissue or large volume organ samples, which are embedded in low-melting-point agarose gel to maintain morphological stability. The cutting adopts a vibration tool, and the vibration frequency is preferably 150 Hz and the driving voltage is 0.4V to obtain stable cutting conditions. The force sensor can adopt a low-range high-precision force sensor of Yuli instrument M3813A type, and the signal is collected through M8128 data acquisition card and stored as a text file. About 25000-30000 points of signal data can be obtained for each cutting, of which 20000 points are intercepted for modeling analysis, and finally a high-quality cutting force sample library can be formed.

[0041] In a micro-optical section tomography system, the high-precision force sensor is used to collect the cutting force signals in the vibration cutting process of biological soft tissue in real time. The signals collected by the sensor include three channels of the feed direction Fx, the vibration direction Fy and the normal direction Fz, and the sampling frequency is, for example, 400 Hz. The original signal can be preferably subjected to zero offset processing to eliminate the baseline drift of the sensor, and then subjected to band-pass filtering (for example, 1-200 Hz) to remove low-frequency drift and high-frequency noise, and then subjected to sliding average to further smooth random fluctuations. Zero-crossing detection is performed on the preprocessed Fx signal to locate the start point of each cutting cycle, and cutting is performed with a fixed length (preferably 20000 points). Each piece of signal after cutting is subjected to standardization processing, thereby ensuring the comparability of features between different cutting segments. Finally, the standardized cutting segments are formed, such as 3 channels x 20000 points. In this process, the tool flank wear width can also be measured by a microscope to obtain the tool flank wear corresponding to each cutting segment, and a sample library with the cutting force as the input and the tool flank wear as the label can be preferably constructed, and divided into training set, validation set and test set.

[0042] The standardized cutting segments are input into a one-dimensional residual network to extract local time domain features through multi-scale convolution kernels (typically 7, 5, 3), and the residual structure is used to relieve the gradient vanishing problem. The output is a multi-channel time domain feature tensor. The above multi-channel features are input into a channel attention module as shown in Figure 3 The channel description vector is obtained by global average pooling, then the channel weight is obtained by two fully connected layers and Sigmoid activation, and finally the original features are weighted channel by channel to strengthen the wear-sensitive direction and suppress the redundant features. Then, the channel re-labeled features can be input into a bidirectional long short-term memory network (preferably two layers, hidden=128-256) to capture short-term dynamics and long-term trends at the same time. The output is a time sequence feature sequence. Then, a time attention mechanism is introduced based on the output of the bidirectional long short-term memory network. In this operation, the importance score of each time step is first calculated, then the time weight is generated by Softmax normalization, and finally the sequence features are weighted and summed to obtain the time domain global representation, and the weight distribution can be output to visualize the attention of the model to the cutting-in, stable cutting and cutting-out stages; in this way, the time sequence features sensitive to wear can be enhanced, and the global features processed by the time attention mechanism are input into the fully connected layer for regression, thereby obtaining the tool flank wear prediction value corresponding to each cutting segment.

[0043] Based on the obtained tool flank wear prediction value, the real-time state of the tool is determined, and when the prediction value reaches or exceeds a preset threshold, the tool is prompted to be replaced.

[0044] If the sample library with the three-channel cutting force original signal as the input and the tool flank wear as the label has been constructed in the foregoing, it can be trained and deployed to the cutting experiment system, real-time input of the cutting force segment collected online, and output of the tool wear prediction value. When the tool wear prediction value is less than the preset wear threshold, it is determined that the tool is normal; when the tool wear prediction value approaches the threshold, it is prompted to be warned; and when the tool wear prediction value is greater than or equal to the threshold, it is prompted to replace the tool. In this way, the warning can be given in advance within 5-10 cutting periods before the tool failure, the rough surface or uneven thickness of the cutting piece is avoided, and the continuity of the micro-optical section imaging and the three-dimensional reconstruction accuracy are ensured.

[0045] In summary, the tool wear state monitoring method according to the present application is overall convenient to operate, can significantly improve the monitoring accuracy and robustness in the soft tissue cutting application scene compared with the prior art, and effectively ensures the high quality of the subsequent micro-optical section imaging, the continuity of the three-dimensional modeling, and the data reliability, and thus is especially suitable for the vibration cutting application occasions of various medical engineering biological soft tissues, and has good practical value and application prospect.

[0046] Those skilled in the art can easily understand that the above description is only a preferred embodiment of the present application, and is not intended to limit the present application, and any modification, equivalent replacement and improvement made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for monitoring the wear condition of vibration-driven cutting tools suitable for biological soft tissues, characterized in that, The method includes the following steps: S1. Perform vibration cutting on biological soft tissue according to the preset length, and establish an independent sample for each cutting segment to acquire three-channel cutting force signals in real time, including the cutting tool feed direction Fx, vibration direction Fy and normal direction Fz. S2. The acquired three-channel cutting force signal is input into a one-dimensional residual network, and the number of corresponding outputs is expanded from three to tens to hundreds of multi-feature channels. S3. Apply a channel attention mechanism to the multi-feature channels output in step S2. In this process, enhance the characteristic channels related to biological soft tissue material features, cutting tool feed direction, vibration direction and normal force, and output the processed feature sequence. S4. Apply a time attention mechanism to the feature sequence processed in step S3. In this process, focus on key time intervals such as tool entry, stable cutting, and tool retraction for each cutting segment to enhance the wear-sensitive temporal features. S5. Regress the feature sequence processed in step S4 to obtain the predicted value of tool flank wear for each cutting segment. S6. Based on the predicted value obtained in step S5, and in conjunction with a preset judgment threshold, the tool status is determined. When the predicted value reaches or exceeds the threshold, a prompt is made to replace the tool.

2. The method for monitoring the wear condition of vibration cutting tools as described in claim 1, characterized in that, After step S3, it is preferable to first input the processed feature sequence into a bidirectional long short-term memory network for processing, and then continue to apply the time attention mechanism based on its output; in this process, short-term fluctuations and the gradual trajectory of tool wear from light to heavy over time can be captured.

3. The method for monitoring the wear condition of vibration cutting tools as described in claim 1 or 2, characterized in that, In step S1, the three-channel cutting force raw signal is acquired by a force sensor. The sampling frequency of the sensor is preferably set in the range of hundreds of hertz to thousands of hertz to adapt to different accuracy and cost requirements. The three-channel cutting force raw signal is preferably subjected to zero bias removal, filtering and light smoothing to suppress sensor drift and background noise.

4. The method for monitoring the wear condition of vibration cutting tools as described in claim 3, characterized in that, In step S1, each cutting segment preferably adopts a uniform length configuration while maintaining a complete cutting cycle, and its temporal segmentation can be enhanced by using methods such as zero-crossing detection of smooth signals, hysteresis threshold, or envelope detection.

5. The method for monitoring the wear condition of vibration cutting tools as described in claim 1 or 2, characterized in that, In step S3, the channel attention mechanism preferably achieves adaptive weighting by feature compression and gating recalibration of each feature channel, and dynamically enhances the characteristic channels related to biological soft tissue material features, cutting tool feed direction, vibration direction and normal force.

6. The method for monitoring the wear condition of vibration cutting tools as described in claim 1 or 2, characterized in that, In step S4, the time attention mechanism preferably constructs weights using a learnable similarity metric, and then normalizes and weights the importance of each moment in the feature sequence, thereby achieving focus on the key time interval; in addition, a visualized time weight distribution is generated for result tracing and interpretability analysis.

7. The method for monitoring the wear condition of vibration cutting tools as described in claim 1 or 2, characterized in that, In step S6, the determination threshold can be preset according to different application scenarios. Specifically, when the application scenario is single neuron morphological reconstruction, vascular reconstruction, synaptic connection tracking, subcellular structure observation or other similar scenarios, the determination threshold is set to a lower first threshold to ensure slice continuity and imaging accuracy. When the application scenario is overall brain region partitioning, macroscopic three-dimensional modeling or other similar scenarios, the determination threshold is set to a higher second threshold, allowing for greater cutting redundancy to improve tool utilization.

8. The method for monitoring the wear condition of vibration cutting tools as described in claim 1 or 2, characterized in that, In step S6, the real-time status of the tool is preferably divided into several states: normal, initial wear, moderate wear, severe wear, and tool breakage.

9. The method for monitoring the wear condition of vibration cutting tools as described in claim 1 or 2, characterized in that, Preferably, a sample library is established in step S1. The sample library uses the original three-channel cutting force signal as input, the tool flank wear amount as label, and is divided into training set, validation set and test set, and keeps them from overlapping.

10. The method for monitoring the wear condition of vibration cutting tools as described in claim 9, characterized in that, The sample library is trained using end-to-end deep learning through steps S2 to S4, thereby obtaining a predicted value of the tool flank wear, and triggering an early warning to replace the tool based on the predicted value.

Images