Intelligent cutter wear monitoring method and related equipment

Abstract

The invention discloses an intelligent tool wear monitoring method and related equipment. The method comprises the following steps: acquiring a cutting force typical feature set and a vibration typical feature set in the working process of a tool; inputting the cutting force typical feature set and the vibration typical feature set into a pre-trained double-branch feature fusion neural network model; wherein the dual-branch feature fusion neural network model is configured to process the cutting force typical feature set through a first branch to extract a first local feature, perform frequency domain information enhancement and time domain long-range dependency relationship capture on the first local feature in parallel, and output a first fusion enhancement feature; processing the vibration typical feature set through a second branch to extract a second local feature, performing frequency domain information enhancement and time domain long-range dependency relationship capture on the second local feature in parallel, and outputting a second fusion enhancement feature; splicing the first fusion enhancement feature and the second fusion enhancement feature to obtain a spliced feature; performing multi-scale context information extraction and mixed attention mechanism weighting on the spliced features to obtain fusion features; and outputting a predicted value of the tool wear state according to the fused features. The purpose of the invention is to meet high-precision and high-generalization monitoring requirements under complex working conditions and adapt to complex and changeable milling machining scenes.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of tool wear prediction, and particularly relates to a tool wear intelligent monitoring method and related equipment. BACKGROUND

[0002] In the field of mechanical manufacturing, as the core executive component directly involved in cutting processing, the service state of the tool directly determines the machining precision, surface quality and overall operation efficiency of the machine tool. In the process of high-efficiency milling, the tool needs to withstand high-frequency cyclic cutting force, mechanical impact caused by intermittent cutting, sharply changing thermal load in the cutting area, and complex dynamic stress caused by fluctuation of processing parameters. The coupling of multiple factors makes the tool wear mechanism more diversified, and the wear rate significantly accelerates. If the tool wear state cannot be effectively monitored and warned in time, it is easy to lead to degradation of processing quality, scrap of workpieces, and even shutdown of equipment. Therefore, real-time and accurate monitoring of tool wear in milling has become a key technical problem to be solved in the process of intelligent upgrading of manufacturing industry.

[0003] The existing tool wear monitoring method cannot meet the high-precision and high-generalization requirements under complex working conditions. The traditional monitoring method based on fixed time threshold or cumulative processing time does not consider the nonlinear characteristics of dynamic processing conditions and actual tool wear, and cannot effectively fuse the complementary value of multi-source and multi-channel information such as milling force and vibration, which is easily affected by noise interference and feature redundancy, and the monitoring precision and generalization ability are limited. Chinese patent CN120645038A discloses a monitoring scheme based on Gaussian regression model combined with grid search optimization parameters, but it does not fully exploit the complementary information of multi-source signals, lacks long-range dependence modeling capability, and has insufficient adaptability to cross-condition scenarios; Chinese patent CN120431338A uses Mel spectrogram conversion and ResNet-Transformer structure for feature extraction and fusion, but it relies on a complex image conversion process, which makes it difficult to retain the physical meaning of the original signal. The two have the following problems: 1. The fusion effect of multi-source information is not good or there is information loss in the conversion process, which cannot fully represent the tool wear state; 2. The model generalization ability and working condition adaptability are limited, and it is difficult to adapt to complex and variable milling scenarios. SUMMARY

[0004] In view of the problems existing in the prior art, the present application provides a tool wear intelligent monitoring method and related equipment, which aims to meet the monitoring requirements of high precision and high generalization under complex working conditions, and adapt to complex and variable milling scenarios.

[0005] In order to solve the above technical problems, the present application is realized by the following technical scheme: According to a first aspect of the present application, a tool wear intelligent monitoring method is provided, comprising: obtain a cutting force typical feature set and a vibration typical feature set in a tool working process; input the cutting force typical feature set and the vibration typical feature set into a pre-trained double-branch feature fusion neural network model; wherein the double-branch feature fusion neural network model is configured to: process the cutting force typical feature set through a first branch to extract first local features, and perform frequency domain information enhancement and time domain long-range dependence capture on the first local features in parallel, and output first fusion enhanced features; process the vibration typical feature set through a second branch to extract second local features, and perform frequency domain information enhancement and time domain long-range dependence capture on the second local features in parallel, and output second fusion enhanced features; splice the first fusion enhanced features and the second fusion enhanced features to obtain spliced features; perform multi-scale context information extraction and mixed attention mechanism weighting on the spliced features to obtain fusion features; output a predicted value of a tool wear state according to the fusion features.

[0006] In a possible implementation manner of the first aspect, the obtaining of the cutting force typical feature set and the vibration typical feature set in the tool working process comprises: real-time acquisition of a cutting force signal and a vibration signal in a tool working process; preprocessing of the cutting force signal and the vibration signal to eliminate invalid data segments and smooth abnormal values, and extraction of effective cutting force signals and effective vibration signals representing a stable cutting process; feature extraction on the effective cutting force signals and the effective vibration signals respectively to obtain a cutting force initial feature set and a vibration initial feature set; based on the correlation degree of each feature and the tool wear state, high-correlation features are selected from the cutting force initial feature set and the vibration initial feature set respectively, and standardized processing is performed to construct a cutting force standard feature vector set and a vibration standard feature vector set; based on the cutting state reflected by the cutting force standard feature vector set and the vibration standard feature vector set, data division is performed on both, and typical feature data representing different cutting states are selected therefrom to form a cutting force typical feature set and a vibration typical feature set.

[0007] In a possible implementation manner of the first aspect, the frequency domain information enhancement on the first local features / second local features comprises: performing fast Fourier transform on the time dimension of the first local features / second local features to obtain first frequency domain features / second frequency domain features; Calculate the power spectrum of the first frequency domain feature / second frequency domain feature, and based on the power spectrum of the first frequency domain feature / second frequency domain feature, generate a first binary mask / second binary mask through a learnable threshold; Based on the first binary mask / second binary mask, the first frequency domain feature / second frequency domain feature is denoised to obtain the denoised first frequency domain feature / second frequency domain feature. The first frequency domain features / second frequency domain features after denoising are filtered using a globally learnable filter and a locally learnable filter to obtain the filtered global first frequency domain features / second frequency domain features and local first frequency domain features / second frequency domain features. The filtered global first frequency domain features / second frequency domain features and local first frequency domain features / second frequency domain features are integrated and then subjected to inverse fast Fourier transform to obtain the first frequency domain enhanced features / second frequency domain enhanced features.

[0008] In one possible implementation of the first aspect, temporal long-range dependency capture of the first local feature / second local feature includes: The first local feature / second local feature is input into a long short-term memory network, and the gating mechanism of the long short-term memory network is used to capture the long-range dependency relationship between the first local feature / second local feature in the time dimension.

[0009] In one possible implementation of the first aspect, the step of extracting multi-scale contextual information and weighting the spliced ​​features using a hybrid attention mechanism to obtain fused features includes: Multi-scale context extraction step: The spliced ​​features are processed using multiple parallel dilated convolution branches, where the dilated convolution of each branch has a different dilation rate to synchronously capture context information at different time scales, and the multi-scale features output by each branch are fused to obtain multi-scale fused features. Hybrid attention weighting step: The multi-scale fusion feature is input into the hybrid attention module, which simultaneously learns the importance weights in the feature channel dimension and the time series dimension, and performs dual-dimensional enhancement on the multi-scale fusion feature, outputting the fusion feature; The hybrid attention module achieves dual-dimensional enhancement in the following way: Based on the multi-scale fusion features, a channel attention weight vector is generated; Based on the multi-scale fusion features, a pixel attention weight map is generated; The multi-scale fusion features are weighted along the channel dimension using the channel attention weight vector, and then weighted along the temporal dimension using the pixel attention weight map to obtain the final enhanced fusion features.

[0010] In one possible implementation of the first aspect, the step of outputting a predicted value of the tool wear state based on the fused features specifically includes: After the fused features are subjected to adaptive average pooling, they are input into an interpretable Kolmogorov-Arnold network to output a predicted value of the tool wear state.

[0011] According to a second aspect of the present invention, a smart tool wear monitoring device is provided, comprising: The acquisition module is used to acquire typical feature sets of cutting forces and vibrations during the tool's operation. The wear condition prediction module is used to input the typical feature set of cutting force and the typical feature set of vibration into a pre-trained dual-branch feature fusion neural network model; wherein, the dual-branch feature fusion neural network model is configured as follows: The first branch processes the typical feature set of cutting force to extract the first local feature, and performs frequency domain information enhancement and time domain long-range dependency capture on the first local feature in parallel to output the first fused enhanced feature; The vibration typical feature set is processed by the second branch to extract the second local feature, and the second local feature is enhanced in the frequency domain and captured in the time domain long-range dependency in parallel to output the second fused enhanced feature; The first fusion enhancement feature and the second fusion enhancement feature are concatenated to obtain the concatenated feature; The spliced ​​features are weighted by multi-scale contextual information extraction and a hybrid attention mechanism to obtain fused features; Based on the fusion characteristics, a predicted value of the tool wear condition is output.

[0012] According to a third aspect of the present invention, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the aforementioned intelligent tool wear monitoring method.

[0013] According to a fourth aspect of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium storing a computer program, which, when executed by a processor, implements the aforementioned intelligent tool wear monitoring method.

[0014] According to a fifth aspect of the present invention, a computer program product is provided that, when executed by a processor, implements the aforementioned intelligent tool wear monitoring method.

[0015] Compared with the prior art, the present invention has at least the following beneficial effects: This invention provides an intelligent tool wear monitoring method. By processing typical feature sets of cutting force and vibration in parallel, and performing frequency domain information enhancement and time domain long-range dependency capture in two branches respectively, it fully explores the inherent correlation and complementary value of multi-source signals, avoiding information loss caused by single signal sources or simple feature splicing in traditional methods, thus more comprehensively characterizing the dynamic evolution of tool wear. Through multi-scale context information extraction and hybrid attention mechanism, it effectively captures the local and global features of tool wear under different machining conditions, reducing the impact of noise interference and feature redundancy, enabling the model to adapt to complex and changing milling conditions. Through dual-branch feature fusion and enhancement mechanism, combined with neural network structure, it can quickly output predicted values ​​of tool wear status during dynamic machining, promptly identify abnormal wear, and effectively prevent machining quality degradation and equipment failure. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the specific embodiments of the present invention, the drawings used in the description of the specific embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0017] Figure 1 This is a flowchart of a method for intelligent monitoring of tool wear according to the present invention.

[0018] Figure 2 This is a schematic diagram of the tool wear monitoring signal acquisition system in the embodiment.

[0019] Figure 3a This is the original signal diagram in the embodiment; Figure 3b This is a preprocessed signal diagram from an example.

[0020] Figure 4a This is a graph showing the absolute value of the Spearman correlation coefficient between each channel feature and wear value as a function of the features in the example. Figure 4b The image shows a heatmap illustrating the correlation between the characteristics of each channel and the wear value in the embodiment.

[0021] Figure 5a The example shows the t-SNE two-dimensional distribution of the cutting signal after clustering in the X direction; Figure 5b This example shows a comparison of the cutting signal before and after dimensionality reduction in the X direction.

[0022] Figure 6 This is a schematic diagram of the dual-branch network structure in the embodiment.

[0023] Figure 7This is a schematic diagram of the ResDualFT Encoder structure in the embodiment.

[0024] Figure 8 This is a schematic diagram of the MSD-CPA Decoder structure in the embodiment.

[0025] Figure 9 This is a comparison chart of the RMSE of the proposed model with other models in the embodiments. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] like Figure 1 As shown, this invention provides an intelligent tool wear monitoring method, which specifically includes the following steps: S1. Obtain the typical characteristic set of cutting force and vibration during the tool's working process.

[0028] In one feasible approach, the specific method for obtaining the typical characteristic sets of cutting forces and vibrations during the tool's operation is as follows: S10. Real-time acquisition of cutting force and vibration signals during tool operation.

[0029] Specifically, a triaxial piezoelectric cutting force measuring instrument and a triaxial acceleration vibration sensor are installed on a CNC milling machine to simultaneously collect cutting force signals and vibration signals in the X, Y, and Z directions during the cutting process.

[0030] S11. The cutting force signal and vibration signal are preprocessed to remove invalid data segments and smooth outliers, and the effective cutting force signal and effective vibration signal characterizing the stable cutting process are extracted.

[0031] Optionally, the cutting force signal and vibration signal are preprocessed, specifically as follows: The invalid data of the cutting force signal and vibration signal are truncated by the interquartile range method, and the outlier of the truncated cutting force signal and vibration signal is removed by the Hampel filtering algorithm to extract the valid cutting force signal and valid vibration signal.

[0032] Specifically, the preprocessing steps include invalid data truncation and outlier handling, and the detailed implementation process is as follows (for simplicity, the cutting force signal or vibration signal will be referred to as the cutting signal, and the effective cutting force signal or effective vibration signal will be referred to as the effective cutting signal): (1) The specific steps for invalid data truncation are as follows: The interquartile range is calculated using the interquartile range method. The calculation formula is as follows:

[0033] In the formula, It is the first quartile of the cutting signal sequence. It is the third quartile of the cutting signal sequence.

[0034] Based on interquartile range Set the effective cutting signal threshold The calculation formula is as follows:

[0035] In the formula, This is an empirical coefficient.

[0036] Locate the starting point of the effective cutting signal, starting from the first sampling point of the cutting signal sequence. Begin iterating backwards until the first satisfied condition is found. The sampling point is denoted as index . ,delete to All data; Locate the end point of the valid cutting signal, starting from the last sampling point of the cutting signal sequence. Begin traversing forward until the first satisfied condition is found. The sampling point is denoted as index . ,delete to All data; reserve to The data serves as an effective cutting signal. .

[0037] (2) The specific steps for handling abrupt outliers are as follows: The Hampel filtering algorithm is used to process the effective cutting signal. Processing is performed on each sampling point in the valid cutting signal sequence. In its size Calculate the local median within the sliding window. Absolute deviation from the median The calculation formula is as follows:

[0038]

[0039] In the formula, For sampling points within the window, This indicates the operation of taking the median.

[0040] If sampling points A value is considered an outlier if it meets the following conditions, calculated using the following formula:

[0041] In the formula, For threshold coefficient, 1.4826 is the value under a normal distribution. The conversion factor to the standard deviation ensures the unbiasedness of the estimate.

[0042] The sampling points identified as outliers will be determined by the local median of their corresponding window. The replacement is performed to obtain the denoised effective cutting signal. This is the final effective cutting signal.

[0043] S12. Extract features from the effective cutting force signal and the effective vibration signal respectively to obtain the initial feature set of cutting force and the initial feature set of vibration.

[0044] Optionally, feature extraction is performed on the effective cutting force signal and the effective vibration signal respectively, specifically as follows: S120. Perform time-series segmentation processing on the effective cutting force signal and the effective vibration signal respectively to obtain several segments of effective cutting force signal and effective vibration signal.

[0045] S121. Extract the multidimensional time-domain features and statistical features of each segment of effective cutting force signal and effective vibration signal to obtain the initial feature set of cutting force and the initial feature set of vibration.

[0046] Specifically, the effective cutting signal According to a fixed time window Divided into Each data segment. The time-domain characteristics and statistical features of the cutting force and vibration are calculated respectively, including absolute mean, standard deviation, root mean square value, peak value, peak-to-peak value, peak factor, skewness, kurtosis, impulse factor, waveform factor and margin factor, to obtain the initial feature set of cutting force and the initial feature set of vibration.

[0047] S13. Based on the correlation between each feature and the tool wear state, highly correlated features are selected from the initial feature set of cutting force and the initial feature set of vibration, respectively, and standardized to construct the standard feature vector set of cutting force and the standard feature vector set of vibration.

[0048] In detail, the standard eigenvector set of cutting force and the standard eigenvector set of vibration are constructed as follows: For each valid cutting signal channel, the Spearman correlation coefficients are calculated between all extracted time-domain features and statistical features and the tool state (target variable). The Spearman correlation coefficients are calculated by converting the feature values ​​into ranks and then calculating the Spearman correlation coefficients between the ranks. The calculation formula is:

[0049] In the formula, For the difference in eigenvalue levels, This represents the number of samples.

[0050] For each channel, features are sorted according to the absolute value of the Spearman correlation coefficient. The three features with the highest correlation to tool wear state are then selected from the initial feature sets of cutting force and vibration. The selected features from all channels are then aggregated to form the high-correlation feature sets of cutting force and vibration.

[0051] Finally, the StandardScaler method is used to standardize the high-correlation feature sets of cutting force and vibration. Specifically, for each feature dimension, its mean is subtracted and divided by its standard deviation, ensuring that all features have zero mean and unit variance, thus eliminating the influence of dimensions between different features. The calculation formula is as follows:

[0052] In the formula, The mean of a certain feature across all data segments. This represents the standard deviation of the feature.

[0053] S14. Based on the cutting state reflected by the standard feature vector set of cutting force and the standard feature vector set of vibration, the data of the two are divided respectively, and typical feature data representing different cutting states are selected to form a typical feature set of cutting force and a typical feature set of vibration.

[0054] Optionally, based on the cutting state reflected by the cutting force standard feature vector set and the vibration standard feature vector set, the data of the two are divided separately, specifically as follows: The HDBSCAN algorithm is used to cluster the standard feature vector set of cutting force and the standard feature vector set of vibration respectively to obtain categories reflecting different cutting states, thereby realizing data partitioning.

[0055] Similarly, for the sake of brevity, the standard eigenvector set of cutting force and the standard eigenvector set of vibration are integrated into the standard eigenvector set of cutting.

[0056] Specifically, the HDBSCAN algorithm is used to cluster the standard feature vector set of cutting data, including: First, calculate the mutual reachability distance between sample points using the following formula:

[0057] In the formula, For Euclidean distance, For point The core distance, that is, the distance from this point to its th core. The distance between nearest neighbors.

[0058] Secondly, based on the mutual reachability distance, a minimum spanning tree (MST) is constructed, and a clustering tree structure is formed according to the hierarchical decomposition method.

[0059] Finally, the final set of pattern categories is determined through cluster stability analysis, noise points are removed, and the clustering results are obtained. The calculation formula is as follows:

[0060] In the formula, The number of pattern categories obtained from clustering. For the first A sample set of classes.

[0061] Specifically, typical characteristic data representing different cutting states are selected to form typical characteristic sets of cutting forces and vibrations, including: For the first Class Pattern Categories The proportion of its sample size to the total sample size The sampling quota is allocated proportionally, and the calculation formula is as follows:

[0062] in, The target total number of samples is set. This is the rounding operation.

[0063] Within each pattern category, calculate the first... Each sample in the class With this type of cluster center The Euclidean distance is calculated using the following formula:

[0064] in, This serves as the cluster center. (According to...) Sort by size from smallest to largest, prioritizing the smallest distance. Using a sample as a representative sample, we obtain the corresponding typical feature set of cutting force and typical feature set of vibration.

[0065] S2. Input the typical feature set of cutting force and the typical feature set of vibration into a pre-trained dual-branch feature fusion neural network model; wherein, the dual-branch feature fusion neural network model is configured as follows: The first branch processes the typical feature set of cutting force to extract the first local feature, and performs frequency domain information enhancement and time domain long-range dependency capture on the first local feature in parallel to output the first fused enhanced feature; The vibration typical feature set is processed by the second branch to extract the second local feature, and the second local feature is enhanced in the frequency domain and captured in the time domain long-range dependency in parallel to output the second fused enhanced feature; The first fusion enhancement feature and the second fusion enhancement feature are concatenated to obtain the concatenated feature; The spliced ​​features are weighted by multi-scale contextual information extraction and a hybrid attention mechanism to obtain fused features; Based on the fusion characteristics, a predicted value of the tool wear condition is output.

[0066] In one feasible approach, frequency domain information enhancement is performed on the first local feature / second local feature, specifically including: a. Perform a Fast Fourier Transform on the time dimension of the first local feature / second local feature to obtain the first frequency domain feature / second frequency domain feature.

[0067] b. Calculate the power spectrum of the first frequency domain feature / second frequency domain feature, and generate a first binary mask / second binary mask based on the power spectrum of the first frequency domain feature / second frequency domain feature through a learnable threshold.

[0068] c. Based on the first binary mask / second binary mask, denoise the first frequency domain feature / second frequency domain feature to obtain the denoised first frequency domain feature / second frequency domain feature.

[0069] d. Using a globally learnable filter and a locally learnable filter, the denoised first frequency domain features / second frequency domain features are filtered to obtain the filtered global first frequency domain features / second frequency domain features and local first frequency domain features / second frequency domain features.

[0070] e. Integrate the filtered global first frequency domain features / second frequency domain features and local first frequency domain features / second frequency domain features, and obtain the first frequency domain enhanced features / second frequency domain enhanced features through inverse fast Fourier transform.

[0071] In one feasible approach, the temporal long-range dependency of the first local feature / second local feature is captured, specifically including: The first local feature / second local feature is input into a long short-term memory network, and the gating mechanism of the long short-term memory network is used to capture the long-range dependency relationship between the first local feature / second local feature in the time dimension.

[0072] In detail, for the sake of brevity, the first local feature / second local feature will be referred to as local features below. (Combined with...) Figure 6 and Figure 7 As shown, each branch of the dual-branch feature fusion neural network model uses the ResDualFT Encoder module for feature extraction. Figure 7 As shown, the ResDualFT Encoder module first uses the ResNet residual network to extract local features, and then adopts a dual-path enhancement strategy: one path is through adaptive spectral enhancement, which adaptively enhances key frequency components and suppresses noise in the frequency domain; the other path is through a long short-term memory network (LSTM network) to capture long-range temporal dependencies, and finally obtains dual-path enhanced features.

[0073] Specifically, each branch uses a ResNet residual network to extract local features, and the calculation formula for the core residual block is as follows:

[0074] In the formula, The input features for the residual block are... For output features, For residual mapping, This refers to network weights.

[0075] Specifically, a dual-path enhancement strategy is adopted, which is implemented as follows: The first enhancement method involves adaptive spectral enhancement of local features, which adaptively enhances key frequency components and suppresses noise in the frequency domain. The specific steps are as follows: The first step is to perform a Fast Fourier Transform on the time dimension of the local features to obtain the frequency domain features. The calculation formula is as follows:

[0076] In the formula, For frequency domain characteristics, For local features, For Fast Fourier Transform, This indicates a transformation along the time axis.

[0077] The second step is to calculate the power spectrum of the frequency domain characteristics. Through learnable thresholds Generate binary mask The calculation formula is as follows:

[0078] In the formula, The network is trained to adaptively learn and is used to filter out low-power noise high-frequency components.

[0079] The third step involves applying globally learnable filters and locally learnable filters to the denoised frequency domain features, respectively. The filtered global frequency domain features are obtained. and local frequency domain features The calculation formula is as follows:

[0080]

[0081] In the formula, For global filters, It is a local filter.

[0082] The fourth step is to integrate the two filtering results and transform them back to the time domain using an inverse fast Fourier transform to obtain the frequency domain enhancement features. The calculation formula is as follows:

[0083] In the formula, For frequency domain enhancement features, For inverse fast Fourier transform, This indicates a transformation along the time axis.

[0084] The second path enhancement involves inputting the same local features in parallel into a long short-term memory network, utilizing its gating mechanism to capture long-distance dependencies of signals over time. The calculation formula is as follows:

[0085]

[0086] in, For the Gate of Oblivion For input gate, In cellular state, For output gate, For LSTM Step input, , These represent the hidden state and the cell state from the previous step, respectively. , For network weights and biases, The Sigmoid activation function is used. The hyperbolic tangent activation function is used. This is element-wise multiplication.

[0087] In one possible implementation, the spliced ​​features are weighted by multi-scale contextual information extraction and a hybrid attention mechanism to obtain fused features, specifically: Multi-scale context extraction step: The spliced ​​features are processed using multiple parallel dilated convolution branches, where the dilated convolution of each branch has a different dilation rate to synchronously capture context information at different time scales, and the multi-scale features output by each branch are fused to obtain multi-scale fused features. Hybrid attention weighting step: The multi-scale fusion feature is input into the hybrid attention module, which simultaneously learns the importance weights in the feature channel dimension and the time series dimension, and performs dual-dimensional enhancement on the multi-scale fusion feature, outputting the fusion feature; The hybrid attention module achieves dual-dimensional enhancement in the following way: Based on the multi-scale fusion features, a channel attention weight vector is generated; Based on the multi-scale fusion features, a pixel attention weight map is generated; The multi-scale fusion features are weighted along the channel dimension using the channel attention weight vector, and then weighted along the temporal dimension using the pixel attention weight map to obtain the final enhanced fusion features.

[0088] In detail, combined Figure 6 and Figure 8 As shown, the concatenated features are input into the MSD-CPA Decoder module for fusion. The MSD-CPA Decoder module includes multi-scale dilated convolution and a hybrid attention mechanism to obtain the fused features, as detailed below: Multiple parallel dilated convolutional branches are used to process the concatenated features, with different dilation rates for the convolutional kernels. Taking values ​​of [1, 3, 5] respectively, the calculation formula is as follows:

[0089] In the formula, expansion rate Corresponding multi-scale features, For splicing features, This represents the kernel size.

[0090] This design expands the receptive field as the expansion rate increases without increasing the number of parameters. The calculation formula is as follows:

[0091] In the formula, This forms the previous receptive field, thereby simultaneously capturing contextual information at different time scales.

[0092] After fusing multi-scale features, the input is fed into the hybrid attention module, specifically implemented as follows: The channel attention submodule performs global average pooling (GAP) on the multi-scale fused features, and then generates channel weights through a small multilayer perceptron (MLP). The calculation formula is as follows:

[0093] In the formula, ( (time step) ( (Number of channels), used to enhance feature channels related to wear conditions.

[0094] The pixel attention submodule performs average pooling on the multi-scale fused features along the channel dimension, and then generates pixel (time step) weights through convolutional layers. The calculation formula is as follows:

[0095] In the formula, It is used to focus on key time points in a signal sequence.

[0096] Dual-dimensional enhancement applies channel weights and pixel weights separately to the multi-scale fusion feature, resulting in the final enhanced fusion feature. The calculation formula is as follows:

[0097] in, Weighting by channel dimension, Weighted by the pixel dimension.

[0098] In one possible implementation, based on the fused features, a predicted value of the tool wear state is output, specifically as follows: After the fused features are subjected to adaptive average pooling, they are input into an interpretable Kolmogorov-Arnold network (KAN) to output a predicted value of the tool wear state.

[0099] Specifically, the fused features are processed by adaptive average pooling and then input into an interpretable Kolmogorov-Arnold network to achieve accurate prediction of tool wear conditions. The specific implementation is as follows: The enhanced fused features are subjected to average pooling along the time dimension, and converted into a fixed-length one-dimensional feature vector. The calculation formula is as follows:

[0100] In the formula, This achieves both data dimensionality reduction and makes the model robust to changes in the length of the input signal.

[0101] By inputting a one-dimensional feature vector into an interpretable Kolmogorov-Arnold network, tool wear condition prediction can be achieved. The formula for the KAN network is as follows:

[0102] In the formula, This is an approximation function for the weighted sample data. For the KAN network The number of nodes in the layer To connect the KAN network The first layer The first neuron and the first The first layer Activation function of a neuron For the KAN network The number of nodes in the layer This represents the number of nodes in the second layer of the KAN network. To connect the second layer of the KAN network The first neuron and the third layer of the third layer Activation function of a neuron This represents the number of nodes in the first layer of the KAN network. To connect the first layer of the KAN network The first neuron and the second layer of the second layer Activation function of a neuron This represents the number of nodes in layer 0 of the KAN network. To connect the 0th layer of the KAN network The first neuron and the first layer of layer 1 Activation function of a neuron For the weighted sample data of the first Each component.

[0103] Example: like Figure 2The diagram shows a schematic of the tool wear monitoring signal acquisition system in this embodiment, mainly including a worktable 1, a force gauge 2, a workpiece 3, an accelerometer 4, a cutting tool 5, a microscope 6, a preamplifier 7, a data acquisition card 8, and an industrial computer 9. The force gauge 2 is a Kistler 9265B three-axis force gauge, and the data acquisition card 8 is an NI 9205 data acquisition card; together, they form the cutting force signal acquisition system. The accelerometer 4 is a PCB 356A44 miniature three-axis accelerometer mounted on the back of the workpiece, which, together with the NI 9234 data acquisition card, forms the vibration signal acquisition system. During milling, the two signal acquisition systems synchronously acquire cutting force signals and vibration signals in real time.

[0104] In this embodiment, the workpiece is a Ti6Al4V additive part with dimensions of 150 mm × 100 mm × 50 mm. The cutting tool used is a 20 mm diameter four-flute end mill with AlCrSiN coated inserts. A total of three tools were used, each milling 100 times. The experiment was conducted under side milling conditions, with a spindle speed of 8000 rpm, a feed rate of 1280 mm / min, a radial depth of cut of 0.2 mm, and an axial depth of cut of 4 mm. A special fixture was used to ensure a tight fit between the accelerometer and the workpiece. The sampling frequency of all signal channels was set to 5000 Hz. During the experiment, machining was paused after each cut, and the tool flank wear (VB) was measured offline using a Zhiqi ZQ-601 electron microscope. The average wear value was taken as the true tool wear value for that cutting cycle.

[0105] Step 1: During the milling process, cutting force and vibration signals in the X, Y, and Z directions are simultaneously acquired. The interquartile range method is used to truncate invalid data from the original signals, effectively eliminating signals from non-cutting periods such as tool feed and retraction. Furthermore, the Hampel filtering algorithm (sliding window length) is employed. =5000, threshold coefficient =3) Outlier detection and replacement are performed on the truncated signal to effectively suppress abrupt anomalies caused by electromagnetic interference, instantaneous impacts, etc., during signal acquisition, resulting in a smooth and effective cutting signal, such as... Figure 3a and Figure 3b As shown.

[0106] Step 2: Perform time-series segmentation on the effective cutting signal, dividing it into segments with a fixed time window of 200 ms (i.e., 1000 sampling points). Extract 11-dimensional time-domain features (such as mean, variance, kurtosis, margin, etc.) from each data segment, and standardize all features to eliminate the influence of dimensions, forming a high-dimensional initial feature set. Calculate the correlation between the features of each channel and the tool wear based on the Spearman correlation coefficient, and select the three features with the highest correlation from each channel signal of each type of sensor to construct a highly correlated feature set (standardized tool state feature vector). Figure 4a The figure shown is a graph illustrating the absolute value of the Spearman correlation coefficient between the characteristics and wear values ​​of each channel as a function of the characteristics. Figure 4b A heatmap showing the correlation between the characteristics of each channel and the wear value.

[0107] Step 3: Cluster the highly correlated feature set based on the HDBSCAN algorithm, set the minimum number of samples to 5, identify the features reflecting different cutting states, select representative samples from each class and splice them together to form a dimensionality-reduced signal, that is, to form the typical feature set of cutting force and the typical feature set of vibration. Figure 5a This is a two-dimensional t-SNE distribution diagram of the cutting signal after clustering in the X direction. Figure 5b This is a comparison diagram of the cutting signal before and after dimensionality reduction in the X direction.

[0108] Step 4: Input the typical feature sets of cutting force and vibration (the dimensionality-reduced cutting force signal and vibration signal) into the two branches of the dual-branch feature fusion neural network model, respectively. Figure 6 This is a schematic diagram of the dual-branch feature fusion neural network model structure, which processes mechanical signals and vibration signals respectively. Each branch is first processed by the ResDualFT Encoder module, where the ResNet structure extracts local spatiotemporal features. Then, the dual-enhancement branch one path enhances the frequency band related to tool wear and suppresses noise through the adaptive spectrum enhancement module, and the other path captures the temporal dependence through LSTM. Figure 7 This is a schematic diagram of the ResDualFTEncoder module structure.

[0109] Step 5: Input the enhanced features from the dual-branch output into the MSD-CPA Decoder module. The MSD-CPA Decoder module uses multi-scale dilated convolution to extract spatial features under different receptive fields, and then performs weighted fusion of the features through a hybrid attention mechanism. Figure 8 This is a schematic diagram of the MSD-CPA Decoder module structure.

[0110] Step 6: After compressing the fused features into a one-dimensional vector using adaptive average pooling, input it into the prediction module built based on the interpretable Kolmogorov-Arnold network, and output the predicted value of tool wear. Figure 9The chart shows a comparison of the RMSE of the proposed method with other models. The average root mean square error of the proposed method is 0.18, which is the best result among all models.

[0111] As can be seen from the above embodiments, the present invention fully explores the sensitive features related to tool wear in cutting signals through multi-source signal collaborative analysis, adaptive clustering dimensionality reduction, dual-path feature enhancement and deep fusion, and achieves high-precision prediction of tool wear state by combining interpretable KAN network.

[0112] In another embodiment of the present invention, a smart tool wear monitoring device is provided, comprising: The acquisition module is used to acquire typical feature sets of cutting forces and vibrations during the tool's operation.

[0113] The wear condition prediction module is used to input the typical feature set of cutting force and the typical feature set of vibration into a pre-trained dual-branch feature fusion neural network model; wherein, the dual-branch feature fusion neural network model is configured as follows: The first branch processes the typical feature set of cutting force to extract the first local feature, and performs frequency domain information enhancement and time domain long-range dependency capture on the first local feature in parallel to output the first fused enhanced feature; The vibration typical feature set is processed by the second branch to extract the second local feature, and the second local feature is enhanced in the frequency domain and captured in the time domain long-range dependency in parallel to output the second fused enhanced feature; The first fusion enhancement feature and the second fusion enhancement feature are concatenated to obtain the concatenated feature; The spliced ​​features are weighted by multi-scale contextual information extraction and a hybrid attention mechanism to obtain fused features; Based on the fusion characteristics, a predicted value of the tool wear condition is output.

[0114] All relevant content regarding the steps involved in the aforementioned embodiment of the intelligent tool wear monitoring method can be referenced from the functional description of the corresponding functional module of the intelligent tool wear monitoring device in this embodiment of the invention, and will not be repeated here. The module division in this embodiment of the invention is illustrative and only represents a logical functional division. In actual implementation, there may be other division methods. Furthermore, the functional modules in the various embodiments of the invention can be integrated into a processor, exist as separate physical entities, or two or more modules can be integrated into one module. The integrated module can be implemented in hardware or as a software functional module.

[0115] In another embodiment of the present invention, a computer device is provided, comprising a processor and a memory. The memory stores a computer program, which includes program instructions. The processor executes the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions in the computer storage medium to achieve a corresponding method flow or corresponding function. The processor described in this embodiment of the present invention can be used in the operation of a tool wear intelligent monitoring method.

[0116] In another embodiment of the present invention, a storage medium is provided, specifically a computer-readable storage medium (Memory), which is a memory device in a computer device used to store programs and data. It is understood that the computer-readable storage medium here can include both the built-in storage medium in the computer device and extended storage media supported by the computer device. The computer-readable storage medium provides storage space that stores the terminal's operating system. Furthermore, the storage space also stores one or more instructions suitable for loading and execution by a processor. These instructions can be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here can be Random Access Memory (RAM) or non-volatile memory, such as at least one disk storage device. The processor can load and execute one or more instructions stored in the computer-readable storage medium to implement the corresponding steps of the intelligent tool wear monitoring method in the above embodiments.

[0117] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, optical storage, etc.) containing computer-usable program code.

[0118] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure One One or more processes and / or boxes Figure One A device that provides the functions specified in one or more boxes.

[0119] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure One One or more processes and / or boxes Figure One The function specified in one or more boxes.

[0120] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure One One or more processes and / or boxes Figure One Figure One The steps of the function specified in one or more boxes.

[0121] This invention also provides a computer program product, which is used to execute any of the above-described intelligent tool wear monitoring methods. Since the computer program product provided by this invention belongs to the same inventive concept as the intelligent tool wear monitoring method described above, it possesses all the advantages of the aforementioned intelligent tool wear monitoring method. Therefore, the beneficial effects of the computer program product provided by this invention will not be elaborated upon here.

[0122] In this invention, the terms "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to a specific feature, structure, material, or characteristic described in connection with that embodiment or example, which is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0123] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit them. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the scope of the technology disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention.

Claims

1. A method for intelligent monitoring of tool wear, characterized in that, include: Obtain typical characteristic sets of cutting force and vibration during the tool's operation; The typical feature set of cutting force and the typical feature set of vibration are input into a pre-trained dual-branch feature fusion neural network model; wherein, the dual-branch feature fusion neural network model is configured as follows: The first branch processes the typical feature set of cutting force to extract the first local feature, and performs frequency domain information enhancement and time domain long-range dependency capture on the first local feature in parallel to output the first fused enhanced feature; The vibration typical feature set is processed by the second branch to extract the second local feature, and the second local feature is enhanced in the frequency domain and captured in the time domain long-range dependency in parallel to output the second fused enhanced feature; The first fusion enhancement feature and the second fusion enhancement feature are concatenated to obtain the concatenated feature; The spliced ​​features are weighted by multi-scale contextual information extraction and a hybrid attention mechanism to obtain fused features; Based on the fusion characteristics, a predicted value of the tool wear condition is output.

2. The intelligent tool wear monitoring method according to claim 1, characterized in that, The acquisition of typical cutting force characteristic sets and typical vibration characteristic sets during the tool's working process includes: Real-time acquisition of cutting force and vibration signals during tool operation; The cutting force signal and vibration signal are preprocessed to remove invalid data segments and smooth outliers, thereby extracting the effective cutting force signal and effective vibration signal that characterize the stable cutting process. Feature extraction is performed on the effective cutting force signal and the effective vibration signal respectively to obtain the initial feature set of cutting force and the initial feature set of vibration. Based on the correlation between each feature and the tool wear state, highly correlated features are selected from the initial feature set of cutting force and the initial feature set of vibration, respectively, and then standardized to construct the standard feature vector set of cutting force and the standard feature vector set of vibration. Based on the cutting state reflected by the standard feature vector set of cutting force and the standard feature vector set of vibration, the data of the two are divided respectively, and typical feature data representing different cutting states are selected to form a typical feature set of cutting force and a typical feature set of vibration.

3. The intelligent tool wear monitoring method according to claim 1, characterized in that, Frequency domain information enhancement is performed on the first local feature / second local feature, including: Perform a Fast Fourier Transform on the time dimension of the first local feature / second local feature to obtain the first frequency domain feature / second frequency domain feature; Calculate the power spectrum of the first frequency domain feature / second frequency domain feature, and based on the power spectrum of the first frequency domain feature / second frequency domain feature, generate a first binary mask / second binary mask through a learnable threshold; Based on the first binary mask / second binary mask, the first frequency domain feature / second frequency domain feature is denoised to obtain the denoised first frequency domain feature / second frequency domain feature. The first frequency domain features / second frequency domain features after denoising are filtered using a globally learnable filter and a locally learnable filter to obtain the filtered global first frequency domain features / second frequency domain features and local first frequency domain features / second frequency domain features. The filtered global first frequency domain features / second frequency domain features and local first frequency domain features / second frequency domain features are integrated and then subjected to inverse fast Fourier transform to obtain the first frequency domain enhanced features / second frequency domain enhanced features.

4. The intelligent tool wear monitoring method according to claim 1, characterized in that, Capturing long-range temporal dependencies of the first local features / second local features, including: The first local feature / second local feature is input into a long short-term memory network, and the gating mechanism of the long short-term memory network is used to capture the long-range dependency relationship between the first local feature / second local feature in the time dimension.

5. The intelligent tool wear monitoring method according to claim 1, characterized in that, The step of extracting multi-scale contextual information and weighting the spliced ​​features using a hybrid attention mechanism to obtain fused features includes: Multi-scale context extraction step: The spliced ​​features are processed using multiple parallel dilated convolution branches, where the dilated convolution of each branch has a different dilation rate to synchronously capture context information at different time scales, and the multi-scale features output by each branch are fused to obtain multi-scale fused features. Hybrid attention weighting step: The multi-scale fusion feature is input into the hybrid attention module, which simultaneously learns the importance weights in the feature channel dimension and the time series dimension, and performs dual-dimensional enhancement on the multi-scale fusion feature, outputting the fusion feature; The hybrid attention module achieves dual-dimensional enhancement in the following way: Based on the multi-scale fusion features, a channel attention weight vector is generated; Based on the multi-scale fusion features, a pixel attention weight map is generated; The multi-scale fusion features are weighted along the channel dimension using the channel attention weight vector, and then weighted along the temporal dimension using the pixel attention weight map to obtain the final enhanced fusion features.

6. The intelligent tool wear monitoring method according to claim 1, characterized in that, The step of outputting a predicted value of the tool wear state based on the fused features is specifically as follows: After the fused features are subjected to adaptive average pooling, they are input into an interpretable Kolmogorov-Arnold network to output a predicted value of the tool wear state.

7. A smart tool wear monitoring device, characterized in that, include: The acquisition module is used to acquire typical feature sets of cutting forces and vibrations during the tool's operation. The wear condition prediction module is used to input the typical feature set of cutting force and the typical feature set of vibration into a pre-trained dual-branch feature fusion neural network model; wherein, the dual-branch feature fusion neural network model is configured as follows: The first branch processes the typical feature set of cutting force to extract the first local feature, and performs frequency domain information enhancement and time domain long-range dependency capture on the first local feature in parallel to output the first fused enhanced feature; The vibration typical feature set is processed by the second branch to extract the second local feature, and the second local feature is enhanced in the frequency domain and captured in the time domain long-range dependency in parallel to output the second fused enhanced feature; The first fusion enhancement feature and the second fusion enhancement feature are concatenated to obtain the concatenated feature; The spliced ​​features are weighted by multi-scale contextual information extraction and a hybrid attention mechanism to obtain fused features; Based on the fusion characteristics, a predicted value of the tool wear condition is output.

8. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements a tool wear intelligent monitoring method as described in any one of claims 1 to 6.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements a tool wear intelligent monitoring method as described in any one of claims 1 to 6.

10. A computer program product, characterized in that, When the computer program is executed by the processor, it implements a tool wear intelligent monitoring method as described in any one of claims 1 to 6.

Images