A machine vision-based numerical control machine tool tool wear identification method and system

Abstract

The application discloses a kind of based on machine vision's numerical control machine tool tool wear identification method and system, it is related to machine vision technical field.Sequence of two-dimensional image of numerical control machine tool tool is collected, high-precision point cloud is obtained after three-dimensional reconstruction and calibration, and high-precision tool reference three-dimensional point cloud is obtained;Function area division rule is constructed, and tool function area is segmented according to high-precision tool reference three-dimensional point cloud, and tool area structured reference data is generated after geometric morphological quantitative description algorithm analysis;Area structured data of different processing periods of tool is collected, and tool appearance evolution trajectory is obtained after space-time registration;Tool wear identification model is constructed based on time series convolution network, tool appearance evolution trajectory is input, and tool wear type is output;Tool historical case data is collected to construct wear mechanism knowledge base, tool wear type is matched and mapped with knowledge base, and tool wear identification result is generated.Based on the wear identification result, the identification of numerical control machine tool tool wear is realized.

Description

Technical Field

[0001] This invention relates to the field of machine vision technology, specifically to a method and system for identifying wear of CNC machine tool cutting tools based on machine vision. Background Technology

[0002] As CNC machining technology continues to advance towards higher speeds and precision, the condition of cutting tools, as key components directly involved in cutting, directly determines machining quality and production efficiency. During the cutting process, cutting tools gradually wear down due to mechanical friction, high temperature, and high pressure. This wear manifests in various forms, including edge dulling, flank wear, micro-crack propagation, and even chipping. Failure to accurately and promptly monitor tool wear can lead not only to workpiece dimensional deviations and surface quality deterioration but also to sudden tool failure, causing machining interruptions or even equipment damage. Therefore, achieving real-time, accurate monitoring and intelligent identification of tool wear is a crucial technological step in ensuring the safe, stable, and efficient operation of automated machining systems.

[0003] Currently, the most commonly used method for CNC machine tool wear detection in industry is a non-contact detection method based on two-dimensional vision. This method acquires two-dimensional images of the tool surface using an industrial camera, analyzes the images using image processing algorithms, and indirectly assesses the tool's wear state by extracting features such as brightness, contrast, and texture, or by extracting the tool's contour geometric features using edge detection algorithms. However, the above method can only obtain two-dimensional projection information of the tool from a single viewpoint and cannot reconstruct the true three-dimensional geometric shape of the tool surface. Therefore, it is difficult to accurately quantify and assess key three-dimensional wear features such as changes in cutting edge curvature. Consequently, the detection results are easily affected by factors such as lighting conditions, tool posture, and surface oil contamination, and the recognition accuracy is insufficient to meet the needs of CNC machine tool manufacturing scenarios. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a CNC machine tool tool wear identification method and system based on machine vision, which solves the problem that the true three-dimensional geometric morphology of the tool surface cannot be restored, making it difficult to accurately quantify and evaluate key three-dimensional wear characteristics such as changes in cutting edge curvature.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method and system for identifying CNC machine tool wear based on machine vision, comprising the following steps:

[0006] Step S1: Acquire a two-dimensional image sequence of the CNC machine tool tool, process the tool two-dimensional image sequence through a multi-view geometric three-dimensional reconstruction algorithm to obtain a tool reference three-dimensional point cloud, and calibrate the tool reference three-dimensional point cloud through an iterative nearest point ICP algorithm to obtain a high-precision tool reference three-dimensional point cloud;

[0007] Step S2: Construct functional region division rules, segment the high-precision tool reference 3D point cloud according to the functional region division rules to obtain tool functional regions; analyze the tool functional regions through geometric morphology quantization description algorithm to generate tool region structured reference data;

[0008] Step S3: Collect structured data of the tool area after the tool has undergone different machining cycles to obtain structured time-series data of the tool area; based on the structured reference data of the tool area, align and stitch the structured time-series data of the tool area using a spatiotemporal registration algorithm to generate the tool morphology evolution trajectory;

[0009] Step S4: Based on a temporal convolutional network, construct a tool wear recognition model, input the tool morphology evolution trajectory into the tool wear recognition model, and output the tool wear type;

[0010] Step S5: Collect historical case data of cutting tools and build a wear mechanism knowledge base; match and map the tool wear type with the wear mechanism knowledge base to generate tool wear identification results.

[0011] Preferably, a sequence of two-dimensional images of the CNC machine tool is acquired, and the sequence is processed using a multi-view geometric 3D reconstruction algorithm to obtain a reference 3D point cloud of the tool, including:

[0012] By using a stereo vision imaging unit fixed in the working cavity of a CNC machine tool, when the machine tool spindle is stationary or rotating at low speed, the spindle is controlled to rotate in preset angle steps. At each controlled angle position, the left and right cameras in the stereo vision imaging unit are synchronously triggered to acquire a sequence of two-dimensional images of the tool covering the key parts of the tool cutting edge.

[0013] The process of processing the two-dimensional image sequence of the cutting tool using a multi-view geometric three-dimensional reconstruction algorithm is as follows:

[0014] The core of multi-view geometric 3D reconstruction algorithms for processing image sequences is the application of triangulation principles to inversely calculate a 3D point cloud from a sequence of 2D tool images. Consider images taken from two different viewpoints, with projection matrices of... and For a feature point on the left image Based on its matching point on the right image The following equation can be established:

[0015]

[0016] in, The three-dimensional homogeneous coordinates of the point to be determined are: That is, its three-dimensional spatial coordinates; These are the projection matrices corresponding to the left and right cameras, respectively. Indicates scale equivalence relations;

[0017] This equation can be transformed into a linear equation, and the spatial points can be obtained by solving a least-squares problem. 3D coordinates Perform this operation on all successfully matched feature points to generate a large number of discrete feature sets. The initial three-dimensional point cloud composed of three-dimensional points of coordinates is denoted as the tool reference three-dimensional point cloud. .

[0018] Preferably, obtaining a high-precision tool reference 3D point cloud by using the Iterative Closest Point (ICP) algorithm on the tool reference 3D point cloud includes:

[0019] The Iterative Closest Point (ICP) algorithm is used for accurate determination. The goal of the ICP algorithm is to find an optimal rigid body transformation that results in a point cloud that... With reference point cloud To minimize the overall distance, the problem can be formulated as minimizing the following objective function:

[0020]

[0021] in, It is the objective function, which is the sum of squared Euclidean distances between all corresponding pairs of points; These are the rigid body transformation parameters to be determined. Indicates the starting index for the summation, i.e., from... =1 to start the summation; N represents the endpoint index of the summation, i.e., the total number of points in the point cloud. Indicates the first point in the registration point cloud One point, Represents the first point in the target point cloud One point;

[0022] The Iterative Closest Point (ICP) algorithm solves for the optimal solution through two steps: iterative data association and motion estimation. and ;

[0023] First, perform data association for each point in Q. Find the nearest neighbor in P and establish a correspondence; then perform motion estimation, and based on the current correspondence, solve for the optimal R and t using the SVD decomposition method; iterate until the error is reached. It converges to the minimum value;

[0024] The optimal transformation (R,t) obtained is used to perform a spatial transformation on the tool reference 3D point cloud, resulting in a high-precision point cloud that is precisely aligned with the reference coordinate system. This high-precision tool reference 3D point cloud is denoted as the high-precision tool reference 3D point cloud. .

[0025] Preferably, a functional region division rule is constructed, and the high-precision tool reference 3D point cloud is segmented according to the functional region division rule to obtain the tool functional region including:

[0026] To achieve refined analysis of tool damage, functional area subdivision rules are first constructed based on tool failure mechanics and cutting mechanics mechanisms. These rules are defined based on the mechanical stress and thermal load experienced by different tool faces during the cutting process and the resulting typical damage modes. The specific functional area division rules are as follows:

[0027] The main cutting area is the primary cutting region involved in material removal; the flank face area is the region that contacts the surface to be machined; the rake face area is the surface through which the cutting flow passes; the tool tip transition area is the transition region of the main cutting edge; and the auxiliary functional areas vary depending on the tool type.

[0028] Based on the above functional area division rules, the high-precision tool reference three-dimensional point cloud obtained by S1 is processed. , The segmentation process is achieved through the following steps:

[0029] First, feature calculation is performed. By analyzing the spatial distribution characteristics of each point in the point cloud, its normal vector and curvature features are calculated.

[0030] The normal vector calculation uses a local plane fitting method, first calculating the covariance matrix of the local neighborhood:

[0031]

[0032] in, The calculated 3x3 covariance matrix describes the distribution direction and dispersion of points within the local neighborhood. This represents the total number of neighboring points of point k; Let i represent the i-th neighboring point of point k. represents the centroid of the local neighborhood of point k; T represents the transpose of a vector; It's an outer product, and the result is a 3x3 matrix that describes the 6th position of point k. Neighboring points Relative to the local neighborhood centroid The distribution relationship of displacement in various directions;

[0033] Next, the covariance matrix Eigenvalue decomposition yields three eigenvalues. and the corresponding feature vectors ,point normal vector It is the smallest eigenvalue corresponding feature vector ;

[0034] Curvature calculation is based on the eigenvalue decomposition of the local covariance matrix:

[0035]

[0036] in, This represents the curvature estimate of the local neighborhood of the k-th point in the point cloud. These are the three eigenvalues ​​obtained after eigenvalue decomposition of the covariance matrix formed by the local neighborhood point set, which satisfy... ;

[0037] Based on the obtained point cloud features, a regional growth process is carried out, with the highest point of the spike being used as the starting point. As the starting point for growth, neighboring points are gradually gathered according to three criteria: consistency of normal vector, continuity of curvature, and spatial proximity, to form the initial region;

[0038] It should be noted that normal vector consistency requires the angle between the normal vectors of adjacent points to be less than 15 degrees to ensure surface smoothness; curvature continuity controls the curvature difference between adjacent points to be within 0.05 to ensure the coherence of geometric features; spatial proximity ensures the connectivity of the region through Euclidean distance constraints.

[0039] Finally, a deep learning-based segmentation network is used to semantically annotate the initial region, ultimately dividing the point cloud into disjoint subsets to determine the tool's functional regions.

[0040]

[0041] in, , , , , These are point cloud subsets corresponding to the main cutting area, the back cutting face area, the front cutting face area, the cutting tip arc area, and the auxiliary function area, respectively.

[0042] Preferably, the analysis of the tool functional area using a geometric morphology quantization description algorithm to generate structured benchmark data for the tool area includes:

[0043] To transform geometric information into computable wear evaluation indicators, it is necessary to process a subset of the point cloud for each region. The analysis is performed using a geometric morphology quantification algorithm, as follows:

[0044] Macroscopic geometric parameters are extracted to obtain the variance of the surface normal vector direction distribution for each region. Contour curvature radius The radius of curvature is the reciprocal of the curvature. = ; and regional macroeconomic characteristic indicators ;

[0045] It should be noted that the variance of the surface normal vector direction distribution The calculation formula is:

[0046]

[0047] in, It is the variance of the surface normal vector direction distribution, used to quantify the degree of dispersion of the surface normal vector direction within the tool area; It is the surface normal vector of the k-th point cloud; n refers to the surface normal vector; k is the index of a point in the point cloud; This represents the total number of point clouds within the current tool area; It is the mean vector of the normal vectors of all points within the region; It is the Euclidean norm of a vector;

[0048] For micro-morphological statistics, the skewness of the point cloud elevation distribution is calculated for wear-sensitive areas. and kurtosis Skewness characterizes the asymmetry of wear, while kurtosis reflects the degree of wear concentration.

[0049] The formula for calculating skewness is:

[0050]

[0051] in, This represents the calculated skewness value, quantifying the asymmetry of the worn surface; Indicates the first Elevation values ​​of each data point; This represents the arithmetic mean of the elevation values ​​of all data points.

[0052] The formula for calculating kurtosis is:

[0053]

[0054] in, The calculated kurtosis value quantifies the degree of concentration of the morphological distribution on the worn surface. The standard deviation of the elevation values ​​for all data points measures the dispersion of elevation values, i.e., the overall level of surface roughness.

[0055] Organize the above parameters into a feature vector And generate reusable tool area structured reference data. :

[0056]

[0057] Among them, edge is used to identify the main cutting edge area, highlighting its core position as the cutting edge; flank is used to identify the flank face area, which is derived from the traditional terminology in mechanical engineering that specifically refers to the flank face of a tool; face is used to identify the front face area, which is in line with the convention of representing the main working face; corner is used to identify the tip arc area, accurately describing the geometric characteristics of the tip angle; aux is used to identify the auxiliary functional area, reflecting its auxiliary function positioning; the finally generated feature vector and the regional point cloud together constitute the tool region structured reference data.

[0058] Preferably, the structured data of the tool region after the tool has undergone different machining cycles is collected to obtain the structured time-series data of the tool region, including:

[0059] After the tool has undergone different machining cycles, steps S1 and S2 are repeated to collect the structured data of the tool area at the corresponding time. in ,and As the initial reference time, each A subset of the point cloud containing each functional area at that moment and its corresponding structured description fields .

[0060] Preferably, based on the structured reference data of the tool region, the structured time-series data of the tool region is aligned and stitched together using a spatiotemporal registration algorithm to generate the tool morphology evolution trajectory, including:

[0061] For any time Tool area structured data The iterative nearest point algorithm is used to solve for its relationship with the reference time. data Optimal rigid body transformation between This makes the transformed topographic anchor point group With reference topography anchor point group Minimize alignment error:

[0062]

[0063] in, Indicates time The A shape anchor point, Given the total number of anchor points, apply the obtained transformation parameters ( , Spatial transformation is performed on the subsets of the regional point cloud at each time point to obtain the registered point cloud. Simultaneously, the structured description fields are recalculated based on the registered point cloud. ;

[0064] Organize all registered data in chronological order to construct the tool morphology evolution trajectory. :

[0065]

[0066] in, , These are the times after registration. Functional area Point cloud and feature fields.

[0067] Preferably, a tool wear recognition model is constructed based on a temporal convolutional network. The tool morphology evolution trajectory is input into the tool wear recognition model, and the output tool wear types include:

[0068] A tool wear recognition model is constructed based on a temporal convolutional network. The construction process of this model is as follows:

[0069] Input the tool morphology evolution trajectory generated in step S3 The time-series feature matrix is ​​extracted from it. The extracted feature matrix is ​​as follows:

[0070]

[0071] in, Indicates time The feature vector, its dimension Corresponding to the geometric parameters and statistical characteristics extracted from the structured reference data T of the tool region;

[0072] We collected evolution trajectory data of various cutting tools under different processing conditions as a training set to ensure a balanced distribution of samples of stable wear and abnormal wear, so as to avoid model bias.

[0073] The TCN network architecture is constructed by stacking extended causal convolutional layers, and its core computation is as follows:

[0074]

[0075] in, As the expansion factor, The kernel size is used to ensure that the receptive field grows exponentially with the network depth, thereby capturing long-term dependencies in the evolutionary trajectory. s represents the time step index of the output sequence. The hierarchical output of TCN is passed to the output layer after residual connection and regularization.

[0076] A fully connected layer is used to map the temporal features extracted by TCN to the class space, and a probability distribution is generated through the Softmax function:

[0077]

[0078] in, This is the hidden state of the last layer of TCN. and For output layer parameters, output This indicates the probability that the input trajectory belongs to each category;

[0079] The tool morphology evolution trajectory generated by S3 Input the trained tool wear recognition model and obtain the probability output. Set confidence threshold The decision-making rules are as follows:

[0080]

[0081] The final output is a tool wear type label. .

[0082] Preferably, the feature is that collecting historical tool case data and constructing a wear mechanism knowledge base includes:

[0083] Historical case data for cutting tools was collected in two different environments:

[0084] In a laboratory environment, accelerated wear tests were conducted using the controlled variable method to record the evolution of characteristic parameters under different historical working conditions, and to simultaneously record the remaining life data, the machining time from the initial state to complete failure of the tool, and the triggering factor data, so as to accurately record the threshold values ​​of key parameters that lead to abnormal wear.

[0085] In actual production environments, long-term monitoring is used to collect complete data trajectories of characteristic parameter changes, normal wear, and abnormal wear within different tool life cycles, as well as historical working conditions and manifestations of various abnormal failure modes.

[0086] Based on historical tool case data, a knowledge base for quantitative characteristic wear mechanisms is constructed:

[0087]

[0088] in, For feature vectors; Indicates the type of wear mechanism; This is the threshold condition vector for determining this mechanism; This represents typical remaining lifetime data under this mechanism model; This is the dataset of triggering factors for this mechanism.

[0089] This invention provides a machine vision-based method for identifying tool wear in CNC machine tools, involving machine learning and deep learning technologies, which has the following beneficial effects:

[0090] (1) This CNC machine tool wear identification method based on machine vision upgrades tool detection from traditional two-dimensional image analysis to three-dimensional point cloud model analysis through stereo vision reconstruction technology. This makes wear assessment no longer dependent on easily disturbed image features such as brightness and texture, but can directly and accurately quantify key three-dimensional geometric parameters such as cutting edge curvature, wear band width, and chipping volume, fundamentally overcoming the problem of inaccurate measurement caused by insufficient information dimensions in traditional methods.

[0091] (2) This CNC machine tool wear identification method based on machine vision can continuously track the initiation and development of wear by constructing the tool morphology evolution trajectory. This not only identifies the current wear state, but also reveals the trend of wear rate change, providing a solid data foundation for predicting the remaining tool life and realizing predictive maintenance, thereby effectively avoiding sudden failures.

[0092] (3) This machine vision-based CNC machine tool wear identification method can associate wear phenomena with specific physicochemical mechanisms by matching and mapping detection data with a wear mechanism knowledge base. This makes the diagnostic results interpretable and can directly guide maintenance personnel to take targeted process optimization or maintenance measures, rather than simply replacing the tool, thus improving the accuracy and efficiency of maintenance. Attached Figure Description

[0093] Figure 1 This is a flowchart of a CNC machine tool tool wear identification method based on machine vision proposed in this invention.

[0094] Figure 2 The present invention proposes a machine vision-based method for identifying CNC machine tool wear to obtain a hierarchical diagram of the tool's functional areas.

[0095] Figure 3 This is a hierarchy diagram of the tool wear identification results obtained in the CNC machine tool tool wear identification method based on machine vision proposed in this invention. Detailed Implementation

[0096] The technical solutions of the embodiments 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, and 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.

[0097] Please see Figure 1-3This invention provides a technical solution: a method and system for identifying CNC machine tool wear based on machine vision. Specifically, the method and system for identifying CNC machine tool wear based on machine vision are provided below. Figure 1 The method includes the following steps:

[0098] Step S1: Acquire a two-dimensional image sequence of the CNC machine tool cutting tool, process the two-dimensional image sequence of the cutting tool cutting tool through a multi-view geometric three-dimensional reconstruction algorithm to obtain a tool reference three-dimensional point cloud, and obtain a high-precision tool reference three-dimensional point cloud by applying the iterative nearest point ICP algorithm to the tool reference three-dimensional point cloud.

[0099] A stereo vision imaging unit, fixed within the working cavity of a CNC machine tool, controls the spindle to rotate in preset angular steps when the spindle is stationary or rotating at low speed. At each controlled angular position, the left and right cameras in the stereo vision imaging unit are synchronously triggered to acquire a sequence of two-dimensional images of the tool covering key areas such as the cutting edge. It should be noted that the imaging unit is equipped with an active uniform light source to suppress ambient light interference and enhance surface texture features, providing high-quality image input for subsequent 3D reconstruction.

[0100] The process of processing the two-dimensional image sequence of the cutting tool using a multi-view geometric three-dimensional reconstruction algorithm is as follows:

[0101] The core of multi-view geometric 3D reconstruction algorithms for processing image sequences is the application of triangulation principles to inversely calculate a 3D point cloud from a sequence of 2D tool images. Consider images taken from two different viewpoints, with projection matrices of... and For a feature point on the left image Based on its matching point on the right image The following equation can be established:

[0102]

[0103] in, The three-dimensional homogeneous coordinates of the point to be determined are: That is, its three-dimensional spatial coordinates; These are the projection matrices corresponding to the left and right cameras, respectively. Indicates scale equivalence relations;

[0104] This equation can be transformed into a linear equation, and the spatial points can be obtained by solving a least-squares problem. 3D coordinates Perform this operation on all successfully matched feature points to generate a large number of discrete feature sets. The initial three-dimensional point cloud composed of three-dimensional points of coordinates is denoted as the tool reference three-dimensional point cloud. .

[0105] To eliminate the pose deviation introduced by repeated tool clamping, the obtained tool reference 3D point cloud is... With the pre-stored zero-time reference point cloud Perform precise registration.

[0106] The Iterative Closest Point (ICP) algorithm is used for accurate determination. The goal of the ICP algorithm is to find an optimal rigid body transformation that results in a point cloud that... With reference point cloud To minimize the overall distance, the problem can be formulated as minimizing the following objective function:

[0107]

[0108] in, It is the objective function, which is the sum of squared Euclidean distances between all corresponding pairs of points; These are the rigid body transformation parameters to be determined. Indicates the starting index for the summation, i.e., from... =1 to start the summation; N represents the endpoint index of the summation, i.e., the total number of points in the point cloud. Indicates the first point in the registration point cloud One point, Represents the first point in the target point cloud One point;

[0109] The Iterative Closest Point (ICP) algorithm solves for the optimal solution through two steps: iterative data association and motion estimation. and ;

[0110] First, perform data association for each point in Q. Find the nearest neighbor in P and establish a correspondence; then perform motion estimation, and based on the current correspondence, solve for the optimal R and t using the SVD decomposition method; iterate until the error is reached. The solution converges to the minimum value. Finally, the optimal transformation (R,t) is used to perform a spatial transformation on the tool reference 3D point cloud, resulting in a high-precision point cloud that is precisely aligned with the reference coordinate system. This high-precision tool reference 3D point cloud is denoted as the high-precision tool reference 3D point cloud. It is also composed of precise three-dimensional points. constitute.

[0111] This step marks the data perception starting point of the method. By controlling the indexing rotation of the machine tool spindle, a sequence of two-dimensional images of the tool is acquired from multiple angles using a stereo vision system. Subsequently, based on the principle of multi-view geometric triangulation, the two-dimensional images are reconstructed into an initial three-dimensional point cloud. To eliminate position and orientation deviations caused by repeated tool clamping, the Iterative Closest Point (ICP) algorithm is employed to precisely register the measured initial point cloud with a pre-stored zero-time reference point cloud, ultimately generating a high-precision tool reference three-dimensional point cloud that can serve as a unique spatial reference.

[0112] Step S2: Construct functional region division rules, segment the high-precision tool reference 3D point cloud according to the functional region division rules to obtain tool functional regions; analyze the tool functional regions through geometric morphology quantization description algorithm to generate tool region structured reference data.

[0113] To achieve refined analysis of tool damage, functional area subdivision rules are first constructed based on tool failure mechanics and cutting mechanics mechanisms. These rules are defined based on the mechanical stress and thermal load experienced by different tool faces during the cutting process and the resulting typical damage modes. The specific functional area division rules are as follows:

[0114] The main cutting area is the primary cutting zone involved in material removal, and it bears the greatest shear force and temperature, making it prone to crescent-shaped wear and groove-edge wear.

[0115] The back face area is the region that contacts the surface to be machined, and it mainly experiences uniform wear or grooved wear.

[0116] The rake face area is the surface through which the cutting flow passes, and it is prone to crescent-shaped wear due to friction and abrasion.

[0117] The tool tip transition zone is the transition area of ​​the main cutting edge, which is subject to complex forces and is prone to chipping or plastic deformation.

[0118] The auxiliary function area is supplemented according to the type of tool, such as the chip groove of the milling cutter and the chisel edge of the drill bit.

[0119] Based on the above functional area division rules, the high-precision tool reference three-dimensional point cloud obtained by S1 is processed. , The segmentation process is achieved through the following steps:

[0120] First, feature calculation is performed. By analyzing the spatial distribution characteristics of each point in the point cloud, its normal vector and curvature features are calculated.

[0121] The normal vector calculation uses a local plane fitting method, first calculating the covariance matrix of the local neighborhood:

[0122]

[0123] in, The calculated 3x3 covariance matrix describes the distribution direction and dispersion of points within the local neighborhood. This represents the total number of neighboring points of point k; Let i represent the i-th neighboring point of point k. represents the centroid of the local neighborhood of point k; T represents the transpose of a vector; It's an outer product, and the result is a 3x3 matrix that describes the 6th position of point k. Neighboring points Relative to the local neighborhood centroid The distribution of displacement in various directions.

[0124] Next, the covariance matrix Eigenvalue decomposition yields three eigenvalues. (satisfy and the corresponding feature vectors ,point normal vector It is the smallest eigenvalue corresponding feature vector eigenvalues Reflects the point set in The direction with the smallest distribution variance means that the direction in which the point set is the "flattest" is the normal direction;

[0125] Curvature calculation is based on the eigenvalue decomposition of the local covariance matrix:

[0126]

[0127] in, This represents the curvature estimate of the local neighborhood of the k-th point in the point cloud. These are the three eigenvalues ​​obtained after eigenvalue decomposition of the covariance matrix formed by the local neighborhood point set, which satisfy... .

[0128] Based on the obtained point cloud features, a regional growth process is carried out, with the highest point of the spike being used as the starting point. As the starting point for growth, neighboring points are gradually gathered according to three criteria: consistency of normal vector, continuity of curvature, and spatial proximity, to form the initial region;

[0129] It should be noted that normal vector consistency requires the angle between the normal vectors of adjacent points to be less than 15 degrees to ensure surface smoothness; curvature continuity controls the curvature difference between adjacent points to be within 0.05 to ensure the continuity of geometric features; spatial proximity ensures the connectivity of the region through Euclidean distance constraints.

[0130] The basis for setting the normal vector consistency threshold angle to <15° is that this value (15°) is an empirical standard based on point cloud processing. In most 3D reconstruction and segmentation algorithms, 15 degrees is a commonly used threshold that can balance surface smoothness and detail preservation. The basis for setting the curvature continuity control of the curvature difference between adjacent points to within 0.05 is that if the threshold is too large, it will cause the merging of regions with different curvatures (such as flat back face and high curvature cutting edge); if it is too small, it may produce fragmentation due to measurement error. 0.05 is the optimization result based on prior knowledge of tool geometry.

[0131] Spatial proximity ensures the connectivity of regions through Euclidean distance constraints. Specifically, during region growing, a candidate point is selected only if its three-dimensional Euclidean distance to any existing point within the current region is less than or equal to a set threshold. Only when a certain condition is met can a region be included. Threshold The point cloud density is adaptively determined, and its calculation formula is as follows: ,in The average point spacing of the point cloud. This is a proportionality coefficient, ranging from 2.0 to 3.0. In a preferred embodiment of the present invention, it is taken as... This value has been experimentally verified to effectively ensure the integrity of the physical connectivity of the region while avoiding excessive growth.

[0132] Finally, the initial region is semantically annotated using the deep learning-based segmentation network PointNet++, ultimately dividing the point cloud into disjoint subsets to determine the tool's functional regions.

[0133]

[0134] in, , , , , These are point cloud subsets corresponding to the main cutting area, the back cutting face area, the front cutting face area, the cutting tip arc area, and the auxiliary function area, respectively.

[0135] To transform geometric information into computable wear evaluation indicators, it is necessary to process a subset of the point cloud for each region. The analysis is performed using a geometric morphology quantification algorithm, as follows:

[0136] Macroscopic geometric parameters are extracted to obtain the variance of the surface normal vector direction distribution for each region. Contour curvature radius The radius of curvature is the reciprocal of the curvature. = ; and regional macroeconomic characteristic indicators (in (These are the eigenvalues ​​of the covariance matrix).

[0137] It should be noted that the variance of the surface normal vector direction distribution The calculation formula is:

[0138]

[0139] in, It is the variance of the surface normal vector direction distribution, used to quantify the degree of dispersion of the surface normal vector direction within the tool area; n refers to the surface normal vector. is the surface normal vector of the k-th point cloud; k is the index of a point in the point cloud; This represents the total number of point clouds within the current tool area; It is the mean vector of the normal vectors of all points within the region; It is the Euclidean norm of the vector (i.e., the magnitude of the vector).

[0140] For micro-morphological statistics, the skewness of the point cloud elevation distribution is calculated for wear-sensitive areas (such as the main cutting edge region). and kurtosis Skewness characterizes the asymmetry of wear, while kurtosis reflects the degree of wear concentration.

[0141] The formula for calculating skewness is:

[0142]

[0143] in, This represents the calculated skewness value, quantifying the asymmetry of the worn surface; Indicates the first Elevation values ​​of each data point; This represents the arithmetic mean of the elevation values ​​of all data points.

[0144] The formula for calculating kurtosis is:

[0145]

[0146] in, The calculated kurtosis value quantifies the degree of concentration of the morphological distribution on the worn surface. The standard deviation of the elevation values ​​for all data points measures the dispersion of elevation values, i.e., the overall level of surface roughness.

[0147] Organize the above parameters into a feature vector And generate reusable tool area structured reference data. :

[0148]

[0149] In this system, "edge" identifies the main cutting edge area, highlighting its central role as the cutting edge; "flank" identifies the flank face area, a term derived from the mechanical engineering tradition of using "flankface" specifically to refer to the flank face of a cutting tool; "face" identifies the rake face area, consistent with the convention of characterizing the main working face; "corner" identifies the tip arc area, accurately describing the geometric characteristics of the tip angle; and "aux" identifies the auxiliary functional area, reflecting its auxiliary functional positioning. The resulting feature vectors, together with the regional point cloud, constitute the structured benchmark data for the tool region, providing a complete morphological benchmark for subsequent wear evolution analysis. This benchmark data includes information on macroscopic geometric feature changes and sensitively captures early wear signs through microscopic statistical parameters, establishing a bridge from three-dimensional point cloud to quantitative wear evaluation.

[0150] Step S3: Collect structured data of the tool area after the tool has undergone different machining cycles to obtain structured time-series data of the tool area; based on the structured reference data of the tool area, align and stitch the structured time-series data of the tool area using a spatiotemporal registration algorithm to generate the tool morphology evolution trajectory.

[0151] The tool undergoes different machining cycles (referred to as a time series). Then, steps S1 and S2 are repeated to collect the tool area structured data at the corresponding time points. in ,and This is the initial reference time. Each A subset of the point cloud containing each functional area at that moment and its corresponding structured description fields ;

[0152] To achieve accurate comparison of data at different times, spatiotemporal registration is required. A group of topographic anchor points is introduced as the registration reference. The topographic anchor point group A is defined as a set of significant geometric feature points (such as tool holder positioning surface features, specific marker points, etc.) on the tool surface that are relatively stable and not easily worn during the machining process. It can be extracted from the structured reference data T of the tool area.

[0153] For any time Tool area structured data The iterative closest point (ICP) algorithm is used to solve for its distance from the reference time. data Optimal rigid body transformation between This makes the transformed topographic anchor point group With reference topography anchor point group Minimize alignment error:

[0154]

[0155] in, Indicates time The A shape anchor point, Given the total number of anchor points, apply the obtained transformation parameters ( , Spatial transformation is performed on the subsets of the regional point cloud at each time point to obtain the registered point cloud. Simultaneously, the structured description fields are recalculated based on the registered point cloud. (Calculation process and steps) Consistent).

[0156] Organize all registered data in chronological order to construct the tool morphology evolution trajectory. :

[0157]

[0158] in, , These are the times after registration. Functional area The point cloud and feature fields of the trajectory fully record the changes in the shape of each functional area of ​​the tool with machining time, providing a data basis for wear condition monitoring and prediction.

[0159] The core of this step is to introduce a time dimension to achieve dynamic tracking. After the tool has undergone different machining cycles, S1 and S2 are repeated to acquire structured data at a series of time points. To address the slight pose differences in each clamping, a group of topographic anchor points is introduced as a stable benchmark. A spatiotemporal registration algorithm is used to precisely align the data from all time points to the same coordinate system. Finally, the data across time are stitched together to form a continuous and quantifiable tool topographic evolution trajectory, which intuitively shows how the wear of each functional area develops over time. This serves as the direct basis for the machine learning model in S4 to identify wear types and distinguish between stable wear and abnormal abrupt changes.

[0160] Step S4: Based on a temporal convolutional network, construct a tool wear recognition model, input the tool morphology evolution trajectory into the tool wear recognition model, and output the tool wear type.

[0161] A tool wear recognition model is constructed based on a Temporal Convolutional Network (TCN). The construction process of this model is as follows:

[0162] Input the tool morphology evolution trajectory generated in step S3 The time-series feature matrix is ​​extracted from it. The extracted feature matrix is ​​as follows:

[0163]

[0164] in, Indicates time The feature vector, its dimension Corresponding to the geometric parameters (such as curvature) extracted from the tool region structured reference data T. Changes in normal vector ) and statistical characteristics (such as the standard deviation of elevation distribution) wait.

[0165] We collected evolution trajectory data of various cutting tools under different processing conditions as a training set to ensure a balanced distribution of samples of stable wear and abnormal wear, so as to avoid model bias.

[0166] The TCN network architecture is constructed by stacking extended causal convolutional layers, and its core computation is as follows:

[0167]

[0168] in, As the expansion factor, The kernel size is set to ensure that the receptive field grows exponentially with network depth, thereby capturing long-term dependencies in the evolution trajectory. The layer outputs of the TCN are passed to the output layer after residual connections and regularization. The kernel size V is set to 5 to ensure effective perception of local temporal patterns. The dilation factor d increases exponentially with the number of network layers; specifically, for a 4-layer TCN module, the dilation factors for each layer are set to 1, 2, 4, and 8 respectively. This parameter combination has been experimentally verified to achieve a good balance between computational complexity and model performance, effectively capturing short-term fluctuations and long-term trends in the tool morphology evolution trajectory.

[0169] A fully connected layer is used to map the temporal features extracted by TCN to the class space, and a probability distribution is generated through the Softmax function:

[0170]

[0171] in, This is the hidden state of the last layer of TCN. and For output layer parameters, output This represents the probability that the input trajectory belongs to each category.

[0172] The model training process aims to minimize the weighted cross-entropy loss, as shown in the following formula:

[0173]

[0174] Among them, weight and The settings are based on the training set distribution to mitigate the class imbalance problem.

[0175] The evolution trajectory of the tool (to be tested) morphology generated by S3. Input the trained tool wear recognition model and obtain the probability output. Set confidence threshold The decision-making rules are as follows:

[0176]

[0177] The final output is a tool wear type label. This provides a basis for subsequent wear mechanism analysis;

[0178] It should be noted that the confidence threshold θ is set to 0.7. The determination of this value is based on the following: First, given the extremely high reliability requirements of CNC machine tool processing, it is necessary to minimize unplanned downtime losses caused by false alarms. Second, performance analysis of the model output on historical datasets revealed that when θ=0.7, an accuracy rate exceeding 95% can be achieved, meaning that the abnormal alarms issued by the model have extremely high reliability. Finally, this threshold setting conforms to the conservative decision-making principle of this invention, that is, the model only outputs a clear classification when it has sufficient confidence in the judgment result (confidence ≥70%); otherwise, it is judged as 'pending' to initiate manual review or additional testing, thereby optimizing the overall efficiency and safety of the production process while ensuring identification accuracy.

[0179] For example, the final output of the model If [0.85, 0.15] is a probability distribution, it means that the tool wear identification model believes that the current tool wear trajectory has an 85% probability of being stable wear and a 15% probability of being abnormal wear. Only when the model has sufficient confidence (conf>= Only then can a clear judgment be given on the stable wear type or the abnormal wear type.

[0180] This step is the intelligent identification and classification stage. It constructs a tool wear identification model based on a temporal convolutional network (TCN). This model excels at processing time-series data and can automatically learn the dynamic characteristics and evolution patterns of wear from the morphological evolution trajectory provided by S3. The model ultimately outputs a probabilistic classification result, determining whether the current tool wear state belongs to the stable wear type or the abnormal wear type.

[0181] Step S5: Collect historical case data of cutting tools and build a wear mechanism knowledge base; match and map the tool wear type with the wear mechanism knowledge base to generate tool wear identification results.

[0182] Historical case data for cutting tools was collected in two different environments:

[0183] In a laboratory environment, accelerated wear tests were conducted using the controlled variable method to record the evolution of characteristic parameters under different historical working conditions (cutting speed, feed rate, material hardness), and simultaneously record the remaining life data, the machining time from the initial state to complete failure of the tool, and trigger factor data to accurately record the threshold values ​​of key parameters (such as critical temperature and vibration intensity) that lead to abnormal wear.

[0184] In actual production environments, long-term monitoring is used to collect complete data trajectories of characteristic parameter changes, normal wear, and abnormal wear within different tool life cycles, as well as historical working conditions and manifestations of various abnormal failure modes.

[0185] Based on historical tool case data, a knowledge base for quantitative characteristic wear mechanisms is constructed:

[0186]

[0187] in, For feature vectors; Indicates the type of wear mechanism; This is the threshold condition vector for determining this mechanism; This represents typical remaining lifetime data under this mechanism model; The dataset of triggering factors for this mechanism, based on a knowledge base built with sufficient experimental and field data, lays a solid foundation for accurate matching of subsequent mechanisms;

[0188] The wear type output by S4 is matched with the morphological evolution trajectory feature H of S3 to construct a query vector. Matching is performed using weighted Euclidean distance:

[0189]

[0190] in, It is the query vector of the current tool. With the knowledge base Feature vector of a wear mechanism case The weighted Euclidean distance between them. The smaller this value, the more similar the two are; It is the first The weights for each feature dimension are assigned based on the importance of that feature in distinguishing different wear mechanisms. Important features (such as those sensitive to thermal softening) are given higher weights. It is a query vector In the Eigenvalues ​​in the dimension. It is the first in the knowledge base Feature vector of a wear mechanism case In the Eigenvalues ​​in the dimension.

[0191] After obtaining the matching results, the corresponding identification and judgment branch is entered according to the wear type, and the stable wear type and the abnormal wear type are judged respectively.

[0192] For stable wear types, the system accurately matches normal wear mechanism patterns (such as uniform wear and oxidative wear) in the knowledge base and calculates the remaining lifetime based on feature evolution trends.

[0193]

[0194] in, This indicates the calculated remaining tool life. It refers to the estimated time after which the tool, under its current wear condition, is expected to continue operating normally, usually measured in hours or minutes. This indicates the expected total lifespan of this type of cutting tool under specific working conditions, representing the total machining time a new tool can withstand from the time it is put into use until it is completely worn out and scrapped. This indicates the current measured amount of wear. This represents the wear life threshold, which is the critical wear level at which the tool is deemed to have failed or needs replacement. When the wear level of the tool reaches this threshold, its lifespan is considered exhausted.

[0195] The complete identification and judgment result is: {Normal uniform wear, remaining life estimation} Hour}.

[0196] For wear anomaly types, select the one with the smallest distance from the knowledge base. The first i abnormal mechanisms constitute a candidate set, and the triggering conditions of each mechanism are verified to meet the threshold requirements as follows:

[0197]

[0198] The final output of the recognition result includes complete information on the main triggering factors and their confidence levels:

[0199]

[0200] Based on the tool wear identification results above, the process of generating a targeted adaptive early warning strategy is as follows:

[0201] For stable wear types, the warning not only provides remaining tool life information but also outputs specific suggestions for optimizing machining parameters based on the current tool wear status and machining history data. These suggestions include measures to extend tool life, such as adjusting cutting speed and optimizing feed rate. It comprehensively considers tool material properties, machining parameters, and economic factors to provide the best parameter adjustment scheme, ensuring maximum tool utilization while maintaining machining quality.

[0202] For abnormal wear types, the warning system generates tiered warnings and handling suggestions based on the specific triggering factors identified. When temperature-related anomalies such as thermal softening are detected, specific measures such as reducing cutting speed and increasing coolant flow are recommended. When mechanical overload anomalies are identified, adjustments to the depth of cut or feed rate are suggested. The warning system also automatically triggers different levels of response mechanisms based on the severity and confidence level of the anomaly, ranging from simple parameter adjustment suggestions to emergency shutdown checks, forming a complete closed-loop control system of detection, diagnosis, decision-making, and execution. Simultaneously, the system records the processing effect of each warning and continuously optimizes the warning thresholds and handling strategies through machine learning algorithms, achieving continuous improvement in warning capabilities.

[0203] This step serves as the decision-making and output endpoint of the entire method, aiming to achieve interpretable root cause analysis. First, a wear mechanism knowledge base is constructed, containing data on various wear mechanisms, characteristics, thresholds, remaining life, and triggering factors. Then, the wear types identified by the model are matched and mapped with the current tool's feature data to the knowledge base. For stable wear, the corresponding conventional mechanism and remaining life prediction are output; for abnormal wear, the root cause, i.e., the abnormal wear triggering factors, is diagnosed, and specific warnings and remedial suggestions are generated.

[0204] This technical solution first transforms the two-dimensional image sequence of CNC machine tool cutting tools into a high-precision three-dimensional point cloud of tool reference using stereo vision imaging and multi-view geometric 3D reconstruction algorithms, establishing an accurate geometric reference for wear identification. Then, based on the tool failure mechanism, functional area division rules are constructed, and the 3D point cloud is semantically segmented and the geometric morphology quantification features of each region are extracted, generating structured reference data of tool regions with clear engineering significance, realizing the transformation from 3D geometry to computable semantics. On this basis, a spatiotemporal registration algorithm is used to align and stitch together the structured data of tool regions from different machining cycles, constructing a complete reference data that reflects wear. The tool wear morphology evolution trajectory during the wear development process has achieved a leap from static detection to dynamic tracking. Subsequently, a tool wear identification model is constructed based on a temporal convolutional network to intelligently analyze the morphology evolution trajectory and accurately output the type of stable wear or abnormal wear. Finally, by matching and mapping the wear type with the constructed wear mechanism knowledge base, not only is the wear state identified, but the underlying physicochemical mechanism and triggering factors are further revealed. Explainable decision results containing remaining life prediction and targeted maintenance strategies are generated, thus forming a complete technical closed loop from three-dimensional perception, dynamic tracking, intelligent identification to explainable decision-making.

[0205] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, the phrase "comprising an element defined as..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0206] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their likenesses.

Claims

1. A machine vision-based tool wear recognition method for a numerical control machine tool, characterized in that, The method comprises the following steps: Step S1: Collecting a two-dimensional image sequence of a tool of a numerical control machine tool, processing the two-dimensional image sequence of the tool by a multi-view geometric three-dimensional reconstruction algorithm to obtain a tool reference three-dimensional point cloud, and calibrating the tool reference three-dimensional point cloud by an iterative closest point (ICP) algorithm to obtain a high-precision tool reference three-dimensional point cloud; Step S2: Constructing a functional area division rule, segmenting the high-precision tool reference three-dimensional point cloud according to the functional area division rule to obtain a tool functional area, and analyzing the tool functional area by a geometric morphological quantitative description algorithm to generate tool area structured reference data; Step S3: Collecting tool area structured data after the tool undergoes different machining cycles to obtain tool area structured time series data, aligning and splicing the tool area structured time series data by a space-time registration algorithm based on the tool area structured reference data to generate a tool morphology evolution trajectory; Step S4: Constructing a tool wear identification model based on a time series convolution network, inputting the tool morphology evolution trajectory into the tool wear identification model, and outputting a tool wear type; Step S5: Collecting tool historical case data to construct a wear mechanism knowledge base, and matching and mapping the tool wear type with the wear mechanism knowledge base to generate a tool wear identification result.

2. The machine vision-based tool wear recognition method for a CNC machine tool according to claim 1, characterized in that, Collecting a two-dimensional image sequence of a tool of a numerical control machine tool, processing the two-dimensional image sequence of the tool by a multi-view geometric three-dimensional reconstruction algorithm to obtain a tool reference three-dimensional point cloud, comprising: Through a stereo vision imaging unit fixed in a working cavity of the numerical control machine tool, when a main shaft of the machine tool is in a stationary or low-speed rotating state, the main shaft is controlled to rotate by indexing at a preset angle step, and at each controlled angle position, left and right cameras in the stereo vision imaging unit are synchronously triggered to collect a two-dimensional image sequence of the tool covering key parts of a tool edge; The process of processing the two-dimensional image sequence of the tool by the multi-view geometric three-dimensional reconstruction algorithm is as follows: The core of multi-view geometry three-dimensional reconstruction algorithm is to apply triangulation principle to inversely calculate two-dimensional image sequence of the tool into three-dimensional point cloud. Assuming that two images are taken from two view angles, the projection matrices are and For a feature point on the left image , according to the matching point on the right image, the following equation can be established: ； wherein, is the three-dimensional homogeneous coordinate of the spatial point to be solved, i.e. its three-dimensional spatial coordinate; are the projection matrices corresponding to the left and right cameras, respectively; denotes the dimensional equivalence relation; This equation can be converted into a linear one, and the three-dimensional coordinates of the space points can be obtained by solving a least squares problem . This operation is performed for all successfully matched feature points, and an initial three-dimensional point cloud composed of a large number of discrete three-dimensional points with coordinates is generated, denoted as the tool reference three-dimensional point cloud . .

3. The machine vision-based tool wear recognition method for a CNC machine tool according to claim 2, characterized in that, Calibrating the tool reference three-dimensional point cloud by an iterative closest point (ICP) algorithm to obtain a high-precision tool reference three-dimensional point cloud, comprising: An Iterative Closest Point (ICP) algorithm is used for fine registration. The goal of the ICP algorithm is to find an optimal rigid body transformation that minimizes the overall distance between the transformed point cloud and the reference point cloud The problem is formulated as minimizing the following objective function: ； wherein, is the objective function, i.e. the sum of squared Euclidean distances between all corresponding point pairs; is the rigid body transformation parameter to be solved for, denotes the starting index of the summation, i.e. from =1 the summation is started; N denotes the end index of the summation, i.e. the total number of points in the point cloud, denotes the -th point in the registered point cloud, denotes the -th point in the target point cloud; The iterative closest point (ICP) algorithm solves the optimal transformation between two point clouds by iteratively performing data association and motion estimation and ; Firstly, data association is performed, for each point in Q The nearest neighbor in P is found, and the correspondence is established; then, motion estimation is performed, and based on the current correspondence, the optimal R and t are solved by the SVD decomposition method; iteration is performed until the error converges to the minimum value; The solved optimal transformation (R, t) is used to perform spatial transformation on the tool reference three-dimensional point cloud, i.e. a high-precision point cloud that is accurately aligned with the reference coordinate system is obtained, denoted as a high-precision tool reference three-dimensional point cloud .

4. The machine vision-based tool wear recognition method for a CNC machine tool according to claim 3, characterized in that, Constructing a functional area division rule, segmenting the high-precision tool reference three-dimensional point cloud according to the functional area division rule to obtain a tool functional area, comprising: To achieve fine analysis of tool damage, first, a functional area subdivision rule is constructed according to the mechanism of tool failure mechanics and cutting mechanics, which is defined based on mechanical stress, thermal load and typical damage modes caused by different tool surfaces in the cutting process. The specific functional area division rule is as follows: The main tool area is the main cutting area participating in material removal; the relief surface area is the area in contact with the machined surface; the rake surface area is the surface through which the cutting flow passes; the tool tip transition area is the transition area of the main cutting edge; and the auxiliary functional area is according to the tool type; According to the above function region division rule, the high-precision tool reference three-dimensional point cloud obtained by S1 is segmented , The segmentation process is realized by the following steps: First, feature calculation is performed, the spatial distribution characteristics of each point in the point cloud are analyzed, and the normal vector and curvature features thereof are calculated; The normal vector calculation adopts a method based on local plane fitting, and first, the covariance matrix of the local neighborhood is calculated: ； in, The calculated 3x3 covariance matrix describes the distribution direction and dispersion of points in the local neighborhood. This represents the total number of neighboring points of point k; Let i represent the i-th neighboring point of point k. represents the centroid of the local neighborhood of point k; T represents the transpose of a vector; It's an outer product, and the result is a 3x3 matrix that describes the 6th position of point k. Neighboring points Relative to the local neighborhood centroid The distribution relationship of displacement in various directions; Next, the covariance matrix Eigenvalue decomposition yields three eigenvalues. and the corresponding feature vectors ,point normal vector It is the smallest eigenvalue Corresponding feature vector ; The curvature calculation is based on feature decomposition of the local covariance matrix: ； wherein, represents the curvature estimate of the local neighborhood of the kth point in the point cloud, These are three eigenvalues obtained by eigen decomposition of the covariance matrix of the local neighborhood point set, satisfying ; On the basis of the point cloud features, a region growing process is carried out to the highest point of the sharp knife As a growth starting point, adjacent points are gradually gathered according to three criteria of normal vector consistency, curvature continuity and spatial proximity to form an initial region; The normal vector consistency requires that the included angle of normal vectors of adjacent points is less than 15 degrees, ensuring the smoothness of the surface; the curvature continuity controls the difference of curvatures of adjacent points to be within 0.05, ensuring the coherence of geometric features; and the spatial proximity ensures the connectivity of the region through the Euclidean distance constraint; Finally, the initial region is semantically labeled by a segmentation network based on deep learning, and the point cloud is finally divided into mutually disjoint subsets to determine the tool functional region: ； wherein, , , , , respectively correspond to the point cloud subsets of the main flank region, the relief region, the rake region, the nose region and the auxiliary function region.

5. The machine vision-based tool wear identification method for a CNC machine tool according to claim 4, characterized in that, The tool functional region is analyzed by a geometric morphological quantitative description algorithm to generate tool region structured datum, including: To transform geometric information into computable wear evaluation indicators, it is necessary to process a subset of the point cloud for each region. The analysis is performed using a geometric morphology quantification algorithm, as follows: extracting macro-geometric parameters to obtain a surface normal direction distribution variance of each region , contour curvature radius , curvature radius is the inverse of curvature = ; and a region macro-feature index ; Surface normal vector direction distribution variance The formula for calculating the surface normal vector direction distribution variance is: ； wherein, is the variance of the surface normal direction distribution, which is used to quantify the degree of the direction dispersion of the surface normal vectors within the tool region; is the surface normal vector of the kth point cloud; n denotes the surface normal vector; k is the index of the point in the point cloud; is the total number of point clouds within the current tool region; is the mean vector of the surface normal vectors of all points within the region; is the Euclidean norm of the vector; For micro-topography statistics, skewness and kurtosis of the elevation distribution of the point cloud are calculated for the wear-sensitive regions , where skewness characterizes the asymmetry of wear and kurtosis reflects the concentration of wear. ​ The skewness calculation formula is: ； wherein, represents a calculated skewness value quantifying the asymmetry of the wear surface; represents the elevation value of the th data point; ; The kurtosis calculation formula is: ； wherein, represents the calculated kurtosis value, quantifying the concentration of the wear surface topography distribution; represents the standard deviation of the elevation values of all data points, measuring the dispersion of the elevation values, i.e. the overall level of surface roughness; The above parameters are organized as a feature vector and generate reusable tool region structured reference data : ； Wherein, edge is used to identify the main tool area, highlighting its core position as the cutting edge; flank is used to identify the flank face area, which is derived from the traditional terminology of flank face in mechanical engineering, which specifically refers to the tool flank face; face is used to identify the rake face area, which conforms to the habit of representing the main working surface; corner is used to identify the tool tip arc area, which accurately describes the tool tip corner geometric feature; aux is used to identify the auxiliary functional area, which embodies its auxiliary function positioning; and the finally generated feature vector and the region point cloud together constitute the tool region structured datum.

6. The machine vision-based tool wear identification method for a CNC machine tool according to claim 5, wherein, Tool region structured data after the tool experiences different machining cycles is collected to obtain tool region structured time series data, including: After the tool undergoes different machining cycles, steps S1 and S2 are repeatedly executed to collect the tool region structured data at corresponding time points wherein , and is an initial reference time point, each contains a point cloud subset of each functional region at the time point and its corresponding structured description field .

7. The machine vision-based tool wear identification method for a CNC machine tool according to claim 6, characterized in that, Based on the tool region structured datum, the tool region structured time series data is aligned and spliced by a space-time registration algorithm to generate a tool morphology evolution trajectory, including: For any time instant of the tool region structuring data , the optimal rigid body transformation between it and the reference time instant data is solved by Iterative Closest Point algorithm , so that the transformed topography anchor point group has the minimum alignment error with the reference topography anchor point group : ； wherein, represents the time of the th appearance anchor point, is the total number of anchor points, the obtained transformation parameters (, , ) are applied to the region point cloud subsets at each time to obtain the registered point cloud At the same time, the structured description field is recalculated based on the registered point cloud ; The data at all time instants after registration are organized in time sequence to construct a tool topography evolution trajectory : ； wherein, , are the point cloud and feature field of the registered post-time functional region .

8. The machine vision-based tool wear identification method for a CNC machine tool according to claim 7, characterized in that, Based on the time series convolution network, a tool wear identification model is constructed, the tool morphology evolution trajectory is input into the tool wear identification model, and the tool wear type is output, including: Based on the time series convolution network, a tool wear identification model is constructed, and the construction process of the model is as follows: The tool topography evolution trajectory generated in the input step S3 The time sequence feature matrix is extracted from the matrix, and the extracted feature matrix is: ； wherein, represents the eigenvector at the time instant whose dimension corresponds to the geometrical parameters and statistical features extracted from the tool area structured reference data T; Evolution trajectory data of various tools under different machining conditions are collected as a training set to ensure balanced distribution of stable wear and wear anomaly samples to avoid model bias; The TCN network architecture is constructed by stacking dilated causal convolution layers, and its core operation is: ； wherein, is an expansion factor, is a convolution kernel size, ensuring receptive field exponentially grows with network depth, thus capturing long-term dependencies in the evolution trajectory, s represents the time step index of the output sequence, the hierarchical output of TCN is passed to the output layer after residual connection and regularization; The time series features extracted by the TCN are mapped to the category space by using a fully connected layer, and a probability distribution is generated by using a Softmax function: ； wherein, is the TCN last layer hidden state, and is the output layer parameter, output denotes the probability that the input track belongs to each category; Evolution trajectory of tool topography generated by S3 Input the trained tool wear recognition model to obtain the probability output , set the confidence threshold , the decision rule is as follows: ； The final output is a tool wear type label .

9. The machine vision-based tool wear identification method for a CNC machine tool according to claim 8, wherein, Tool historical case data are collected to construct a wear mechanism knowledge base, including: Tool historical case data are collected in two environments: In the laboratory environment, accelerated wear tests are carried out by the control variable method, the evolution of characteristic parameters under different working conditions is recorded, and the remaining life data are recorded simultaneously, the machining time of the tool from the initial state to complete failure is recorded, and the trigger factor data are recorded to accurately record the key parameter threshold that causes abnormal wear; In the actual production environment, the complete data trajectory of the characteristic parameters of different tools in the life cycle is collected, including normal wear and abnormal wear, and the historical working conditions and forms corresponding to various abnormal failure modes are collected; Based on the tool historical case data, a quantitative feature wear mechanism knowledge base is constructed: ； wherein, is a feature vector; represents a wear mechanism type; is a decision threshold condition vector for the mechanism; is typical residual life data under the mechanism mode; is a trigger factor data set for the mechanism.

10. A machine vision based tool wear recognition system for a CNC machine tool comprising a memory, a processor and a computer program stored on the memory, characterized in that, The processor executes the computer program to implement the steps of the method of any one of claims 1-9.

Images