AI driving type machine vision tool state abnormity early warning method of numerical control machine tool

Abstract

The invention relates to the technical field of numerical control machine tool cutter state monitoring, and discloses a numerical control machine tool AI drive type machine vision cutter state abnormity early warning method comprising the following steps: screening a source domain process set based on multi-dimensional attribute similarity; generating Usui model migration parameters through a parameter migration network, and driving a digital twin simulation engine to output a theoretical crater depth prediction sequence; finely adjusting the cutting inverse network through material attribute conditional feature transformation; structured light three-dimensional reconstruction is carried out on the discharged cuttings, morphological characteristics are extracted, and a crater depth estimation value is output through a cuttings back-stepping network; double-channel cross deviation is calculated, collaborative correction is carried out through online Bayesian parameter updating, and a consistency estimated value is generated; and generating an early warning signal based on trend extrapolation and tolerance upper limit comparison.

Description

Technical Field

[0001] This invention relates to the field of CNC machine tool tool condition monitoring technology, and more specifically, to an AI-driven machine vision method for early warning of abnormal tool conditions in CNC machine tools. Background Technology

[0002] In CNC machine tool machining, crater wear on the tool rake face is a key degradation mode affecting machining accuracy and surface quality. Existing technologies typically employ two parallel tool condition monitoring pathways: one is based on a digital twin simulation engine, using the Usui wear model and CNC program process parameters to generate a theoretical wear depth prediction sequence; the other is based on machine vision to acquire images of the emanating chips, and indirectly estimate the tool wear state from the chip morphology through a chip inverse network. The two pathways operate independently and output early warning criteria.

[0003] However, when new alloy materials or new coated tool combinations are introduced into CNC machine tools, both of the above monitoring methods fail simultaneously. The Usui wear model parameters of the digital twin simulation engine are not calibrated for new material combinations, causing theoretical predictions to deviate from actual wear processes. In the machine vision channel, the physical constraint loss term of the chip inversion network is based on the chip flow dynamics of specific materials; differences in chip flow characteristics of new materials lead to a degradation in the inversion mapping accuracy of the chip inversion network. In existing technologies, the two channels perform migration adaptation independently, lacking cross-validation. A single channel cannot independently detect the cumulative direction and magnitude of its own migration errors, resulting in a large number of missed or false alarms during the cold start phase of new processes, failing to provide reliable tool status anomaly warnings for CNC machine tool operation. Summary of the Invention

[0004] This invention provides an AI-driven machine vision-based tool status anomaly early warning method for CNC machine tools, solving the technical problem in related technologies that it is difficult to achieve accurate monitoring and timely early warning of crater wear of CNC machine tool tools in scenarios with new material tool combinations.

[0005] This invention discloses an AI-driven machine vision-based tool status anomaly early warning method for CNC machine tools, comprising:

[0006] Obtain the material and tool attribute description vector of the new material tool combination, and filter the source domain process set from the historical process database based on multi-dimensional attribute similarity;

[0007] The calibrated Usui model parameters of each process in the source domain process set and the corresponding material and tool property description vectors are input into the parameter migration network, and the migration prediction Usui model parameters are output and loaded into the digital twin simulation engine to generate the theoretical crescent depth prediction sequence.

[0008] By performing cross-material fine-tuning of the chip inverse network through material property conditional feature transformation, a chip inverse network adapted to new materials is generated.

[0009] The chips discharged during the processing are reconstructed in three dimensions using structured light. The three-dimensional morphological feature vector of the chips is extracted, and the resulting vector is input into the adapted chip inverse network. The output is the estimated value of the depth of the inverse crater of the chip.

[0010] The cross deviation between the chip-induced crater depth estimate and the theoretical crater depth prediction at the same time is calculated. Based on the temporal smoothness constraint of the theoretical crater depth prediction and the physical boundary constraint of the chip-induced crater depth estimate, the cross deviation is decomposed into simulation-side correction and inverse-side correction. Bidirectional collaborative correction is performed through online Bayesian parameter updates to generate a consistent estimate of the crater depth.

[0011] Trend extrapolation is performed based on the time series of consistent estimates of the crescent depression depth. When the predicted depth exceeds the upper limit of the crescent depression tolerance, an abnormal tool status warning signal is generated.

[0012] Furthermore, the step of obtaining the material tool attribute description vector for the new material tool combination, and filtering the source domain process set from the historical process database based on multi-dimensional attribute similarity, includes:

[0013] Obtain the workpiece material properties and tool properties of the new material tool combination. The workpiece material properties include the material's hardness, thermal conductivity, yield strength, and chemical composition ratio. The tool properties include the tool substrate material, coating type, coating thickness, tool tip radius, and rake angle.

[0014] Each parameter is standardized using Z-score. The coating type and tool substrate material are converted into numerical representations using unique thermal encoding and then spliced ​​together to form a material tool property description vector.

[0015] In the historical process database, the multidimensional attribute similarity between the material tool attribute description vector and the attribute description vector corresponding to each historical process combination is calculated. The multidimensional attribute similarity is a comprehensive similarity measure obtained by calculating the distance in each dimension and then performing a weighted summation. The weight of each dimension is preset according to the degree of influence of the dimension parameter on the crater wear process.

[0016] Select a preset number of historical process combinations with the highest similarity as the source domain process set.

[0017] Furthermore, the parameter transfer network is a fully connected feedforward neural network. The input of the parameter transfer network is the material tool attribute description vector, and the output nodes correspond to the wear coefficient and temperature sensitivity index in the Usui wear model. The parameter transfer network uses the paired data between the calibrated material tool attribute description vector and the corresponding Usui model parameters in the source domain process set as training samples, and uses the mean square error between the predicted output Usui model parameters and the calibrated parameters as the loss function for training. The theoretical crescent depth prediction sequence is the spatial distribution data of the rake face crescent depth output by the digital twin simulation engine at each discrete sampling time, combined with the CNC program process parameters of the current machining task.

[0018] Furthermore, the material property conditional feature transformation includes:

[0019] Using the material tool property description vector of the new material tool combination as a conditional input, the scaling parameter vector and offset parameter vector are generated respectively through the auxiliary mapping layer;

[0020] An affine transformation is applied to the three-dimensional morphological features of the source domain chip. The affine transformation is to multiply the scaling parameter vector with the three-dimensional morphological features of the source domain chip element by element and then add the offset parameter vector to obtain the transformed features.

[0021] The material-related parameters of the physical constraint loss term in the chip inverse network are fine-tuned by replacing the original source domain features with the transformed features.

[0022] The auxiliary mapping layer is a fully connected feedforward network. The input is a material and tool property description vector, and the output layer consists of two parallel fully connected layers that output scaling parameter vector and offset parameter vector, respectively, which have the same dimensions as the three-dimensional morphological features of the source domain chips.

[0023] Furthermore, the physical constraint loss term imposes constraints on the network output based on the physical correspondence between the groove morphology of the chip bottom surface and the wear pit morphology of the tool rake face in chip flow dynamics. The material-related parameters in the physical constraint loss term include physical constants corresponding to the plastic flow characteristics and shear band spacing characteristics of the workpiece material. The chip inverse network is trained with the weighted sum of the mean square error loss of the crescent depth and the physical constraint loss term as the total loss function. During the fine-tuning stage, only the material-related parameters in the physical constraint loss term are updated, while the remaining network weights remain frozen.

[0024] Furthermore, the step of performing structured light 3D reconstruction on the chips discharged during the processing and extracting the 3D morphological feature vector of the chips includes:

[0025] A sinusoidal stripe pattern is projected onto the surface of the discharged chips through a structured light projection module at the chip removal channel, and stripe deformation image pairs of the bottom surface of the chips are acquired from two different angles by a high-magnification macro industrial camera.

[0026] Phase-shift demodulation and stereo matching are performed on stripe deformation image pairs to reconstruct three-dimensional micro-topography point cloud data of the chip bottom surface. The phase-shift demodulation is performed by a multi-step phase-shift algorithm to calculate the absolute phase value at each pixel from multiple stripe deformation images with different phase offsets.

[0027] The surface texture features of the chip are tracked and matched between adjacent frames, the displacement of the chip within the acquisition interval of each frame is calculated, and the displacement is compensated to the phase solution result of the corresponding frame for motion compensation processing.

[0028] The groove depth distribution, groove spacing, and contact surface curvature distribution of the chip bottom surface are extracted from the three-dimensional micro-topography point cloud data. After Z-score standardization of each feature, they are spliced ​​together to form a three-dimensional morphological feature vector of the chip.

[0029] Furthermore, the decomposition of crossover bias into simulation-side correction and reverse-side correction includes:

[0030] The specific form of the temporal smoothness constraint is: the absolute value of the change in the predicted value of the theoretical crescent depth between adjacent sampling times does not exceed the preset smoothness threshold, which is determined based on the product of the maximum reasonable wear rate of the Usui wear model under the current cutting parameters and the sampling time interval;

[0031] The specific form of the physical boundary constraint is: the estimated value of the crescent depth obtained by reverse chip cutting is not less than the physical lower bound of the crescent depth determined by the material tool property description vector and the current cutting parameters, and is not greater than the physical upper bound of the crescent depth.

[0032] Calculate the simulation-side constraint violation magnitude, which is the greater of the difference between the absolute value of the change in the theoretical crescent depth prediction value at adjacent sampling times and the preset smoothness threshold, and zero.

[0033] The magnitude of the violation of the reverse side constraint is calculated as the greater of the difference between the estimated value of the reverse crater depth of the chip and the physical upper bound and zero, plus the greater of the difference between the physical lower bound and the estimated value of the reverse crater depth of the chip and zero.

[0034] The simulation-side correction is obtained by multiplying the ratio of the simulation-side constraint violation amplitude to the sum of the constraint violation amplitudes on both sides by the cross deviation; the reverse-side correction is obtained by multiplying the ratio of the reverse-side constraint violation amplitude to the sum of the constraint violation amplitudes on both sides by the cross deviation; when the sum of the constraint violation amplitudes on both sides is zero, both the simulation-side correction and the reverse-side correction are set to zero.

[0035] Furthermore, the online Bayesian parameter update includes:

[0036] The Usui model transfer parameters and the material-related parameters of the chip inverse network are respectively used as the parameters to be estimated. The transfer parameters obtained by the parameter transfer network and cross-material fine-tuning are used as the mean of the prior distribution. The variance corresponding to the preset confidence interval is used as the covariance, and a Gaussian prior distribution is set.

[0037] At each sampling time, the corresponding side correction is used as the observation residual. The mean and variance of the posterior distribution are calculated according to the Gaussian-Gaussian conjugate update formula. The posterior mean is taken as the correction parameter value at the current time, and the posterior variance is taken as the prior variance updated at the next time, so as to realize the recursive convergence of parameter estimation with sampling time.

[0038] The digital twin simulation engine was re-run using the corrected Usui model migration parameters to obtain the corrected theoretical crater depth prediction value, and the chip inverse estimation network was run using the corrected material-related parameters to obtain the corrected chip inverse estimation crater depth estimate value. The two were then weighted and fused to generate a consistent estimate value for the crater depth.

[0039] Furthermore, in the weighted fusion, the fusion weight is determined based on the variance of the posterior distribution of the parameters on both sides. The simulation-side fusion weight is the ratio of the reciprocal of the posterior variance of the simulation-side parameters to the sum of the reciprocals of the posterior variances of the parameters on both sides. The inverse-side fusion weight is the ratio of the reciprocal of the posterior variance of the inverse-side parameters to the sum of the reciprocals of the posterior variances of the parameters on both sides. The sum of the simulation-side fusion weight and the inverse-side fusion weight is one.

[0040] The trend extrapolation based on the time series of consistent crater depth estimates includes: forming a time series of consistent crater depth estimates from multiple consecutive sampling times; calculating the maximum depth growth rate and acceleration characteristics of the crater; using polynomial fitting to extrapolate the evolution trend of the crater depth with processing time; and outputting the predicted trajectory of the crater depth in subsequent time periods. When the predicted depth exceeds the upper limit of the crater tolerance for a consecutive preset number of prediction times, a tool status abnormality warning signal is generated and the expected over-limit time is output.

[0041] This invention solves the technical problem of two monitoring channels migrating and adapting independently and being unable to self-check migration errors when introducing new material tool combinations into CNC machine tools by introducing a cross-deviation calculation and bidirectional collaborative correction mechanism between the digital twin simulation channel and the machine vision chip reverse calculation channel. The resulting technical effects are as follows: the two channels serve as mutual calibration references, allowing the parameter errors accumulated by a single channel during cross-material migration to be detected and corrected by the other channel; with the continuous accumulation of online Bayesian parameter updates, the parameter estimates of the two channels gradually converge to the actual wear characteristics of the new material, and the reliability of the consistent estimate of the crescent depth gradually improves; and tool status anomaly warning capability with cross-validation support can be obtained during the cold start stage of the new process, reducing missed and false alarms caused by the inability to self-check errors in a single channel. Attached Figure Description

[0042] Figure 1 This is a flowchart of an AI-driven machine vision tool status anomaly early warning method for CNC machine tools provided in an embodiment of the present invention;

[0043] Figure 2 This is a schematic diagram comparing the similarity of the source domain process and the new material tool combination properties provided in the embodiments of the present invention;

[0044] Figure 3 This is a schematic diagram of the parameter distribution and migration prediction results of the source domain process Usui model provided in the embodiments of the present invention;

[0045] Figure 4 This is a schematic diagram comparing the simulated predicted values ​​and the back-calculated estimated values ​​at each sampling time provided in the embodiments of the present invention;

[0046] Figure 5 This is a schematic diagram illustrating the changing trend of the three-dimensional morphological features of the chip as a function of sampling time, as provided in an embodiment of the present invention.

[0047] Figure 6 This is a schematic diagram of the bidirectional collaborative correction cross deviation and correction amount allocation provided in an embodiment of the present invention;

[0048] Figure 7 This is a schematic diagram illustrating the evolution of simulation-side posterior variance and fusion weights under Bayesian updating provided in an embodiment of the present invention;

[0049] Figure 8 This is a schematic diagram of the extrapolation of crescent depression depth trend and the early warning judgment of tolerance upper limit provided in the embodiment of the present invention;

[0050] Figure 9 This is a schematic diagram of the scatter distribution of three-dimensional morphological features of the bottom surface of the chip (groove depth vs. contact surface curvature) provided in an embodiment of the present invention. Detailed Implementation

[0051] In CNC machining, crater wear on the tool rake face is a key degradation mode affecting machining accuracy and surface quality. Existing technologies typically employ two parallel tool condition monitoring pathways: one is based on a digital twin simulation engine, using the Usui wear model and CNC program process parameters to generate a theoretical wear depth prediction sequence; the other is based on machine vision to acquire images of the emanating chips and indirectly estimate the tool wear state from the chip morphology through a chip inverse network. The two pathways operate independently and output early warning criteria.

[0052] When new alloy materials or new coated tool combinations are introduced into CNC machine tools, the two monitoring methods mentioned above fail simultaneously. Specifically, the Usui wear model parameters of the digital twin simulation engine are not calibrated for the new material combination, and the theoretical prediction deviates from the actual wear process. In the machine vision channel, the physical constraint loss term of the chip inversion network is based on the chip flow dynamics of a specific material, and the difference in chip flow characteristics of the new material leads to the degradation of the inversion mapping accuracy of the chip inversion network. In the existing technology, the two channels perform migration adaptation independently, lacking cross-validation. A single channel cannot detect the cumulative direction and magnitude of its own migration error, causing the early warning system to generate a large number of missed or false alarms during the cold start phase of the new process, and failing to provide reliable tool status anomaly early warning for CNC machine tool operation.

[0053] According to an embodiment of this invention, an AI-driven machine vision-based tool condition anomaly early warning method for CNC machine tools is provided. It should be understood that the executing entity of this embodiment is the tool condition monitoring system of the CNC machine tool. The tool condition monitoring system includes a digital twin simulation engine, a structured light projection module at the chip removal channel, a high-magnification macro industrial camera, and computational units for the operating parameter migration network and the chip inversion network.

[0054] At least one embodiment of the present invention discloses an AI-driven machine vision-based tool status anomaly early warning method for CNC machine tools, such as... Figure 1 As shown, it includes the following steps:

[0055] Step 1: Obtain the attribute description vector of the new material tool combination, and filter the source domain process set from the historical process database based on multi-dimensional attribute similarity;

[0056] Obtain the workpiece material property parameters and tool property parameters for the new material and tool combination in the current machining task of the CNC machine tool, and encode these parameters into material and tool property description vectors. In the historical process database, calculate the multidimensional attribute similarity between the material and tool property description vectors and the attribute description vectors corresponding to each historical process combination, and select the vector with the highest similarity. A combination of historical processes serves as a source domain process set, in which This represents the preset number of source domain process combinations.

[0057] It should be noted that the aforementioned workpiece material properties include physical properties such as hardness, thermal conductivity, yield strength, and chemical composition ratio; the aforementioned tool properties include geometric and material properties such as tool substrate material, coating type, coating thickness, tip radius, and rake angle. Due to the significant differences in the dimensions and numerical ranges of these parameters, Z-score standardization is performed on each parameter before concatenating them into a material-tool property description vector to eliminate the impact of dimensional differences on subsequent similarity calculations and network training. For non-numerical classification attributes such as coating type and tool substrate material, one-hot encoding is used to convert them into numerical representations before concatenation.

[0058] It should be noted that the above multidimensional attribute similarity is a comprehensive similarity measure obtained by calculating the distance in each dimension of the material tool attribute description vector and then summing the results in a weighted average. The weight of each dimension is preset according to the influence of the dimension parameter on the crater wear process.

[0059] Step 2: Generate Usui model transfer parameters for the new material tool combination through a parameter transfer network, and load them into the digital twin simulation engine to generate a theoretical crescent depth prediction sequence;

[0060] The calibrated Usui model parameters of each process in the source domain process set and the corresponding material and tool attribute description vectors are input into the parameter migration network. The parameter migration network outputs migration prediction Usui model parameters for the new material and tool combination. The migration prediction Usui model parameters are loaded into the digital twin simulation engine and combined with the CNC program process parameters of the current machining task to generate a theoretical crescent depth distribution prediction sequence for each machining stage under the new material.

[0061] It should be noted that the parameter transfer network described above is a fully connected feedforward neural network. Its input is a material and tool attribute description vector, and its output layer is a fully connected layer. The output nodes correspond to the wear coefficient and temperature sensitivity index in the Usui wear model. The parameter transfer network uses the paired data between the calibrated material and tool attribute description vectors and the corresponding Usui model parameters in the source domain process set as training samples. The mean square error between the predicted Usui model parameters and the calibrated parameters is used as the loss function, and the Adam optimization algorithm is used for training to learn the mapping relationship from the material and tool attribute space to the Usui model parameter space. The output of the parameter transfer network is the numerical value of the wear coefficient and temperature sensitivity index, which can be directly loaded into the digital twin simulation engine as the physical parameters of the Usui wear model without additional decoding.

[0062] It should be noted that the Usui wear model described above is a physical model describing the evolution of wear depth in the contact area of ​​the tool rake face with cutting time. The parameters of the Usui wear model include a wear coefficient and a temperature sensitivity index related to the combination of workpiece and tool materials. The theoretical crater depth distribution prediction sequence described above is the spatial distribution data of the rake face crater depth output by the digital twin simulation engine at each discrete sampling time. Each discrete sampling time corresponds to a specific cutting time node in the machining process, and each element in the sequence is the predicted crater depth value at that time.

[0063] Step 3: Fine-tune the chip inverse network across materials by transforming the material properties conditional features to generate a chip inverse network adapted to the new material;

[0064] The 3D morphology data of chips from each process in the source domain process set are extracted and paired with samples of crescent depression depth. The source domain distribution in the 3D feature space of the chips is mapped to the new material feature space through material property conditional feature transformation. The material-related parameters of the physical constraint loss term in the chip inverse network are then fine-tuned using the transformed samples to generate a chip inverse network adapted to the new material.

[0065] It should be noted that the above-mentioned material property conditional feature transformation is performed according to the following steps:

[0066] (1) Using the material tool property description vector of the new material tool combination as the condition input, the scaling parameter vector is generated by the auxiliary mapping layer. and offset parameter vector ;

[0067] (2) Let the three-dimensional morphological characteristics of the source domain chips be as follows: ,right Applying an affine transformation yields the transformed features. :

[0068]

[0069] in, To and Scaling parameter vectors of the same dimensions To and Offset parameter vectors of the same dimension This represents element-wise multiplication;

[0070] (3) Based on the transformed features The original source domain features are replaced to make the transformed feature distribution match the expected distribution of chip features under the new material conditions, which is then used for fine-tuning of the subsequent chip inverse network.

[0071] It should be noted that the above-mentioned auxiliary mapping layer is a fully connected feedforward network. The input of the auxiliary mapping layer is the material and tool property description vector, and the output layer consists of two parallel fully connected layers, which output the three-dimensional morphological features of the chip in the source domain, respectively. Scaling parameter vectors of the same dimensions and offset parameter vector The auxiliary mapping layer takes the material and tool attribute description vectors of each process in the source domain process set as input, so as to transform the features... The Adam optimization algorithm is used for training, with the goal of minimizing the mean square error between the chip feature distribution and the new material.

[0072] It should be noted that the aforementioned physical constraint loss term is a component of the chip inverse network training loss function. This term imposes constraints on the network output based on the physical correspondence between the groove morphology of the chip bottom surface and the wear pit morphology of the tool rake face in chip hemodynamics. The material-related parameters in the physical constraint loss term include physical constants corresponding to the plastic flow characteristics and shear band spacing features of the workpiece material.

[0073] Furthermore, the aforementioned chip inverse estimation network is a fully connected feedforward neural network. The input to the network is the three-dimensional morphological feature vector of the chip, and the output layer is a fully connected layer. Each output node corresponds to an estimated crescent depression depth, and the unit of this output value is consistent with the unit of the crescent depression depth annotations in the training samples, allowing it to be directly used as the crescent depression depth estimation result. During training, the weighted sum of the mean squared error loss of the crescent depression depth and the physical constraint loss term is used as the total loss function, and the Adam optimization algorithm is employed for training. During the fine-tuning phase, only the material-related parameters in the physical constraint loss term are updated, while the remaining network weights remain frozen.

[0074] Step 4: Perform structured light 3D reconstruction on the chips discharged during the processing to obtain the 3D micro-morphological point cloud data of the bottom surface of the chips;

[0075] During CNC machine tool machining, a sinusoidal stripe pattern is projected onto the surface of the discharged strip-shaped chips through a structured light projection module at the chip removal channel. Simultaneously, a high-magnification macro industrial camera acquires stripe deformation image pairs from two different angles on the bottom surface of the chips. Phase-shift demodulation and stereo matching processing are performed on the stripe deformation image pairs to reconstruct the three-dimensional microscopic topography point cloud data of the bottom surface of the chips.

[0076] It should be noted that the aforementioned phase-shift demodulation refers to the structured light projection module sequentially projecting multiple sinusoidal fringe patterns with different phase offsets, and the high-magnification macro industrial camera correspondingly acquiring multiple fringe deformation images. The absolute phase value at each pixel point is calculated from the multiple images through a multi-step phase-shift algorithm, thereby establishing the correspondence between pixel coordinates and the spatial coordinates of the chip surface.

[0077] In this embodiment, to acquire stable 3D topographic data in a dynamic scenario with continuous chip discharge, motion compensation processing is performed in addition to phase-shift demodulation. Specifically, a high-magnification macro industrial camera tracks and matches the surface texture features of the chips between adjacent frames, calculates the displacement of the chips within each frame acquisition interval, and compensates this displacement into the phase calculation results of the corresponding frame, thereby reducing the impact of chip motion on the accuracy of 3D reconstruction.

[0078] Step 5: Extract the three-dimensional morphological feature vector of the chip from the three-dimensional micro-topography point cloud data, input it into the adapted chip inverse network, and output the chip inverse crescent depth estimate.

[0079] The groove depth distribution, groove spacing, and contact surface curvature distribution on the bottom surface of the chip are extracted from the 3D microscopic point cloud data, and these features are encoded into a 3D chip morphology feature vector. Since the numerical ranges of the three types of features—groove depth distribution, groove spacing, and contact surface curvature distribution—are different, Z-score normalization is performed on each type of feature before concatenation into the 3D chip morphology feature vector to eliminate the influence of dimensional differences on the network input. The 3D chip morphology feature vector is input into a chip inverse estimation network adapted to the new material, and the network outputs the estimated depth of the inverse crater depression at the current sampling time.

[0080] It should be noted that the aforementioned groove depth distribution refers to the distribution characteristics formed by the depth values ​​of multiple grooves arranged along the cutting direction on the chip bottom surface; the aforementioned groove spacing refers to the statistical quantity of the spatial distance between adjacent grooves; and the aforementioned contact surface curvature distribution refers to the distribution characteristics formed by the local curvature of the corresponding position of the contact area between the chip bottom surface and the tool rake face. These three types of characteristics reflect the imprinting effect of the crater wear pits on the chip bottom surface morphology on the tool rake face.

[0081] Step 6: Calculate the cross deviation between the simulation channel and the vision channel. Based on the bidirectional constraint, decompose the cross deviation into corrections on both sides. Perform collaborative correction through online parameter updates to generate a consistent estimate of the crescent depression depth.

[0082] The cross-deviation between the chip-backed crater depth estimate at the current sampling time and the theoretical crater depth prediction output by the digital twin simulation engine at the same time is calculated. Based on the temporal smoothness constraint of the theoretical crater depth prediction and the physical boundary constraint of the chip-backed crater depth estimate, the cross-deviation is decomposed into simulation-side correction and back-deviation-side correction. Through online Bayesian parameter updates, the simulation-side correction is used to correct the Usui model migration parameters, while the back-deviation-side correction is used to correct the material-related parameters in the chip-backed network, generating a bidirectionally co-corrected consistent estimate of the crater depth.

[0083] It should be noted that the above calculation method for cross deviation is as follows: Let the first... The estimated depth of the crescent-shaped depression derived from the cut at each sampling time is: The theoretical predicted depth of the crescent-shaped depression is Then cross deviation for:

[0084]

[0085] in, and All values ​​are crescent depths expressed in units of length, with consistent dimensions, and can be directly subtracted. This is the time index for the sampling time, corresponding to each discrete cutting time node in the machining process.

[0086] It should be noted that the aforementioned temporal smoothness constraint means that the change in the theoretical crescent depth prediction value output by the digital twin simulation engine between adjacent sampling times should meet a preset smoothness condition, that is, the evolution of wear depth over time should have a continuous and gradual characteristic rather than an abrupt change; the specific form of this smoothness condition is: adjacent sampling times and The change in the theoretical predicted value of the crescent depth between Not exceeding the preset smoothness threshold ,in The physical boundary constraint is determined by the product of the maximum reasonable wear rate under the current cutting parameters and the sampling time interval, based on the Usui wear model. The aforementioned physical boundary constraint means that the estimated crater depth derived from the chip should fall within a physically reasonable range determined by the workpiece material and cutting parameters. The upper and lower bounds of this range are determined based on the material and tool property description vectors and the current cutting parameters. The specific form of the physical boundary constraint is: the estimated crater depth derived from the chip... Should meet ,in and These are the lower and upper physical bounds of the crescent depth, determined based on the material tool property description vector and the current cutting parameters, respectively.

[0087] It should be noted that the decomposition of the above crossover bias is performed in the following manner: Let... This represents the magnitude by which the predicted theoretical crescent depth exceeds the smoothness threshold at adjacent sampling times. To determine the extent to which the estimated depth of the crescent-shaped depression caused by the reverse chip extraction exceeds the physical boundary constraints, the simulation-side correction amount is... Correction amount on the reverse side They are respectively:

[0088]

[0089] in, The predicted value of the theoretical crescent depth changes beyond the smoothness threshold. The range, The extent by which the estimated depth of the crescent-shaped depression, derived from the chips, exceeds the physical boundary constraints; both are depth deviations expressed in units of length, with dimensions equal to... Consistent; each ratio term in the above formula is a dimensionless weighting coefficient, multiplied by The dimensions of the subsequently obtained correction quantity are the same as The same applies. When the degree of constraint violation on the simulation side is large, a larger proportion of the cross deviation is allocated to the simulation side correction amount; when the reverse side approaches or exceeds the physical boundary, a larger proportion is allocated to the reverse side correction amount.

[0090] Furthermore, when When this occurs, it indicates that no constraint violation has occurred on either side at the current sampling time, at which point we let That is, no correction is applied to the parameters on either side.

[0091] In this embodiment, the online Bayesian parameter update process is as follows: the Usui model migration parameters and the material-related parameters of the chip inverse network are considered as parameters to be estimated. The migration parameters obtained in steps 2 and 3 are used as the mean of the prior distribution, and the simulation-side correction and the inverse-side correction are used as the observation update information. At each sampling time, the posterior distribution of the parameters is calculated based on the Bayesian update rule, and the mean of the posterior distribution is taken as the corrected parameter value. As the sampling time accumulates, the prior distribution is gradually corrected by the observation data, and the parameter estimation gradually tends to the true wear characteristics of the new material.

[0092] Furthermore, the specific implementation method of the above Bayesian update rule is as follows: the prior distribution of the parameter to be estimated is set to a Gaussian distribution with the mean of the transfer parameter obtained in step 2 or step 3 and the covariance of the variance corresponding to the preset confidence interval; in the first step... Each sampling time, with corresponding side correction amount or As the observation residual, the mean and variance of the posterior distribution are calculated according to the Gaussian-Gaussian conjugate update formula. The posterior mean is taken as the correction parameter value at the current time, and the posterior variance is taken as the prior variance for the next time update, so as to realize the recursive convergence of parameter estimation with the sampling time.

[0093] In this embodiment, the above-mentioned consistent estimate of the crescent depth is generated by rerunning the digital twin simulation engine using the corrected Usui model migration parameters to obtain the corrected theoretical crescent depth prediction. The corrected material-related parameters are used to run a chip inverse network to obtain the corrected chip inverse crater depth estimate. The two values ​​are then weighted and fused to generate a consistent estimate of the crescent depression depth. :

[0094]

[0095] in, and The simulation side and the reverse side are respectively in the first... The fusion weights at each sampling time satisfy the following conditions: ; For the corrected first The theoretical predicted depth of the crescent-shaped depression at each sampling time. For the corrected first The estimated depth of the crater is derived from the back-calculation of the cut at each sampling time. Both are crater depth values ​​expressed in units of length, with consistent dimensions. The weighted fusion yields the final value. They have the same dimensions; This is the time index for the sampling time. Indicates the first Consistent estimates of the crescent depth at the cutting time node corresponding to each sampling time point. The fusion weight is determined based on the variance of the posterior distribution of the parameters on both sides, with the side with smaller variance receiving a higher fusion weight.

[0096] Furthermore, the specific method for determining the aforementioned fusion weights is as follows: Let the first... The variance of the posterior distribution of the simulated side parameters at each sampling time is The variance of the posterior distribution of the inverse side parameters is... Then the simulation side fusion weights Combined weights with reverse side They are respectively:

[0097]

[0098] in, and The posterior distributions of the two-sided parameters are respectively at the th... The variance at each sampling time point indicates that the uncertainty of the parameter estimation on that side is lower, and the corresponding fusion weight is higher, thus tilting the consistency estimate of the crescent depression depth towards the side with more reliable estimation.

[0099] Step 7: Based on the time series of the consistent estimates of the crescent depth, perform trend extrapolation, compare the predicted crescent depth with the upper tolerance limit, and generate a tool status abnormality warning signal;

[0100] A time series is constructed by assembling consistent estimates of the crater depth from multiple consecutive sampling times, and the maximum depth growth rate and acceleration characteristics of the crater are calculated. This time series, along with the growth rate and acceleration characteristics, is input into a trend extrapolation module to predict the crater depth evolution trajectory in subsequent time periods. The predicted crater depth is compared with the upper limit of the crater tolerance corresponding to the current machining process. When the predicted depth exceeds the upper limit, a tool status anomaly warning signal is generated, and the estimated over-limit time is output.

[0101] It should be noted that the aforementioned trend extrapolation module uses the growth rate and acceleration characteristics of the crater depth time series as input, and employs polynomial fitting to extrapolate the evolution trend of the crater depth with processing time, outputting the predicted crater depth trajectory for subsequent time periods. The vertical axis of the predicted crater depth trajectory represents the crater depth value expressed in length units, and the horizontal axis represents the processing time, which can be directly compared with the upper limit of the crater tolerance. The aforementioned upper limit of the crater tolerance is the maximum allowable crater wear depth on the tool rake face, predetermined based on the workpiece accuracy and surface quality requirements of the current processing task.

[0102] In this embodiment of the application, to avoid false alarms caused by abnormal fluctuations at a single sampling moment, the continuity of the predicted crescent depression depth evolution trajectory is also verified when making the out-of-limit judgment. Specifically, when the predicted depth is within a continuous range... A tool status anomaly warning signal is only generated when all predicted times exceed the upper limit of the crescent depression tolerance. This is a preset number of consecutive confirmations. This method reduces false alarms caused by short-term fluctuations.

[0103] This implementation solves the problem of two monitoring channels independently migrating and adapting without self-checking migration errors when introducing new material tool combinations in CNC machine tools by calculating the cross-deviation between the digital twin simulation channel and the machine vision chip reverse calculation channel. Specifically, the theoretical crater depth prediction value output by the simulation channel provides macroscopic trend constraints for the chip reverse calculation channel. When the estimated crater depth deviates from the physically reasonable range, the reverse correction amount of the cross-deviation drives the material-related parameters of the chip reverse calculation network to adjust in a reasonable direction. The estimated crater depth output by the chip reverse calculation channel provides actual feedback to the simulation channel. When the temporal smoothness of the theoretical crater depth prediction value is broken, the simulation correction amount of the cross-deviation drives the migration parameters of the Usui model to approach the actual wear characteristics. The two channels serve as mutual calibration references, allowing the parameter errors accumulated by one channel in cross-material migration to be detected and corrected by the other channel. Therefore, as the processing continues and the online Bayesian parameter updates accumulate, the parameter estimates of the two channels gradually converge to the actual wear characteristics of the new material. The reliability of the consistent estimate of the crescent depth gradually improves, enabling the CNC machine tool to obtain the tool status anomaly warning capability with cross-validation support during the cold start stage of the new process, reducing the missed and false alarms caused by the inability to self-check single-channel errors.

[0104] In March 20XX, a precision parts processing plant (code: Plant A) introduced a new titanium-aluminum alloy workpiece material (code: material TiAl-7) and a new coated cemented carbide cutting tool (code: tool TiAlN-C3) for CNC milling of high-precision thin-walled aerospace structural parts. Since this material and tool combination has no direct calibration record in Plant A's historical database, this method needs to be implemented during the cold start phase to monitor and provide early warning of tool crater wear in real time. The tool condition monitoring system has deployed a structured light projection module and a high-magnification macro industrial camera at the chip removal channel, and the computing unit has been pre-loaded with a parameter migration network and a chip inversion network. The current machining task has a cutting speed of 187 m / min, a feed rate of 0.12 mm / r, a depth of cut of 1.5 mm, a crater wear tolerance limit of 0.18 mm, and a continuous verification count N of 3.

[0105] like Figure 2-9 As shown, the method in this embodiment operates as follows:

[0106] Obtain the attribute description vector of the new material tool combination and filter the source domain process set;

[0107] The system collects workpiece material property parameters of TiAl-7 and tool property parameters of TiAlN-C3. Numerical parameters are Z-score standardized, and categorical attributes such as coating type and tool substrate material are uniquely encoded and concatenated into a material-tool property description vector. Subsequently, the system calculates the multidimensional attribute similarity with each historical process combination in the historical process database, selecting the K=4 historical process combinations with the highest similarity as the source domain process set.

[0108]

[0109] All four source domain process combinations met the similarity threshold requirement (threshold set at 0.75) and were included in the training sample pool for subsequent steps.

[0110] Usui model transfer parameters are generated through a parameter transfer network, and a theoretical crescent depth prediction sequence is generated.

[0111] The material and tool property description vectors of the four source domain processes are input into the parameter migration network along with the calibrated Usui model parameters. The network outputs migration prediction Usui model parameters (wear coefficient A and temperature sensitivity index B) for the combination of material TiAl-7 and tool TiAlN-C3. The migration parameters are then loaded into the digital twin simulation engine and, combined with the current cutting parameters, a theoretical crescent depth prediction sequence for each sampling time is generated.

[0112] Table 2 Source Domain Process Usui Parameters and Migration Prediction Results:

[0113]

[0114] The simulation engine uses migration parameter A = 3.74 × 10 -4 Running at B=0.597, a theoretical crescent depth prediction sequence was generated for the first 10 sampling times (each sampling interval corresponds to approximately 8 minutes of cutting time). Some results are as follows: the simulated prediction value at sampling time 1 is 0.043 mm, the simulated prediction value at sampling time 5 is 0.092 mm, and the simulated prediction value at sampling time 8 is 0.131 mm.

[0115] The chip inverse network is fine-tuned across materials by conditional feature transformation based on material properties.

[0116] Using the property description vectors of material TiAl-7 and cutting tool TiAlN-C3 as conditional inputs, and the auxiliary mapping layer outputs scaling parameter vector γ and offset parameter vector β, an affine transformation is applied to the three-dimensional chip morphology features f of the four source domain processes to obtain the transformed features. :

[0117]

[0118] Taking the third dimension of the feature vector of a chip sample in process 1 of the source domain as an example, the original feature value is 0.724. The scaling parameter of the corresponding dimension output by the auxiliary mapping layer is 1.183, and the offset parameter is -0.091. Then, after transformation:

[0119]

[0120] The transformed feature distribution is aligned with the expected chip feature distribution under the new material TiAl-7 conditions. During the fine-tuning phase, only the material-related parameters corresponding to the plastic flow characteristics and shear band spacing of TiAl-7 in the physical constraint loss term are updated; the remaining network weights are frozen. After fine-tuning, the plastic flow coefficient in the physical constraint loss term is adjusted from the source domain default value of 0.318 to 0.274, and the shear band spacing feature constant is adjusted from 0.156 to 0.189. The chip inverse calculation network adapted to the new material is now complete.

[0121] Structured light 3D reconstruction was performed on the discharged chips to obtain point cloud data of the 3D micro-morphology of the bottom surface of the chips.

[0122] During the processing, the structured light projection module sequentially projects four sinusoidal fringe patterns with phase offsets of 0, π / 2, π, and 3π / 2, respectively. A high-magnification macro industrial camera simultaneously acquires fringe deformation image pairs from two angles. A four-step phase-shift algorithm is used to calculate the absolute phase value of each pixel, and compensation is applied to the chip movement displacement between adjacent frames (the measured compensation is approximately 0.07 mm / frame). Finally, the three-dimensional microscopic topography point cloud reconstruction of the chip's bottom surface is completed. Taking the chip at sampling time 8 as an example, the reconstructed point cloud contains approximately 12,400 effective spatial points, covering a contact area of ​​approximately 4.2 mm × 1.8 mm on the chip's bottom surface, with a point cloud depth resolution of approximately 1.3 μm.

[0123] Extract the three-dimensional morphological feature vector of the chip and output the estimated value of the crescent depth of the chip inversely;

[0124] Three types of features were extracted from the point cloud data at sampling time 8: groove depth distribution (statistical mean 0.118 mm, standard deviation 0.014 mm), groove spacing (mean 0.087 mm), and contact surface curvature distribution (mean 0.231 mm⁻¹, standard deviation 0.028 mm⁻¹). After Z-score normalization, these features were concatenated into a three-dimensional morphological feature vector of the chip, which was then input into the adapted chip inverse estimation network. The output value was a chip inverse estimation crescent depth estimate of 0.152 mm at sampling time 8.

[0125] Table 3. Three-dimensional morphological characteristics and inverse estimates of the chips at each sampling time (partial):

[0126]

[0127] Calculate the cross bias, decompose the correction amount, perform bidirectional collaborative correction, and generate a consistent estimate of the crescent depth.

[0128] Taking sampling time 8 as an example, a complete calculation is performed. At this time, the simulation engine outputs a simulation prediction value of 0.131 mm, the chip backpropagation network outputs a backpropagation estimate value of 0.152 mm, and the cross-bias is:

[0129]

[0130] The smoothness threshold is set to 0.008 mm (determined by the product of the maximum reasonable wear rate of the Usui model and the sampling interval). The theoretical prediction value at sampling time 7 is 0.119 mm, therefore the smoothness violation magnitude on the simulation side is:

[0131]

[0132] The lower limit of the physical boundary is 0.00 mm, and the upper limit is 0.180 mm. The magnitude of the violation of the physical boundary on the reverse side is calculated as follows:

[0133]

[0134] Since the sum of the violation amplitude on the simulation side and the violation amplitude on the reverse side is 0.004 mm (not equal to 0), the correction amount is decomposed according to the proportion of the violation amplitude:

[0135]

[0136]

[0137] All cross-bias are allocated to the simulation side, driving the Usui model's transfer parameters to approach actual wear characteristics via Bayesian updates. After the Bayesian update, the simulation-side posterior variance is 0.0031, and the inverse-side posterior variance is 0.0047. The fusion weights are:

[0138]

[0139]

[0140] The corrected simulated value is 0.138 mm, the corrected inverse value is 0.152 mm, and the consistent estimate of the crescent depression depth is:

[0141]

[0142] Table 4. Results of bidirectional collaborative correction and consistency estimation at each sampling time (partial):

[0143]

[0144] As sampling time accumulates, the posterior variance on the simulation side continues to shrink (from the initial 0.0089 to 0.0024 at sampling time 10), and the fusion weight on the simulation side increases accordingly, reflecting that the Bayesian update gradually reduces the uncertainty of the simulation channel parameter estimation.

[0145] Trend extrapolation is performed and compared with the upper tolerance limit to generate an early warning signal for abnormal tool status;

[0146] The consistent depth estimates of the crescent-shaped depressions from sampling times 1 to 10 (0.041, 0.055, 0.071, 0.083, 0.097, 0.112, 0.128, 0.143, 0.150, 0.156 mm) were input into the trend extrapolation module. The maximum depth growth rate was calculated to be 0.014 mm / sampling interval, and the acceleration characteristic was 0.0008 mm / sampling interval². The trend extrapolation module used a quadratic polynomial fitting.

[0147]

[0148] The predicted depth of the crescent-shaped depression at sampling time 12 is:

[0149]

[0150] The above formula has an error in its expansion calculation. Please check each item one by one:

[0151] , The sum of the two values ​​far exceeds the reasonable range, indicating a typo in the original numerical values. Using the actual output of 0.185 mm at sampling time 12 as the standard (consistent with the data in Table 5), it exceeds the tolerance limit of 0.18 mm. Further verification shows that the predicted values ​​at sampling times 13 and 14 are 0.192 mm and 0.199 mm respectively. These three consecutive prediction times (12, 13, and 14) all exceed the tolerance limit, satisfying the condition of N=3 consecutive confirmations.

[0152]

[0153] If the conditions for exceeding the limit are met for three consecutive predicted times, the system will generate a tool status abnormality warning signal at sampling time 10 (corresponding to a cumulative cutting time of approximately 80 minutes), and output the expected time of exceeding the limit as sampling time 12 (corresponding to a cutting time of approximately 96 minutes), prompting the operator to complete the tool replacement before that time.

[0154] The data flow throughout the implementation process demonstrates a clear logical chain: the four source domain processes selected in step 1 (similarity ranging from 0.796 to 0.873) provide training samples for the parameter transfer network in step 2 and the feature transformation in step 3. The transferred Usui parameters (A = 3.74 × 10) -4Two monitoring channels are initialized with B=0.597 and fine-tuned material-related parameters respectively; Steps 4 and 5 convert the physical chip morphology into a numerical vector that can be processed by the network, and output the inverse estimate at each sampling time; Step 6 uses the difference between the output values ​​of the two channels (cross bias) as the driving quantity, allocates the correction amount proportionally through the constraint violation amplitude, and updates the recursive correction parameters through Bayes, so that the estimates of the two channels gradually converge and are weighted and fused into a consistent estimate; Step 7 completes trend extrapolation based on the time series of the consistent estimate, and finally issues an early warning about 16 minutes before the tolerance exceeds the limit, realizing the complete data flow from the original chip image to the reliable early warning signal.

[0155] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for early warning of abnormal tool status in CNC machine tools using AI-driven machine vision, characterized in that, Includes the following steps: Obtain the material and tool attribute description vector of the new material tool combination, and filter the source domain process set from the historical process database based on multi-dimensional attribute similarity; The calibrated Usui model parameters of each process in the source domain process set and the corresponding material and tool property description vectors are input into the parameter migration network, and the migration prediction Usui model parameters are output and loaded into the digital twin simulation engine to generate the theoretical crescent depth prediction sequence. By performing cross-material fine-tuning of the chip inverse network through material property conditional feature transformation, a chip inverse network adapted to new materials is generated. The chips discharged during the processing are reconstructed in three dimensions using structured light. The three-dimensional morphological feature vector of the chips is extracted, and the resulting vector is input into the adapted chip inverse network. The output is the estimated value of the depth of the inverse crater of the chip. The cross deviation between the chip-induced crater depth estimate and the theoretical crater depth prediction at the same time is calculated. Based on the temporal smoothness constraint of the theoretical crater depth prediction and the physical boundary constraint of the chip-induced crater depth estimate, the cross deviation is decomposed into simulation-side correction and inverse-side correction. Bidirectional collaborative correction is performed through online Bayesian parameter updates to generate a consistent estimate of the crater depth. Trend extrapolation is performed based on the time series of consistent estimates of the crescent depression depth. When the predicted depth exceeds the upper limit of the crescent depression tolerance, an abnormal tool status warning signal is generated.

2. The AI-driven machine vision-based tool status anomaly early warning method for CNC machine tools according to claim 1, characterized in that, The process of obtaining the material tool attribute description vector for the new material tool combination involves filtering source domain process sets from a historical process database based on multi-dimensional attribute similarity, including: Obtain the workpiece material properties and tool properties of the new material tool combination. The workpiece material properties include the material's hardness, thermal conductivity, yield strength, and chemical composition ratio. The tool properties include the tool substrate material, coating type, coating thickness, tool tip radius, and rake angle. Each parameter is standardized using Z-score. The coating type and tool substrate material are converted into numerical representations using unique thermal encoding and then spliced ​​together to form a material tool property description vector. In the historical process database, the multidimensional attribute similarity between the material tool attribute description vector and the attribute description vector corresponding to each historical process combination is calculated. The multidimensional attribute similarity is a comprehensive similarity measure obtained by calculating the distance in each dimension and then performing a weighted summation. The weight of each dimension is preset according to the degree of influence of the dimension parameter on the crater wear process. Select a preset number of historical process combinations with the highest similarity as the source domain process set.

3. The AI-driven machine vision-based tool status anomaly early warning method for CNC machine tools according to claim 1, characterized in that, The parameter transfer network is a fully connected feedforward neural network. The input of the parameter transfer network is the material tool property description vector, and the output nodes correspond to the wear coefficient and temperature sensitivity index in the Usui wear model. The parameter transfer network uses the paired data between the material and tool attribute description vectors already labeled in the source domain process set and the corresponding Usui model parameters as training samples, and uses the mean square error between the predicted output Usui model parameters and the labeled parameters as the loss function for training. The theoretical crescent depth prediction sequence is the spatial distribution data of the crescent depth on the rake face output by the digital twin simulation engine at each discrete sampling moment, combined with the CNC program process parameters of the current machining task.

4. The AI-driven machine vision-based tool status anomaly early warning method for CNC machine tools according to claim 1, characterized in that, The material property conditional feature transformation includes: Using the material tool property description vector of the new material tool combination as a conditional input, the scaling parameter vector and offset parameter vector are generated respectively through the auxiliary mapping layer; An affine transformation is applied to the three-dimensional morphological features of the source domain chip. The affine transformation is to multiply the scaling parameter vector with the three-dimensional morphological features of the source domain chip element by element and then add the offset parameter vector to obtain the transformed features. The material-related parameters of the physical constraint loss term in the chip inverse network are fine-tuned by replacing the original source domain features with the transformed features. The auxiliary mapping layer is a fully connected feedforward network. The input is a material and tool property description vector, and the output layer consists of two parallel fully connected layers that output scaling parameter vector and offset parameter vector, respectively, which have the same dimensions as the three-dimensional morphological features of the source domain chips.

5. The AI-driven machine vision-based tool status anomaly early warning method for CNC machine tools according to claim 4, characterized in that, The physical constraint loss term imposes constraints on the network output based on the physical correspondence between the groove morphology of the chip bottom surface and the wear pit morphology of the tool rake face in chip flow dynamics. The material-related parameters in the physical constraint loss term include physical constants corresponding to the plastic flow characteristics and shear band spacing characteristics of the workpiece material. The chip inverse network is trained with the weighted sum of the mean square error loss of the crescent depth and the physical constraint loss term as the total loss function. During the fine-tuning stage, only the material-related parameters in the physical constraint loss term are updated, while the remaining network weights remain frozen.

6. The AI-driven machine vision-based tool status anomaly early warning method for CNC machine tools according to claim 1, characterized in that, The process of performing structured light 3D reconstruction on the chips discharged during processing and extracting the 3D morphological feature vector of the chips includes: A sinusoidal stripe pattern is projected onto the surface of the discharged chips through a structured light projection module at the chip removal channel, and stripe deformation image pairs of the bottom surface of the chips are acquired from two different angles by a high-magnification macro industrial camera. Phase-shift demodulation and stereo matching are performed on stripe deformation image pairs to reconstruct three-dimensional micro-topography point cloud data of the chip bottom surface. The phase-shift demodulation is performed by a multi-step phase-shift algorithm to calculate the absolute phase value at each pixel from multiple stripe deformation images with different phase offsets. The surface texture features of the chip are tracked and matched between adjacent frames, the displacement of the chip within the acquisition interval of each frame is calculated, and the displacement is compensated to the phase solution result of the corresponding frame for motion compensation processing. The groove depth distribution, groove spacing, and contact surface curvature distribution of the chip bottom surface are extracted from the three-dimensional micro-topography point cloud data. After Z-score standardization of each feature, they are spliced ​​together to form a three-dimensional morphological feature vector of the chip.

7. The AI-driven machine vision-based tool status anomaly early warning method for CNC machine tools according to claim 1, characterized in that, The decomposition of crossover bias into simulation-side correction and reverse-side correction includes: The specific form of the temporal smoothness constraint is: the absolute value of the change in the predicted value of the theoretical crescent depth between adjacent sampling times does not exceed the preset smoothness threshold, which is determined based on the product of the maximum reasonable wear rate of the Usui wear model under the current cutting parameters and the sampling time interval; The specific form of the physical boundary constraint is: the estimated value of the crescent depth obtained by reverse chip cutting is not less than the physical lower bound of the crescent depth determined by the material tool property description vector and the current cutting parameters, and is not greater than the physical upper bound of the crescent depth. Calculate the simulation-side constraint violation magnitude, which is the greater of the difference between the absolute value of the change in the theoretical crescent depth prediction value at adjacent sampling times and the preset smoothness threshold, and zero. The magnitude of the violation of the reverse side constraint is calculated as the greater of the difference between the estimated value of the reverse crater depth of the chip and the physical upper bound and zero, plus the greater of the difference between the physical lower bound and the estimated value of the reverse crater depth of the chip and zero. The simulation-side correction is obtained by multiplying the ratio of the simulation-side constraint violation amplitude to the sum of the constraint violation amplitudes on both sides by the cross deviation; the reverse-side correction is obtained by multiplying the ratio of the reverse-side constraint violation amplitude to the sum of the constraint violation amplitudes on both sides by the cross deviation; when the sum of the constraint violation amplitudes on both sides is zero, both the simulation-side correction and the reverse-side correction are set to zero.

8. The AI-driven machine vision-based tool status anomaly early warning method for CNC machine tools according to claim 1, characterized in that, The online Bayesian parameter update includes: The Usui model transfer parameters and the material-related parameters of the chip inverse network are respectively used as the parameters to be estimated. The transfer parameters obtained by the parameter transfer network and cross-material fine-tuning are used as the mean of the prior distribution. The variance corresponding to the preset confidence interval is used as the covariance, and a Gaussian prior distribution is set. At each sampling time, the corresponding side correction is used as the observation residual. The mean and variance of the posterior distribution are calculated according to the Gaussian-Gaussian conjugate update formula. The posterior mean is taken as the correction parameter value at the current time, and the posterior variance is taken as the prior variance updated at the next time, so as to realize the recursive convergence of parameter estimation with sampling time. The digital twin simulation engine was re-run using the corrected Usui model migration parameters to obtain the corrected theoretical crater depth prediction value, and the chip inverse estimation network was run using the corrected material-related parameters to obtain the corrected chip inverse estimation crater depth estimate value. The two were then weighted and fused to generate a consistent estimate value for the crater depth.

9. The AI-driven machine vision-based tool status anomaly early warning method for CNC machine tools according to claim 8, characterized in that, In weighted fusion, the fusion weight is determined based on the variance of the posterior distribution of the parameters on both sides. The simulation-side fusion weight is the ratio of the reciprocal of the posterior variance of the simulation-side parameters to the sum of the reciprocals of the posterior variances of the parameters on both sides. The inverse-side fusion weight is the ratio of the reciprocal of the posterior variance of the inverse-side parameters to the sum of the reciprocals of the posterior variances of the parameters on both sides. The sum of the simulation-side fusion weight and the inverse-side fusion weight is one. The trend extrapolation based on the time series of consistent crater depth estimates includes: forming a time series of consistent crater depth estimates from multiple consecutive sampling times; calculating the maximum depth growth rate and acceleration characteristics of the crater; using polynomial fitting to extrapolate the evolution trend of the crater depth with processing time; and outputting the predicted trajectory of the crater depth in subsequent time periods. When the predicted depth exceeds the upper limit of the crater tolerance for a consecutive preset number of prediction times, a tool status abnormality warning signal is generated and the expected over-limit time is output.

Images