An artificial intelligence-based automatic detection and compensation method for wear of a numerical control machine tool

Abstract

The application provides an artificial intelligence-based automatic detection and compensation method for wear of a numerical control machine tool, comprising: collecting cutting force data of a high-curvature area in real time through multi-sensor fusion to obtain cutting force fluctuation characteristics; monitoring and obtaining cutting temperature data of the high-curvature area, extracting a cutting temperature change rate, judging whether the temperature exceeds a preset threshold, and if so, triggering an alarm mechanism and adjusting cutting parameters; combining the cooperative evolution relationship between wear and temperature, online monitoring data, and processed time to predict a tool wear rate, and generating a wear prediction curve for a preset time period; performing trend analysis and feature extraction on the wear prediction curve to obtain wear parameter changes of the tool within the preset time, and if the prediction curve shows that the wear parameter changes at a future time point will exceed a preset critical value, adjusting the cutting parameters to generate a target cutting parameter combination.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of numerical control machine tools, and in particular to a numerical control machine tool wear automatic detection and compensation method based on artificial intelligence. BACKGROUND

[0002] In the numerical control machine tool machining process of the new CVT gearbox thrust wheel, due to the complex thrust wheel groove type surface and the large difference of cutting parameters at different positions, the tool wear presents a special problem of rapid local wear. When continuous cutting is performed in the high-curvature area, the tool is subjected to a complex stress state and thermodynamic environment, which easily accelerates wear. In order to accurately control the machining quality, it is necessary to monitor the cutting process in real time and dynamically adjust the machining parameters. However, due to the geometric characteristics of the high-curvature area, it is difficult for a traditional single sensor to accurately capture the changes in the cutting state. Although multi-sensor fusion technology can provide more comprehensive data, how to effectively integrate and analyze these data to realize real-time evaluation of the tool state is still a challenge. In addition, even if the tool wear trend can be accurately predicted, how to adjust the cutting parameters while ensuring the machining accuracy is also a complex problem. The adjustment of the feed speed and the cutting depth needs to consider the workpiece material characteristics, the current state of the tool and the machining requirements and other factors, and these factors will dynamically change with the progress of the cutting process. Therefore, how to establish an intelligent machining system that can comprehensively consider multi-source information, respond in real time and adaptively adjust, so as to maximize the tool life and stability of the machining quality in the high-curvature area, has become a technical problem to be solved. SUMMARY

[0003] The application provides a numerical control machine tool wear automatic detection and compensation method based on artificial intelligence, mainly comprising:

[0004] The cutting force data of the high-curvature area is collected in real time through multi-sensor fusion to obtain the cutting force fluctuation characteristics;

[0005] The cutting temperature data of the high-curvature area is monitored and obtained, the cutting temperature change rate is extracted, it is judged whether the temperature exceeds the preset threshold, if yes, an alarm mechanism is triggered and the cutting parameters are adjusted, the cutting parameters including the tool feed speed and the cutting depth;

[0006] The vibration signal in the cutting process is collected in real time, and the vibration frequency and amplitude change of the tool are obtained through frequency spectrum analysis, the stress distribution is obtained by combining the cutting force fluctuation characteristics and the cutting temperature data and using the finite element analysis method, the relationship model between the cutting force, the cutting temperature, the tool vibration signal and the stress distribution is established, and the stress concentration position of the high-curvature area is determined;

[0007] If the stress concentration position is continuously cut for a long time, the cutting force and vibration signal data evaluation of off-line and on-line monitoring can evaluate the local tool wear rate and wear morphology change, and according to the historical cutting force and vibration signal and the current cutting temperature data, the synergistic evolution relationship between wear and temperature is analyzed;

[0008] Combining the synergistic evolution relationship between wear and temperature, the on-line monitoring data and the processed time, the tool wear rate is predicted, and the wear prediction curve in the preset time period is generated;

[0009] The trend analysis and feature extraction are performed on the wear prediction curve to obtain the wear parameter change of the tool in the preset time, and if the prediction curve shows that the wear parameter change at a future time point will exceed the preset critical value, the cutting parameters are adjusted to generate a target cutting parameter combination.

[0010] According to the target cutting parameter combination, the wear compensation area is identified according to the set control period, the wear state of the tool and the machining quality of the workpiece are evaluated, so as to calculate the compensation amount of the cutting parameter, and the tool movement trajectory of the wear compensation area is adjusted in real time according to the compensation amount.

[0011] Further, the cutting force data of the high curvature area is collected in real time by multi-sensor fusion to obtain the cutting force fluctuation characteristics, including: obtaining the sensor coordinate matrix according to the arrangement position relationship of the multi-sensor array on the cutting surface, extracting the force signal sensitivity coefficient from the sensor calibration data record, generating the sensor data acquisition time stamp sequence for the sampling frequency, and using the coordinate transformation matrix to perform spatial mapping calibration on the sensor original data. According to the sensor acquisition period and the data cache threshold, the data block is divided, the force signal time stamp identifier is extracted from the data block, the force signal amplitude noise is eliminated by using the sliding mean filter, and the multi-group sensor data is time-aligned based on the time stamp identifier. The curvature threshold is set according to the curvature gradient value of the cutting surface, the cutting surface is divided into high and low curvature regions by using the curvature threshold, the corresponding sensor data sequence is extracted from the divided region, and the wavelet coefficient of the force signal amplitude sequence is calculated by using wavelet transform. According to the arrangement position of the sensor, the data fusion weight coefficient is determined, the wavelet coefficients of multiple sensors are weighted and summed by using the weight coefficient, and the fused force signal sequence is reconstructed from the weighted result. The local variance value of the force signal sequence is calculated by using the sliding window to obtain the cutting force fluctuation rate curve, the clustering center is set based on the mean and standard deviation of the fluctuation rate curve, the fluctuation rate data is segmented and clustered by using the Euclidean distance measurement, and the cutting force fluctuation feature vector is extracted from the clustering result.

[0012] Further, the cutting temperature data of the high curvature area is monitored and acquired, the cutting temperature change rate is extracted, it is judged whether the temperature exceeds the preset threshold, if yes, the alarm mechanism is triggered and the cutting parameters are adjusted, the cutting parameters include the tool feed speed and the cutting depth, including: the machining area is uniformly grid divided according to the curvature gradient value of the cutting surface, the thermocouple sensor array is arranged at the grid nodes to collect temperature data, the grid temperature field distribution map is generated by using cubic spline interpolation, and the temperature field time sequence is calculated through the temperature sampling rate setting. The temperature field data is noise-removed by using the Bayesian filtering state equation, the measurement error is compensated by using the temperature sensor calibration curve, the temperature field gradient distribution is calculated by using the central difference, and the high curvature area boundary is judged based on the gradient threshold. The temperature change rate of adjacent sampling points is calculated according to the temperature field time sequence, the temperature mutation point is detected by using the sliding variance window, the temperature change rate curve is obtained from the mutation point, and the temperature fluctuation characteristics are extracted for the change rate curve. The abnormal point coordinates are located when the temperature fluctuation amplitude exceeds the threshold range by comparing the temperature fluctuation characteristics with the preset temperature threshold, and the temperature overrun area range is determined based on the abnormal point coordinates. The feed speed adjustment interval is divided according to the temperature overrun area range, the cutting depth constraint condition is established by using the temperature change rate, and the optimal parameter combination is calculated based on the linear programming solver. The feed speed is adjusted in sections by using the parameter combination, the feed amount is limited by the cutting depth constraint, and the tool motion trajectory is obtained from the adjusted parameter sequence.

[0013] Further, the vibration signal in the cutting process is collected in real time, and the vibration frequency and amplitude change of the tool are obtained by frequency spectrum analysis, the stress distribution is obtained by using the finite element analysis method in combination with the cutting force fluctuation characteristics and the cutting temperature data, the relationship model between the cutting force, the cutting temperature, the tool vibration signal and the stress distribution is established, and the stress concentration position of the high curvature area is determined, including: the acceleration sensor data is collected according to the vibration sampling frequency, the vibration signal envelope is obtained by Hilbert transform, the vibration frequency and amplitude peak value are extracted from the frequency spectrum, and the vibration signal is decomposed into a low-frequency fundamental wave band and a high-frequency harmonic wave band by using wavelet packet decomposition. The boundary conditions are constructed by using the force signal peak points in the cutting force fluctuation characteristic curve and the temperature field change rate extreme points, the tetrahedral grid element is constructed according to the grid size parameters, the stress tensor boundary value is given to the grid nodes, and the element stress component is calculated based on the linear elastic constitutive equation. The root mean square value of the vibration signal is calculated according to the vibration signal envelope curve, the force signal standard deviation is calculated by the cutting force fluctuation characteristic curve, the temperature gradient vector is calculated by using the temperature field change rate data, and the stress influence factor matrix is constructed by using the three groups of data. The stress influence factor matrix is mapped by using the neural network, the stress distribution predictor is trained and established, the stress field distribution function is obtained from the predictor output, and the stress concentration position is determined by using the stress field gradient.

[0014] Further, if the stress concentration position is continuously cut for a long time, the cutting force, vibration signal data evaluation of off-line and on-line monitoring, local tool wear rate and wear morphology change, and according to the historical cutting force and vibration signal and the current cutting temperature data, the cooperative evolution relationship between wear and temperature is analyzed, including: according to the cutting time record to set the fixed acquisition period, the time series of force signal and vibration signal in the historical cutting process is obtained through the off-line database, the real-time cutting force and vibration signal is collected from the on-line sensor, and the high-frequency noise is removed by using wavelet multi-scale decomposition. A grid monitoring area is constructed for the stress concentration position, the wear initial position is marked by using the cutting force fluctuation degree peak point, the wear contour boundary is identified by using the vibration amplitude change curve, and the wear area value is calculated according to the boundary contour. The wavelet energy coefficient is used to calculate the local wear rate, the wear morphology characteristics are analyzed based on the vibration signal envelope spectrum, the wear evolution speed is calculated by using the cumulative wear amount, and the wear development trend is obtained from the wear speed curve. The long short-term memory network is used to carry out feature decomposition on the historical cutting force signal sequence, the vibration signal time sequence features are extracted through the self-attention mechanism, and the temperature change function is constructed by using the temperature acquisition time sequence. According to the cutting parameter matrix, the feature space is constructed, the support vector regressor is used to establish the mapping relationship between the wear rate and the temperature change function, and the wear temperature coupling change is predicted by using the recurrent neural network. Based on the wear temperature coupling change rule, the state transition matrix is established, the Markov chain is used to describe the wear and temperature cooperative evolution process, and the evolution model parameters are obtained from the state probability distribution.

[0015] Further, the historical cutting force and vibration signal data are acquired, the historical cutting force values and vibration values are extracted, the current cutting temperature of the current monitoring system is collected, the current temperature value is recorded, the regression analysis method is adopted, the relationship model between the cutting force, the vibration value and the temperature value is established, if the fitting degree of the relationship model meets the preset threshold value, the cooperative evolution relationship between the wear amount and the temperature value is determined, including: according to the historical data period division sampling time period, the force value peak-valley ratio, the force value standard deviation and the force value root mean square characteristics are extracted from the cutting force sequence, the amplitude mean, the amplitude variance and the amplitude frequency characteristics are calculated through the vibration value sequence, the characteristic data are time-aligned by using linear interpolation. Real-time temperature numerical values are collected for the temperature monitoring points, data recording periods are set based on the temperature collection frequency, the temperature average, the temperature fluctuation amplitude and the temperature change slope are calculated from the temperature data sequence, the temperature data are denoised by using the db4 wavelet base function. The least square method is used to establish a multivariate linear equation group for the cutting force characteristic value, the vibration amplitude characteristic and the temperature characteristic, the equation fitting degree is calculated by the R square value, and the feature correlation matrix is constructed by using the Pearson correlation coefficient. According to the feature correlation matrix, the features with a significant coefficient greater than a preset value are screened, the feature mapping function is established by using the polynomial regression, the polynomial order is determined by the significance test, and the variable weight coefficient is obtained from the mapping function. The linear combination of the wear amount and the temperature value is calculated based on the variable weight coefficient, the time series prediction function is established by using the recurrent neural network, if the prediction function R square value exceeds the preset threshold value, the neural network weight matrix is obtained. The state transition equation is constructed by using the neural network weight matrix, and the evolution function is generated according to the cooperative change rule of the wear amount and the temperature value, and the coupling coefficient matrix is obtained from the evolution function.

[0016] Further, the tool wear rate is predicted by combining the synergistic evolution relationship of wear and temperature, online monitoring data and machining time, and a wear prediction curve for a preset time period is generated, including: the online collected data is time-sequentially segmented according to a monitoring data cycle, the historical machining amount is counted from a machining time period, the data sequence is block-processed by using a fixed sampling interval, and the temperature data is linearly interpolated to obtain an equal-interval sampling sequence. The prediction time span is discretized by using a uniform time step, the root mean square value of cutting force, the vibration amplitude and the temperature gradient feature are extracted for the discretized time points, the feature sequence is standardized based on a long short-term memory network. The feature sequence is correlation-screened by using a Pearson correlation coefficient, the change rate of the feature sequence is calculated by using a central difference, the wear rate prediction value is calculated according to a synergistic evolution model parameter, and the trend feature is extracted from the wear rate sequence, wherein the feature change rate calculated by using the central difference reflects the dynamic characteristics of the wear process. A cubic spline function is established for the wear rate trend feature, the spline coefficients are determined by using a least square method, and a continuous wear prediction curve is generated according to the spline function. The wear prediction curve is noise-filtered based on a recurrent neural network, the wear rate average value is calculated by using a sliding mean window, and a warning judgment threshold is set according to the wear rate standard deviation. The prediction error statistics are calculated by using the historical wear data, the prediction confidence interval is established by using an error distribution function, and the wear prediction accuracy evaluation index is obtained.

[0017] Further, trend analysis and feature extraction are performed on the wear prediction curve to obtain the wear parameter variation of the tool within a preset time. If the prediction curve shows that the wear parameter variation at a future time point will exceed a preset critical value, the cutting parameters are adjusted to generate a target cutting parameter combination, including: dividing the wear prediction curve into segments according to the wear prediction length, determining the trend inflection point position using the zero point of the second derivative, calculating the segmented linear regression coefficient to obtain the wear slope value using the least squares method, and calculating the wear fluctuation degree sequence by a fixed length sliding window. Fourier transform is performed on the wear fluctuation degree sequence to extract the main frequency component, the wear amount per unit time is calculated from the wear prediction curve segment, and the wear feature spectrum is extracted based on the convolutional neural network. If the peak value of the feature spectrum exceeds a preset critical threshold, it is determined as a critical wear point. In another embodiment, the extracted wear features can also include the wear rate variation trend, the average wear rate, and the wear rate fluctuation. A three-layer parameter classification structure of feed speed, spindle speed, and cutting depth is constructed using a decision tree, the parameter optimization order is determined by the maximum information gain criterion, the weight coefficient is set according to the parameter sensitivity, and the parameter adjustment amount is calculated from the weight coefficient matrix. The parameter adjustment amount is limited by the cutting parameter constraint function, and the parameter combination is optimized using the particle swarm algorithm. The parameter sequence that meets the constraint condition is obtained from the optimization result. A cutting process evaluation function is constructed based on the parameter sequence, the parameter combination fitness is calculated by weighted summation, the optimal parameter combination is selected according to the fitness ranking, and the target cutting parameter is generated from the optimal combination. The stability of the target cutting parameter is verified, and the parameter perturbation is simulated using the Monte Carlo method. The parameter stability interval is obtained from the simulation result.

[0018] Further, the current wear rate is calculated, a time series analysis algorithm is used, historical wear data is combined to construct a prediction curve model, a wear rate prediction value at a future preset time point is calculated, if the prediction value exceeds a preset critical value, a parameter optimization algorithm is called to generate a plurality of sets of cutting parameter combinations, and a target cutting parameter combination is searched within a preset parameter range, including: sampling the wear data in a fixed time window according to the historical data period, extracting a training data sequence from the historical data length, establishing a wear rate prediction model using a long short-term memory network, and calculating the prediction fitting accuracy by root mean square error. The time domain statistics of the wear rate sequence is calculated, the frequency domain characteristic coefficients are extracted by wavelet transform, and the time series state equation is constructed based on the recurrent neural network. The wear rate prediction value at the preset time point is calculated from the state equation. The warning level is calculated according to the difference between the wear rate prediction value and the critical wear value, and if the warning level exceeds the preset threshold, the parameter optimization link is triggered. Based on the preset parameter range, the feed speed interval, the cutting depth interval, and the spindle speed interval are constructed, the parameter interval is discretized using the uniform quantization method, and the parameter combinations are screened according to the machining process constraints. The parameter combinations are evaluated using a cutting load constraint function, the upper limit of the parameters is set according to the cutting temperature limit, and the lower limit of the parameters is determined by the machining accuracy requirement, and the feasible parameter space is obtained from the constraint conditions. The feasible parameter space is searched using a genetic algorithm, the parameter combination mode is controlled by the crossover probability, the parameter variation amplitude is adjusted based on the mutation probability, and the optimal cutting parameter combination is obtained from the search results.

[0019] Further, according to the target cutting parameter combination, the wear area to be compensated is identified, the wear state of the tool and the machining quality of the workpiece are evaluated, and the compensation amount of the cutting parameter is calculated, and the tool motion trajectory of the wear area to be compensated is adjusted in real time according to the compensation amount, including: segmenting the target cutting parameter combination according to the control period length, arranging the wear detection array in a grid manner, extracting the wear profile features from the detection image using a convolutional neural network, and measuring the roughness value of the machined surface using a laser profilometer. Geometric parameter sets are extracted from the wear profile features, including wear area, wear depth, and wear bandwidth, and the wear parameters are dynamically estimated using Kalman filtering, and a wear state vector is constructed from the estimation results. The parameter compensation increment is calculated based on the wear state vector, the feed speed correction amount and the cutting depth correction amount are generated using a linear compensation function, and the optimal compensation parameter combination is obtained by a quadratic programming solver. The compensation parameters are smoothed using cubic spline interpolation, the feed speed correction amount is used to adjust the feed axis speed command, and the cutting depth correction amount is used to calculate the tool position compensation amount. A compensation trajectory equation is established for the tool position compensation amount, a five-order polynomial is used to fit the trajectory curve, and a position compensation instruction sequence is generated from the fitted curve. The compensation instruction sequence is executed by a position feedforward controller, the actual compensation error is calculated using servo feedback signals, and the compensation gain coefficient is adjusted from the error signal.

[0020] The technical scheme provided by the embodiment of the present application can include the following beneficial effects:

[0021] The application discloses an automatic detection and compensation method for wear of a numerical control machine tool based on artificial intelligence. In view of the problem that tool wear is difficult to accurately control in the cutting process of a high-curvature area, the application collects cutting force, temperature and vibration data in real time through multi-sensor fusion, establishes a relationship model of cutting parameters and stress distribution, and locates the stress concentration position. In combination with online monitoring data and historical data, a cooperative evolution model of wear and temperature is constructed, tool wear rate is predicted, and a wear prediction curve is generated. Based on the prediction result, the application dynamically adjusts the cutting parameters, and compensates the tool movement track in real time, so as to realize intelligent control of tool wear in the high-curvature area. The method can effectively improve the machining precision of the high-curvature area, prolong the service life of the tool, reduce the machining error, and improve the machining efficiency and product quality. BRIEF DESCRIPTION OF DRAWINGS

[0022] Fig. 1 A flowchart of the automatic detection and compensation method for wear of a numerical control machine tool based on artificial intelligence.

[0023] Fig. 2 A schematic diagram of the automatic detection and compensation method for wear of a numerical control machine tool based on artificial intelligence. DETAILED DESCRIPTION

[0024] The technical scheme in the embodiments of the present application will be described clearly and in detail below with reference to the drawings in the embodiments of the present application. The described embodiments are only some of the embodiments of the present application.

[0025] As Figs. 1-2 , the automatic detection and compensation method for wear of a numerical control machine tool based on artificial intelligence specifically can include:

[0026] S101, cutting force data of a high-curvature area is collected in real time through multi-sensor fusion technology, and cutting force fluctuation characteristics are extracted; a coordinate matrix is obtained according to the spatial arrangement of a sensor array; spatial mapping calibration is performed on the collected data by using coordinate transformation, and noise is eliminated; subsequently, the area is divided based on a curvature threshold value, and force signal sequences are reconstructed through wavelet analysis and weighted fusion; a fluctuation rate curve is calculated to determine the cutting force fluctuation characteristics.

[0027] S1011、In the cutting process in the high-curvature area, the multi-sensor array is used to collect cutting force data and generate a sensor coordinate matrix, wherein the sensors are uniformly distributed along the edge of the cutting surface to form a ring-shaped monitoring network, the sensitivity coefficients of each sensor are extracted according to the calibration test and the spatial position is recorded, the original data is mapped to a unified coordinate system through coordinate transformation, then the mapped data is noise-removed by using a sliding mean filter and a time sequence alignment of multiple groups of data is completed according to the time stamp.

[0028] S1012、According to the curvature gradient of the cutting surface, a threshold value is set and the high and low curvature areas are divided, the force signal sequence is extracted from the high curvature area, the wavelet coefficients are calculated by using wavelet transform, then the fusion weight coefficients are determined according to the sensor positions and the wavelet coefficients are weighted and summed to reconstruct the force signal sequence, the local variance is calculated by using a sliding window to generate a cutting force fluctuation rate curve, and the clustering centers are set based on the statistical characteristics of the fluctuation rate curve and the cutting force fluctuation feature vector is extracted by using the Euclidean distance measurement.

[0029] In the embodiment of the application, the change of the cutting force is monitored in real time by using the multi-sensor array, and the comprehensiveness of the data in the high-curvature area is ensured. The sensor array is usually composed of 8 piezoelectric force sensors, which are distributed at an angle of 45 degrees along the cutting surface, the sampling frequency is set to 1000 Hz, and each data block contains 512 sampling points. The sensitivity coefficient ranges from 0.95 to 1.05 in the calibration, the filter uses a 15-point sliding mean processing, and the alignment accuracy is better than 0.1 ms. The wavelet transform uses db4 basis function for 4-layer decomposition, the weight coefficient is assigned according to the distance between the sensor and the high-curvature area, the maximum value is 0.3, and the minimum value is 0.05. Further, 3 clustering centers are set according to the mean and standard deviation of the fluctuation rate curve, which respectively represent stable, transition and unstable cutting states, and the feature vector is extracted by using clustering analysis. The experiment shows that the fluctuation rate in the high-curvature area is significantly higher than that in the low-curvature area, and the feature has repeatability, which reflects the influence of curvature on cutting stability and provides a reliable data basis for subsequent wear evaluation. The implementation mode of S101 is not limited to the specific parameters described above, and can be adjusted according to the actual machining scene, such as the number of sensors, the sampling frequency or the filter window size, etc., to adapt to the needs of different working conditions.

[0030] S102、Monitoring and acquiring cutting temperature data in the high-curvature area and extracting temperature change rate features, judging whether the temperature exceeds the preset threshold value and triggering an alarm and parameter adjustment when the threshold value is exceeded, acquiring temperature data by grid division and thermocouple array, generating temperature field distribution and performing noise processing, then calculating temperature gradient and mutation features, and optimizing cutting parameters based on linear programming to control tool feed speed and cutting depth.

[0031] S1021, uniformly grid the machining area according to the cutting surface curvature gradient value and arrange a thermocouple sensor array at the nodes to collect temperature data, use cubic spline interpolation technology to spatially reconstruct the discrete temperature data to generate a continuous grid temperature field distribution map, calculate the temperature field time series by setting the sampling frequency and eliminate measurement noise by using the Bayesian filtering state equation, and at the same time, combine the sensor calibration curve to compensate for errors to improve data accuracy.

[0032] S1022, calculate the temperature gradient distribution by the central difference method for the grid temperature field distribution map and determine the high curvature area boundary, extract the temperature change rate of adjacent sampling points from the temperature field time series and use a sliding variance window to detect temperature abrupt points, generate a temperature change rate curve according to the abrupt points and extract the fluctuation characteristics, then compare the fluctuation characteristics with the preset temperature threshold to locate the temperature overrun area and divide the feed speed adjustment interval.

[0033] S1023, establish a cutting depth constraint condition based on the temperature overrun area range and combine the temperature change rate characteristics, use a linear programming solver to calculate the optimal cutting parameter combination and adjust the feed speed in sections, and at the same time, limit the feed amount by cutting depth constraint and generate the adjusted tool movement trajectory.

[0034] In the embodiment of the application, the cutting temperature data is monitored in real time to provide a basis for process parameter adjustment. For the complex thermodynamic environment of the high curvature area, a grid division scheme with a 10mm spacing is adopted, a K-type thermocouple sensor is arranged at each node, the temperature measurement range covers 0 to 800 degrees Celsius, and the sampling frequency is set to 100Hz. The temperature field is reconstructed using cubic spline interpolation, experimental verification shows that the interpolation error is less than 2 degrees Celsius, ensuring the reliability of the distribution map. The Bayesian filter describes the temperature evolution through the state equation, and combines an observation model with a noise standard deviation of 0.5 degrees Celsius for filtering. The calibration curve is obtained by piecewise fitting of a constant temperature oil tank, and the error compensation accuracy is better than 1 degree Celsius in the full temperature range.

[0035] Further, the temperature gradient is calculated by 5-point central difference, and the threshold is set to 50 degrees Celsius per millimeter to identify the high-curvature boundary. The temperature change rate is calculated by the difference between adjacent points in the time series, and a 32-point sliding variance window is used to detect the mutation point. When the variance exceeds 10, it is determined to be abnormal, and the generated change rate curve shows that the maximum curvature area can reach 120 degrees Celsius per second, which is much higher than the 40 degrees Celsius per second in the flat area. If the temperature fluctuation amplitude exceeds the preset threshold, the abnormal coordinates are located and the out-of-limit area is divided, and the feed speed adjustment interval is set to 500 to 2000 mm / min, and the cutting depth is constrained to 0.1 to 2 mm. The linear programming solver calculates the parameter combination with temperature stability as the target and speed continuity constraint. In the area with a curvature radius less than 5 mm, experiments show that reducing the feed speed to 40% of the initial value and adjusting the cutting depth to below 0.3 mm can effectively suppress temperature rise. The adjustment process uses a segmented smoothing strategy, with a speed change rate limited to within 200 mm / min to avoid parameter mutations causing unstable machining. The cutting depth is adjusted in real time through a compensation quantity positively correlated with the temperature change rate, and the optimized tool path significantly reduces the heat accumulation in the high-curvature area, improving the machining quality and tool life. For real-time processing of temperature out-of-limit, an alarm mechanism is triggered immediately after detecting the mutation point, and the optimized parameters are transmitted to the numerical control system through a preset control cycle to ensure timely adjustment of the machining process. Experimental data show that the temperature control effect in the high-curvature area is better than that of the traditional static parameter scheme, and the temperature distribution uniformity is improved by about 15%, providing reliable support for subsequent wear prediction and compensation.

[0036] S103, real-time acquisition of vibration signals in the cutting process and extraction of tool vibration frequency and amplitude change characteristics through frequency spectrum analysis, combination of cutting force fluctuation data and cutting temperature information, construction of stress distribution model using finite element analysis, establishment of relationship between cutting force, temperature and vibration and stress and determination of stress concentration position in high-curvature area, optimization of stress field calculation through signal processing and deep learning technology to improve positioning accuracy.

[0037] In the embodiment of the application, a piezoelectric acceleration sensor is used to collect vibration signals in real time during the cutting process at a sampling frequency of 20000 Hz. The envelope curve of the vibration signal is generated by Hilbert transform, and the frequency spectrum data is calculated by fast Fourier transform. The frequency and amplitude characteristics of the fundamental frequency below 500 Hz and the high-frequency harmonic in the range of 2000 to 5000 Hz are extracted, and then the signal is divided into low-frequency fundamental wave and high-frequency harmonic two frequency bands by wavelet packet decomposition to enhance the feature resolution. The periodic peak information of the cutting force fluctuation data is provided by the force signal sequence obtained in the previous step, and the temperature data reflects the heat source distribution in the form of gradient vector, which are used as the input basis for stress analysis.

[0038] S1031、According to the vibration signal envelope curve, the root mean square value is calculated, and the standard deviation is extracted from the cutting force fluctuation characteristic curve, and at the same time, the temperature field time sequence is used to generate a temperature gradient vector, and the stress influence factor matrix is constructed by the three groups of data and is standardized to a unified dimension, and then the boundary conditions are given to the tetrahedral grid element based on the linear elastic constitutive equation, and the element stress component is calculated to generate the initial stress distribution data.

[0039] S1032、The stress influence factor matrix is input into the neural network model based on deep learning, the stress field distribution function is generated by feature mapping, and the stress concentration position of the high curvature area is determined, the principal stress size and direction are calculated by using the stress field gradient, and the concentration degree is evaluated by combining the stress intensity criterion, then the positioning accuracy of the stress singular point is optimized by the grid encryption technology, and the stress hot spot area is constructed to analyze the crack propagation trend.

[0040] In the embodiment of the application, the calculation reliability of stress distribution is improved by multi-source data fusion. The root mean square value of the vibration signal represents the overall strength, which is maintained at 0.5 to 1.5g during stable cutting, the cutting force standard deviation reflects the discreteness, which is in the range of 50 to 80N, and the temperature gradient vector changes with the cutting direction, and the angle between the cutting direction and the high curvature area is close to 90 degrees. After these data are integrated in the form of a matrix, they are input as boundary conditions to the finite element model. The tetrahedral grid adopts an adaptive division strategy, and the size of the stress concentration area is reduced to 0.1 millimeter to ensure accuracy, and the element stress component is calculated by the linear elastic equation, considering the material stiffness and deformation characteristics.

[0041] Further, a 5-layer convolutional neural network is used to extract features from the matrix, and a stress distribution predictor is generated by training 3000 sets of experimental data. The stress field distribution function output by the predictor shows that the angle between the principal stress direction of the high curvature area and the cutting direction is 75 to 85 degrees, and the principal stress value is significantly higher than that of the flat area, about 2 to 3 times higher. The stress intensity factor is close to 80% of the material fracture toughness at the singular point, indicating a high risk of crack propagation. Therefore, the area with a curvature change rate greater than a certain threshold is grid-encrypted, the element size is reduced to 0.05 millimeters, the singular point coordinates are accurately positioned to the micron level by the stress field gradient, and the crack propagation rate is calculated based on fracture mechanics to evaluate the stability of the area.

[0042] In practical applications, it is found that the coupling of temperature gradient and stress direction significantly affects the stress distribution in high-curvature regions, especially when the radius of curvature is less than 5 mm. This coupling effect is quantified through the construction of stress hot spot regions, and the stress concentration degree of the hot spot region is closely related to the cutting conditions. To improve the analysis accuracy, the wavelet packet decomposition uses a 4-layer structure to divide the signal into 16 frequency bands, and the extracted harmonic features further enrich the expression ability of the vibration data. Experiments show that the stress model combined with multi-source data can better reflect the changes in the machining state than single signal analysis.

[0043] S104, if the stress concentration position of the high-curvature region experiences long-time continuous cutting, the local wear rate and morphological changes of the tool are evaluated by combining the historical cutting force and vibration signals in the offline database with the real-time data monitored online, a wear and temperature co-evolution model is constructed through multi-scale signal processing and feature extraction to reveal the coupling law, and a neural network is used to predict the wear trend to provide a basis for subsequent compensation.

[0044] In the embodiments of the present application, the force signals and vibration signals during cutting are collected at a fixed period, the offline data are derived from a database storing more than 500 sets of machining records, and the real-time data are collected by online sensors at a frequency of 1000 Hz. For high-frequency interference in the signal, 4-layer wavelet multi-scale decomposition is used to remove noise above 2000 Hz, and the main frequency band reflecting the cutting state is retained. Subsequently, a grid monitoring area with a spacing of 0.1 mm is constructed at the stress concentration position, the wear initial point is marked by the cutting force fluctuation peak value, and the wear boundary is identified by the vibration amplitude change curve, the wear area is calculated and the evolution trend is analyzed. In the experiment, the cutting force peak fluctuation is 350-400 N, the vibration amplitude increases from 0.5 g to 2.5 g, and the wear area can reach 0.8 square millimeters.

[0045] S1041, feature data is extracted from the denoised cutting force and vibration signal time series, the signal characteristics are represented by indicators such as force value peak-to-valley ratio, standard deviation and amplitude mean, time alignment is achieved by linear interpolation, then db4 wavelet basis function is used for further denoising and calculation of temperature change slope and average, a multivariate linear regression model is established based on the least squares method, and the feature mapping function is optimized through significance test to improve the fitting accuracy.

[0046] S1042, the local wear rate is calculated for the wear boundary points, and the morphology features are extracted by the vibration signal envelope spectrum, the wavelet energy system is used to quantify the frequency band distribution change, and the long short-term memory network and self-attention mechanism are used to decompose the time sequence features of the historical signal, then the mapping relationship between wear rate and temperature is established based on the support vector regressor and recurrent neural network, and the coupling change trend is predicted.

[0047] S1043、According to the prediction result, a state transition matrix is constructed and a Markov chain is used to describe the synergistic evolution process of wear and temperature, and the mutual influence of wear evolution and temperature accumulation is quantified, thereby providing a dynamic basis for tool state monitoring.

[0048] In the embodiment of the present application, cutting force features are extracted from historical data, such as the force value peak-to-valley ratio remaining 1.2 to 1.5 in the stable state, the standard deviation fluctuating between 50 to 80 N, the vibration signal amplitude mean value rising from 0.5 g to 2.0 g, and the frequency feature peak value changing between 2000 to 3000 Hz. Linear interpolation is used to align the time axis of multi-source data. Temperature data is collected at a frequency of 5 Hz, with an average value rising from 200 to 550 degrees Celsius, and the change slope showing a phased jump. The multiple linear regression model contains 9 variables, with an R square value of 0.92. The Pearson correlation coefficient shows that the correlation between cutting force and temperature is as high as 0.85, and the significance threshold is set to 0.75. Five key feature pairs are selected. The mapping function uses a 3rd order polynomial regression, and the weight coefficient shows that the cutting force impact accounts for 0.45, reflecting its dominant role in wear.

[0049] Further, the sideband modulation in the 2000 to 3000 Hz frequency band is analyzed using vibration envelope spectrum, and the modulation degree increases with the deepening of wear, with the high-frequency energy proportion rising from 15% to 35%. The local wear rate is initially 0.02 mm / min and then increases exponentially to 0.08 mm / min. The long short-term memory network is configured with 128 hidden layer nodes, combined with the self-attention mechanism to extract long-range dependence features. The support vector regressor uses a radial basis kernel function to map the relationship between wear and temperature, and the recurrent neural network predicts that the wear rate increases by 0.015 mm / min for every 100 degrees Celsius increase in temperature. The state transition matrix defines four states: normal, accelerated, severe, and failure. The Markov chain calculation shows that the transition probability is positively correlated with the temperature change rate. Experimental verification shows that when the temperature exceeds 550 degrees Celsius, the tool enters the severe wear stage, revealing the synergistic evolution mechanism of wear and temperature in the high curvature region. Long-term cutting experiments show that when machining exceeds 30 minutes, the vibration frequency band energy proportion exceeds 40%, the cutting force peak value reaches 450 N, and the remaining life is less than 5 minutes. The prediction error is controlled within 10%. This dynamic modeling method is more adaptable to complex working conditions than traditional static analysis, laying a foundation for real-time monitoring and optimal adjustment of tool wear.

[0050] In practical applications, the collection period and model parameters can be adjusted according to the machining material and cutting conditions, such as shortening the period to 5 minutes for hard alloy machining, or increasing the hidden layer nodes to improve the prediction accuracy, ensuring the adaptability of the model to different scenarios. The construction of the evolution model can also provide feedback for tool design to optimize the cutting edge geometry to slow down the wear rate.

[0051] S105, input the online monitoring data and the machining time into the synergistic evolution model of wear and temperature to predict the tool wear rate and generate a wear prediction curve for a preset time period, extract a feature sequence through time segmentation processing, and optimize the prediction result by using a neural network.

[0052] In the embodiment of the application, online monitoring data is collected every 30 seconds, each cycle contains 600 sampling points, and the machining time statistics show that the cumulative cutting time of a single tool is up to 120 minutes, of which the high-curvature area accounts for about 35%. The data is processed by time segmentation, the data is blocked by fixed sampling interval, and the temperature sequence is resampled to 1Hz interval data by linear interpolation to ensure the consistency of the time axis. The prediction time span is set to 30 minutes, which is discretized by 1 minute step, the root mean square value of cutting force, vibration amplitude and temperature gradient features are extracted, which respectively reflect the energy of force signal, vibration intensity and heat distribution characteristics, and then the long short-term memory network is standardized to eliminate the dimension difference.

[0053] S1051, extract the historical machining amount statistics from the segmented data and calculate the root mean square value of cutting force, vibration amplitude and temperature gradient feature sequence, analyze the coupling relationship between the features by using the Pearson correlation coefficient, and generate the change rate sequence by central difference, then predict the wear rate based on the synergistic evolution model parameters and extract the trend feature to quantify the wear evolution process.

[0054] S1052, construct a cubic spline function for the wear rate trend feature and optimize the fitting coefficient by least squares method to generate a continuous prediction curve, then use a recurrent neural network filter to suppress noise and calculate the average wear rate sequence by a sliding mean window, and set an early warning threshold based on the standard deviation to realize real-time monitoring.

[0055] In the embodiment of the application, the extracted root mean square value of cutting force remains at 300-400N during normal cutting, the vibration amplitude ranges from 0.5 to 2.0g, and the temperature gradient in the high-curvature area can reach 80 degrees Celsius per millimeter. The Pearson correlation coefficient calculation shows that the correlation between the cutting force and the wear rate is 0.85, the correlation between the vibration amplitude is 0.78, and the correlation between the temperature gradient is as high as 0.92, indicating that the temperature has a significant impact on wear. The change rate calculated by the central difference reveals the dynamic characteristics of the wear acceleration stage, for example, the temperature change rate increases from 0.5 degrees Celsius per minute to 2.0 degrees Celsius per minute, reflecting the catalytic effect of heat accumulation on wear.

[0056] The wear rate curve is fitted using a cubic spline function, with 20 control points and optimized coefficients through least squares method, to control the fitting error within 5%. The prediction curve presents a three-stage characteristic: the initial wear rate is stable at 0.02 mm / min, the intermediate period gradually rises to 0.05 mm / min, and the later period shows an exponential growth trend. The recurrent neural network is filtered with a 32-point sliding window to smooth high-frequency fluctuations, and the average sequence is used as the reference, with a 3 times standard deviation as the early warning threshold. Experimental verification shows that when the wear rate exceeds the threshold, the early warning signal can prompt the risk of severe wear 5 to 8 minutes in advance, covering 90% of the test cases.

[0057] Further, based on the analysis of 200 groups of historical data, the prediction error is approximately normally distributed, with a mean of 0 and a standard deviation of 0.008 mm / min. The 95% confidence interval is ±0.016 mm / min of the predicted value, with a short-term prediction accuracy of 95%, a medium-term accuracy of 85%, and a long-term accuracy of more than 75%. This multi-scale prediction capability adapts to the needs of different processing stages and provides a reliable basis for process optimization. In practical applications, the sampling period and feature extraction strategy can be adjusted according to the tool type or workpiece material, such as shortening the step to 15 seconds for hard material processing to capture more subtle changes, or increasing the control points to improve the accuracy of the curve, ensuring the robustness of the prediction model under complex working conditions. By dynamically predicting the wear trend, the tool life is effectively extended and the risk of processing interruption is reduced.

[0058] S106, trend analysis and feature extraction are performed on the wear prediction curve to obtain the change trend of the tool wear parameter within a preset time, and if the prediction shows that the wear parameter exceeds the preset critical value at a future time point, the cutting parameters are optimized through a multi-level algorithm and a target combination is generated, thereby realizing dynamic adjustment of the machining process.

[0059] In the embodiment of the present application, the wear prediction curve is divided into 30-minute prediction periods, with each 5-minute period as an analysis segment. The trend inflection point is located by calculating the second derivative zero point. For example, in the 15 to 20 minute interval, the wear rate increases from 0.02 mm / min to 0.05 mm / min, indicating the start of accelerated wear. The linear regression coefficients of each segment are calculated using the least squares method to quantify the slope change, and the wear amount variance is calculated through a 32-point sliding window to generate a sequence reflecting the fluctuation. Fourier transform is used to extract the main frequency components of the fluctuation sequence, which are concentrated in the 0.05 to 0.1 Hz range, revealing the periodic characteristics of wear. Subsequently, a convolutional neural network is used to process the main frequency data to generate a time-frequency spectrum. When the proportion of high-frequency components increases from 15% to 35%, it is determined that the tool is approaching a critical wear state.

[0060] S1061、According to the time-frequency spectrum, a decision tree classification structure is constructed and the cutting parameter sequence is optimized by the maximum information gain criterion, wherein the decision tree is divided into three layers in turn as the feed speed, the spindle speed and the cutting depth, the weight is allocated by using the parameter sensitivity, and the adjustment amount is calculated combined with the process constraint, and then the parameter combination meeting the condition is generated by optimizing the particle swarm algorithm.

[0061] The training sequence is extracted from the historical data, the long short-term memory network is used to build the wear rate prediction model, 32 hidden layer nodes are configured, the frequency domain features are extracted by wavelet transform, and the future 15-minute wear rate is calculated by recursive neural network, and the prediction value increases from 0.02 mm / min to 0.07 mm / min. The critical value is set to 0.08 mm / min, and the optimization is triggered when the prediction value exceeds 80%. The optimization range includes the feed speed of 500-2000 mm / min, the cutting depth of 0.1-2 mm, and the spindle speed of 2000-8000 rpm, and the parameter interval is quantized by 10 discrete points. The decision tree analysis shows that the weight of the feed speed is 0.5, the weight of the spindle speed is 0.3, and the weight of the cutting depth is 0.2. The constraint condition ensures that the ratio of feed to speed is between 0.1 and 0.4, and the cutting load does not exceed the critical value. The particle swarm algorithm iterates 100 times with 50 particles to generate the optimal combination: feed speed 1200 mm / min, spindle speed 4000 rpm, and cutting depth 0.8 mm.

[0062] S1062、Based on the parameter sequence, a cutting process evaluation function is constructed and the fitness value is calculated by weighted summation, wherein the evaluation indexes include surface roughness, machining efficiency and tool life, and the weights are 0.4, 0.3 and 0.3 respectively, then the parameter stability is verified by Monte Carlo method and the stable interval is extracted from the simulation results to ensure the optimization effect. In the embodiment of the present application, the adaptive adjustment of the parameters is realized by multi-level analysis and optimization. The process evaluation function calculation shows that the fitness of the optimal combination is 0.85, which is significantly improved compared with the initial value 0.65. The Monte Carlo simulation is disturbed for 1000 times, the feed speed is disturbed by 10%, the spindle speed is disturbed by 5%, and the cutting depth is disturbed by 15%, and the index fluctuation is controlled within 5%, which verifies the robustness. Experiments show that the optimized parameters reduce the wear rate by 30%, reduce the fluctuation amplitude by 40%, prolong the tool life by 25%, and maintain the high-frequency component ratio below 20%, ensuring the machining stability.

[0063] S107、According to the target cutting parameter combination, the wear area to be compensated is identified and the tool wear state and workpiece machining quality are evaluated with a fixed control period, and the tool motion trajectory is adjusted in real time by calculating the compensation amount, so as to ensure the machining precision and stability.

[0064] In this embodiment of the invention, the control cycle is set to 1 second, and the target cutting parameters are divided into 50 segments. Wear data is collected through a 5×5 gridded detection array, generating a 64×64 pixel image for each point. A convolutional neural network is used to process the image; the network is configured with three convolutional layers to extract wear contour features and output key point coordinates. Subsequently, a laser profilometer is used to measure surface roughness at a resolution of 10 micrometers, obtaining geometric parameters such as wear area, depth, and bandwidth. In the experiment, the wear area increased from 0.2 square millimeters to 0.8 square millimeters, the depth increased from 0.05 millimeters to 0.15 millimeters, and the bandwidth expanded from 0.3 millimeters to 0.9 millimeters. Kalman filtering is used to dynamically estimate the wear parameters, construct a state vector, and calculate the optimal compensation combination, including feed rate and depth-of-cut corrections, using a quadratic programming solver.

[0065] S1071. Extract wear contour features from the detection image and calculate the set of geometric parameters. Use Kalman filtering to dynamically estimate the wear area and depth to generate a state vector. Then, calculate the feed rate and cutting depth correction based on the linear compensation model and quadratic programming solver to optimize the parameter combination.

[0066] Wear data was processed using Kalman filtering, with second-order state equations describing parameter changes. The standard deviation of observation noise was 0.01 mm, and state noise was 0.005 mm. Compensation increments were calculated based on a linear model; for example, for every 0.1 square millimeter increase in wear area, the feed rate decreased by 50 mm / min, and the depth of cut decreased by 0.05 mm. A quadratic programming solver limited the feed rate correction range to ±200 mm / min and the depth of cut to ±0.2 mm, while also considering machining efficiency constraints to ensure that parameter adjustments balanced quality and efficiency. Cubic spline interpolation was used with four control points to smooth the compensation parameters, ensuring continuity of change.

[0067] S1072. Based on the correction amount, establish the compensation trajectory equation and generate the position compensation command sequence by fitting a fifth-order polynomial. Execute the command using the position feedforward controller and adjust the compensation gain in conjunction with the servo feedback signal to achieve real-time optimization of the tool trajectory.

[0068] In this embodiment of the invention, tool motion is adjusted by compensating for the trajectory equation. A fifth-order polynomial is fitted using the least squares method to ensure acceleration continuity. Command sequences are output to the servo system at 5-millisecond intervals, with the feed axis speed response time controlled within 50 milliseconds and Z-axis positioning accuracy better than 0.005 mm. In the transition region, the compensation amount gradually changes according to a sine function to avoid impact. The initial gain is set to 0.8 and adjusted online based on the error signal, with the root mean square error controlled within 0.01 mm. Experiments show that after compensation, the surface roughness stabilizes at 1.2 micrometers, dimensional accuracy is improved by 40%, wear rate is reduced by 35%, and tool life is doubled.

[0069] The above description is merely a specific implementation of this specification. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, modules, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. It should be understood that the scope of protection of this specification is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this specification, and these modifications or substitutions should all be covered within the scope of protection of this specification.

Claims

1. An automatic wear detection and compensation method for CNC machine tools based on artificial intelligence, characterized in that, The method includes: Cutting force fluctuation characteristics are obtained by real-time acquisition of cutting force data in high curvature regions through multi-sensor fusion. Monitor and acquire cutting temperature data in high curvature areas, extract the cutting temperature change rate, determine whether the temperature exceeds the preset threshold, and if so, trigger an alarm mechanism and adjust the cutting parameters, including tool feed rate and depth of cut. Vibration signals during the cutting process are acquired in real time, and spectrum analysis is performed to obtain the vibration frequency and amplitude changes of the tool. Combined with the characteristics of cutting force fluctuation and cutting temperature data, the stress distribution is obtained using the finite element analysis method. A relationship model between cutting force, cutting temperature, tool vibration signal and stress distribution is established to determine the stress concentration location in the high curvature region. If continuous cutting is performed at stress concentration locations for an extended period, the local tool wear rate and wear morphology changes are assessed by combining offline and online monitoring data of cutting force and vibration signals. Furthermore, the co-evolution relationship between wear and temperature is analyzed based on historical cutting force and vibration signals as well as current cutting temperature data. By combining the co-evolution relationship between wear and temperature, online monitoring data, and the tool wear rate predicted by the machining time, a wear prediction curve for a preset time period is generated. Trend analysis and feature extraction are performed on the wear prediction curve to obtain the changes in tool wear parameters within a preset time. If the prediction curve shows that the wear parameter changes will exceed the preset critical value at a certain time point in the future, the cutting parameters are adjusted to generate the target cutting parameter combination. Based on the target cutting parameter combination and according to the set control cycle, the wear compensation area is identified, the tool wear condition and workpiece machining quality are evaluated, and the compensation amount of the cutting parameters is calculated. The tool movement trajectory in the wear compensation area is then adjusted in real time according to the compensation amount.

2. The method according to claim 1, characterized in that, The method of obtaining cutting force fluctuation characteristics by real-time acquisition of cutting force data in high curvature regions through multi-sensor fusion includes: The sensor coordinate matrix is ​​obtained based on the positional relationship of the sensor on the cutting surface, and the sensor coordinate matrix is ​​used to perform spatial mapping calibration on the sensor-collected data through coordinate transformation. A moving average filter is used to eliminate noise in the sensor data after spatial mapping calibration, and time-series alignment is performed on multiple sets of filtered sensor data using timestamp identifiers. Based on the time-aligned sensor data, high and low curvature regions are divided according to the curvature threshold of the cutting surface, and wavelet coefficients are obtained by extracting the force signal amplitude sequence. The data fusion weighting coefficients are determined based on the sensor placement. The wavelet coefficients are then weighted and summed using these weighting coefficients to reconstruct the force signal sequence. The force signal sequence is then calculated using a sliding window to obtain the cutting force fluctuation characteristics.

3. The method according to claim 1, characterized in that, The process involves monitoring and acquiring cutting temperature data in high curvature regions, extracting the rate of change of cutting temperature, determining whether the temperature exceeds a preset threshold, and if so, triggering an alarm mechanism and adjusting cutting parameters, including tool feed rate and depth of cut. The machining area is divided into grids based on the curvature gradient value of the cutting surface. Temperature data collected by the thermocouple sensor array is obtained from the grid nodes. Cubic spline interpolation is used to obtain the grid temperature field distribution map. The Bayesian filtering state equation is used to eliminate noise in the grid temperature field distribution map, and the temperature field gradient distribution map is obtained by central difference calculation. Temperature field time series data are obtained from the temperature field gradient distribution map, and temperature abrupt points in the time series data are detected by a sliding variance window. Temperature fluctuation features are extracted for the temperature abrupt points. The temperature fluctuation characteristics are compared with the preset temperature threshold. If the temperature fluctuation amplitude exceeds the threshold range, the temperature over-limit area is determined. The feed rate adjustment interval is divided for the temperature over-limit area, and the optimal parameter combination is calculated using a linear programming solver.

4. The method according to claim 1, characterized in that, The system acquires vibration signals during the real-time cutting process and performs spectral analysis to obtain the vibration frequency and amplitude changes of the tool. Combined with cutting force fluctuation characteristics and cutting temperature data, the stress distribution is obtained using finite element analysis. A relationship model is established between cutting force, cutting temperature, tool vibration signals, and stress distribution to determine the stress concentration locations in high-curvature regions, including: The vibration signal envelope curve is obtained by Hilbert transform based on the vibration sampling data, and the vibration signal envelope curve is then subjected to fast Fourier transform to obtain vibration spectrum data. Boundary condition parameters are constructed based on the vibration spectrum data and the cutting force fluctuation characteristic curve. The element stress components are calculated from the linear elastic constitutive equation based on the boundary condition parameters. The stress influence factor matrix is ​​calculated using the vibration signal envelope curve and the cutting force fluctuation characteristic curve. The stress influence factor matrix includes the root mean square value of the vibration signal and the standard deviation of the force signal. The stress field distribution function is obtained by using neural network feature mapping based on the stress influence factor matrix. The stress field distribution function is used to determine the stress concentration location by the output of the deep learning training model.

5. The method according to claim 1, characterized in that, If continuous cutting is performed at the stress concentration location for an extended period, the local tool wear rate and wear morphology changes are assessed by combining offline and online monitoring data of cutting force and vibration signals. Furthermore, based on historical cutting force and vibration signals and current cutting temperature data, the co-evolution relationship between wear and temperature is analyzed, including: The cutting force signal and vibration signal time series are obtained from historical cutting process records, and the cutting force signal and vibration signal time series are denoised by wavelet multi-scale decomposition. The initial wear position is determined by using the peak point of cutting force fluctuation for the denoised signal. The initial wear position is obtained by using the vibration amplitude variation curve to obtain the wear contour boundary point. The local wear rate is calculated using the wear profile boundary points, and the wear morphology characteristic value is obtained by using the vibration signal envelope spectrum. Based on the wear morphology feature values, the historical cutting force signal sequence is decomposed, and the feature decomposition obtains the vibration signal temporal features through a self-attention mechanism. Based on the temporal characteristics of the vibration signal, the wear temperature coupling change is predicted; Based on the predicted wear temperature coupling changes, a Markov chain is used to describe the co-evolution process of wear and temperature.

6. The method according to claim 5, characterized in that, Also includes: Historical cutting force and vibration signal data are acquired, and historical cutting force and vibration values ​​are extracted. The current cutting temperature of the monitoring system is collected and recorded. A regression analysis method is used to establish a relationship model between cutting force, vibration values, and temperature values. If the model fit meets a preset threshold, the co-evolutionary relationship between wear and temperature is determined, specifically including: The peak-to-valley ratio and standard deviation of force values ​​are obtained from the cutting force sequence, and the mean amplitude and frequency of amplitude are obtained from the vibration sequence. Time-aligned feature data are obtained through linear interpolation. The time-aligned feature data is denoised using wavelet basis functions. The average temperature and temperature change slope features are calculated from the denoised data sequence. A system of multiple linear equations is established using the least squares method. The feature correlation matrix is ​​obtained using the Pearson correlation coefficient. If the significance coefficient in the feature correlation matrix is ​​greater than a preset value, then a feature mapping function is established using multinomial regression, a state prediction function is obtained through a recurrent neural network, and the neural network weight matrix is ​​obtained from the prediction function.

7. The method according to claim 1, characterized in that, The method combines the co-evolution relationship between wear and temperature, online monitoring data, and the tool wear rate predicted by the machining time to generate a wear prediction curve for a preset time period, including: The online collected data is processed into time-series segments according to the monitoring data period, and historical processing volume statistics are obtained from the time-series segmented data. The root mean square value of cutting force, vibration amplitude and temperature gradient feature sequences are extracted from the historical machining volume statistics, and standardized feature data are obtained through a long short-term memory network. The Pearson correlation coefficient is calculated using the standardized feature data, and the feature change rate sequence is obtained by filtering from the correlation coefficient. If the feature change rate sequence meets the preset threshold requirement, a recurrent neural network is used to filter the wear prediction curve to obtain the wear prediction curve for the preset time period.

8. The method according to claim 1, characterized in that, The wear prediction curve is analyzed for trends and features are extracted to obtain the changes in tool wear parameters within a preset time period. If the prediction curve shows that the wear parameter changes will exceed a preset critical value at a future time point, the cutting parameters are adjusted to generate a target cutting parameter combination, including: The inflection point of the trend is determined based on the wear prediction curve, and the inflection point of the trend is calculated from the zero point of the second derivative. For the inflection point of the trend, the piecewise linear regression coefficient is calculated using the least squares method, and the wear slope value is obtained from the piecewise linear regression coefficient. A Fourier transform operation is performed on the wear slope value, and the wear feature main frequency component is extracted from the Fourier transform operation. The wear feature main frequency component is then processed by a convolutional neural network to obtain the feature time spectrum. A decision tree classification structure is constructed based on the time-frequency spectrum of the features. The decision tree classification structure includes a feed rate layer, a spindle speed layer, and a cutting depth layer. The parameter optimization sequence is obtained from the decision tree classification structure through the maximum information gain criterion. A cutting process evaluation function is constructed for the parameter optimization sequence. The cutting process evaluation function obtains the fitness value of the parameter combination through weighted summation. The target cutting parameter combination is generated from the fitness value of the parameter combination.

9. The method according to claim 8, characterized in that, Also includes: The current wear rate is calculated using a time series analysis algorithm, combined with historical wear data to construct a prediction curve model. This model calculates the predicted wear rate at a preset time point. If the predicted value exceeds a preset critical value, a parameter optimization algorithm is invoked to generate several sets of cutting parameter combinations. A target cutting parameter combination is then searched within a preset parameter range. Specifically, this includes: Based on the wear data sequence, a fixed time window is used for sampling, and a long short-term memory network model is constructed from the sampled data to obtain the predicted wear rate; Wavelet transform is performed on the predicted wear rate to obtain frequency domain feature coefficients. The frequency domain feature coefficients are used to establish a recurrent neural network state equation, and the wear rate at a preset time point is calculated from the state equation. If the difference between the wear rate and the preset critical wear value exceeds the warning threshold, the cutting parameter optimization is triggered, and the feed rate range, cutting depth range and spindle speed range are established based on the cutting parameter optimization. The discrete parameter combination is obtained by uniform quantization of the parameter range. The feasible parameter space is obtained from the constraints. The genetic algorithm is used to search the feasible parameter space and obtain the optimal cutting parameter combination from the search results.

10. The method according to claim 1, characterized in that, The process involves identifying the wear-compensated area based on the target cutting parameter combination and a set control cycle, evaluating the tool wear state and workpiece machining quality, calculating the compensation amount for the cutting parameters, and adjusting the tool movement trajectory in the wear-compensated area in real time based on the compensation amount. This includes: A convolutional neural network is used to obtain wear contour features from the detected image, and wear area and wear depth parameters are extracted based on the wear contour features. Perform Kalman filtering on the wear area and wear depth parameters to obtain the wear state vector; The optimal compensation parameter combination is calculated using a quadratic programming solver based on the wear state vector. The optimal compensation parameter combination includes feed rate correction and depth of cut correction. A compensation trajectory equation is established for the feed rate correction and the depth of cut correction. The position compensation command sequence is obtained by fitting the compensation trajectory equation with a fifth-order polynomial. The position compensation command sequence is used to control the tool position compensation.

Images