Tool wear self-adaptive compensation system based on multi-dimensional sensor data fusion

Abstract

This invention relates to the field of adaptive control technology, specifically disclosing a tool wear adaptive compensation system based on multi-dimensional sensor data fusion. It acquires force, vibration, and acoustic emission signals in real time to obtain a fused feature sequence; extracts the dynamic force gradient and calculates the fractal dimension, decomposing it into long-trend and short-period components to generate a coupled modulation sequence, establishing a nonlinear mapping with two latent states: tool wear amount and workpiece hardness; employs a dual-state unscented Kalman filter to jointly estimate wear amount and hardness, using a cross-covariance matrix to decouple their contributions to the sensor signals; calculates dimensional deviations based on wear estimates, and uses hardness estimates to eliminate hardness fluctuation interference, forming a compensation amount specific to wear, which is then written into the CNC system. Simultaneously, it monitors signal changes after compensation to generate confidence levels, and uses feedback to adjust the filter covariance for iterative optimization. This invention can separate the interference of material hardness fluctuations on wear estimation, avoid erroneous compensation, and improve batch processing accuracy.

Description

Technical Field

[0001] This invention relates to the field of adaptive control technology, and more specifically to a tool wear adaptive compensation system based on multi-dimensional sensor data fusion. Background Technology

[0002] In CNC milling, turning, and other machining processes, tool wear is a key factor affecting machining accuracy, surface quality, and production efficiency. As tool wear increases, the actual cutting point shifts relative to the theoretical tool tip, leading to dimensional deviations. To reduce the impact of wear on machining accuracy, current technologies often employ a combination of online monitoring and compensation control. This involves real-time acquisition of sensor signals such as force, vibration, and acoustic emission during the cutting process, establishing a mapping model between signal characteristics and tool wear, predicting the current wear value, and automatically adjusting the tool compensation parameters of the CNC system.

[0003] In batch machining processes, how to decouple the sensor signal drift caused by the cumulative evolution of tool wear from the sudden change in sensor signal caused by batch fluctuations in workpiece material hardness online, so as to avoid misjudging the increase in cutting force caused by hardness change as accelerated tool wear and thus avoid overcompensation or undercompensation. Summary of the Invention

[0004] The purpose of this invention is to provide an adaptive tool wear compensation system based on multi-dimensional sensor data fusion to solve the problems mentioned above.

[0005] The objective of this invention can be achieved through the following technical solutions: The tool wear adaptive compensation system based on multi-dimensional sensor data fusion includes a signal acquisition and fusion module that collects force signals, vibration signals, and acoustic emission signals during the tool cutting process in real time, and performs synchronization and preprocessing on the collected signals to obtain a fused feature sequence. The hidden state setting mapping module sets the tool wear amount and the workpiece surface equivalent hardness coefficient as two coupled time-varying hidden states, and establishes a nonlinear mapping relationship between the fused feature sequence and the two hidden states, using the nonlinear mapping relationship as the input for joint estimation. The dual-state joint estimation module uses a dual-state unscented Kalman filter to perform online joint estimation of the two hidden states. It uses the cross covariance matrix to decouple and fuse the contributions of tool wear and hardness change in the feature sequence, and outputs the tool wear estimate and hardness estimate at the current time. The compensation calculation module calculates the machining dimension deviation caused by wear based on the tool wear estimate, and uses the hardness estimate to correct the cutting force baseline to eliminate the interference of hardness fluctuation on wear estimation, thus forming a compensation amount only for wear. The compensation execution and feedback module writes the compensation amount into the tool compensation parameters of the CNC system in real time, monitors the signal changes after compensation, and feeds the new data back to the update process of the dual-state unscented Kalman filter to iteratively optimize the joint estimation at the next moment.

[0006] As a further aspect of the present invention: establishing a nonlinear mapping relationship between the fused feature sequence and the two hidden states, and using the nonlinear mapping relationship as input for joint estimation, specifically includes: The dynamic force gradient within each cutting stroke is extracted from the fused feature sequence, and the fractal dimension of the dynamic force gradient is calculated to obtain the fractal feature curve. The fractal characteristic curve is decomposed into a long-term trend component and a short-period fluctuation component. The long-term trend component represents the cumulative evolution of tool wear, while the short-period fluctuation component represents the instantaneous disturbance of the equivalent hardness coefficient of the workpiece surface. The long-term trend component and the short-period fluctuation component are multiplied point-to-point to generate a coupled modulation sequence, which is then used as the output of the nonlinear mapping relationship.

[0007] As a further aspect of the present invention: obtaining the fractal characteristic curve specifically includes: The force signal in each cutting stroke is differentially processed according to the sampling time sequence to obtain the instantaneous gradient sequence; The instantaneous gradient sequence is mapped to a grid space of different scales, and the total number of grids covering the gradient waveform at each scale is counted. A linear fit is performed on the logarithmic relationship between the grid scale and the total number of grids. The negative value of the fitted slope is taken as the fractal dimension of the cutting stroke. The fractal dimensions of each cutting stroke are then connected in sequence to form a fractal characteristic curve.

[0008] As a further aspect of the present invention: the output of the current tool wear estimate and hardness estimate specifically includes: Based on the tool wear estimate and hardness estimate output from the previous moment and their joint covariance, a set of symmetrically distributed sampling points are generated, each of which contains both wear and hardness components. Substitute each sampling point into the nonlinear mapping relationship to obtain the corresponding predicted observation point, and obtain the predicted observation sequence through weighted fusion; The residual sequence between the predicted observation sequence and the actual fused feature sequence is calculated. The update weights of wear and hardness are dynamically adjusted using the cross covariance matrix between the sampling points and the predicted observation points. This ensures that fluctuations caused by hardness in the residual sequence are preferentially absorbed by the hardness estimate, and drifts caused by wear are preferentially absorbed by the wear estimate. The decoupled tool wear estimate and hardness estimate are then output.

[0009] As a further aspect of the present invention: the method of obtaining the predicted observation sequence through weighted fusion specifically includes: Each sampling point is combined with its corresponding predicted observation point into a point pair. The diagonal length of the recursion graph in phase space for each point pair is calculated to obtain the recursion intensity value. The reciprocal of the recursive intensity value of each sampling point is multiplied by the reciprocal of the covariance between the sampling point and the mean of all sampling points to generate the fusion weight of each sampling point. The fusion weights are logarithmically compressed and normalized, then multiplied by the predicted observation points corresponding to each sampling point and summed to output the predicted observation sequence.

[0010] As a further aspect of the present invention: the formation of the compensation amount specifically for wear includes: The estimated tool wear value is decomposed into axial offset and radial offset along the tool axis, and the initial value of machining dimension deviation is synthesized. Using the hardness estimate as a normalization coefficient, recursively median filtering is applied to the historical sliding window of the real-time cutting force baseline to extract the pure wear force residual after removing hardness fluctuations. The linear projection coefficient is calculated by combining the residual of pure wear force with the initial value of machining dimension deviation. The final compensation amount is then output by multiplying the linear projection coefficient with the initial value of machining dimension deviation.

[0011] As a further aspect of the present invention: the initial value of the synthetic processing dimensional deviation specifically includes: Obtain the unit direction vector of the axis corresponding to the current tool posture, and project the estimated tool wear value onto the direction vector to obtain the axial offset component. After subtracting the axial offset component from the estimated tool wear value, the tool is orthogonally decomposed into the radial plane of the tool to obtain the radial offset component with the spindle rotation center as the origin. Based on the angle between the radial offset component and the normal of the machined surface, the radial offset component is projected onto the normal and tangential directions respectively. Then, the normal projection is combined with the square root of the axial component to output the initial value of the machining dimension deviation.

[0012] As a further aspect of the present invention: the step of feeding new data back to the update process of the two-state unscented Kalman filter to iteratively optimize the joint estimate at the next time step specifically includes: Force and vibration signals are captured during three consecutive cutting strokes after the compensation amount is written. The power spectrum energy difference before and after compensation is calculated, and the compensation response characteristic quantity is generated. The compensation response features are time-aligned and waveform-matched with the preset compensation effect template to output the compensation confidence level. The diagonal element value of the process noise covariance in the dual-state unscented Kalman filter is dynamically adjusted based on the compensation confidence level, so that the filter update process is biased towards historical data with low confidence or new data with high confidence, thus completing iterative optimization.

[0013] As a further aspect of the present invention: the generation of the compensation response feature quantity specifically includes: Take the force signal and vibration signal of the last complete cutting stroke before writing the compensation amount, multiply them point by point to form a composite reference sequence, and calculate the autocorrelation spectrum energy of the composite reference sequence. After writing the compensation amount, the force signal and vibration signal of each of the three consecutive cutting strokes are taken and the sliding inner product is performed with the composite reference sequence to obtain three differential energy sequences. For each differential energy sequence, sort the local peaks, take the sum of the top three peaks in each sequence as the differential feature of the cutting stroke, and then sum the differential features of the three cutting strokes according to exponential decay weighting to output the compensation response feature quantity.

[0014] The beneficial effects of this invention are: (1) By setting the tool wear amount and the equivalent hardness coefficient of the workpiece surface as two coupled time-varying hidden states, and using dual-state unscented Kalman filtering for online joint estimation, the contributions of the two to multi-dimensional sensing signals such as force, vibration, and acoustic emission are decoupled using the cross-covariance matrix. This effectively distinguishes between signal changes caused by normal tool wear and signal changes caused by batch fluctuations in workpiece material hardness during the cutting process. Compared with methods that only estimate a single wear amount, this invention avoids misjudging the increase in cutting force caused by sudden changes in hardness as severe tool wear, thereby preventing the adaptive compensation system from generating excessive tool bias or feed rate reduction commands, reducing the risk of overcutting or scrapping of parts due to miscompensation, and improving the consistency of batch processing.

[0015] (2) By capturing the force and vibration signals during three consecutive cutting strokes after compensation, the difference in power spectrum energy before and after compensation is calculated, and the compensation response characteristic is generated. Then, it is matched with the preset compensation effect template to obtain the compensation confidence. Based on this confidence, the diagonal element value of the process noise covariance in the dual-state unscented Kalman filter is dynamically adjusted, so that the filter update process can adaptively bias towards historical data or newly acquired data according to the confidence level of the current compensation effect. When the compensation effect meets expectations, the filter trusts the new data more to speed up the tracking speed; when the compensation effect is disturbed and deviates from expectations, the filter relies more on historical data to maintain the estimation stability. This feedback mechanism enhances the robustness of the system under non-stationary working conditions such as changes in coolant state and intermittent cutting impact, and avoids the estimation drift caused by a single fixed filter parameter when the working conditions change abruptly. Attached Figure Description

[0016] The invention will now be further described with reference to the accompanying drawings.

[0017] Figure 1 This is a system block diagram of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1 As shown, this invention is a tool wear adaptive compensation system based on multi-dimensional sensor data fusion, comprising: The signal acquisition and fusion module acquires force signals, vibration signals, and acoustic emission signals during the cutting process in real time, and performs synchronization and preprocessing on the acquired signals to obtain a fused feature sequence. The hidden state setting mapping module sets the tool wear amount and the workpiece surface equivalent hardness coefficient as two coupled time-varying hidden states, and establishes a nonlinear mapping relationship between the fused feature sequence and the two hidden states, using the nonlinear mapping relationship as the input for joint estimation. The dual-state joint estimation module uses a dual-state unscented Kalman filter to perform online joint estimation of the two hidden states. It uses the cross covariance matrix to decouple and fuse the contributions of tool wear and hardness change in the feature sequence, and outputs the tool wear estimate and hardness estimate at the current time. The compensation calculation module calculates the machining dimension deviation caused by wear based on the tool wear estimate, and uses the hardness estimate to correct the cutting force baseline to eliminate the interference of hardness fluctuation on wear estimation, thus forming a compensation amount only for wear. The compensation execution and feedback module writes the compensation amount into the tool compensation parameters of the CNC system in real time, monitors the signal changes after compensation, and feeds the new data back to the update process of the dual-state unscented Kalman filter to iteratively optimize the joint estimation at the next moment.

[0020] In the signal acquisition and fusion module, force signals, vibration signals, and acoustic emission signals during the tool cutting process are acquired in real time. The acquired signals are synchronized and preprocessed to obtain a fused feature sequence, which specifically includes: During the cutting process, force signals are collected in real time by a piezoelectric triaxial force sensor installed between the spindle and the tool holder, vibration signals are collected by a piezoelectric accelerometer magnetically fixed to the spindle housing, and acoustic emission signals are collected by a broadband acoustic emission sensor attached to the outside of the tool holder. All three sensors are connected to the same multi-channel synchronous data acquisition card, which is triggered by an external clock to ensure that the start times of the three signals are perfectly aligned.

[0021] The acquired force, vibration, and acoustic emission signals were preprocessed separately. For the force and vibration signals, a low-pass filter with a cutoff frequency of 10 kHz was first used to eliminate high-frequency noise, and then a moving average method was used to remove the DC drift trend term from the signal. For the acoustic emission signal, a band-pass filter was first used to retain the effective frequency band with a center frequency of 300 kHz to 500 kHz, and then its root mean square (RMS) value was calculated. The RMS value sequence was then downsampled to 20 kHz to give it the same time resolution as the force and vibration signals. After the above preprocessing was completed, each sampling point of the three signals was aligned in time to form a set of multidimensional data tuples. Each tuple contained the preprocessed force signal amplitude, the preprocessed vibration signal amplitude, and the preprocessed acoustic emission RMS value at the same time.

[0022] The aforementioned multidimensional data tuples are arranged sequentially according to the cutting order to form a two-dimensional array, which serves as the fusion feature sequence. Each column of this fusion feature sequence corresponds to a sampling time. The first row stores the preprocessed force signal, the second row stores the preprocessed vibration signal, and the third row stores the preprocessed root mean square value of acoustic emission.

[0023] In the hidden state setting mapping module, tool wear and the equivalent hardness coefficient of the workpiece surface are set as two coupled time-varying hidden states, and a nonlinear mapping relationship between the fused feature sequence and the two hidden states is established. The nonlinear mapping relationship is used as the input for joint estimation, specifically including: After obtaining the fused feature sequence, the cutting stroke is defined. Each cutting stroke refers to a complete and continuous cutting process from the moment the tool enters the workpiece to the moment it exits. Its start and end points are determined by whether the force signal amplitude exceeds a preset threshold, which is set to three times the average force amplitude during no-load cutting. For each cutting stroke, the corresponding force signal data segment is extracted. The force signal amplitudes of two adjacent sampling points are subtracted according to the sampling time sequence, i.e., the force signal amplitude of the previous sampling point is subtracted from the force signal amplitude of the later sampling point, yielding the instantaneous gradient value at that sampling moment. All instantaneous gradient values ​​are arranged in chronological order to form the instantaneous gradient sequence of that cutting stroke.

[0024] The fractal dimension of the instantaneous gradient sequence for each cutting stroke is calculated. A set of different grid scales is defined, for example, scales of 1, 2, 3, 4, 5, 6, 7, 8, 9, and 10, each scale corresponding to a grid side length. The instantaneous gradient sequence of this stroke is plotted on a two-dimensional plane, with the sampling time sequence number on the x-axis and the instantaneous gradient value on the y-axis. For each scale, a square grid with a side length of that scale is used to cover the entire two-dimensional plane, and the number of grids containing at least one waveform point is counted. Then, the logarithm of each scale is calculated, along with the logarithm of the corresponding number of grids. A linear fit is performed with the logarithm of the scale as the x-axis and the logarithm of the number of grids as the y-axis to obtain a straight line. The slope of this line is negative, and this negative value is taken as the fractal dimension of the cutting stroke. The above operation is performed sequentially for each cutting stroke to obtain the fractal dimension of each stroke, and these fractal dimensions are connected sequentially according to the cutting order to form a fractal characteristic curve.

[0025] After obtaining the fractal characteristic curve, it is decomposed into components. A low-pass filter with a cutoff frequency of 0.5 times the cutting stroke frequency is used to filter the fractal characteristic curve. The absolute value corresponding to this cutoff frequency is typically 0.1 Hz in actual machining (when the cutting stroke frequency is about 0.2 Hz), and is used to extract the long-term trend component, which reflects the cumulative evolution of tool wear with the cutting stroke. Subtracting the long-term trend component from the original fractal characteristic curve yields the short-period fluctuation component, which reflects the instantaneous disturbance of the workpiece surface equivalent hardness coefficient between adjacent strokes.

[0026] The long-trend component and the short-period fluctuation component are multiplied point-to-point. Specifically, at the same stroke index position, the value of the long-trend component is multiplied by the value of the short-period fluctuation component to obtain the coupling value at that position. The same multiplication operation is performed on all stroke index positions to obtain a new sequence, called the coupled modulation sequence. This coupled modulation sequence, as the output of the nonlinear mapping relationship, is used in the subsequent joint estimation process.

[0027] In the above fractal dimension calculation process, for the first cutting stroke of a new workpiece or a new tool, if historical data is insufficient, the fractal dimension is assumed to be 1.5. This value is derived from the statistical average of offline pre-experimental tests on standard specimens of the same type of material. The cutoff frequency of the decomposition filter for the long-term trend component and the short-period fluctuation component is set to 0.1 Hz based on the slow change characteristics of tool wear in actual machining. The change frequency corresponding to the short-period fluctuation is usually higher than 0.5 Hz, and the two can be effectively separated by a cutoff frequency of 0.2 Hz (this value is the dividing point between low-pass and high-pass filters, different from the application of the 0.1 Hz low-pass filter mentioned above). Before the point-to-point multiplication operation, both the long-term trend component and the short-period fluctuation component need to be normalized to the 0 to 1 interval. The normalization method is: subtract the minimum value of the component in the entire stroke from the original value of each component, and then divide by the difference between the maximum and minimum values ​​of the component.

[0028] In the dual-state joint estimation module, a dual-state unscented Kalman filter is used to perform online joint estimation of the two hidden states. The contributions from tool wear and hardness changes in the feature sequence are decoupled and fused using the cross-covariance matrix, respectively, to output the current tool wear and hardness estimates, specifically including: Before performing the two-state unscented Kalman filtering, the initial states are first set. Before the start of the first cutting stroke, the tool wear estimate is set to zero millimeters, and the workpiece surface equivalent hardness coefficient estimate is set to 1.0 (representing standard hardness). The initial value of the joint covariance matrix of these two estimates is set to a second-order diagonal matrix, where the variance of the tool wear estimate is set to 0.01 square millimeters, the variance of the hardness estimate is set to 0.04, and the covariance between the two is set to zero.

[0029] Based on the tool wear estimate, hardness estimate, and their joint covariance from the previous time step, a set of symmetrically distributed sampling points is generated. Specifically, the two-dimensional joint covariance matrix is ​​decomposed using Cholliski decomposition to obtain a lower triangular matrix. Each column of this lower triangular matrix is ​​multiplied by a scaling factor, and then added to and subtracted from the state mean, resulting in five sampling points. The first sampling point is the state mean itself, and the other four sampling points correspond to the addition and subtraction of the decomposition results from each column of the mean. Each sampling point contains two components: a wear component and a hardness component.

[0030] Substituting each sampling point into the previously established nonlinear mapping relationship yields the corresponding predicted observation point. This nonlinear mapping relationship is expressed in the form of the aforementioned coupled modulation sequence. Specifically, the wear and hardness components in the sampling point are used as inputs, and the corresponding predicted observation point is calculated using the same mapping parameters as the training data. This predicted observation point is a three-dimensional vector, corresponding to the predicted force signal, vibration signal, and root mean square value of acoustic emission, respectively. After obtaining the predicted observation points for all sampling points, these predicted observation points need to be weighted and fused to obtain the predicted observation sequence.

[0031] The nonlinear mapping relationship is implemented using a three-layer backpropagation neural network. The input layer has 2 nodes, which correspond to tool wear and workpiece hardness coefficient, respectively. The output layer has 3 nodes, which are the predicted values ​​of stress signal, vibration signal and acoustic emission root mean square value, respectively. The hidden layer has 8 nodes, and the activation function is the Sigmoid function.

[0032] The training data was collected as follows: using standard specimens made of the same material as the workpiece to be processed, and under the same machine tool, cutting tool and cutting parameters, 5 different tool wear values ​​(0 mm, 0.05 mm, 0.10 mm, 0.15 mm, 0.20 mm) and 3 different workpiece hardness values ​​(0.8 times, 1.0 times, 1.2 times the standard hardness) were set, for a total of 15 combinations. Multidimensional sensor signals of 5 cutting strokes were continuously collected under each combination, and the average value was taken as a training sample.

[0033] The training process for the mapping parameters is as follows: supervised training is performed using the Levenberg-Marquardt algorithm, with the loss function being the mean squared error between the predicted observations and the actual fused feature sequences. Training is iterated until the error is less than 1e-4 or until 500 iterations are reached. After training, the network weights and biases are fixed in the system. When used online, the wear and hardness components of the sampling points can be directly substituted to obtain the corresponding predicted observation points.

[0034] The specific process of weighted fusion is as follows: Each sampling point is combined with its corresponding predicted observation point into a point pair, and the diagonal length of the recurrence graph in phase space for each point pair is calculated. The diagonal length of the recurrence graph reflects the degree of deterministic evolution of the system state in phase space; a longer length indicates stronger predictability of the trajectory segment where the sampling point is located. Its reciprocal serves as a reliability measure, used to reduce the contribution of unreliable sampling points to the fusion result. The reciprocal of the covariance is used to balance the dispersion of the sampling points. The combination of these two factors can effectively suppress the negative impact of outlier observations in UKF and improve the robustness of the fusion weights.

[0035] The phase space construction method is as follows: Each pair of points (wear component, hardness component, predicted force signal, and predicted vibration signal) is treated as a point in four-dimensional space. For all pairs of points formed by sampling points, the Euclidean distance between each pair is calculated. If this distance is less than a preset threshold (set to 30% of the average distance between all pairs), the two pairs are considered to constitute a recursion. The lengths of the line segments formed by all consecutive recursive points are counted, and the length of the longest diagonal is taken as the recursion intensity value of that sampling point. Then, the reciprocal of the recursion intensity value of each sampling point is taken, and this reciprocal is multiplied by the reciprocal of the covariance between that sampling point and the mean of all sampling points. The reciprocal of the covariance is calculated by first calculating the average value of each component of all sampling points, then calculating the deviation vector of each sampling point from the average value, then calculating the covariance matrix of all deviation vectors, and finally inverting the covariance matrix and taking the reciprocal of the trace of the inverse matrix (i.e., the sum of the diagonal elements). The result of the above multiplication is used as the fusion weight of each sampling point.

[0036] Before constructing the phase space, the embedding dimension *m* and the time delay *τ* need to be determined. In this system, the embedding dimension is determined using the pseudo-nearest neighbor method: for each pair of points, the embedding dimension is gradually increased, and the rate of change of distance between nearest neighbors is calculated. The dimension corresponding to the rate of change being less than 10% is the optimal embedding dimension. In this embodiment, *m* = 4 is chosen. The time delay is determined using the mutual information method: the mutual information value between the original sequence and its delayed sequence is calculated, and the time delay corresponding to the first minimum point of the mutual information curve is taken. In this embodiment, *τ* = 2 sampling periods is chosen.

[0037] The above parameters may change under different machining conditions (such as different tool materials and different cutting parameters). The system allows users to recalibrate m and τ according to the actual working conditions during the initialization phase, and the calibration method is the same as above. If the user does not provide calibration data, the system defaults to m=4 and τ=2.

[0038] After obtaining the fusion weights, each weight is first compressed by taking the base-10 logarithm. Then, the compressed weights are divided by the sum of all the logarithmic values ​​of the weights to obtain the normalized fusion weights. Finally, the normalized fusion weights are multiplied by the predicted observation points corresponding to each sampling point, and all the weighted predicted observation points are accumulated component by component to output the weighted fusion predicted observation sequence. This predicted observation sequence is a three-dimensional vector sequence with the same dimension and temporal resolution as the subsequently acquired fusion feature sequence.

[0039] The residual sequence between the predicted observation sequence and the actual fused feature sequence is calculated, i.e., the predicted observation sequence is subtracted from the actual fused feature sequence, resulting in a three-dimensional residual vector at each sampling time. The update weights for wear and hardness are dynamically adjusted using the cross-covariance matrix between the sampling points and the predicted observation points. The cross-covariance matrix is ​​calculated by outer productting the state deviation (sampling point minus state mean) of each sampling point with its corresponding predicted observation deviation (predicted observation point minus predicted observation mean), and then summing the results over all sampling points. This cross-covariance matrix reflects the sensitivity of the wear and hardness components to each dimension of the observed residual.

[0040] The Kalman gain matrix is ​​calculated based on the cross-covariance matrix and the covariance matrix of the predicted observation sequence. Then, the residual sequence is multiplied by the Kalman gain matrix to obtain the state correction. This state correction is a two-dimensional vector, corresponding to the correction values ​​for wear and hardness, respectively. Since the cross-covariance matrix already includes the different effects of wear and hardness on the observations, during the update process, fluctuations in the residual sequence caused by hardness are preferentially corrected for the hardness estimate through the hardness-observation correlation channel in the cross-covariance matrix, while drifts caused by wear are preferentially corrected for the wear estimate through the wear-observation correlation channel. The state estimate from the previous time step is added to the current correction to obtain the decoupled tool wear and hardness estimates for the current time step, which are then output for subsequent compensation calculations. Simultaneously, the joint covariance matrix is ​​updated for filtering iteration in the next time step.

[0041] In the compensation calculation module, the machining dimension deviation caused by wear is calculated based on the tool wear estimate, and the cutting force baseline is corrected using the hardness estimate to eliminate the interference of hardness fluctuations on the wear estimate, thus forming a compensation amount specifically for wear, including: After obtaining the current tool wear estimate (in millimeters) and hardness estimate (dimensionless), the initial values ​​of the machining dimensional deviations are first synthesized. The tool wear estimate is a scalar representing the average wear width of the tool's flank face, but its impact on the machining dimensions depends on where the wear occurs on the tool and the tool's current orientation. Therefore, the unit direction vector of the current tool axis in the workpiece coordinate system needs to be obtained. This vector is directly read from the tool orientation data of the CNC system, and its three components represent the direction cosines of the axis on the X, Y, and Z axes of the workpiece coordinate system, respectively. The tool wear estimate is then projected onto this unit direction vector using a scalar method, i.e., the wear estimate is multiplied by each component of the unit direction vector to obtain the three coordinate values ​​of the axial offset component. The axial offset component represents the displacement of the tool tip along the axial direction due to axial shortening or elongation of the tool.

[0042] Handling Radial Offset. Treat the tool wear estimate as a vector, equal in magnitude to the wear estimate, directed along the tool axis. Subtract the obtained axial offset component vector from this vector to obtain the residual vector, which lies within the tool's radial plane. Orthogonally decompose the residual vector onto the radial plane with the spindle rotation center as the origin, preserving its projection onto the radial plane to obtain the radial offset component. The direction of the radial offset component is determined by the asymmetry of tool wear; for example, greater wear on one side of the tool will result in a radial offset pointing towards that side.

[0043] After obtaining the radial offset component, it needs to be converted to the normal and tangential directions of the machined surface. First, obtain the surface normal unit vector and tangential unit vector of the current machining point. These two vectors are also provided in real-time by the CNC system based on the tool path and workpiece geometry. Calculate the angle between the radial offset component and the surface normal vector. The cosine of the angle is equal to the dot product of the radial offset component and the normal vector divided by the modulus of the radial offset component. Multiply the modulus of the radial offset component by this cosine value to obtain the projection component of the radial offset in the normal direction. Simultaneously, the projection component of the radial offset component in the tangential direction is obtained in a similar manner but is not involved in subsequent synthesis. The final initial value of the machining dimension deviation is determined by the modulus of the axial offset component and the normal projection component, specifically synthesized using the square root of the sum of squares. The calculation formula is as follows: ; in, This indicates the initial value of the machining dimensional deviation (unit: millimeters). The modulus (in millimeters) represents the axial offset component, which is the square root of the sum of the squares of the three coordinate components of the axial offset component. This represents the projected component of the radial offset onto the normal direction of the machined surface (unit: mm). This formula quantifies the combined effect of axial and radial wear on the final dimension into an initial scalar deviation value.

[0044] While obtaining the initial value of the machining dimensional deviation, it is necessary to extract the pure wear force residual after removing hardness fluctuations. To this end, a historical sliding window with a fixed length of ten cutting strokes is constructed. This window stores the real-time cutting force baseline value at each sampling moment of the most recent ten strokes. The cutting force baseline is defined as the average value of the cutting force signal within each cutting stroke, obtained by summing the force signal amplitudes of all sampling points within the stroke and dividing by the total number of sampling points. The current hardness estimate is used as a normalization coefficient to perform recursive median filtering on all baseline values ​​within the historical sliding window. The specific process of recursive median filtering is as follows: the ten baseline values ​​within the sliding window are sorted in ascending order, and the value at the middle position is taken as the median; then, this median is multiplied by the hardness estimate (i.e., the normalization coefficient) to obtain the hardness contribution baseline for the current stroke; finally, the actual cutting force baseline of the current stroke is subtracted from this hardness contribution baseline to obtain the pure wear force residual. This pure wear force residual is a scalar, with the same unit as the force signal (kilonewtons), representing the increase in cutting force caused solely by tool wear after removing hardness disturbances.

[0045] The final compensation amount is determined by calculating the linear projection coefficient between the pure wear force residual and the initial value of the machining dimensional deviation. The principle of linear projection coefficient calculation is to treat the pure wear force residual as a one-dimensional vector and the initial value of the machining dimensional deviation as another one-dimensional vector. The projection coefficient between the two is calculated by the following formula: ; in, This represents the cutting stiffness coefficient (unit: kN / mm). This represents the projection matching coefficient (dimensionless). Represents the residual of pure wear force (unit: kN). This indicates the initial value of the machining dimensional deviation (unit: millimeters). This represents a very small positive number, taking the value of 10 to the power of -6, to avoid the case where the denominator is zero. The calculated projection matching coefficients... Initial value of machining dimension deviation Multiplying these components yields the matched projected component, which is the final compensation amount in millimeters. This compensation amount only reflects the effect of tool wear on dimensions, while force changes caused by hardness fluctuations have been completely eliminated.

[0046] It should be noted that before formal machining, a standard specimen made of the same material as the workpiece to be machined is used, and an offline calibration experiment is conducted under the same machine tool, cutting tool, and cutting parameters. A set of standard cutting tools with different known wear amounts (e.g., 0 mm, 0.05 mm, 0.10 mm, 0.15 mm, 0.20 mm) are prepared. Cutting strokes are performed on the standard specimens, and the average cutting force in each stroke is recorded. The machining dimensional deviation caused by tool wear is also measured. The increment of the cutting force under each wear condition relative to the initial sharp tool is linearly fitted with the measured dimensional deviation. The slope of the fitted line is the cutting stiffness coefficient. Its unit is kilonewtons per millimeter. Once calibrated, this coefficient can be used as a fixed parameter in the same batch of processing.

[0047] The final compensation value is output to the subsequent CNC system tool compensation parameter writing step. If the compensation value is positive, it indicates that the tool length compensation value or radius compensation value needs to be increased to counteract the overcutting trend; if it is negative, it indicates that the compensation value needs to be decreased to avoid undercutting. The entire calculation process is executed once after each cutting stroke to ensure the real-time performance and accuracy of the compensation value.

[0048] In the compensation execution and feedback module, the compensation amount is written into the tool compensation parameters of the CNC system in real time, and the signal changes after compensation are monitored. The new data is fed back to the update process of the dual-state unscented Kalman filter to iteratively optimize the joint estimate at the next moment. Specifically, this includes: After the compensation amount is written into the tool compensation parameters of the CNC system, the system begins to capture the force and vibration signals during the three consecutive cutting strokes following the writing of the compensation amount, while retaining the force and vibration signals from the last complete cutting stroke before the compensation amount is written as a reference. The last complete cutting stroke before the compensation amount is written refers to a complete cutting stroke completed before the compensation command takes effect, and its force and vibration signal data have been fully recorded. The force signal amplitude at each sampling moment within this stroke is multiplied point by point with the vibration signal amplitude at the same moment, i.e., the force signal value at that moment is multiplied by the vibration signal value, to obtain a new sequence, called the composite reference sequence. The length of the composite reference sequence is equal to the number of sampling points within this stroke. Subsequently, the autocorrelation spectral energy of the composite reference sequence is calculated: first, the composite reference sequence is subjected to a Fourier transform to obtain the spectrum, then the amplitude of each frequency component in the spectrum is squared and summed to obtain the autocorrelation spectral energy value, denoted as the reference energy.

[0049] For the first, second, and third cutting strokes after the compensation amount is written, the following operations are performed: The force signal and vibration signal within this stroke are captured and multiplied point-by-point to obtain the composite sequence for that stroke. This composite sequence is then subjected to a sliding inner product with the previously obtained composite reference sequence. Starting from the beginning of the composite sequence, a segment of equal length to the composite reference sequence is taken each time, and the sum of the products at corresponding positions within this segment is calculated. The window is then moved forward by one sampling point, and the above calculation is repeated until the entire composite sequence is covered. This yields a differential energy sequence whose length is equal to the length of the composite sequence minus the length of the composite reference sequence plus one. The above sliding inner product is then performed on the first, second, and third strokes respectively to obtain three differential energy sequences.

[0050] For each differential energy sequence, extract its local peaks. A local peak is defined as a point in the differential energy sequence whose value is greater than the values ​​of its three left-hand and three right-hand adjacent points, and whose value is at least twice the sequence average. Sort all local peaks in descending order of value, take the top three peaks, and calculate the algebraic sum of these three peaks as the differential feature for that cutting stroke. Thus, each cutting stroke yields a differential feature value, denoted as the first differential feature, the second differential feature, and the third differential feature.

[0051] To generate the final compensated response characteristic, the difference characteristics of the three cutting strokes need to be summed using an exponentially decaying weighted sum. The decay coefficient for the exponential decay is set to 0.5, meaning that the more recent the stroke, the greater its weight. The specific steps are as follows: multiply the third difference characteristic (latest stroke) by a coefficient of 1.0, add the second difference characteristic multiplied by a coefficient of 0.5, and add the first difference characteristic multiplied by a coefficient of 0.25. Sum these three terms, then divide by (1.0 + 0.5 + 0.25) = 1.75, and normalize to obtain the compensated response characteristic. This characteristic is a dimensionless positive number, reflecting the significance of the cutting process response after the compensation is applied.

[0052] The calculated compensation response characteristic is time-aligned and waveform-matched with the compensation effect template pre-stored in the system. The compensation effect template is a reference curve obtained by averaging a series of typical response characteristic quantities collected after applying a standard compensation amount to a standard tool under known wear conditions through multiple experiments. The horizontal axis of this curve represents the stroke number, and the vertical axis represents the response characteristic quantity. Time alignment refers to comparing the currently obtained compensation response characteristic quantity (a numerical value) with the value at the corresponding stroke position in the template. Waveform matching refers to calculating the ratio of the current characteristic quantity to the template value. If the ratio is between 0.8 and 1.2, the match is successful. The compensation confidence score is output, which is defined as how close the matching ratio is to 1.0. Specifically, if the ratio is less than 1.0, the confidence score is equal to the ratio; if the ratio is greater than 1.0, the confidence score is equal to the reciprocal of the ratio. The confidence score ranges from 0 to 1, with a value closer to 1 indicating that the compensation effect is more in line with expectations.

[0053] The method for establishing the compensation effect template is as follows: For each type of tool (such as ball end mill, flat end mill, turning tool), each workpiece material (such as 45# steel, aluminum alloy, titanium alloy), and each combination of cutting parameters (such as cutting speed, feed rate, and depth of cut), an independent template is created.

[0054] For each set of working conditions, a set of standard tools (with known wear amounts of 0.05 mm, 0.10 mm, and 0.15 mm) are used to apply corresponding standard compensation amounts (80%, 100%, and 120% of the dimensional deviation) to a workpiece with standard hardness. Force and vibration signals are collected for three consecutive cutting strokes after compensation, and a set of response curves are obtained according to the calculation method of compensation response characteristics.

[0055] The experiment was repeated 5 times under each working condition. The response curves obtained from the 5 experiments were dynamically time-normalized and aligned on the time axis, and the average value was taken point by point to form a reference template for that working condition.

[0056] Template generalization capability boundary: When the actual machining parameters (such as cutting speed change exceeding ±20%, feed rate change exceeding ±15%, workpiece hardness change exceeding ±30%) exceed the template calibration range, the system will automatically prompt the user to recalibrate the template, or use the closest existing template and reduce the upper limit of the compensation confidence to 0.7.

[0057] The diagonal elements of the process noise covariance in the two-state unscented Kalman filter are dynamically adjusted based on the compensated confidence level. The process noise covariance is a two-dimensional diagonal matrix, with its diagonal elements corresponding to the noise variances of the tool wear estimate and the hardness estimate, respectively. The adjustment rules are as follows: if the compensated confidence level is higher than 0.8, indicating that the newly acquired data is reliable, both diagonal element values ​​are multiplied by 0.9 to make the filter update process more reliant on the new data; if the compensated confidence level is lower than 0.3, indicating that the new data is significantly affected by interference, both diagonal element values ​​are multiplied by 1.2 to make the filter more dependent on historical data; if the confidence level is between 0.3 and 0.8, the diagonal element values ​​remain unchanged. After adjustment, the updated process noise covariance is used for the next iteration of the two-state unscented Kalman filter, thereby achieving continuous optimization of the joint estimation.

[0058] The working principle of this invention is as follows: Force, vibration, and acoustic emission signals during the cutting process are acquired in real time using a piezoelectric triaxial force sensor, a piezoelectric accelerometer, and a broadband acoustic emission sensor. After synchronization and preprocessing, a fused feature sequence is obtained. Then, the cutting stroke is defined, and the dynamic force gradient within each cutting stroke is extracted from the fused feature sequence. A fractal feature curve is obtained through fractal dimension calculation and decomposed into a long-term trend component characterizing the cumulative evolution of tool wear and a short-period fluctuation component characterizing the instantaneous disturbance of the workpiece surface equivalent hardness coefficient. The two components are then multiplied point-to-point to generate a coupled modulation sequence as the output of a nonlinear mapping relationship. Tool wear and the workpiece surface equivalent hardness coefficient are set as two coupled time-varying hidden states. A nonlinear mapping relationship is established between the fused feature sequence and the two hidden states. A dual-state unscented Kalman filter is used to perform online joint estimation of the two hidden states. The cross-covariance matrix is ​​used to decouple the fused feature sequence. The contribution of tool wear and hardness variation to the feature sequence is calculated, and the estimated tool wear and hardness values ​​at the current moment are output. Then, based on the estimated tool wear, it is decomposed into axial offset and radial offset along the tool axis and synthesized into the initial value of machining dimension deviation. At the same time, the hardness estimate is used to perform recursive median filtering on the historical cutting force baseline to eliminate hardness fluctuation interference, extract the pure wear force residual, and then calculate the final compensation amount for wear only through linear projection coefficients. The compensation amount is written into the tool compensation parameters of the CNC system in real time, and the force and vibration signals of three consecutive cutting strokes after compensation are captured. The power spectrum energy difference before and after compensation is calculated to generate the compensation response feature quantity. It is matched with the preset compensation effect template to obtain the compensation confidence. Based on the confidence, the diagonal element value of the process noise covariance in the dual-state unscented Kalman filter is dynamically adjusted to iteratively optimize the joint estimation at the next moment, thereby realizing adaptive compensation for tool wear.

[0059] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A tool wear adaptive compensation system based on multi-dimensional sensor data fusion, characterized in that, include: The signal acquisition and fusion module acquires force signals, vibration signals, and acoustic emission signals during the cutting process in real time, and performs synchronization and preprocessing on the acquired signals to obtain a fused feature sequence. The hidden state setting mapping module sets the tool wear amount and the workpiece surface equivalent hardness coefficient as two coupled time-varying hidden states, and establishes a nonlinear mapping relationship between the fused feature sequence and the two hidden states, using the nonlinear mapping relationship as the input for joint estimation. The dual-state joint estimation module uses a dual-state unscented Kalman filter to perform online joint estimation of the two hidden states. It uses the cross covariance matrix to decouple and fuse the contributions of tool wear and hardness change in the feature sequence, and outputs the tool wear estimate and hardness estimate at the current time. The compensation calculation module calculates the machining dimension deviation caused by wear based on the tool wear estimate, and uses the hardness estimate to correct the cutting force baseline to eliminate the interference of hardness fluctuation on wear estimation, thus forming a compensation amount only for wear. The compensation execution and feedback module writes the compensation amount into the tool compensation parameters of the CNC system in real time, monitors the signal changes after compensation, and feeds the new data back to the update process of the dual-state unscented Kalman filter to iteratively optimize the joint estimation at the next moment.

2. The tool wear adaptive compensation system based on multi-dimensional sensor data fusion according to claim 1, characterized in that, The establishment of a nonlinear mapping relationship between the fused feature sequence and the two hidden states, using this nonlinear mapping relationship as input for joint estimation, specifically includes: The dynamic force gradient within each cutting stroke is extracted from the fused feature sequence, and the fractal dimension of the dynamic force gradient is calculated to obtain the fractal feature curve. The fractal characteristic curve is decomposed into a long-term trend component and a short-period fluctuation component. The long-term trend component represents the cumulative evolution of tool wear, while the short-period fluctuation component represents the instantaneous disturbance of the equivalent hardness coefficient of the workpiece surface. The long-term trend component and the short-period fluctuation component are multiplied point-to-point to generate a coupled modulation sequence, which is then used as the output of the nonlinear mapping relationship.

3. The tool wear adaptive compensation system based on multi-dimensional sensor data fusion according to claim 2, characterized in that, The acquisition of the fractal characteristic curve specifically includes: The force signal in each cutting stroke is differentially processed according to the sampling time sequence to obtain the instantaneous gradient sequence; The instantaneous gradient sequence is mapped to a grid space of different scales, and the total number of grids covering the gradient waveform at each scale is counted. A linear fit is performed on the logarithmic relationship between the grid scale and the total number of grids. The negative value of the fitted slope is taken as the fractal dimension of the cutting stroke. The fractal dimensions of each cutting stroke are then connected in sequence to form a fractal characteristic curve.

4. The tool wear adaptive compensation system based on multi-dimensional sensor data fusion according to claim 1, characterized in that, The output of the current tool wear estimate and hardness estimate specifically includes: Based on the tool wear estimate and hardness estimate output from the previous moment and their joint covariance, a set of symmetrically distributed sampling points are generated, each of which contains both wear and hardness components. Substitute each sampling point into the nonlinear mapping relationship to obtain the corresponding predicted observation point, and obtain the predicted observation sequence through weighted fusion; The residual sequence between the predicted observation sequence and the actual fused feature sequence is calculated. The update weights of wear and hardness are dynamically adjusted using the cross covariance matrix between the sampling points and the predicted observation points. This ensures that fluctuations caused by hardness in the residual sequence are preferentially absorbed by the hardness estimate, and drifts caused by wear are preferentially absorbed by the wear estimate. The decoupled tool wear estimate and hardness estimate are then output.

5. The tool wear adaptive compensation system based on multi-dimensional sensor data fusion according to claim 4, characterized in that, The process of obtaining the predicted observation sequence through weighted fusion specifically includes: Each sampling point is combined with its corresponding predicted observation point into a point pair. The diagonal length of the recursion graph in phase space for each point pair is calculated to obtain the recursion intensity value. The reciprocal of the recursive intensity value of each sampling point is multiplied by the reciprocal of the covariance between the sampling point and the mean of all sampling points to generate the fusion weight of each sampling point. The fusion weights are logarithmically compressed and normalized, then multiplied by the predicted observation points corresponding to each sampling point and summed to output the predicted observation sequence.

6. The tool wear adaptive compensation system based on multi-dimensional sensor data fusion according to claim 1, characterized in that, The compensation amount formed is only for wear and tear, specifically including: The estimated tool wear value is decomposed into axial offset and radial offset along the tool axis, and the initial value of machining dimension deviation is synthesized. Using the hardness estimate as a normalization coefficient, recursively median filtering is applied to the historical sliding window of the real-time cutting force baseline to extract the pure wear force residual after removing hardness fluctuations. The linear projection coefficient is calculated by combining the residual of pure wear force with the initial value of machining dimension deviation. The final compensation amount is then output by multiplying the linear projection coefficient with the initial value of machining dimension deviation.

7. The tool wear adaptive compensation system based on multi-dimensional sensor data fusion according to claim 6, characterized in that, The initial value of the composite processing dimensional deviation specifically includes: Obtain the unit direction vector of the axis corresponding to the current tool posture, and project the estimated tool wear value onto the direction vector to obtain the axial offset component. After subtracting the axial offset component from the estimated tool wear value, the tool is orthogonally decomposed into the radial plane of the tool to obtain the radial offset component with the spindle rotation center as the origin. Based on the angle between the radial offset component and the normal of the machined surface, the radial offset component is projected onto the normal and tangential directions respectively. Then, the normal projection is combined with the square root of the axial component to output the initial value of the machining dimension deviation.

8. The tool wear adaptive compensation system based on multi-dimensional sensor data fusion according to claim 1, characterized in that, The process of feeding new data back to the two-state unscented Kalman filter update to iteratively optimize the joint estimate at the next time step specifically includes: Force and vibration signals are captured during three consecutive cutting strokes after the compensation amount is written. The power spectrum energy difference before and after compensation is calculated, and the compensation response characteristic quantity is generated. The compensation response features are time-aligned and waveform-matched with the preset compensation effect template to output the compensation confidence level. The diagonal element value of the process noise covariance in the dual-state unscented Kalman filter is dynamically adjusted based on the compensation confidence level, so that the filter update process is biased towards historical data with low confidence or new data with high confidence, thus completing iterative optimization.

9. The tool wear adaptive compensation system based on multi-dimensional sensor data fusion according to claim 8, characterized in that, The generated compensation response feature specifically includes: Take the force signal and vibration signal of the last complete cutting stroke before writing the compensation amount, multiply them point by point to form a composite reference sequence, and calculate the autocorrelation spectrum energy of the composite reference sequence. After writing the compensation amount, the force signal and vibration signal of each of the three consecutive cutting strokes are taken and the sliding inner product is performed with the composite reference sequence to obtain three differential energy sequences. For each differential energy sequence, sort the local peaks, take the sum of the top three peaks in each sequence as the differential feature of the cutting stroke, and then sum the differential features of the three cutting strokes according to exponential decay weighting to output the compensation response feature quantity.

Images