A CNC machining center online precision compensation method and system

Abstract

The application discloses a CNC machining center online precision compensation method and system, and relates to the technical field of numerical control machining and intelligent control. The method synchronously collects temperature, vibration and displacement multi-channel sensing signals, and obtains normalized feature vectors of each data source through time sequence alignment, denoising and feature standardization processing; sub-indexes are calculated from three dimensions to obtain comprehensive reliability through nonlinear product fusion; dynamic fusion weights are generated through proportional normalization, time smoothing memory and forced weight reduction; independent error mapping is established for each data source to obtain independent estimators, and fusion error estimation is obtained through weighted fusion and converted into compensation amount of each motion axis to inject into the numerical control system; model parameters and historical feature mean values are updated online according to the measured machining error, and the next compensation cycle is executed in a loop. The application solves the problem of insufficient robustness of fixed weight fusion, and effectively improves the machining precision and compensation stability of the CNC machining center under complex working conditions.

Description

Technical Field

[0001] This application relates to the field of numerical control machining and intelligent control technology, specifically to an online accuracy compensation method and system for CNC machining centers. Background Technology

[0002] In the field of high-end equipment manufacturing, the machining accuracy of Computerized Numerical Control (CNC) directly affects part quality and production efficiency. With the increasing demand for complex surface machining and high-precision manufacturing, error control during the machining process has become a critical issue. Currently, the main sources of error affecting CNC machining accuracy include thermal deformation errors, structural geometric errors, and dynamic errors generated during the machining process. These errors exhibit significant time-varying and coupling characteristics.

[0003] Existing technologies typically employ error compensation methods based on a single data source, such as thermal error modeling using temperature sensors or static compensation using probes to acquire geometric errors. While these methods are effective under specific operating conditions, the complexity and multi-factor coupling of error sources in actual processing environments mean that a single data source cannot fully reflect the characteristics of error changes, leading to lags or biases in the compensation results. To improve compensation accuracy, some existing technologies attempt to introduce multiple sensor data for joint modeling, but most use fixed weights or simple superposition methods for fusion, failing to fully consider the differences in contributions from different data sources at different processing stages. When a data source experiences noise or failure, it still participates in the compensation calculation, which can actually reduce the overall compensation accuracy. Summary of the Invention

[0004] The purpose of this application is to provide an online accuracy compensation method and system for CNC machining centers to solve the problems mentioned in the background art.

[0005] In one aspect, an embodiment of this application provides an online accuracy compensation method for CNC machining centers. The method includes: during machining, simultaneously acquiring multiple sensor signals reflecting different error sources, including temperature signals reflecting thermal deformation errors, vibration signals reflecting dynamic errors, and displacement signals reflecting geometric errors; performing time alignment and denoising on each sensor signal, extracting the corresponding feature quantities for each sensor signal, standardizing each feature quantity to obtain a normalized feature vector for each data source; based on the normalized feature vectors of each data source and the sensor's operating status, calculating sub-indicators from three dimensions: data statistical stability, consistency with historical feature distributions, and the sensor's own health status; fusing the three sub-indicators using a non-linear product to obtain the comprehensive reliability of each data source; generating dynamic fusion weights based on the comprehensive reliability of each data source, and applying these weights to the overall data source. The initial weights are obtained by proportionally normalizing the overall credibility. Time smoothing memory is introduced to smooth the initial weights, and data sources with an overall credibility lower than a preset threshold are subject to forced weight reduction. Based on the normalized feature vectors of each data source, a local error mapping is independently established for each data source, mapping the normalized feature vectors of each data source into independent estimates of the current machining error. The independent estimates of each data source are weighted and fused based on the dynamic fusion weights of each data source to obtain a fusion error estimate. The fusion error estimate is converted into the compensation amount of each motion axis and injected into the CNC system for real-time online compensation of the machining error. The measured value of the current machining error is obtained, and the model parameters and historical feature mean of the local error mapping of each data source are updated online based on the measured value. The updated model parameters and historical feature mean are returned to the step of collecting sensor signals, and the next compensation cycle is executed cyclically.

[0006] In conjunction with the first aspect, in some implementations of the first aspect, based on the normalized feature vectors of each data source and the sensor's operating status, sub-indicators are calculated from three dimensions: data statistical stability, consistency with historical feature distribution, and the sensor's own health status. These three sub-indicators are then fused using a non-linear product to obtain the comprehensive reliability of each data source. This includes: calculating a data statistical stability sub-indicator based on the fluctuation range of the normalized feature vectors of each data source within a preset time window; the larger the fluctuation range, the smaller the data statistical stability sub-indicator. Calculating a historical feature distribution consistency sub-indicator based on the deviation between the current normalized feature vector and the historical feature mean of the data source; the larger the deviation, the smaller the historical feature distribution consistency sub-indicator. Calculating a sensor health status sub-indicator based on the data packet loss rate of the corresponding sensor within a preset time window; the higher the data packet loss rate, the smaller the sensor health status sub-indicator. Finally, the data statistical stability sub-indicator, the historical feature distribution consistency sub-indicator, and the sensor health status sub-indicator are fused using a power product to obtain the comprehensive reliability of each data source.

[0007] In conjunction with the first aspect, in some implementations of the first aspect, dynamic fusion weights are generated based on the comprehensive credibility of each data source. The comprehensive credibility is proportionally normalized to obtain the initial weights. Time smoothing memory is introduced to smooth the initial weights, and forced weight reduction is applied to data sources with comprehensive credibility below a preset threshold. This includes: proportionally normalizing the comprehensive credibility of each data source to obtain the initial weights of each data source; weighting the initial weights of each data source with the dynamic fusion weights of the previous compensation period using a preset smoothing coefficient to obtain the smoothed weights of each data source, with the initial weights used as the dynamic fusion weights of the previous compensation period in the first compensation period; obtaining the data statistical stability sub-indicator and sensor health status sub-indicator of each data source, and when any sub-indicator is below the corresponding preset threshold, multiplying the smoothed weight of the data source by a preset penalty coefficient for forced weight reduction, where the penalty coefficient is less than one; and re-normalizing the smoothed weights of all data sources after forced weight reduction to obtain the dynamic fusion weights of each data source.

[0008] In conjunction with the first aspect, in some implementations of the first aspect, the independent estimates of each data source are weighted and fused based on the dynamic fusion weights of each data source to obtain the fusion error estimate. This includes: multiplying the dynamic fusion weights of each data source with the independent estimates of the corresponding data source and summing the results to obtain the global error estimate; and performing low-pass filtering on the global error estimate to obtain the fusion error estimate.

[0009] In conjunction with the first aspect, in certain implementations of the first aspect, the fusion error estimate is converted into compensation amounts for each motion axis and injected into the CNC system to perform real-time online compensation for machining errors. This includes: converting the fusion error estimate into compensation amounts corresponding to each motion axis through a preset compensation mapping matrix, wherein the compensation mapping matrix is ​​determined by the machine tool geometric parameters and the error propagation relationship; applying an upper bound constraint to the variation range of the compensation amounts of each motion axis during adjacent compensation cycles, and when the variation range exceeds the upper bound constraint, truncating the excess portion and accumulating it to subsequent compensation cycles for gradual compensation; and injecting the compensation amounts that satisfy the upper bound constraint into the CNC system through at least one of coordinate correction or tool compensation correction methods to complete the real-time online compensation for the current machining error.

[0010] In conjunction with the first aspect, in certain implementations of the first aspect, the steps of obtaining the measured value of the current processing error, updating the model parameters and historical feature mean of the local error mapping of each data source online based on the measured value, and returning the updated model parameters and historical feature mean to the acquired sensor signal, and then cyclically executing the next compensation cycle, include: obtaining the measured value of the current processing error through online probe or sampling measurement; updating the model parameters of the local error mapping of each data source online using a recursive least squares algorithm with a forgetting factor based on the measured value, where the forgetting factor is used to control the weight decay rate of historical data in updating model parameters; updating the historical feature mean of each data source online using an exponentially weighted moving average method based on the current normalized feature vector of each data source, where the updated historical feature mean is used to calculate the historical feature distribution consistency sub-index; and returning the updated model parameters and historical feature mean to the acquired sensor signal, and then cyclically executing the next compensation cycle.

[0011] In conjunction with the first aspect, in certain implementations of the first aspect, time alignment and denoising processing are performed on each sensing signal, and the corresponding feature quantities of each sensing signal are extracted. Each feature quantity is then standardized to obtain a normalized feature vector for each data source. This includes: resampling each sensing signal to a unified target frequency using a resampling method to obtain an alignment sequence for each data source; applying filtering processing to the alignment sequence to obtain an alignment-filtered sequence for each data source; extracting corresponding feature quantities from the alignment-filtered sequences of each data source, wherein the first-order time difference is extracted as the thermal deformation rate feature quantity for the temperature signal, the root mean square value within a preset time window is extracted as the vibration intensity feature quantity for the vibration signal, and the offset relative to the reference position is extracted as the geometric deviation feature quantity for the displacement signal; and standardizing the thermal deformation rate feature quantity, vibration intensity feature quantity, and geometric deviation feature quantity respectively to eliminate dimensional differences between different physical quantities, thereby obtaining a normalized feature vector for each data source.

[0012] In conjunction with the first aspect, in some implementations of the first aspect, the standardization process adopts a zero-mean unit variance standardization method based on historical mean and historical standard deviation. The historical mean and historical standard deviation are obtained from historical data statistics in the offline calibration stage and are continuously corrected in the online update step.

[0013] In conjunction with the first aspect, in some implementations of the first aspect, a local error mapping is independently established for each data source, and the normalized feature vector of each data source is mapped to an independent estimate of the current processing error. This includes: establishing an independent local error mapping model for each data source, wherein the local error mapping model adopts a linear mapping model or a polynomial mapping model, and the independent estimate of each data source has a unified error displacement dimension; the initial values ​​of the model parameters of the local error mapping model are obtained through offline calibration and continuously corrected by an online update step during the processing operation.

[0014] Secondly, one embodiment of this application provides an online precision compensation system for a CNC machining center. This system includes: a signal acquisition module, used to synchronously acquire multiple sensor signals reflecting different error sources during machining. These multiple sensor signals include a temperature signal reflecting thermal deformation error, a vibration signal reflecting dynamic error, and a displacement signal reflecting geometric error; a preprocessing module, connected to the signal acquisition module, used to perform time-series alignment and denoising on each sensor signal, extract the corresponding feature quantities of each sensor signal, and standardize each feature quantity to obtain a normalized feature vector for each data source; a reliability assessment module, connected to the preprocessing module, used to calculate sub-indicators based on the normalized feature vectors of each data source and the sensor's operating status, from three dimensions: data statistical stability, consistency with historical feature distribution, and the sensor's own health status. The three sub-indicators are then fused using a non-linear multiplication method to obtain the comprehensive reliability of each data source; and a weight generation module, connected to the reliability assessment module, used to generate dynamic fusion weights based on the comprehensive reliability of each data source, and to proportionally normalize the comprehensive reliability to obtain initial weights. The system employs a time-smoothing memory approach to smooth the initial weights and applies forced weight reduction to data sources with a comprehensive credibility score below a preset threshold. An error mapping module, connected to the preprocessing module, independently establishes local error mappings for each data source based on its normalized feature vectors, mapping each data source's normalized feature vector into an independent estimate of the current machining error. A fusion estimation module, connected to both the weight generation and error mapping modules, weights and fuses the independent estimates of each data source based on dynamic fusion weights to obtain a fusion error estimate. A compensation injection module, connected to the fusion estimation module, converts the fusion error estimate into compensation values ​​for each motion axis and injects them into the CNC system for real-time online compensation of machining errors. An online update module, connected to both the error mapping and credibility assessment modules, acquires the measured value of the current machining error, updates the model parameters and historical feature mean values ​​of the local error mappings for each data source online based on the measured values, and feeds the updated model parameters and historical feature mean values ​​back to the error mapping and credibility assessment modules for use in the next compensation cycle.

[0015] Compared with the prior art, the beneficial effects of this application are: 1. This application independently calculates sub-indicators from three dimensions: data statistical stability, consistency with historical feature distribution, and the sensor's own health status. It obtains the comprehensive reliability through a power product nonlinear fusion method, so that a severe deterioration in any dimension can independently drive a sharp drop in the comprehensive reliability without being diluted by the high values ​​of other dimensions. This fundamentally prevents the failure of data sources from contaminating the fusion results and significantly improves the robustness and accuracy of multi-source fusion error compensation.

[0016] 2. This application generates dynamic fusion weights through the synergistic effect of proportional initialization, time smoothing memory, and forced weight reduction, enabling a rapid response to persistent severe anomalies. The two mechanisms complement each other, giving the dynamic fusion weights both temporal continuity and anomaly robustness, effectively ensuring the stability and reliability of online accuracy compensation. Attached Figure Description

[0017] Figure 1 A flowchart illustrating an embodiment of the online accuracy compensation method for CNC machining centers provided in this application; Figure 2 This is a flowchart illustrating the process of calculating sub-indicators based on normalized feature vectors from various data sources and sensor operating status, from three dimensions: data statistical stability, consistency with historical feature distribution, and sensor health status, according to an embodiment of this application. Figure 3 This application provides a schematic diagram of a process for generating dynamic fusion weights based on the comprehensive credibility of each data source, proportionally normalizing the comprehensive credibility to obtain initial weights, introducing time smoothing memory to smooth the initial weights, and applying forced weight reduction to data sources with comprehensive credibility below a preset threshold. Figure 4 This is a schematic diagram illustrating the process of converting the fusion error estimate into compensation amounts for each motion axis and injecting it into the CNC system to perform real-time online compensation for machining errors, as provided in an embodiment of this application. Figure 5 This is a schematic diagram of the structure of an online precision compensation system for a CNC machining center provided in an embodiment of this application. Detailed Implementation

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

[0019] Figure 1This is a flowchart illustrating an embodiment of the online accuracy compensation method for CNC machining centers provided in this application. Figure 1 As shown, an embodiment of this application provides an online accuracy compensation method for CNC machining centers, which includes the following steps: Step 101: During the processing, multiple sensor signals reflecting different error sources are collected simultaneously.

[0020] It should be understood that multi-channel sensor signals refer to the time-series physical quantity signals synchronously output by multiple types of sensors arranged in different parts of the machine tool. Each signal is independently acquired and transmitted in parallel, collectively forming the original data source for error modeling. Multi-channel sensor signals include temperature signals reflecting thermal deformation errors, vibration signals reflecting dynamic errors, and displacement signals reflecting geometric errors.

[0021] Thermal deformation error refers to the thermal expansion or deformation of heat-sensitive components such as machine tool spindles, lead screws, and columns due to changes in the temperature field, resulting in relative displacement deviation between the tool and the workpiece; dynamic error refers to the machine tool structure vibration caused by dynamic factors such as cutting force excitation and spindle speed fluctuations during machining, causing the actual movement trajectory of the tool to deviate from the theoretical trajectory; geometric error refers to the inherent geometric deviations of each motion axis of the machine tool, such as positioning deviation, straightness error, and perpendicularity error, caused by manufacturing, assembly, and wear factors.

[0022] Step 102: Perform time-series alignment and denoising on each sensor signal, extract the feature quantities corresponding to each sensor signal, standardize each feature quantity, and obtain the normalized feature vector of each data source.

[0023] It should be understood that timing alignment refers to using a resampling method to unify sensor signals with different sampling frequencies to the same time reference, thereby eliminating timing deviations caused by differences in sampling frequencies.

[0024] Characteristic quantities refer to physical quantities extracted from various sensor signals that can sensitively reflect their contribution to processing errors. Specifically, they include thermal deformation rate characteristic quantities of temperature signals, vibration intensity characteristic quantities of vibration signals, and geometric deviation characteristic quantities of displacement signals.

[0025] Step 103: Based on the normalized feature vectors of each data source and the sensor's working status, calculate sub-indicators from three dimensions: data statistical stability, consistency with historical feature distribution, and the sensor's own health status. Then, fuse the three sub-indicators in a non-linear product manner to obtain the comprehensive credibility of each data source.

[0026] It should be understood that the sensor's operating status refers to the hardware-level operating status of the corresponding sensor within a preset time window, such as the data packet loss rate, which is statistically output in real time by the signal acquisition module.

[0027] Step 104: Generate dynamic fusion weights based on the overall credibility of each data source, normalize the overall credibility proportionally to obtain the initial weights, introduce time smoothing memory to smooth the initial weights, and apply forced weight reduction to data sources with overall credibility below a preset threshold.

[0028] It should be understood that dynamic fusion weight refers to the weighting coefficient that is adjusted in real time according to the current comprehensive credibility of each data source in each compensation cycle, and is used to control the contribution ratio of each data source in subsequent error fusion.

[0029] Step 105: Based on the normalized feature vectors of each data source, establish local error mappings independently for each data source, and map the normalized feature vectors of each data source into independent estimates of the current processing error.

[0030] It should be understood that local error mapping refers to a mathematical model that is independently established in the feature space of each data source and maps the normalized feature vector to the error displacement. It adopts a linear mapping model or a polynomial mapping model, and there is no cross-source information coupling between the data sources.

[0031] Step 106: Based on the dynamic fusion weights of each data source, the independent estimates of each data source are weighted and fused to obtain the fusion error estimate.

[0032] Step 107: Convert the fusion error estimate into compensation amounts for each motion axis and inject it into the CNC system to perform real-time online compensation for machining errors.

[0033] Step 108: Obtain the measured value of the current processing error, update the model parameters and historical feature mean of the local error mapping of each data source online based on the measured value, and return the updated model parameters and historical feature mean to the step of collecting sensor signals, and repeat the next compensation cycle.

[0034] It should be understood that the measured value refers to the true value of the current machining error obtained through online probe or sampling measurement, which serves as a supervision label for online updating of model parameters.

[0035] The historical feature mean refers to the long-term weighted statistical mean of the normalized feature vectors of each data source. It is continuously updated using an exponentially weighted moving average method and serves as the benchmark for calculating the historical feature distribution consistency sub-index.

[0036] The compensation cycle refers to the time unit corresponding to the system completing one complete process from sensor signal acquisition to compensation injection. After each compensation cycle, the system returns to step 101 with the updated model parameters and the historical feature mean.

[0037] This embodiment synchronously collects multiple sensor signals of temperature, vibration, and displacement. After normalized feature extraction, comprehensive reliability assessment through nonlinear fusion of multi-dimensional sub-indices, dynamic fusion weight generation with time smoothing memory and forced weight reduction, and independent local error mapping and weighted fusion of each data source, the fusion error estimate is converted into compensation amount for each motion axis and injected into the CNC system in real time. Based on the measured value of machining error, the model parameters and historical feature mean are updated online and iterated cyclically. This achieves multi-source adaptive online accuracy compensation for thermal deformation error, dynamic error, and geometric error, effectively improving the machining accuracy and robustness of the CNC machining center while ensuring compensation stability.

[0038] Figure 2 This is a flowchart illustrating the process of calculating sub-indicators based on normalized feature vectors from various data sources and sensor operating status, from three dimensions: data statistical stability, consistency with historical feature distribution, and the sensor's own health status, as provided in one embodiment of this application. Figure 2 As shown in the figure, an embodiment of this application provides an online accuracy compensation method for CNC machining centers. Based on the normalized feature vectors of each data source and the working status of the sensor, sub-indicators are calculated from three dimensions: data statistical stability, consistency with historical feature distribution, and the health status of the sensor itself. The three sub-indicators are then fused using a nonlinear product to obtain the comprehensive reliability of each data source. The method includes the following steps: Step 201: Calculate the data statistical stability sub-index based on the fluctuation range of the normalized feature vectors of each data source within a preset time window. The larger the fluctuation range, the smaller the data statistical stability sub-index.

[0039] Specifically, the data statistical stability sub-indicator Reflecting data source The statistical stability of the normalized feature vector within a preset time window. If the variance of the normalized feature vector of a data source increases abnormally within a short period of time, it usually means that the sensor is subject to external interference or is unstable in its operation, and the reliability of its current output should be reduced. The data statistical stability sub-index is calculated according to the following formula: ; in, For data source The normalized eigenvector of length is the normalized eigenvector of length . The mean of the variances of each dimension within the preset time window. This is a proportional adjustment parameter used to control the sensitivity to stability degradation. The value range of this sub-index is... The more stable the normalized eigenvector, the better. The closer the normalized eigenvector is to 1, the more drastic the fluctuation. The closer it is to 0, the smaller the statistical stability sub-index of the data becomes, thus achieving a mapping relationship that the greater the fluctuation range, the smaller the statistical stability sub-index.

[0040] Step 202: Based on the degree of deviation between the current normalized feature vector and the historical feature mean of the data source, calculate the historical feature distribution consistency sub-index. The greater the degree of deviation, the smaller the historical feature distribution consistency sub-index.

[0041] Specifically, the historical characteristic distribution consistency sub-index This reflects the degree of deviation between the current normalized feature vector and the historical normal feature distribution of the data source. Under normal operating conditions, each data source should output a normalized feature vector that is similar to the historical pattern under similar processing conditions. If the Euclidean distance between the current normalized feature vector and the historical feature mean is too large, it indicates that the characteristics of the data source have undergone abnormal drift, and the reliability should be reduced accordingly. The consistency sub-index of historical feature distribution is calculated using the following formula: ; in, For data source The historical feature mean of the normalized feature vector is continuously updated by the online update step; The tolerance bandwidth parameter controls the sensitivity to the decrease in historical consistency. The range of values ​​is The value is adjusted based on the historical distribution variance of the feature space of each data source, and can be taken as the mean of the square of the magnitude of the historical feature vector. The value range of this sub-index is... The greater the deviation, the better. The smaller the value, the smaller the mapping relationship between the greater the deviation and the smaller the historical feature distribution consistency sub-index. This sub-index can effectively identify abnormal situations such as sensor zero-point drift, gain drift, and feature distribution migration caused by switching operating conditions.

[0042] Step 203: Calculate the sensor health status sub-index based on the data packet loss rate of the corresponding sensor within a preset time window. The higher the data packet loss rate, the smaller the sensor health status sub-index.

[0043] Specifically, sensor health status sub-indicators This assessment directly evaluates the physical operating status of the corresponding sensor at the communication and hardware level, independent of the numerical content of the normalized feature vector, serving as a supplement to the first two types of sub-indicators. The sensor health status sub-indicator is calculated using the following formula: ; in, For data source The corresponding data packet loss rate of the sensor within a preset time window, with a value range of [value missing]. The higher the data packet loss rate, the better. The smaller the value, the higher the data packet loss rate, and the smaller the value of the sensor health status sub-index. This sub-index is used to identify hardware-level failure modes such as communication interruptions and data acquisition board malfunctions.

[0044] Step 204: The data statistical stability sub-indicator, the historical feature distribution consistency sub-indicator, and the sensor health status sub-indicator are fused by power product to obtain the comprehensive credibility of each data source.

[0045] Specifically, the three types of sub-indicators obtained in steps 201 to 203 are non-linearly fused in the form of a power product, and the overall credibility of each data source is calculated according to the following formula. : ; in, As the trade-off parameters for each sub-indicator, the relative contributions of the data statistical stability sub-indicator, the historical characteristic distribution consistency sub-indicator, and the sensor health status sub-indicator to the overall reliability are adjusted respectively. The typical value range for each of the three is [missing information]. In actual deployment, the settings can be adjusted based on the characteristics of the machine tool's operating conditions and prior knowledge of sensor reliability. As a normalization factor, ensure A severe deterioration in any sub-indicator will lead to a sharp drop in overall reliability, without being diluted by high values ​​in other sub-indicators. For example, when the sensor health status sub-indicator... At that time, regardless of the values ​​of the data statistical stability sub-indicator and the consistency sub-indicator of historical characteristic distribution, All values ​​approach 0, fundamentally preventing abnormal outputs from faulty data sources from contaminating subsequent fusion results.

[0046] Figure 3 This application provides a schematic diagram illustrating a process for generating dynamic fusion weights based on the comprehensive credibility of various data sources, proportionally normalizing the comprehensive credibility to obtain initial weights, introducing time-smoothing memory to smooth the initial weights, and applying forced weight reduction to data sources with comprehensive credibility below a preset threshold. Figure 3 As shown, an embodiment of this application provides an online accuracy compensation method for CNC machining centers. This method generates dynamic fusion weights based on the comprehensive reliability of each data source, proportionally normalizes the comprehensive reliability to obtain initial weights, introduces time-smoothing memory to smooth the initial weights, and applies forced weight reduction to data sources with comprehensive reliability below a preset threshold. The method includes the following steps: Step 301: Normalize the overall credibility of each data source proportionally to obtain the initial weight of each data source.

[0047] Specifically, the overall credibility of the data sources obtained in step 204 is... Normalize proportionally to form initial weights. : ; This step transforms the relative magnitude of overall credibility into a relative allocation of weights, ensuring that data sources with higher overall credibility receive greater initial weights in subsequent fusion.

[0048] Step 302: The initial weights of each data source and the dynamic fusion weights of the previous compensation period are weighted and averaged according to a preset smoothing coefficient to obtain the smoothing weights of each data source. In the first compensation period, the initial weights are used as the dynamic fusion weights of the previous compensation period.

[0049] Specifically, based on the proportional initialization, a time-smoothing memory is introduced, which combines the initial weights of the current compensation cycle with the dynamically fused weights of the previous compensation cycle according to a preset smoothing coefficient. A weighted average is performed to form a smooth weight with a time memory effect. : ; in, The preset smoothing coefficient has a range of values. Typical value ; The smaller the value, the smoother the change in the smoothing weight, the stronger the resistance to short-term anomalies, but the more delayed the response to persistent changes. The larger the smoothing weight, the more timely it tracks changes in overall credibility, but the relatively weaker its ability to resist transient interference. This is the dynamic fusion weight from the previous compensation cycle. When the first compensation cycle is executed, the dynamic fusion weight from the previous compensation cycle does not yet exist; therefore, the current initial weight is used. Substitute the values ​​into the calculation. Time-smoothing memory ensures that occasional data anomalies within a single compensation cycle do not cause abrupt changes in the smoothing weights, thereby avoiding unexpected jumps in the fusion error estimation and ensuring the stability of the CNC system's compensation injection.

[0050] Step 303: Obtain the data statistical stability sub-indicator and sensor health status sub-indicator of each data source. When any sub-indicator is lower than the corresponding preset threshold, the smoothing weight of the data source is multiplied by a preset penalty coefficient to force a weight reduction. The penalty coefficient is less than one.

[0051] Specifically, a forced weight reduction is introduced. When a severe anomaly is detected in the data source, a forced scaling is actively applied to its smoothing weights to achieve a rapid response to failed data sources. The data statistical stability sub-indicator obtained in step 201 is then acquired. Compared with the sensor health status sub-indicators obtained in step 203 ,when or At that time, a forced reduction in the smoothing weights of the data source is applied: ; in, The preset threshold for the data statistical stability sub-indicator, with a typical value range of [value range missing]. ; The preset threshold values ​​for the sensor health status sub-indicators are typically within the range of [value range missing]. This means that when the data packet loss rate exceeds 10% to 30% within a preset time window, a forced demotion is triggered. The preset penalty coefficient has a range of values. The typical value is 0.1, and the value is much less than 1 to ensure that the impact of the data source on the fusion result is reduced to a negligible level after forced weight reduction. The forced weight reduction mentioned above complements the time smoothing memory in step 302: the time smoothing memory ensures the stability of the smoothed weights under normal working conditions, while the forced weight reduction ensures timely response under abnormal working conditions. The two work together to overcome the shortcomings of existing methods, such as severe weight fluctuations and untimely weight reduction of abnormal data sources.

[0052] Step 304: Normalize the smoothing weights of all data sources after forced weight reduction to obtain the dynamic fusion weights of each data source.

[0053] Specifically, after completing the forced weight reduction process in step 303, the smoothing weights of all data sources are renormalized to obtain the dynamic fusion weights of each data source. : ; This normalization operation ensures that the sum of the dynamic fusion weights of all data sources is always equal to 1, satisfying the weight constraints of subsequent weighted fusion. The dynamic fusion weights generated through steps 301 to 304... It must simultaneously meet the following three conditions: first, it must be positively correlated with the current overall credibility; second, it must have temporal continuity and not change abruptly due to occasional disturbances in a single compensation cycle; and third, it must be able to proactively respond to severe anomalies.

[0054] Furthermore, based on the dynamic fusion weights obtained in step 304 Independent estimators from each data source The global error estimate at the current time is obtained by performing global weighted fusion according to the following formula. : ; The global fusion result comprehensively reflects the contribution of multi-source information such as temperature, vibration, and displacement to the current processing error. Moreover, the contribution ratio of each data source has been dynamically adjusted according to its current comprehensive credibility and historical status. Therefore, it has stronger robustness and accuracy than any single-source independent estimator.

[0055] Further suppressing residual high-frequency jitter in the global error estimate, for Applying a first-order exponential low-pass filter yields the fusion error estimate. : ; in, This is the filtering smoothing coefficient, and its value range is... Typical value ; The smaller the value, the stronger the filtering effect, but the more delayed the response to changes in error. The larger the value, the more timely the response, but the weaker the resistance to high-frequency jitter. In practical applications, Tuning is performed by combining the compensation period with the characteristic frequency of error dynamic changes to achieve a balance between smoothness and timeliness. Fusion error estimation. This is the optimal estimate of the processing error at the current moment in this application, which will be used as the input for generating the compensation amount.

[0056] Figure 4 This is a schematic diagram illustrating a process for real-time online compensation of machining errors, provided in one embodiment of this application, by converting the fusion error estimate into compensation amounts for each motion axis and injecting it into the CNC system. Figure 4 As shown in the figure, an embodiment of this application provides an online accuracy compensation method for CNC machining centers, which converts the fusion error estimate into compensation amounts for each motion axis and injects them into the CNC system to perform real-time online compensation for machining errors, including the following steps: Step 401: The fusion error estimate is converted into the compensation amount corresponding to each motion axis through a preset compensation mapping matrix. The compensation mapping matrix is ​​determined by the machine tool geometric parameters and the error propagation relationship.

[0057] Specifically, through a preset compensation mapping matrix Estimating fusion error Compensation amount mapped to each motion axis : ; in, ( (where is the number of machine tool motion axes, usually 3 to 5) is the compensation mapping matrix, which reflects the distribution of the fusion error estimate in each motion axis direction and is determined by the machine tool geometric parameters and the error propagation relationship. The initial values ​​are obtained through the machine tool's factory precision calibration and can be recalibrated during periodic precision calibration. When multiple error sources have significant projections in multiple motion axis directions, the model can be extended to a multi-dimensional case. The corresponding extension is a vector. Expand to matrix.

[0058] Step 402: Apply an upper bound constraint to the change range of each motion axis compensation amount during adjacent compensation cycles. When the change range exceeds the upper bound constraint, the excess part is truncated and accumulated to subsequent compensation cycles for gradual compensation.

[0059] Specifically, an upper bound constraint is applied to the periodic variation range of the compensation amount: ; in, The maximum allowable single-step compensation change ranges from [value range missing]. Typical values ​​range from 1µm to 5µm, determined based on a combination of machine tool servo response bandwidth and machining accuracy requirements. When the calculated... When the upper limit is exceeded, the excess portion will be truncated to the allowable range, and the excess amount will be... The compensation is accumulated gradually over the compensation cycle to avoid causing disturbance to the processing by injecting too large a compensation amount at once.

[0060] Step 403: The compensation amount that satisfies the upper limit constraint is injected into the CNC system through at least one of coordinate correction method or tool compensation correction method to complete the real-time online compensation for the current machining error.

[0061] Specifically, the following two injection methods are supported, and one or a combination can be selected according to the characteristics of the machine tool CNC system interface. The first method is the coordinate correction method: directly modifying the target point coordinates of the CNC interpolator, and superimposing the compensation amount onto the CNC command target point in real time. : ; The first method is suitable for CNC systems that support real-time coordinate offset interfaces, minimizing compensation injection delay. The second method is tool compensation correction: error compensation is achieved by adjusting the tool radius or length compensation parameters built into the CNC system. This method is suitable for CNC systems that do not have a real-time coordinate offset interface but have an open tool compensation parameter writing interface. The two methods have equivalent compensation effects, both aiming to modify the effective position of the tool relative to the workpiece; the only difference lies in the injection channel.

[0062] Furthermore, the measured value of the current machining error is obtained through online probes or sampling measurement methods. The online probe solution can automatically measure key features after each workpiece is processed, enabling piece-by-piece feedback; the sampling measurement solution can perform precision measurements on selected workpieces at a set frequency during batch production, reducing the time overhead of online measurement. Both methods can provide error truth labels for subsequent model updates.

[0063] Based on measured values, a recursive least squares algorithm with a forgetting factor is used to update the model parameters of the local error mapping for each data source online. The forgetting factor is used to control the weight decay rate of historical data in updating the model parameters. Specifically, for each data source... Local error mapping model parameters Update online using the following recursive formula: ; ; ; in, The covariance matrix is ​​the recursive estimate. For the recursive gain vector, This is the forgetting factor, used to control the rate at which weights decay due to historical data in updating model parameters. A typical value range is... ; The smaller the value, the higher the weight the algorithm assigns to recent data, the faster the model tracks, but the lower its robustness to measurement noise. The closer the value is to 1, the more stable the model, but the slower its response to changes in machine tool conditions. Bias term. It can be extended to the parameter vector and updated uniformly in the same recursive manner.

[0064] Based on the current normalized feature vectors of each data source, the historical feature mean of each data source is updated online using an exponentially weighted moving average method. The updated historical feature mean is used to calculate the historical feature distribution consistency sub-index. Specifically, the historical feature mean of each data source... Update continuously using the following formula: ; in, This is the historical mean decay coefficient, with a value range of [value range missing]. Typical value The settings are adjusted based on the machine tool's thermal balance cycle. The larger the value, the slower the historical characteristic mean updates to new observations, and the stronger its adaptability to long-term slowly changing operating conditions. The smaller the value, the faster the historical feature mean tracks new operating conditions, but the higher the sensitivity to short-term fluctuations. This update ensures that the reference benchmark for the historical feature distribution consistency sub-index in step 202 can adapt to long-term changes in machine tool operating conditions, without continuously misjudging the overall reliability due to the normal evolution of machine tool conditions.

[0065] The step of returning the acquired sensor signal with the updated model parameters and the historical feature mean is repeated for the next compensation cycle.

[0066] Furthermore, a resampling method is used to resample each sensor signal to a unified target frequency, obtaining an aligned sequence for each data source. Specifically, linear interpolation or cubic spline interpolation methods are used to resample the original sampling sequences of each data source. Resampling to a unified target frequency The target frequency is selected based on the principle of being no less than twice the lowest effective information frequency among all data sources to satisfy the basic requirements of the Nyquist sampling theorem, while not exceeding the lowest inherent sampling frequency to avoid invalid upsampling. After resampling, all data sources are fully aligned on the time axis, forming an aligned sequence under a unified time base.

[0067] A filter is applied to the alignment sequence to obtain the alignment filter sequence for each data source. Specifically, a filter of length [missing information] is applied to the alignment sequence. Moving average filtering suppresses high-frequency random noise: ; For broadband signals such as vibration signals, a bandpass filter can be used to retain dynamic information within the effective frequency band. The specific filtering parameters are determined based on the sensor's frequency response characteristics and the processing conditions.

[0068] For each data source's aligned and filtered sequence, corresponding feature quantities are extracted. Specifically, the first-order time difference is extracted from the temperature signal as the thermal deformation rate feature quantity; the root mean square value within a preset time window is extracted from the vibration signal as the vibration intensity feature quantity; and the offset relative to the reference position is extracted from the displacement signal as the geometric deviation feature quantity. Specifically, the three types of feature quantities are calculated using the following formulas: For the temperature signal, the first-order time difference is extracted as a feature of the thermal deformation rate. This reflects the dynamic trend of the machine tool's thermal state: ; For vibration signals, the root mean square value within a preset time window is extracted as the vibration intensity feature. This reflects the dynamic excitation amplitude of the cutting action at the current moment: ; For the displacement signal, the offset relative to the reference position is extracted as a geometric deviation feature. This reflects the deviation between the actual position and the theoretical position of the current motion axis: ; in, This serves as the reference position for the machine tool, obtained through factory calibration or periodic calibration.

[0069] The thermal deformation rate characteristic, vibration intensity characteristic, and geometric deviation characteristic are standardized to eliminate dimensional differences between different physical quantities, resulting in normalized feature vectors for each data source. Specifically, the standardization process employs a zero-mean, unit-variance standardization method based on historical mean and historical standard deviation to standardize the extracted feature components of each data source. ; in, For data source No. dimensional original eigenvalues, and These represent the historical mean and historical standard deviation for the corresponding dimension, obtained from historical data statistics during the offline calibration phase and continuously corrected during the online update step. After standardization, the normalized feature vectors of each data source are uniformly represented as: ; Heterogeneous signals from different types of sensors are mapped to a unified feature space, providing a basis for direct comparison and fusion.

[0070] Furthermore, independent local error mapping models are established for each data source. These models employ either linear or polynomial mapping, ensuring that the independent estimators for each data source have a unified dimension of error displacement. Specifically, a linear mapping model is used as the basic form for each data source. Independent estimators Calculate using the following formula: ; in, For the mapping coefficient vector, These bias terms together constitute the data source. The local error mapping model parameters. For scenarios with significant error nonlinearity, such as the nonlinear relationship between thermal error and temperature difference, the model can be extended to a polynomial mapping form without affecting the validity of the overall methodological framework. Independent estimators for each data source. All have a unified dimension of error displacement, ensuring the physical rationality of weighted fusion.

[0071] The initial values ​​of the model parameters for the local error mapping model are obtained through offline calibration and continuously corrected by an online update step during machining operations. Specifically, in the offline calibration stage, a large number of real error labels under various working conditions are collected using high-precision measuring equipment. Combined with synchronously collected normalized feature vector data, the initial values ​​of the model parameters are solved by least squares regression. During machining operations, the model parameters are continuously updated by a recursive least squares algorithm with a forgetting factor, achieving adaptive tracking of long-term changes in the machine tool's condition.

[0072] Figure 5 This is a schematic diagram of the structure of an online accuracy compensation system for a CNC machining center provided in one embodiment of this application. Figure 5 As shown in one embodiment of this application, an online accuracy compensation system for a CNC machining center is provided. The system includes: The signal acquisition module 501 is used to simultaneously acquire multiple sensor signals reflecting different error sources during the processing. The multiple sensor signals include temperature signals reflecting thermal deformation errors, vibration signals reflecting dynamic errors, and displacement signals reflecting geometric errors.

[0073] The preprocessing module 502 is connected to the signal acquisition module 501. It receives the multi-channel sensor signals output by the signal acquisition module 501 and performs time alignment and noise reduction processing on each channel sensor signal. It also extracts the feature quantities corresponding to each channel sensor signal, performs standardization processing on each feature quantity, and obtains the normalized feature vector of each data source.

[0074] The credibility assessment module 503 is connected to the preprocessing module 502 and the signal acquisition module 501 respectively. It receives the normalized feature vectors of each data source output by the preprocessing module 502 and the sensor operating status output by the signal acquisition module 501. Based on the normalized feature vectors of each data source and the sensor operating status, it calculates sub-indices from three dimensions: data statistical stability, consistency with historical feature distribution, and sensor health status. The three sub-indices are fused in a non-linear product manner to obtain the comprehensive credibility of each data source.

[0075] The weight generation module 504 is connected to the credibility assessment module 503. It receives the comprehensive credibility of each data source output by the credibility assessment module 503 and uses it to generate dynamic fusion weights based on the comprehensive credibility of each data source. It performs proportional normalization on the comprehensive credibility to obtain the initial weights, introduces time smoothing memory to smooth the initial weights, and applies forced weight reduction to data sources with comprehensive credibility below a preset threshold.

[0076] Error mapping module 505, connected to preprocessing module 502, receives normalized feature vectors from each data source output by preprocessing module 502. It is used to independently establish local error mappings for each data source based on the normalized feature vectors of each data source, and map the normalized feature vectors of each data source into independent estimates of the current processing error.

[0077] The fusion estimation module 506 is connected to the weight generation module 504 and the error mapping module 505 respectively. It receives the dynamic fusion weights of each data source output by the weight generation module 504 and the independent estimates of each data source output by the error mapping module 505. It is used to perform weighted fusion of the independent estimates of each data source based on the dynamic fusion weights of each data source to obtain the fusion error estimate.

[0078] The compensation injection module 507 is connected to the fusion estimation module 506. It receives the fusion error estimate output by the fusion estimation module 506 and converts the fusion error estimate into the compensation amount of each motion axis, which is then injected into the CNC system to perform real-time online compensation for machining errors.

[0079] The online update module 508 is connected to the error mapping module 505 and the credibility assessment module 503 respectively. It is used to obtain the measured value of the current processing error, update the model parameters and historical feature mean of the local error mapping of each data source online based on the measured value, and feed back the updated model parameters to the error mapping module 505 and the updated historical feature mean to the credibility assessment module 503 for use in the next compensation cycle.

[0080] Through the coordinated operation of various signal modules, a system was constructed that integrates synchronous acquisition of multi-channel sensor signals, normalized feature extraction, comprehensive reliability assessment through nonlinear fusion of multi-dimensional sub-indicators, dynamic fusion weight generation, independent local error mapping and weighted fusion of each data source, and real-time injection of compensation into the CNC system and online iterative updating of model parameters. This system achieves multi-source adaptive online accuracy compensation for thermal deformation error, dynamic error and geometric error, effectively improving the machining accuracy and system robustness of CNC machining centers under complex working conditions.

[0081] It will be apparent to those skilled in the art that this application is not limited to the details of the exemplary embodiments described above, and that this application can be implemented in other specific forms without departing from the spirit or essential characteristics of this application. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of this application is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within this application. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A method for online accuracy compensation in CNC machining centers, characterized in that, include: During the processing, multiple sensor signals reflecting different error sources are collected simultaneously. These multiple sensor signals include temperature signals reflecting thermal deformation errors, vibration signals reflecting dynamic errors, and displacement signals reflecting geometric errors. The time-series alignment and denoising of each sensor signal are performed, and the feature quantities corresponding to each sensor signal are extracted. The feature quantities are then standardized to obtain the normalized feature vectors of each data source. Based on the normalized feature vectors of each data source and the sensor's working status, sub-indicators are calculated from three dimensions: data statistical stability, consistency with historical feature distribution, and the sensor's own health status. The three sub-indicators are then fused in a non-linear product manner to obtain the comprehensive credibility of each data source. Dynamic fusion weights are generated based on the overall credibility of each data source. The overall credibility is proportionally normalized to obtain the initial weights. Time smoothing memory is introduced to smooth the initial weights. Data sources with overall credibility below a preset threshold are subject to forced weight reduction. Based on the normalized feature vectors of each data source, local error mappings are independently established for each data source, and the normalized feature vectors of each data source are mapped to independent estimates of the current processing error. The independent estimators from each data source are weighted and fused based on the dynamic fusion weights of each data source to obtain the fusion error estimate; The fusion error estimate is converted into compensation amounts for each motion axis and injected into the CNC system to perform real-time online compensation for machining errors. The steps of obtaining the measured value of the current processing error, updating the model parameters and historical feature mean of the local error mapping of each data source online based on the measured value, returning the updated model parameters and historical feature mean to the collected sensor signal, and repeating the process for the next compensation cycle.

2. The method according to claim 1, characterized in that, Based on the normalized feature vectors of each data source and the sensor's operating status, sub-indicators are calculated from three dimensions: data statistical stability, consistency with historical feature distribution, and the sensor's own health status. These three sub-indicators are then fused using a non-linear product to obtain the comprehensive reliability of each data source, including: Based on the fluctuation range of the normalized feature vectors of each data source within a preset time window, a data statistical stability sub-index is calculated. The larger the fluctuation range, the smaller the data statistical stability sub-index. Based on the degree of deviation between the current normalized feature vector and the historical feature mean of the data source, a historical feature distribution consistency sub-index is calculated. The greater the degree of deviation, the smaller the historical feature distribution consistency sub-index. Based on the data packet loss rate of the corresponding sensor within a preset time window, a sensor health status sub-index is calculated. The higher the data packet loss rate, the smaller the sensor health status sub-index. The data statistical stability sub-indicator, historical feature distribution consistency sub-indicator, and sensor health status sub-indicator are fused by exponential product to obtain the comprehensive reliability of each data source.

3. The method according to claim 2, characterized in that, The process involves generating dynamic fusion weights based on the overall credibility of each data source, proportionally normalizing the overall credibility to obtain initial weights, introducing time-smoothing memory to smooth the initial weights, and applying forced weight reduction to data sources with overall credibility below a preset threshold. This includes: The overall credibility of each data source is normalized proportionally to obtain the initial weight of each data source; The initial weights of each data source and the dynamic fusion weights of the previous compensation cycle are weighted and averaged according to a preset smoothing coefficient to obtain the smoothed weights of each data source. In the first compensation cycle, the initial weights are used as the dynamic fusion weights of the previous compensation cycle. The statistical stability sub-indicator and sensor health status sub-indicator of each data source are obtained. When any sub-indicator is lower than the corresponding preset threshold, the smoothing weight of the data source is multiplied by a preset penalty coefficient for forced weight reduction. The penalty coefficient is less than one. The smoothed weights of all data sources after forced weight reduction are re-normalized to obtain the dynamic fusion weights of each data source.

4. The method according to claim 1, characterized in that, The weighted fusion of independent estimates from each data source based on dynamic fusion weights to obtain a fusion error estimate includes: The global error estimate is obtained by multiplying the dynamic fusion weights of each data source with the independent estimates of the corresponding data sources and then summing the results. The global error estimate is low-pass filtered to obtain the fused error estimate.

5. The method according to claim 1, characterized in that, The process of converting the fusion error estimate into compensation amounts for each motion axis and injecting them into the CNC system for real-time online compensation of machining errors includes: The fusion error estimate is converted into the compensation amount corresponding to each motion axis through a preset compensation mapping matrix, which is determined by the machine tool geometric parameters and the error propagation relationship. An upper bound constraint is applied to the change range of each motion axis compensation during adjacent compensation cycles. When the change range exceeds the upper bound constraint, the excess portion is truncated and accumulated to subsequent compensation cycles for gradual compensation. The compensation amount that satisfies the upper limit constraint is injected into the CNC system through at least one of coordinate correction method or tool compensation correction method to complete the real-time online compensation of the current machining error.

6. The method according to claim 2, characterized in that, The steps of obtaining the measured value of the current processing error, updating the model parameters and historical feature mean of the local error mapping of each data source online based on the measured value, and returning the updated model parameters and historical feature mean to the acquired sensor signal, and then repeating the next compensation cycle, include: The measured value of the current machining error is obtained by using an online probe or sampling measurement method; Based on the measured values, the model parameters of the local error mapping of each data source are updated online using a recursive least squares algorithm with a forgetting factor. The forgetting factor is used to control the weight decay rate of the historical data for updating the model parameters. Based on the current normalized feature vectors of each data source, the historical feature mean of each data source is updated online using an exponentially weighted moving average method. The updated historical feature mean is used to calculate the historical feature distribution consistency sub-index. The step of returning the acquired sensor signal with the updated model parameters and the historical feature mean is repeated for the next compensation cycle.

7. The method according to claim 1, characterized in that, The process involves performing time-series alignment and denoising on each sensor signal, extracting the corresponding feature quantities for each sensor signal, and standardizing each feature quantity to obtain a normalized feature vector for each data source, including: The resampling method is used to resample each sensor signal to a unified target frequency to obtain the aligned sequence of each data source; The alignment sequence is filtered to obtain the alignment filter sequence for each data source; The corresponding feature quantities are extracted from the aligned filtering sequences of each data source. Specifically, the first-order time difference is extracted from the temperature signal as the thermal deformation rate feature quantity, the root mean square value within the preset time window is extracted from the vibration signal as the vibration intensity feature quantity, and the offset relative to the reference position is extracted from the displacement signal as the geometric deviation feature quantity. The thermal deformation rate characteristic, vibration intensity characteristic, and geometric deviation characteristic are standardized to eliminate the dimensional differences between different physical quantities, thus obtaining the normalized feature vectors of each data source.

8. The method according to claim 7, characterized in that, The standardization process employs a zero-mean unit variance standardization method based on historical mean and historical standard deviation. The historical mean and historical standard deviation are obtained from historical data statistics during the offline calibration phase and are continuously corrected in the online update step.

9. The method according to claim 1, characterized in that, The step of independently establishing local error mappings for each data source, mapping the normalized feature vectors of each data source to independent estimates of the current processing error, includes: An independent local error mapping model is established for each data source. The local error mapping model adopts a linear mapping model or a polynomial mapping model, and the independent estimators of each data source have a unified error displacement dimension. The initial values ​​of the model parameters of the local error mapping model are obtained through offline calibration and continuously corrected by the online update step during the processing and operation.

10. An online precision compensation system for a CNC machining center, used to implement the method according to any one of claims 1 to 9, characterized in that, include: The signal acquisition module is used to simultaneously acquire multiple sensor signals reflecting different error sources during the processing. The multiple sensor signals include temperature signals reflecting thermal deformation errors, vibration signals reflecting dynamic errors, and displacement signals reflecting geometric errors. The preprocessing module, connected to the signal acquisition module, is used to perform time-series alignment and noise reduction on each channel of sensor signals, extract the feature quantities corresponding to each channel of sensor signals, and perform standardization processing on each feature quantity to obtain the normalized feature vector of each data source. The credibility assessment module, connected to the preprocessing module, is used to calculate sub-indicators based on the normalized feature vectors of each data source and the sensor's working status, from three dimensions: data statistical stability, consistency with historical feature distribution, and the sensor's own health status. The three sub-indicators are then fused in a non-linear multiplication manner to obtain the comprehensive credibility of each data source. The weight generation module, connected to the credibility assessment module, is used to generate dynamic fusion weights based on the comprehensive credibility of each data source, perform proportional normalization on the comprehensive credibility to obtain the initial weights, introduce time smoothing memory to smooth the initial weights, and apply forced weight reduction to data sources with comprehensive credibility below a preset threshold. The error mapping module, connected to the preprocessing module, is used to independently establish local error mappings for each data source based on the normalized feature vectors of each data source, and to map the normalized feature vectors of each data source into independent estimates of the current processing error. The fusion estimation module is connected to the weight generation module and the error mapping module respectively, and is used to perform weighted fusion of the independent estimates of each data source based on the dynamic fusion weight of each data source to obtain the fusion error estimate. The compensation injection module, connected to the fusion estimation module, is used to convert the fusion error estimate into compensation amounts for each motion axis and inject them into the CNC system to perform real-time online compensation for machining errors. The online update module is connected to the error mapping module and the reliability assessment module respectively. It is used to obtain the measured value of the current processing error, update the model parameters and historical feature mean of the local error mapping of each data source online based on the measured value, and feed back the updated model parameters and historical feature mean to the error mapping module and the reliability assessment module respectively for use in the next compensation cycle.

Images