An adaptive threshold construction and online updating method for tool condition monitoring

Abstract

The present application belongs to the technical field of intelligent monitoring of numerical control machining, and discloses a self-adaptive threshold construction and online updating method for tool state monitoring, comprising: in the offline stage, a multi-source signal is collected to construct a historical data set, a two-stage segmented positioning key cutting section is performed to extract features, and a threshold model for outputting monitoring thresholds and confidence is trained; in the online stage, real-time data is positioned to extract features, a self-adaptive threshold is obtained by calling the threshold model to realize tool state discrimination; at the same time, a weak supervision label with reliability is generated by accessing field feedback signals, and multiple drift evidences are jointly monitored; when the drift triggering condition is met and verified by a constraint mechanism containing step length, boundary, monotonicity, multiple and rollback, online incremental updating of the threshold model is performed. The present application reduces the manual calibration and failure risk of fixed thresholds, and significantly improves the long-term stability and reliability of tool monitoring under complex working conditions through closed-loop updating with safety constraints.

Description

Technical Field

[0001] This invention belongs to the technical field of intelligent monitoring in CNC machining, and more specifically, relates to an adaptive threshold construction and online update method for tool status monitoring. Background Technology

[0002] Cutting tools are critical consumable components in metal cutting processes. Wear, chipping, and breakage of these tools directly lead to decreased workpiece accuracy and deteriorated surface quality, and in severe cases, result in workpiece scrap, fixture damage, spindle impact, and machine downtime. In large-scale heavy equipment machining, long-cycle processes, and multi-tool collaborative machining scenarios, where individual workpieces are valuable, cycle times are long, and downtime costs are high, tool anomaly monitoring is crucial for ensuring production safety and reducing costs.

[0003] Existing tool monitoring systems typically collect signals such as spindle power, machine tool internal load, and vibration, and generate alarms through filtering and statistical threshold rules. While this type of solution is simple to deploy and has low engineering costs, it suffers from the following significant drawbacks in long-term operation: First, signal distribution is affected by a variety of factors such as material batch, tool batch, clamping rigidity, cutting parameters, machine tool status, and sensor installation position. Fixed thresholds are prone to failure and can lead to false alarms and missed alarms.

[0004] Second, the processing includes non-steady-state stages such as start-up, idling, reversal, cutting in, and cutting out, making it difficult to reliably extract effective processing segments that are strongly correlated with the tool state, resulting in large feature fluctuations and unreliable thresholds.

[0005] Third, the processing site lacks a closed-loop utilization mechanism for feedback information such as alarm confirmation, tool change, and automatic shutdown and tool retraction. Traditional systems cannot use this information to adaptively evolve thresholds.

[0006] Fourth, while some data-driven methods attempt to update online, they lack security constraints and are easily skewed by abnormal samples or human factor adjustments, failing to meet the engineering requirements of "long-term stability, controllability, and rollback capability" in industrial settings.

[0007] Therefore, there is an urgent need for a tool monitoring method and system that can automatically build thresholds and safely update them online during long-term operation. Summary of the Invention

[0008] To address the aforementioned deficiencies or improvement needs of existing technologies, this invention provides an adaptive threshold construction and online update method for tool status monitoring. Its purpose is to achieve long-term stable monitoring of tool wear, chipping, and breakage without altering the existing industrial field data acquisition and linkage control hardware framework. This is achieved through key cutting segment positioning, threshold model construction, field feedback annotation, drift trigger judgment, and constrained online updates. This solves the technical problems of fixed thresholds relying on manual calibration under complex working conditions, difficulty in stably extracting key cutting segments, threshold mismatch caused by changes in working conditions, difficulty in utilizing field feedback in a closed loop, and lack of safety constraints in online updates.

[0009] To achieve the above objectives, according to one aspect of the present invention, an adaptive threshold construction and online update method for tool condition monitoring is provided, comprising the following steps: S1. Offline Modeling: Collect multi-source monitoring signals from the machining process to construct a historical dataset. Preprocess the historical dataset and locate key cutting segments. Extract features from the key cutting segments to form training samples, train a threshold model, and establish an online update strategy. The threshold model is used to output monitoring thresholds and their confidence levels. S2. Online monitoring: Real-time acquisition of multi-source monitoring signals, adaptive segmentation and positioning of the key cutting segment of the current machining based on CNC system instruction information and signals; extraction of features from the key cutting segment, and input into the threshold model trained in step S1 to obtain the adaptive monitoring threshold and confidence level, and outputting the tool status discrimination result based on the adaptive monitoring threshold; S3. Feedback Labeling: Obtain the field feedback signal related to the tool state discrimination result output in step S2, generate a weak supervision label and calculate its reliability; S4. Online Update: Continuously monitor drift triggering conditions; when the drift triggering conditions are met and verified by the preset update constraint mechanism, based on the online update strategy established in step S1, and combined with the weak supervision label generated in step S3 and its reliability, perform online incremental updates on the parameters and threshold output rules of the threshold model, and output the updated monitoring threshold for subsequent online monitoring.

[0010] Preferably, training the threshold model in step S1 specifically includes: S11. Constructing the training dataset: Collect key cutting segment feature vectors under multiple batches, multiple tools, and multiple working conditions during the offline phase. Each feature vector is labeled with a corresponding threshold label, which is determined based on the actual wear or breakage state of the tool. S12. Training the regression sub-model: using feature vectors As input, with threshold label As output, the training sub-model learns the mapping relationship between features and thresholds, and the regression sub-model outputs the regression threshold. ; S13. Training the uncertainty estimation sub-model: using feature vectors As input, output the prediction variance Var, and calculate the confidence level based on the prediction variance. The specific calculation formula is as follows:

[0011] S14. Establish threshold fusion rules: (This involves setting the regression threshold...) Compared with the preset probability threshold By confidence level Weighted fusion is performed to obtain the adaptive monitoring threshold. The specific calculation formula is as follows:

[0012] in, Regarding confidence level The monotonic weight function, This is a conservative threshold obtained based on quantiles or probability inference; when At higher values, the fusion results tend to favor the regression threshold. ,when At lower thresholds, the fusion results tend to favor a conservative threshold. This is to avoid overly aggressive threshold model outputs when data is sparse or operating conditions are abnormal.

[0013] Preferably, the positioning of the critical cutting segment includes two levels of segmentation: The first level uses CNC system instruction information to determine the start and end boundaries of the monitoring window; the second level calculates the short-time energy and its changes within the monitoring window, eliminating unsteady segments including spindle start-up, idling, spindle reversal, and cutting in and out, retaining the stable cutting process as the critical cutting segment; wherein, the sliding window length is W, the sliding step size is P, and the... k The short-time energy of a window is defined as follows: The change in energy is defined as The calculation formulas are as follows:

[0014]

[0015] in, x ′ represents the preprocessed monitoring signal sequence. Indicates the first k Within the first sliding window i The signal amplitude corresponding to each sampling point; k Indicates the sliding window number. iIndicates the sampling point number within the sliding window; Indicates the first k -1 short-term energy of a sliding window, This represents the short-term energy change between adjacent sliding windows; according to Using the sign and amplitude, construct the entry and exit criteria: when Continuous satisfaction And reach the duration When, the corresponding time is determined as the cutting boundary; when Continuous satisfaction And reach the duration When the corresponding moment is determined as the cutting boundary, the critical cutting narrow window is obtained; in, This indicates the threshold for decision-making. This indicates the cut-out threshold. This indicates the number of sliding windows that continuously satisfy the cut-in condition to determine the cut-in boundary. This indicates the number of sliding windows that continuously satisfy the cutting conditions required to determine the cutting boundary; the critical cutting narrow window indicates the stable cutting time interval defined by the entry and exit boundaries.

[0016] Preferably, the CNC system instruction information includes paired process start triggering instructions and process end triggering instructions to define the monitoring window, and the triggering instructions are implemented by M-code or equivalent semantic code.

[0017] Preferably, the extracted features include at least one or more combinations of the following: the root mean square (RMS) value, energy entropy, or waveform factor of the current signal; the mean, variance, quantile, or slope of the load signal; and the time-frequency band energy, spectral peaks, spectral kurtosis, or wavelet packet energy of the vibration signal; wherein the time-frequency band energy, spectral peaks, and spectral kurtosis are calculated based on the time-frequency representation of the vibration signal, which is obtained through short-time Fourier transform, and the specific calculation formula is as follows:

[0018] in, Indicates the vibration signal in time t ,frequency f The short-time Fourier transform result at the point, This represents the preprocessed discrete vibration signal. Indicates t The window function centered on t Indicates a time index. n Indicates the discrete sampling point number. j Represents the imaginary unit. f Represents frequency variables. e Represents the natural constant; The formula for calculating the root mean square (RMS) value of a current signal is as follows:

[0019] in, This represents the preprocessed discrete current signal, where n represents the discrete sampling point number and N represents the number of sample points in the critical cutting section.

[0020] Preferably, the field feedback signals in step S3 include at least one or more of the following: manual confirmation after an alarm, tool change record, automatic control shutdown trigger, and tool life database record; the feedback signals are graded and weighted to generate the weak supervision labels, which include at least true alarm labels, false alarm labels, and uncertain labels; the reliability calculation formula is as follows:

[0021] in, R Indicates the reliability of the feedback signal. M Indicates the number of feedback signal sources. Indicates the first m Preset weights for feedback signals, Indicates the first m The consistency or reliability score between the feedback signal and the current tool state determination result. ε This indicates a small positive number that prevents the denominator from being zero.

[0022] Preferably, the drift triggering conditions in step S4 include at least one or more of feature distribution drift, false positive and false negative statistical drift, and threshold output fluctuation drift; wherein, the feature distribution drift is quantified using distribution drift evidence, the false positive and false negative statistical drift is quantified using statistical drift evidence, and the threshold output fluctuation drift is quantified using threshold fluctuation evidence; the weighted combination of the distribution drift evidence, statistical drift evidence, and threshold fluctuation evidence yields the joint drift amount, and the specific calculation formula is as follows:

[0023]

[0024] Where D represents the joint drift amount, This indicates evidence of characteristic distribution drift. This indicates evidence of statistical drift, including false positives and false negatives. This indicates evidence of threshold output fluctuation drift. , , These represent the weights corresponding to the three types of drift evidence mentioned above; When the combined drift exceeds the preset threshold, or exceeds the warning threshold multiple times consecutively, the online update in step S4 is triggered.

[0025] Preferably, the update constraint mechanism in step S4 includes at least: Update the step size constraint to limit the magnitude of a single parameter change or threshold change to no more than a preset ratio; Threshold boundary constraints restrict the threshold from falling into the safe upper and lower limit range; Monotonic constraint: The constraint threshold changes monotonically with the wear trend within the same tool life stage. Feed rate constraint: When the feed rate or spindle rate deviates from the preset range, the online update will be paused, delayed, or reset. The rollback constraint uses shadow verification to evaluate the updated false positive and false negative metrics. If the performance degradation exceeds the threshold, the system will roll back to the parameters of the previous version.

[0026] According to another aspect of the present invention, an adaptive threshold construction and online update system for tool condition monitoring is provided, which employs the above-described method and includes: The data acquisition and preprocessing module is used to acquire and preprocess multi-source monitoring signals. A critical cutting segment positioning module is used to locate the critical cutting segment from the multi-source monitoring signals; The feature extraction module is used to extract features from the key cutting segment; The threshold model inference module is used to output an adaptive monitoring threshold and confidence level based on the features. The status discrimination and alarm module is used to output the tool status discrimination result according to the adaptive monitoring threshold; The online update module is configured to generate weakly supervised labels by receiving on-site feedback signals, and to perform incremental updates on the threshold model in the threshold model inference module after the drift triggering conditions are met and the update constraint mechanism is verified.

[0027] Preferably, the online update module employs one of the following models: online sequential extreme learning machine, incremental support vector regression, or incremental gradient boosting, to update the parameters of the threshold model and store the updated parameters in a versioned manner to support rollback.

[0028] In summary, compared with the prior art, the adaptive threshold construction and online update method for tool condition monitoring provided by the present invention has the following advantages: 1. By using two-level segmented key cutting segment positioning, non-steady-state interferences such as start-up, idling, cutting in, and cutting out are effectively reduced, improving feature stability and threshold transferability.

[0029] 2. The threshold model outputs both the threshold and the confidence level. When the operating conditions are uncertain or the sample is sparse, it automatically adopts a more conservative strategy to reduce the risk of false alarms and accidental shutdowns.

[0030] 3. Weakly supervised labels are generated using field feedback signals and reliability weights are introduced to reduce labeling costs and suppress noise feedback, thereby achieving continuous adaptive thresholding.

[0031] 4. Multiple drift evidences are combined to trigger updates, and incremental updates are performed by combining step size, boundary, monotonicity, multiplier and backoff constraints to ensure that the online learning process is controllable, auditable and rollbackable, which is suitable for long-term industrial operation. Attached Figure Description

[0032] Figure 1 This is a schematic diagram of the overall process of the method according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the two-stage segmentation of the key cutting section in an embodiment of the present invention; Figure 3 This is a schematic diagram of the threshold model structure in an embodiment of the present invention; Figure 4 This is a schematic diagram of the online update security logic in an embodiment of the present invention; Figure 5 This is an experimental result diagram of the adaptive threshold construction and online update method for tool status monitoring in this embodiment of the invention. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0034] Please see Figure 1 This embodiment provides an adaptive threshold construction and online update method for tool status monitoring, including the following steps: S1. Offline Modeling: Collect multi-source monitoring signals from the machining process to construct a historical dataset. Preprocess the historical dataset and locate key cutting segments. Extract features from the key cutting segments to form training samples, train a threshold model, and establish an online update strategy. The threshold model is used to output monitoring thresholds and their confidence levels. Step S1, training the threshold model, specifically includes: S11. Constructing the training dataset: Collect key cutting segment feature vectors under multiple batches, multiple tools, and multiple working conditions during the offline phase. Each feature vector is labeled with a corresponding threshold label, which is determined based on the actual wear or breakage state of the tool. The specific dataset construction process is as follows: S111, System Deployment and Data Synchronization Acquisition When deployed in an industrial setting, the system synchronously acquires multi-source monitoring signals and operational condition semantic information through sensors and machine tool interfaces. The multi-source monitoring signals include one or more of current signals, vibration signals, and internal machine tool load signals; the operational condition semantic information includes tool number, operation number, spindle speed, feed rate, feed rate, spindle start / stop, tool change events, and M-code inserted by the NC program. Different source signals are aligned using a unified timestamp and compensated during the preprocessing stage. The system is deployed on an edge computing unit and communicates with the CNC system PLC. S112, Signal Preprocessing and Standardization Preprocessing is performed on the acquired raw signals, including filtering, outlier suppression, windowing buffering, and standardization alignment. Standardization uses mean-variance normalization.

[0035] in, Represents the original sampled signal. This represents the standardized signal. Indicates the discrete sampling point number. This represents the mean of the original sampled signal within the selected time period. This represents the standard deviation of the original sampled signal within the selected time period. This indicates a small positive number that prevents the denominator from being zero; S113, Feature Extraction and Labeling Based on the preprocessed signal, key cutting segments are located using a two-level segmentation method, and feature vectors are extracted. Simultaneously, based on the actual wear or breakage state of the tool, a corresponding threshold label is assigned to each feature vector, thus forming a training dataset.

[0036] S12. Training the regression sub-model: using feature vectors As input, with threshold label As output, the training sub-model learns the mapping relationship between features and thresholds, and the regression sub-model outputs the regression threshold. Specifically, feature vectors This represents the multidimensional feature vector extracted from the key cutting segment, with threshold labels. The target monitoring threshold label corresponding to the feature vector, and the regression threshold. This represents the monitoring threshold predicted by the regression sub-model based on the current features; S13. Training the uncertainty estimation sub-model: using feature vectors As input, output the prediction variance Var, and calculate the confidence level based on the prediction variance. The specific calculation formula is as follows:

[0037] Where, the larger Var is, the more uncertain the model's threshold prediction is under the current operating conditions. The smaller.

[0038] S14. Establish threshold fusion rules: (This involves setting the regression threshold...) Compared with the preset probability threshold By confidence level Weighted fusion is performed to obtain the adaptive monitoring threshold. The specific calculation formula is as follows:

[0039] in, Regarding confidence level The monotonic weight function, This is a conservative threshold obtained based on quantiles or probability inference; when At higher values, the fusion results tend to favor the regression threshold. ,when At lower thresholds, the fusion results tend to favor a conservative threshold. This is to avoid overly aggressive threshold model outputs when data is sparse or operating conditions are abnormal.

[0040] Please see Figure 3 It shows the relationship between feature vector input, regression submodel output threshold, uncertainty submodel output confidence, threshold fusion, and state discrimination and alarm.

[0041] Further information regarding the positioning of critical cutting sections can be found in [link to relevant documentation]. Figure 2 The process includes two levels of segmentation: Level 1 uses CNC system instruction information to determine the start and end boundaries of the monitoring window; Level 2 calculates the short-time energy and its changes within the monitoring window, eliminating unsteady segments including spindle start-up, idling, spindle reversal, and cutting in and out, retaining the stable cutting process as the critical cutting segment; where the sliding window length is W, the sliding step size is P, and the... k The short-time energy of a window is defined as follows: The change in energy is defined as The calculation formulas are as follows:

[0042]

[0043] in, x ′ represents the preprocessed monitoring signal sequence. Indicates the first k Within the first sliding window i The signal amplitude corresponding to each sampling point; k Indicates the sliding window number. i Indicates the sampling point number within the sliding window; Indicates the first k -1 short-term energy of a sliding window, This represents the short-term energy change between adjacent sliding windows; according to Using the sign and amplitude, construct the entry and exit criteria: when Continuous satisfaction And reach the duration When, the corresponding time is determined as the cutting boundary; when Continuous satisfaction And reach the duration When the corresponding moment is determined as the cutting boundary, the critical cutting narrow window is obtained; in, This indicates the threshold for decision-making. This indicates the cut-out threshold. This indicates the number of sliding windows that continuously satisfy the cut-in condition to determine the cut-in boundary. This indicates the number of sliding windows that continuously satisfy the cutting conditions required to determine the cutting boundary; the critical cutting narrow window indicates the stable cutting time interval defined by the entry and exit boundaries.

[0044] Furthermore, the CNC system instruction information includes paired process start trigger instructions and process end trigger instructions to define the monitoring window, and the trigger instructions are implemented by M-code or equivalent semantic code.

[0045] Furthermore, the extracted features include at least one or more combinations of the following: the root mean square (RMS) value, energy entropy, or waveform factor of the current signal; the mean, variance, quantile, or slope of the load signal; and the time-frequency band energy, spectral peaks, spectral kurtosis, or wavelet packet energy of the vibration signal. The time-frequency band energy, spectral peaks, and spectral kurtosis are calculated based on the time-frequency representation of the vibration signal, which is obtained through a short-time Fourier transform. The specific calculation formula is as follows:

[0046] in, Indicates the vibration signal in time t ,frequency f The short-time Fourier transform result at the point, This represents the preprocessed discrete vibration signal. Indicates t The window function centered on tIndicates a time index. n Indicates the discrete sampling point number. j Represents the imaginary unit. f Represents frequency variables. e Represents the natural constant; The formula for calculating the root mean square (RMS) value of a current signal is as follows:

[0047] in, This represents the preprocessed discrete current signal, where n represents the discrete sampling point number and N represents the number of sample points in the critical cutting section. In the preset frequency band set Calculate the energy of the frequency band. The specific calculation formula is as follows:

[0048] in, This indicates the vibrational energy within that frequency band. Represents the result of the short-time Fourier transform; frequency band set It can be preset or automatically selected based on the machine tool structure characteristics, the frequency of the cutting teeth, and the resonance range, and can further construct features such as energy ratio and energy entropy to improve interpretability and robustness.

[0049] In this embodiment, in order to reduce online computing overhead, the system maintains a core stable feature subset (current RMS, load mean, vibration band energy) and an enhanced feature subset (the remaining features).

[0050] S2. Online monitoring: Real-time acquisition of multi-source monitoring signals, adaptive segmentation and positioning of the key cutting segment of the current machining based on CNC system instruction information and signals; extraction of features from the key cutting segment, and input into the threshold model trained in step S1 to obtain the adaptive monitoring threshold and confidence level, and outputting the tool status discrimination result based on the adaptive monitoring threshold; S2 specifically includes the following steps: S21: Real-time data acquisition: Real-time acquisition of multi-source monitoring signals, with the sampling rate consistent with the offline stage, and communication with the CNC system PLC to obtain real-time working condition semantic information; S22: Critical cutting segment positioning: A two-level segmentation method is adopted, and the critical cutting segment of the current machining is adaptively segmented and positioned based on the CNC system instruction information and signals. The CNC system instruction information includes paired process start triggering instructions and process end triggering instructions, which are implemented by M-code. S23: Feature extraction: Within the key cutting segment obtained by positioning, feature vectors are extracted using the same feature extraction method as in step S1; S24: Threshold Model Inference: Input the feature vector f into the threshold model trained in step S1, and output the adaptive monitoring threshold and confidence level; S25: Status Judgment and Alarm: Based on Real-time Feature Values s Adaptive monitoring threshold and confidence level The tool status is determined based on the adaptive monitoring threshold. The determination logic is as follows: If real-time feature value And the continuous duration reaches the wear duration threshold. If so, then a "wear warning" will be output; If real-time feature value or short-term impact characteristics And the duration reaches the damage duration threshold. If the output is "damage alarm", a stop and tool retraction signal will be sent to the CNC system or linkage control module. If confidence level Below the preset reliability lower limit If the feed rate or spindle rate is detected to be unstable, the wear duration threshold will be set. The duration was increased from the first duration to the second duration, and the alarm linkage was downgraded to a warning alert; among these... s Represents real-time feature values. Indicates the adaptive monitoring threshold. This represents the damage alarm multiplier. b Indicates short-term impact characteristics. Indicates the threshold value for the characteristic of damage impact. Indicates the wear duration threshold. Indicates the damage duration threshold. This indicates the lower confidence level.

[0051] S3. Feedback Labeling: Obtain the field feedback signal related to the tool state discrimination result output in step S2, generate a weak supervision label and calculate its reliability; S3 specifically includes the following steps: S31: Obtain on-site feedback signals: The system receives one or more of the following feedback signals: manual confirmation after alarm, tool change record, automatic control shutdown trigger, and tool life database record. In this embodiment, all four feedback signals are received simultaneously. S32: Generate weak supervision labels: The feedback signal is weighted hierarchically to generate weak supervision labels, wherein the weak supervision labels include at least a true alarm label (1), a false alarm label (0), and an indeterminate label (1). In this embodiment, the tool change record has the highest weight, followed by automatic shutdown trigger, and manual confirmation has a lower weight.

[0052] S33: Calculate Reliability: The reliability calculation formula is as follows:

[0053] in, R Indicates the reliability of the feedback signal. M Indicates the number of feedback signal sources. Indicates the first m Preset weights for feedback signals, Indicates the first m The consistency or reliability score between the feedback signal and the current tool state determination result. ε This indicates a small positive number that prevents the denominator from being zero.

[0054] S4. Online Update: Continuously monitor drift triggering conditions; when the drift triggering conditions are met and verified by the preset update constraint mechanism, based on the online update strategy established in step S1, and combined with the weak supervision label generated in step S3 and its reliability, perform online incremental updates on the parameters and threshold output rules of the threshold model, and output the updated monitoring threshold for subsequent online monitoring.

[0055] Please see Figure 4 S4 specifically includes the following sub-steps: S41: Monitoring drift trigger conditions: The drift trigger conditions include at least one or more of the following: feature distribution drift, false alarm and false negative statistical drift, and threshold output fluctuation drift; in this embodiment, three types of drift are monitored simultaneously.

[0056] Among them, the distribution drift evidence The statistical drift evidence can be quantified by the relative deviation between the current window feature mean and the historical baseline feature mean, the relative deviation between the current window feature variance and the historical baseline feature variance, or the statistical distance between the current feature distribution and the historical baseline feature distribution; The threshold fluctuation evidence can be determined by the number of false alarms, the number of missed alarms, or their changes within a unit of time. The variance, range, or changes in adjacent thresholds can be adaptively monitored within multiple consecutive monitoring windows.

[0057] In one implementation, the three types of drift evidence described above can be represented as follows:

[0058]

[0059]

[0060] In the formula, This represents the mean of the key features within the current statistical window. Represents the historical baseline characteristic mean. Indicates the number of false alarms. Indicates the number of missed alarms. This indicates the historical statistical baseline alarm deviation. to Indicates continuity The adaptive monitoring threshold output by the threshold model within each monitoring window, where std(·) represents the standard deviation, mean(·) represents the mean, and ε represents a small positive number to prevent the denominator from being zero.

[0061] After normalizing and weighting the above three types of drift evidence, the joint drift D is obtained by the following formula:

[0062] Where D represents the joint drift amount, This indicates evidence of characteristic distribution drift. This indicates evidence of statistical drift, including false positives and false negatives. This indicates evidence of threshold output fluctuation drift. , , These represent the weights corresponding to the three types of drift evidence mentioned above.

[0063] When the combined drift exceeds the preset threshold, or exceeds the warning threshold multiple times consecutively, the online update in step S4 is triggered.

[0064] S42: Update constraint mechanism verification: After an update is triggered, the following constraint mechanism verification is enforced: Update step size constraint: Limit the magnitude of a single parameter change or threshold change to no more than a preset ratio.

[0065] Threshold boundary constraints: restrict the threshold from falling within the safe upper and lower limit range.

[0066] Monotonicity constraint: Within the same tool life stage, the constraint threshold changes monotonically with the wear trend; within the same tool life, the threshold should remain monotonically unchanged as the tool is used, measured by cumulative cutting time or cutting mileage. If an unreasonable decrease occurs, the threshold should not be updated.

[0067] Feed rate constraint: When the feed rate or spindle rate deviates from the preset range, the online update will be paused, delayed, or reset. Rollback constraint: The false positive and false negative metrics after the update are evaluated using shadow verification. If the performance degradation exceeds the threshold, the system will roll back to the previous version parameters. More preferably, after the update, the false positive rate and false negative rate are evaluated using the most recent historical samples in shadow mode. If the false positive rate or false negative rate is worse than before the update and exceeds the allowable range, the system will roll back to the previous version.

[0068] S43: Perform online incremental updates: After constraint verification, use online sequence extreme learning machine (OS-ELM) to incrementally update the regression sub-model of the threshold model; during the update, combine the weak supervision labels generated in step S3 and their reliability, and samples with higher reliability contribute greater update weights.

[0069] S44: Versioned storage and output: Store the updated parameters in a versioned manner (version number, timestamp, parameter file), supporting manual or automatic rollback; after the update is completed, output the updated monitoring threshold for online monitoring in the subsequent step S2.

[0070] This invention also provides an adaptive threshold construction and online update system for tool status monitoring, comprising: The data acquisition and preprocessing module is used to acquire and preprocess multi-source monitoring signals. A critical cutting segment positioning module is used to locate the critical cutting segment from the multi-source monitoring signals; The feature extraction module is used to extract features from the key cutting segment; The threshold model inference module is used to output an adaptive monitoring threshold and confidence level based on the features. The status discrimination and alarm module is used to output the tool status discrimination result according to the adaptive monitoring threshold; The online update module is configured to generate weakly supervised labels by receiving on-site feedback signals, and to perform incremental updates to the threshold model in the threshold model inference module after the drift trigger condition is met and the update constraint mechanism is verified. In this embodiment, the online update module uses one of the following models: online sequential extreme learning machine, incremental support vector regression, or incremental gradient boosting model, to update the parameters of the threshold model, and stores the updated parameters in a versioned manner to support rollback.

[0071] Please see Figure 5 The figure shows the final experimental results of the adaptive threshold construction and online update method for tool condition monitoring. The figure displays the online monitoring curves, real-time alarm information, feedback annotation management module, abnormal state discrimination curves, and online update information during the machining process. It demonstrates the system's adaptive monitoring capability for tool failure states and the effectiveness of the constrained online update and rollback mechanism. It shows the tool condition monitoring system's ability to monitor and alarm abnormal tool failure states during on-site machining and its adaptability to application scenarios, verifying the system's usability and effectiveness in actual production.

[0072] In summary, this invention can reduce the reliance on human experience for fixed thresholds by locating key cutting segments in a two-level segmented manner and outputting adaptive monitoring thresholds and confidence levels based on a threshold model. By generating reliable weakly supervised labels through field feedback signals and performing online incremental updates after multiple drift evidences are triggered and the update constraint mechanism is verified, this invention can improve the stability, reliability, and rollback of tool status monitoring under complex working conditions and long-term operation.

[0073] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for adaptive threshold construction and online updating of tool condition monitoring, characterized in that: Includes the following steps: S1. Offline modeling: Collect multi-source monitoring signals from the machining process to construct a historical dataset, preprocess the historical dataset and locate key cutting segments; extract features based on the key cutting segments to form training samples, train a threshold model and establish an online update strategy; The threshold model is used to output the monitoring threshold and its confidence level; S2. Online monitoring: Real-time acquisition of multi-source monitoring signals, adaptive segmentation and positioning of the key cutting segment of the current machining based on CNC system instruction information and signals; extraction of features from the key cutting segment, and input into the threshold model trained in step S1 to obtain the adaptive monitoring threshold and confidence level, and outputting the tool status discrimination result based on the adaptive monitoring threshold; S3. Feedback Labeling: Obtain the field feedback signal related to the tool state discrimination result output in step S2, generate a weak supervision label and calculate its reliability; S4. Online Update: Continuously monitor drift trigger conditions; When the drift triggering condition is met and verified by the preset update constraint mechanism, the parameters and threshold output rules of the threshold model are updated online incrementally according to the online update strategy established in step S1, combined with the weak supervision label and its reliability generated in step S3, and the updated monitoring threshold is output for subsequent online monitoring.

2. The adaptive threshold construction and online update method for tool condition monitoring as described in claim 1, characterized in that: Step S1, training the threshold model, specifically includes: S11. Constructing the training dataset: Collect key cutting segment feature vectors under multiple batches, multiple tools, and multiple working conditions during the offline phase. Each feature vector is labeled with a corresponding threshold label, which is determined based on the actual wear or breakage state of the tool. S12. Training the regression sub-model: using feature vectors As input, with threshold label As output, the training sub-model learns the mapping relationship between features and thresholds, and the regression sub-model outputs the regression threshold. ; S13. Training the uncertainty estimation sub-model: using feature vectors As input, output the prediction variance Var, and calculate the confidence level based on the prediction variance. The specific calculation formula is as follows: ； S14. Establish threshold fusion rules: (This involves setting the regression threshold...) Compared with the preset probability threshold By confidence level Weighted fusion is performed to obtain the adaptive monitoring threshold. The specific calculation formula is as follows: in, Regarding confidence level The monotonic weight function, This is a conservative threshold obtained based on quantiles or probability inference.

3. The adaptive threshold construction and online update method for tool condition monitoring as described in claim 2, characterized in that: The key cutting segment positioning includes two levels of segmentation: The first level uses CNC system instruction information to determine the start and end boundaries of the monitoring window; the second level calculates the short-time energy and its changes within the monitoring window, eliminating unsteady segments including spindle start-up, idling, spindle reversal, and cutting in and out, retaining the stable cutting process as the key cutting segment; wherein, the sliding window length is W, the sliding step size is P, and the... k The short-time energy of a window is defined as follows: The change in energy is defined as The calculation formulas are as follows: in, x ′ represents the preprocessed monitoring signal sequence. Indicates the first k Within the first sliding window i The signal amplitude corresponding to each sampling point; k Indicates the sliding window number. i Indicates the sampling point number within the sliding window; Indicates the first The short-term energy of a sliding window This represents the short-term energy change between adjacent sliding windows; according to Using the sign and amplitude, construct the entry and exit criteria: when Continuous satisfaction And reach the duration When, the corresponding time is determined as the cutting boundary; when Continuous satisfaction And reach the duration When the corresponding moment is determined as the cutting boundary, the critical cutting narrow window is obtained; in, This indicates the threshold for decision-making. This indicates the cut-out threshold. This indicates the number of sliding windows that continuously satisfy the cut-in condition to determine the cut-in boundary. This indicates the number of sliding windows that continuously satisfy the cutting conditions required to determine the cutting boundary; the critical cutting narrow window indicates the stable cutting time interval defined by the entry and exit boundaries.

4. The adaptive threshold construction and online update method for tool condition monitoring as described in claim 1, characterized in that: The CNC system instruction information includes paired process start trigger instructions and process end trigger instructions to define the monitoring window, and the trigger instructions are implemented by M-code or equivalent semantic code.

5. The adaptive threshold construction and online update method for tool condition monitoring as described in claim 1, characterized in that: The extracted features include at least one or more of the following combinations: the root mean square (RMS) value, energy entropy, or waveform factor of the current signal; the mean, variance, quantile, or slope of the load signal; and the time-frequency band energy, spectral peaks, spectral kurtosis, or wavelet packet energy of the vibration signal. The time-frequency band energy, spectral peaks, and spectral kurtosis are calculated based on the time-frequency representation of the vibration signal, which is obtained through a short-time Fourier transform. The specific calculation formula is as follows: in, Indicates the vibration signal in time t ,frequency f The short-time Fourier transform result at the point, This represents the preprocessed discrete vibration signal. Indicates t The window function centered on t Indicates a time index. n Indicates the discrete sampling point number. j Represents the imaginary unit. f Represents frequency variables. e Represents the natural constant; The formula for calculating the root mean square (RMS) value of a current signal is as follows: in, This represents the preprocessed discrete current signal, where n represents the discrete sampling point number and N represents the number of sample points in the critical cutting section.

6. The adaptive threshold construction and online update method for tool condition monitoring as described in claim 1, characterized in that: The field feedback signals mentioned in step S3 include at least one or more of the following: manual confirmation after an alarm, tool change records, automatic control shutdown triggers, and tool life database records; the feedback signals are graded and weighted to generate the weak supervision labels, which include at least true alarm labels, false alarm labels, and uncertain labels; the reliability calculation formula is as follows: in, R Indicates the reliability of the feedback signal. M Indicates the number of feedback signal sources. Indicates the first m Preset weights for feedback signals, Indicates the first m The consistency or reliability score between the feedback signal and the current tool state determination result. ε This indicates a small positive number that prevents the denominator from being zero.

7. The adaptive threshold construction and online update method for tool condition monitoring as described in claim 1, characterized in that: The drift triggering conditions in step S4 include at least one or more of the following: feature distribution drift, false positive and false negative statistical drift, and threshold output fluctuation drift; wherein, the feature distribution drift is quantified using distribution drift evidence, the false positive and false negative statistical drift is quantified using statistical drift evidence, and the threshold output fluctuation drift is quantified using threshold fluctuation evidence; the weighted combination of the distribution drift evidence, statistical drift evidence, and threshold fluctuation evidence yields the joint drift amount, and the specific calculation formula is as follows: Where D represents the joint drift amount, This indicates evidence of characteristic distribution drift. This indicates evidence of statistical drift, including false positives and false negatives. This indicates evidence of threshold output fluctuation drift. , , These represent the weights corresponding to the three types of drift evidence mentioned above; When the combined drift exceeds the preset threshold, or exceeds the warning threshold multiple times consecutively, the online update in step S4 is triggered.

8. The adaptive threshold construction and online update method for tool condition monitoring as described in claim 1, characterized in that: The update constraint mechanism described in step S4 includes at least the following: Update the step size constraint to limit the magnitude of a single parameter change or threshold change to no more than a preset ratio; Threshold boundary constraints restrict the threshold from falling into the safe upper and lower limit range; Monotonic constraint: The constraint threshold changes monotonically with the wear trend within the same tool life stage. Feed rate constraint: When the feed rate or spindle rate deviates from the preset range, the online update will be paused, delayed, or reset. The rollback constraint uses shadow verification to evaluate the updated false positive and false negative metrics. If the performance degradation exceeds the threshold, the system will roll back to the parameters of the previous version.

9. An adaptive threshold construction and online update system for tool condition monitoring, which employs the method described in any one of claims 1-8, characterized in that: include: The data acquisition and preprocessing module is used to acquire and preprocess multi-source monitoring signals. A critical cutting segment positioning module is used to locate the critical cutting segment from the multi-source monitoring signals; The feature extraction module is used to extract features from the key cutting segment; The threshold model inference module is used to output an adaptive monitoring threshold and confidence level based on the features. The status discrimination and alarm module is used to output the tool status discrimination result according to the adaptive monitoring threshold; The online update module is configured to generate weakly supervised labels by receiving on-site feedback signals, and to perform incremental updates on the threshold model in the threshold model inference module after the drift triggering conditions are met and the update constraint mechanism is verified.

10. The adaptive threshold construction and online update system for tool condition monitoring as described in claim 9, characterized in that: The online update module uses one of the following models: online sequential extreme learning machine, incremental support vector regression, or incremental gradient boosting, to update the parameters of the threshold model and store the updated parameters in a versioned manner to support rollback.

Images