Tool breakage state detection method and system based on multi-dimensional feature fusion

Abstract

The application discloses a kind of tool broken state detection method and system based on multi-dimension feature fusion, method acquisition cutting process multi-channel acoustic emission signal and obtain tool diameter, after framing by sliding window, the fusion feature set obtained by the splicing of CEEMDAN layered IMF statistical characteristics, dimensionless time domain mutation characteristics containing peak ratio and tool diameter working condition coding feature is constructed, input random forest classification model obtains failure probability;According to tool diameter layered decision: when failure probability and peak ratio reach standard threshold simultaneously, determine suspected failure, tool diameter is less than diameter threshold, and low probability, high ratio bottom criterion is additionally set;When a plurality of sliding windows are determined as suspected failure in succession, output broken tool detection result.The application adapts to the baseline difference of different diameter tools, considers large diameter false alarm suppression and micro-diameter broken tool missed detection prevention, and the detection stability and working condition generalization ability are significantly improved.

Description

Technical Field

[0001] This invention belongs to the field of machining tool inspection technology, specifically relating to a tool breakage state detection method and system based on multi-dimensional feature fusion. Background Technology

[0002] Micro-diameter cutting tools are widely used in high-end manufacturing scenarios such as precision mold making, microstructure machining, and micro-hole machining. They are core tools in precision electronics, aerospace, and medical device manufacturing. Micro-diameter tools are characterized by their small diameter, low rigidity, and poor impact resistance. During cutting, they are highly susceptible to sudden tool breakage due to factors such as fluctuating cutting forces and poor chip removal. If tool breakage occurs and is not detected in time, it can lead to serious production accidents such as batch workpiece scrapping, spindle collision damage, and exceeding machining accuracy tolerances, significantly increasing manufacturing costs. Therefore, reliable detection of tool breakage during machining is crucial for ensuring the quality of micro-machining and equipment safety.

[0003] Existing methods for detecting broken cutting tools are mainly classified into three categories: fixed threshold methods, time-frequency analysis methods, and machine learning recognition methods. Traditional fixed threshold methods rely on manually set amplitude thresholds for judgment. However, the baseline amplitude of the cutting signal varies significantly with operating conditions such as tool diameter and cutting parameters. A single threshold cannot adequately cover tools of different diameters: large-diameter tools have higher cutting baseline amplitudes, easily misjudged as faults, resulting in false alarms; micro-diameter tools have extremely low cutting baseline amplitudes, and the absolute amplitude of the broken tool impact is weak, easily missed by the threshold, leading to missed detections. Single time-frequency analysis methods have limited ability to suppress industrial environmental noise, and the transient impact of broken tools is easily confused with normal cutting fluctuations in time-frequency characterization, making reliable differentiation difficult.

[0004] With the development of machine learning technology, tool condition recognition methods based on statistical features and classifier models have been applied, including schemes that combine time-frequency decomposition techniques such as CEEMDAN with classification models such as random forests. These methods are mostly geared towards recognizing slowly changing conditions such as wear on conventional-sized tools. Their feature systems and decision logic have not yet been designed to address the transient impact characteristics of tool breakage and the weak fault characteristics of micro-diameter tools. On the one hand, conventional statistical features are insufficient to fully characterize the abrupt changes in transient impact relative to steady-state cutting; fault features are easily submerged by the steady-state baseline, resulting in weak resistance to fluctuations in operating conditions. On the other hand, the model output often uses a uniform decision threshold, failing to consider the differences in baseline amplitude and tool breakage signal intensity caused by changes in tool diameter. This results in insufficient generalization ability across diameter conditions, and the problem of missed detection of tool breakage on micro-diameter tools remains prominent.

[0005] In summary, how to adapt the tool breakage detection to the differences in cutting baseline and the characteristics of tool breakage faults of different diameter tools, suppress false alarms in large-diameter conditions while reducing missed detections in micro-diameter conditions, and improve the stability and generalization ability of the detection are technical problems that urgently need to be solved in this field. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to overcome the shortcomings of the existing technology for detecting broken tools, which is difficult to take into account the baseline differences of tools of different diameters, the missed detection of micro-diameter tools and the false alarm of large-diameter tools. Therefore, the present invention provides a tool breakage state detection method and system based on multi-dimensional feature fusion.

[0007] A tool breakage detection method based on multi-dimensional feature fusion includes the following steps: Step S1: Acquire multi-channel acoustic emission signals during the cutting process and obtain the current tool diameter; Step S2: Perform sliding window framing processing on the multi-channel acoustic emission signal to obtain the signal of each channel within each sliding window; Step S3: Construct a fusion feature set for each channel signal within each sliding window, obtained by concatenating CEEMDAN hierarchical IMF statistical features, dimensionless temporal abrupt change features, and tool diameter condition coding features; wherein, the CEEMDAN hierarchical IMF statistical features are the temporal statistical features of the high-frequency effective intrinsic mode components obtained by filtering after CEEMDAN decomposition of each channel signal; the dimensionless temporal abrupt change features include peak value ratio, which is the ratio of the peak value of the signal in the current sliding window to the mean of the steady-state baseline before the fault (i.e., the mean of the steady-state cutting signal within the previous preset time length); the tool diameter condition coding features are generated based on the current tool diameter; Step S4: Input the fused feature set into the pre-trained random forest classification model to obtain the failure probability of the current sliding window; Step S5: When the failure probability is not less than the standard probability threshold and the peak multiplier is not less than the standard multiplier threshold, the current sliding window is determined to be a suspected failure window; when the tool diameter is less than the diameter threshold, when the failure probability is not less than the combined probability threshold and the peak multiplier is not less than the micro-diameter multiplier threshold, the current sliding window is also determined to be a suspected failure window; wherein, the combined probability threshold is lower than the standard probability threshold, and the micro-diameter multiplier threshold is higher than the standard multiplier threshold; Step S6: When the number of sliding windows that are continuously identified as suspected fault windows reaches the continuous hit threshold, output the broken tool fault detection result.

[0008] Furthermore, the high-frequency effective intrinsic mode components are obtained by screening in the following way: the effectiveness of each order of intrinsic mode components obtained by CEEMDAN decomposition is judged by kurtosis index, the noise-dominant components with weak impact information are eliminated, and the intrinsic mode components with the highest kurtosis ranking are selected as the high-frequency effective intrinsic mode components. The time-domain statistical features include standard deviation, peak value, RMS value, absolute mean, kurtosis, peak factor, impulse factor, and margin factor.

[0009] Furthermore, the dimensionless time-domain abrupt change characteristics include peak magnitude, maximum rise slope, pulse half width at half maximum (FWHM), rise time, waveform sharpness, fall slope, and baseline drop before and after the peak. The maximum rising slope is an index obtained by normalizing the maximum amplitude increment per unit time within the rising edge range of the impact pulse to the mean of the pre-fault steady-state baseline. The pulse half-width is an index obtained by normalizing the time interval from when the amplitude of the impact pulse rises to half of the peak value to when it falls back to half of the peak value, using the window duration of the current sliding window. The rise time is an index obtained by normalizing the window duration of the current sliding window after the time it takes for the amplitude of the impact pulse to rise from 10% of the peak value to 90% of the peak value. The waveform sharpness is the ratio of the peak amplitude of the impact to the average amplitude of the signal within the impact pulse window; The fallback slope is an index obtained by normalizing the mean of the pre-fault steady-state baseline by the maximum amplitude attenuation per unit time within the falling edge range of the impact pulse. The baseline drop before and after the peak is the relative difference between the mean steady-state baseline after the end of the impact event and the mean steady-state baseline before the impact event.

[0010] Furthermore, the tool diameter condition coding feature is generated in the following way: the tool is divided into four diameter ranges according to its diameter size, including micro diameter range, small diameter range, medium diameter range, and large diameter range; the diameter range to which the current tool diameter belongs is determined, and the tool diameter condition coding feature corresponding to the diameter range is generated using a one-hot coding method; wherein, the diameter threshold is consistent with the minimum diameter boundary of the large diameter range.

[0011] Furthermore, the random forest classification model is obtained through the following training steps: Construct a training dataset, which includes normal cutting samples and tool breakage failure samples, wherein the normal cutting samples are truncated from the steady-state cutting interval; The fault sample is obtained by taking the time corresponding to the peak value of the power channel signal in the multi-channel acoustic emission signal as the fault center point, adaptively determining the expansion range of the fault window according to the tool diameter, and extracting the sample from the fault center point according to the expansion range. The extended range corresponding to a tool whose diameter is not less than the diameter threshold is greater than the extended range corresponding to a tool whose diameter is less than the diameter threshold.

[0012] Furthermore, the training steps also include: using the recall rate of broken tool failures as the optimization objective, determining the hyperparameters of the random forest classification model through grid search combined with cross-validation; and introducing class weights during the training process to balance the class distribution of broken tool failure samples and normal cutting samples, with the class weight of each class being inversely proportional to the number of samples in that class.

[0013] Furthermore, the training steps also include: after training, obtaining the feature importance weights of each feature output by the random forest classification model; summing the feature importance weights of each feature under the same feature category to obtain the feature importance weight sum of that feature category; and determining the ratio of the feature importance weight sum of that feature category to the total feature importance weight sum of all features as the contribution ratio of that feature category to the tool breakage fault judgment; wherein, the feature category includes the feature categories to which the CEEMDAN hierarchical IMF statistical features, the dimensionless temporal abrupt change features, and the tool diameter working condition coding features belong respectively.

[0014] Furthermore, the method also includes: when the number of sliding windows that are consecutively judged as suspected faulty windows does not reach the consecutive hit threshold, it is judged as random noise interference, and the corresponding judgment result is filtered out.

[0015] Furthermore, it also includes: outputting the tool breakage fault detection result to the CNC system in real time through the IO interface to trigger feed hold or emergency stop control.

[0016] A tool breakage detection system based on multi-dimensional feature fusion includes: The acoustic emission signal acquisition module is used to acquire multi-channel acoustic emission signals during the cutting process and obtain the current tool diameter; The frame processing module is used to perform sliding window frame processing on the multi-channel acoustic emission signal to obtain the signal of each channel in each sliding window. The feature construction module is used to construct a fusion feature set for each channel signal within each sliding window, obtained by concatenating CEEMDAN hierarchical IMF statistical features, dimensionless temporal abrupt change features, and tool diameter condition coding features. The CEEMDAN hierarchical IMF statistical features are the temporal statistical features of the high-frequency effective intrinsic mode components obtained by filtering after CEEMDAN decomposition of each channel signal. The dimensionless temporal abrupt change features include a peak value ratio, which is the ratio of the peak value of the signal within the current sliding window to the mean of the steady-state baseline before the fault. The tool diameter condition coding features are generated based on the current tool diameter. The fault classification module is used to input the fused feature set into a pre-trained random forest classification model to obtain the fault probability of the current sliding window. The hierarchical decision module is used to determine the current sliding window as a suspected fault window when the fault probability is not less than a standard probability threshold and the peak multiplier is not less than a standard multiplier threshold; when the tool diameter is less than a diameter threshold, the current sliding window is also determined as a suspected fault window when the fault probability is not less than a combined probability threshold and the peak multiplier is not less than a micro-diameter multiplier threshold; the combined probability threshold is lower than the standard probability threshold, and the micro-diameter multiplier threshold is higher than the standard multiplier threshold. The steady-state verification module is used to output the tool breakage fault detection result when the number of sliding windows that are continuously judged as suspected fault windows reaches the continuous hit threshold.

[0017] Beneficial Effects: This invention constructs a fusion feature set for multi-channel acoustic emission signals, obtained by concatenating CEEMDAN hierarchical IMF statistical features, dimensionless temporal abrupt change features, and tool diameter condition coding features. The multi-channel signals and multiple features complement each other in characterizing tool breakage faults from multiple dimensions, including time-frequency statistics, transient abrupt changes, and condition coding. Specifically, the temporal statistical features of the high-frequency effective intrinsic mode components obtained through CEEMDAN decomposition and filtering retain detailed information about the transient impact of the tool breakage; the dimensionless temporal abrupt change features, including peak value ratios, characterize the degree of abrupt change in impact relative to steady-state cutting using relative quantities such as the ratio of the signal peak value within the window to the mean of the steady-state baseline before the fault, thus eliminating interference from differences in the cutting baseline amplitude of tools with different diameters; the tool diameter condition coding features characterize the baseline condition attributes corresponding to the tool diameter, enabling the input features of the classification model to carry condition difference information. Compared to a single statistical feature system, the aforementioned fusion feature set takes into account both detailed fault impact characterization and cross-diameter condition adaptation, improving the completeness of tool breakage feature characterization and the model's condition generalization ability.

[0018] At the decision-making level, this invention performs hierarchical decision-making based on the relationship between the tool diameter and the diameter threshold. The standard dual-criteria require both the fault probability and peak value multiplier to simultaneously reach their corresponding thresholds. The probability output of the classification model corroborates the transient change characteristics, avoiding false alarms caused by misjudgments from a single decision. For micro-diameter conditions where the tool diameter is less than the diameter threshold, a fallback criterion is added: a joint probability threshold lower than the standard probability threshold and a micro-diameter multiplier threshold higher than the standard multiplier threshold. This specifically captures the hidden tool breakage conditions of micro-diameter tools, characterized by weak fault probability response and extremely strong relative changes, thus compensating for the deficiency of easily missed detections of weak micro-diameter faults under a unified threshold decision. The decision rules adaptively switch according to the range of the tool diameter, achieving a balance between false alarm suppression in large-diameter conditions and missed detection prevention in micro-diameter conditions, significantly improving the adaptability of multi-specification tool breakage detection.

[0019] Meanwhile, this invention uses a continuous hit threshold to determine the number of sliding windows consecutively identified as suspected fault windows as the condition for outputting the broken tool fault detection result. Single-point window decisions that do not form continuous hits do not trigger output, effectively suppressing single-point false alarms caused by random noise and ensuring the stability of the detection results. Through the synergistic effect of feature fusion, hierarchical decision-making, and continuous window verification, this invention can adapt to the differences in cutting baselines and broken tool fault characteristics of tools with different diameters, simultaneously reducing false alarm rates and missed detection rates, and significantly improving detection stability and multi-condition generalization ability. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a schematic diagram of the overall method steps of the present invention; Figure 2 This is a schematic diagram of the hierarchical differentiated fault decision logic of the present invention; Figure 3 This is a schematic diagram of the system structure of the present invention. Detailed Implementation

[0022] To make the above-mentioned objectives, features, and advantages of this application more apparent and understandable, the specific embodiments of this application are described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of this application. However, this application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of this application. Therefore, this application is not limited to the specific embodiments disclosed below.

[0023] Example 1: This embodiment provides a tool breakage state detection method based on multi-dimensional feature fusion. The overall implementation process is as follows: Figure 1 As shown, the hierarchical differentiated decision logic is as follows: Figure 2 As shown. This method collects multi-channel acoustic emission signals from the cutting process and the current tool diameter, constructs a fused feature set after frame segmentation via a sliding window, outputs the fault probability by a random forest classification model, performs hierarchical decision-making based on the relationship between the tool diameter and a diameter threshold, and finally outputs the tool breakage fault detection result after continuous window verification. The specific implementation steps are as follows.

[0024] Step S1: Collect multi-channel acoustic emission signals during the cutting process and obtain the current tool diameter.

[0025] This embodiment employs a broadband acoustic emission sensor to capture the acoustic emission signals excited during micro-cutting. The sensor is mounted on the machine tool table, allowing the transient elastic wave excited at the moment of tool breakage to be transmitted through the machine tool structure to the sensor's mounting location, ensuring that the acquired signal contains complete information about the tool breakage impact. Acoustic emission signals are simultaneously acquired from three monitoring channels: amplitude (amp), power (power), and effective value (rms). These three channels describe the acoustic emission signal characteristics from different physical quantity dimensions such as amplitude, power, and effective value, forming a multi-channel complementary observation of the transient impact of the tool breakage.

[0026] During signal acquisition, the sampling frequency was set to 1MHz to match the high-frequency dynamic response characteristics of transient impacts in micro-machining; the preamplifier gain was set to 40dB to ensure that the weak acoustic emission signal had sufficient amplitude resolution after amplification. The original signal was first subjected to hardware bandpass filtering to remove power frequency interference and low-frequency chassis vibration noise. Hardware bandpass filtering is a conventional signal conditioning method in acoustic emission detection in this field, and its passband can be set according to the frequency band distribution of the effective components of the acoustic emission signal. This embodiment does not specifically limit this.

[0027] The diameter of the current tool can be obtained from the tool information in the CNC machining program, or it can be input by the operator before machining, or it can be obtained by measuring with a tool setter. As long as the diameter information of the tool used for machining can be obtained synchronously during the detection process, it is acceptable.

[0028] Step S2: Perform sliding window framing processing on the multi-channel acoustic emission signal to obtain the signal of each channel within each sliding window.

[0029] A fixed-step sliding window with a window length of 1024 and a sliding step size of 512 is used to process the multi-channel acoustic emission signals frame by frame. This divides the continuous time-domain signal into analysis units based on the sliding window, providing a unified data foundation for subsequent feature construction. The window length and sliding step size can be adjusted according to the characteristics of the acquired signal; this embodiment does not impose any restrictions. Because the sliding step size is smaller than the window length, there is partial overlap between adjacent sliding windows, which avoids missed detections caused by transient impact events falling precisely at the window boundaries.

[0030] Before real-time detection, a random forest classification model needs to be trained offline. The training samples are constructed as follows: normal cutting samples are randomly selected from the steady-state cutting interval at equal intervals; for broken tool failure samples, the fault center point is the time corresponding to the peak value of the power channel signal. The expansion range of the fault window is adaptively determined based on the tool diameter, and the samples are extracted from the fault center point within the expansion range. Specifically, the expansion range corresponds to tools with a diameter not less than the diameter threshold, and to tools with a diameter greater than or less than the diameter threshold. In this embodiment, the fault window expansion is 3 unit windows for tools with a diameter ≥ 2.0 mm, and 1 unit window for micro-diameter tools with a diameter < 2.0 mm. The larger the tool diameter, the longer the impact process triggered by the broken tool. Adaptively setting the expansion range according to the diameter can fully cover the impact interval while avoiding excessive dilution of the transient impact features of the broken tool by too much steady-state cutting baseline, thus ensuring the feature purity of the fault samples.

[0031] The cutting tools are divided into four working conditions according to the processing specifications: micro diameter (diameter < 0.8 mm), small diameter (diameter 0.8 mm ~ 1.5 mm), medium diameter (diameter 1.5 mm ~ 2.0 mm), and large diameter (diameter ≥ 2.0 mm). These four working conditions correspond one-to-one with the working condition coding features constructed in subsequent steps, and also provide a basis for working condition classification for subsequent differentiated judgment.

[0032] Step S3: Construct a fusion feature set for each channel signal within each sliding window, which is obtained by splicing together CEEMDAN hierarchical IMF statistical features, dimensionless time-domain abrupt change features, and tool diameter working condition coding features.

[0033] The CEEMDAN hierarchical IMF statistical features are time-domain statistical features of the high-frequency effective intrinsic mode components obtained by screening after CEEMDAN decomposition of each channel signal; the dimensionless time-domain abrupt change features include peak factor, which is the ratio of the peak value of the signal in the current sliding window to the mean of the steady-state baseline before the fault; the tool diameter condition coding features are generated based on the current tool diameter.

[0034] First, we construct the CEEMDAN hierarchical IMF statistical features. CEEMDAN decomposition overcomes the mode aliasing problem in EMD decomposition by adding adaptive white noise to each order of mode decomposition and taking the ensemble average of multiple decomposition results. Its core decomposition formula is: ; In the formula, This is the original cutting acoustic emission signal. For the added white noise sequence, This is the noise amplitude coefficient. The set mean of the nth decomposition For the i-th eigenmode component, The remaining residual components.

[0035] In this embodiment, CEEMDAN adaptive decomposition is performed on the three channel signals respectively, with the noise amplitude coefficient set to 0.2 and the ensemble averaging set to 100 times. This decomposition yields multiple intrinsic mode components (IMCs). Considering that the broken blade fault manifests as a transient impact component concentrated in the high-frequency band, while some low-order IMCs are noise-dominated components with weak impact information, a kurtosis index is used to determine the validity of each IMC obtained from the CEEMDAN decomposition after decomposition. Noise-dominated components with weak impact information are removed, and IMCs with higher kurtosis ranking are selected as the high-frequency effective IMCs. The kurtosis index is sensitive to the impact components in the signal; components with obvious impact characteristics have larger kurtosis values. Filtering by kurtosis ranking can accurately retain high-frequency IMCs containing broken blade impact information. In this embodiment, the preset number of layers is 4, meaning that a maximum of the top 4 layers of IMCs with higher kurtosis ranking are selected as high-frequency effective IMCs.

[0036] For each layer of high-frequency effective intrinsic mode components (IMMs), eight time-domain statistical features were extracted, including standard deviation, peak value, RMS value, absolute mean, kurtosis, peak factor, impulse factor, and margin factor. These statistical features are commonly used indicators in the time-domain statistical analysis of signals in this field: standard deviation, peak value, RMS value, and absolute mean reflect the dispersion and amplitude level of the signal; kurtosis, peak factor, impulse factor, and margin factor are dimensionless indicators and have good characterization ability for impulse patterns in signals. Thus, each channel yielded 4 layers × 8 dimensions, totaling 32 dimensions of statistical features, and the three signals combined yielded 96 dimensions of CEEMDAN layered IMF statistical features.

[0037] Then, dimensionless time-domain abrupt change characteristics are constructed. For each channel signal, 7-dimensional dimensionless time-domain abrupt change characteristics are calculated, covering the full cycle of the impulse pulse rising edge, peak value, falling edge, and baseline drift, which are peak value multiplier, maximum rise slope, pulse half width at half maximum, rise time, waveform sharpness, fall slope, and baseline drop before and after the peak.

[0038] The peak value multiplier is the ratio of the peak signal value within the current sliding window to the mean of the steady-state baseline before the fault. It represents the degree of abrupt change in the impact amplitude relative to the normal cutting state and is a core fundamental indicator for measuring impact intensity. The calculation formula is as follows: ; In the formula, The maximum amplitude, or peak value, of the signal within a single impact analysis window. The baseline is the average value of the signal amplitude within a steady-state cutting range before the impact event occurs.

[0039] The maximum rising slope is an index obtained by normalizing the maximum amplitude increment per unit time within the rising edge range of the impact pulse to the mean of the pre-fault steady-state baseline, representing the steepness and suddenness of the impact outbreak.

[0040] Let the discrete sampling sequence within the rising edge interval be... The sampling time interval is The calculation formula is: ; In the formula, The amplitude at the i-th sampling point on the rising edge. The signal sampling time interval, This is the mean of the steady-state cutting baseline used for amplitude normalization.

[0041] The pulse half-width at half-maximum (WHM) is an index obtained by normalizing the time interval from when the amplitude of the impact pulse rises to half of its peak value to when it falls back to half of its peak value, using the window duration of the current sliding window. It represents the degree of concentration of impact energy in the time dimension.

[0042] Let the peak amplitude be The half-height times of the rising edge and the falling edge are obtained by linear interpolation. The calculation formula is as follows: ; In the formula, For the rising edge amplitude to reach 0.5 At that moment, The amplitude drops back to 0.5 at the falling edge. At that moment, The total duration of a single impact analysis window, used for time dimension normalization.

[0043] Rise time is an index obtained by normalizing the time taken for the impulse pulse amplitude to rise from 10% of the peak value to 90% of the peak value by the window duration of the current sliding window. It represents the duration of the rapid rise of the impulse amplitude and is a key indicator for measuring the suddenness of the fault.

[0044] The amplitude was obtained by linear interpolation to reach 0.1. The moment and reaching 0.9 The formula for calculating the time is: ; In the formula, For the rising edge amplitude to reach 0.1 At that moment, For the rising edge amplitude to reach 0.9 At that moment, This represents the total duration of the impact analysis window.

[0045] Waveform sharpness is the ratio of the peak amplitude of the impact to the average amplitude of the signal within the impact pulse window. Also known as the pulse shape factor, it is a purely dimensionless index that describes the overall sharpness of the impact waveform. The calculation formula is: ; In the formula, This represents the peak amplitude of the impact pulse. It represents the average signal amplitude within the complete window of a single impulse pulse.

[0046] The fallback slope is an index obtained by normalizing the maximum amplitude attenuation per unit time within the falling edge range of the impact pulse to the mean of the pre-fault steady-state baseline, indicating the speed of impact dissipation.

[0047] Let the discrete sampling sequence within the falling edge interval be... The calculation formula is: ; In the formula, The magnitude of the i-th sampling point along the falling edge. The signal sampling time interval, This represents the mean value of the steady-state cutting baseline.

[0048] The baseline drop before and after the peak is the relative difference between the mean steady-state baseline after the impact event and the mean steady-state baseline before the impact event. It represents the impact of a single impact on the shift of the signal's steady-state baseline and is used to identify irreversible tool degradation. The calculation formula is as follows: ; In the formula, This represents the baseline mean of the steady-state interval preceding the impact event. This represents the baseline mean over a steady-state period following the end of the shock event.

[0049] The aforementioned 7-dimensional abrupt change features were all normalized to become dimensionless quantities, thus eliminating the interference from the differences in cutting baseline amplitudes of tools with different diameters and achieving decoupled characterization of the transient impact characteristics and steady-state cutting characteristics of tool breakage. Each channel yielded 7-dimensional abrupt change features, and the three signals combined yielded 21-dimensional dimensionless time-domain abrupt change features.

[0050] Next, the tool diameter condition coding feature is constructed. Tools are divided into four diameter ranges based on their diameter size: micro-diameter, small-diameter, medium-diameter, and large-diameter. The diameter range to which the current tool diameter belongs is determined, and a one-hot encoding method is used to generate the tool diameter condition coding feature corresponding to that diameter range. In this embodiment, diameters <0.8mm correspond to the micro-diameter range, 0.8mm~1.5mm to the small-diameter range, 1.5mm~2.0mm to the medium-diameter range, and ≥2.0mm to the large-diameter range. One-hot encoding generates a 4-dimensional condition coding feature, representing the baseline condition attributes corresponding to the tool diameter, allowing the model input features to carry the tool's own condition difference information. The diameter threshold is consistent with the minimum diameter boundary of the large-diameter range, i.e., the diameter threshold is 2.0mm.

[0051] The three types of features mentioned above are concatenated according to their dimensions to obtain a 121-dimensional high-dimensional fault feature set, which is then input into the subsequent classification model. This fused feature set is complementary in three dimensions: time-frequency statistics, transient changes, and operating condition coding. It not only preserves the impact details of the broken tool fault but also carries information on diameter operating condition differences, providing a complete input basis for classification decisions under multiple operating conditions.

[0052] Step S4: Input the fused feature set into the pre-trained random forest classification model to obtain the failure probability of the current sliding window.

[0053] The random forest classification model is an ensemble classification model composed of multiple decision trees. For the fused feature set x of the current sliding window, each decision tree independently outputs a broken blade fault classification result. The average of all fault classification outputs from the decision trees is taken as the fault probability, calculated using the following formula: ; In the formula, N is the number of decision trees. Output the fault classification for the i-th decision tree.

[0054] The random forest classification model is trained through the following process: A training dataset is constructed, containing 1173 samples covering seven diameter specifications of cutting tools from 0.6mm to 8.0mm. Among these, 62 samples are positive (broken tool) samples, accounting for 5.29%, exhibiting significant class imbalance. The dataset is divided into a training subset and a validation subset in a 7:3 ratio. Positive samples are broken tool failure samples selected using the diameter adaptive method described in step S2, while negative samples are normal cutting samples selected from the steady-state cutting range.

[0055] During model training, the recall rate of the broken knife fault is optimized as the target. The hyperparameters of the random forest classification model are determined through grid search combined with cross-validation. This embodiment uses 3-fold cross-validation for grid search hyperparameter optimization, with the search range set as follows: number of decision trees n_estimators∈{50,100,200}, maximum decision tree depth max_depth∈{None,8,15}. The optimal hyperparameter combination is ultimately n_estimators=50, and max_depth is unlimited. Recall rate is optimized to ensure the model's sensitivity in detecting positive broken knife samples.

[0056] To address the issue of a low proportion of positive broken tool samples, class weights are introduced during training to balance the class distribution of broken tool failure samples and normal cutting samples. The class weight of each class is inversely proportional to the number of samples in that class, and the calculation formula is as follows: ; In the formula, The total number of samples, This represents the number of samples for the corresponding category. By balancing the category weights, the model bias caused by the low proportion of positive samples of broken blades is mitigated, thus improving the sensitivity to minor faults.

[0057] After training, the feature importance weights of each feature output by the random forest classification model are obtained. The feature importance weights of each feature within the same feature category are summed to obtain the sum of feature importance weights for that feature category. The ratio of this sum of feature importance weights to the total sum of feature importance weights of all features is determined as the contribution percentage of that feature category to the determination of broken blade faults. The calculation formula is as follows: ; ; In the formula, Assign importance weights to individual features. The total weight of features of the same type. The contribution percentage of each feature category is represented. These feature categories include the respective feature categories of CEEMDAN hierarchical IMF statistical features, dimensionless time-domain abrupt change features, and tool diameter condition coding features. Statistical results show that among the various features contributing to fault identification, the statistical features of high-frequency effective intrinsic mode components have the highest contribution percentage and are considered the core fault features. Abrupt change features are auxiliary identification features, and diameter coding features are condition-adaptive features. This quantitative result clarifies the core mechanism of micro-diameter tool breakage identification and provides a quantitative basis for tool fault feature optimization and lightweight model design.

[0058] Step S5: Perform a layered decision on the current sliding window based on the relationship between the current tool diameter and the diameter threshold.

[0059] When the failure probability is not less than a standard probability threshold and the peak multiplier is not less than a standard multiplier threshold, the current sliding window is determined to be a suspected failure window. Similarly, when the tool diameter is less than a diameter threshold, and the failure probability is not less than a combined probability threshold and the peak multiplier is not less than a micro-diameter multiplier threshold, the current sliding window is also determined to be a suspected failure window. The combined probability threshold is lower than the standard probability threshold, and the micro-diameter multiplier threshold is higher than the standard multiplier threshold.

[0060] Specifically, for tools with a diameter ≥ 2.0 mm, i.e., when the tool diameter is not less than the diameter threshold, the standard dual-criteria decision is adopted, and the decision formula is as follows: ; In the formula, The standard probability threshold, K represents the standard magnification threshold, and K represents the peak magnification. In this embodiment... A typical value is 0.40. A typical value is 3.0. When both the fault probability and peak value within the window meet the threshold conditions, the window is determined to be a suspected fault window. This triggering method is defined as standard criterion triggering. The probability output of the classification model and the transient change characteristics corroborate each other, avoiding false alarms caused by misjudgment from a single decision.

[0061] For micro-diameter cutting tools with a diameter < 2.0 mm, i.e., the tool diameter is less than the diameter threshold, a low-probability, high-magnification joint fallback decision is added in addition to the standard dual criteria. The decision formula is as follows: ; In the formula, For the joint probability threshold, This is a micro-diameter magnification threshold specific to the abrupt change characteristics of micro-diameter cutting tools. In this embodiment... A typical value is 0.22. A typical value is 1500.0. Meeting the standard dual criteria, or meeting the high-probability abrupt change condition under low probability conditions, is considered a suspected fault window. The latter is triggered by a combined fallback trigger. Micro-diameter tools have extremely low cutting baseline amplitudes, and tool breakage faults do not exhibit significant probabilistic characteristic responses, but they can produce extremely large relative peak abrupt changes. A dedicated threshold can accurately capture the unique weak faults and strong abrupt tool breakage conditions of micro-diameter tools, compensating for the omissions of conventional criteria.

[0062] It should be noted that the above threshold values ​​are the optimal calibration parameters under the corresponding working conditions, and are only exemplary values, which do not constitute a limitation on the scope of protection of this invention. Those skilled in the art can recalibrate according to the actual processing conditions.

[0063] Step S6: When the number of sliding windows that are continuously identified as suspected fault windows reaches the continuous hit threshold, output the broken tool fault detection result.

[0064] Steady-state verification is performed on the judgment results of consecutive sliding windows. In this embodiment, the consecutive hit threshold is set to 2. When two consecutive sliding windows are both judged as suspected fault windows, the final tool breakage fault judgment result is output. When the number of sliding windows judged as suspected fault windows does not reach the consecutive hit threshold, that is, only a single sliding window is judged as a suspected fault window, it is judged as random noise interference, and the judgment result is filtered out. Steady-state verification of consecutive windows suppresses false alarms caused by single-point spike noise and improves the stability of the detection results.

[0065] The results of the broken tool fault detection can be output to the CNC system in real time through the IO interface, triggering feed hold or emergency stop control, stopping the machining process immediately when the broken tool occurs, and avoiding production accidents such as batch workpiece scrapping and spindle collision damage.

[0066] To verify the detection performance of the method in this embodiment, two independent test sets were constructed for testing. The meanings of each statistic in the evaluation of the test results are shown in Table 1 below: Table 1: Meaning of Statistical Measures for Evaluation of Test Results abbreviation Full English name Chinese name definition The actual meaning of corresponding broken blade detection TP True Positive Real example The actual fault was also predicted by the model. Correctly detected broken blades: Broken blade samples were successfully identified by the model. FN False Negative False negatives The actual condition is a fault, while the model predicts it to be normal. Missed broken tool: The tool actually broke, but the model did not detect it. FP False Positive False positives The actual result is normal, while the model predicts a fault. False alarm / false alarm: The actual cutting is normal, but the model misjudges it as a tool breakage. TN True Negative True negative example The actual result is normal, and the model also predicts it to be normal. Correctly identified normal samples: Normal cutting was correctly identified as normal by the model. The first test set contains 24 samples, covering the full diameter specifications from 0.6mm to 8.0mm, including 6 positive samples of broken tools and 18 samples of normal cutting. The test results show that the true positives TP=6, the false negatives FN=0, the false positives FP=0, and the true negatives TN=18, with both the fault recall rate and precision rate reaching 100%.

[0067] The second set is an expanded test set containing 38 samples, which further enriches the cutting parameters and tool conditions. Among them, there are 12 positive samples of tool breakage and 26 normal cutting samples. The test results show that the true negative cases TP=12, the false negative cases FN=0, the false positive cases FP=0, and the true negative cases TN=26. The fault recall rate and precision rate both reach 100%.

[0068] Among them, the micro-diameter broken tool sample with a diameter of 0.8 mm was successfully identified through the joint fallback criterion, verifying the effectiveness of the micro-diameter fallback mechanism; all normal samples under the large-diameter working condition showed no false alarms or misjudgments. The two sets of independent test results show that the method of this embodiment achieves zero missed detections and zero false alarms for both large-diameter and micro-diameter tool machining conditions, can adapt to the differences in cutting baseline and broken tool fault characteristics of tools with different diameters, and has good multi-condition generalization ability and detection stability.

[0069] Example 2: Reference Figure 3As shown in the figure, this embodiment provides a tool breakage state detection system based on multi-dimensional feature fusion. The system includes an acoustic emission signal acquisition module, a frame processing module, a feature construction module, a fault classification module, a hierarchical decision module, and a steady-state verification module.

[0070] The acoustic emission signal acquisition module is used to acquire multi-channel acoustic emission signals during the cutting process and obtain the current tool diameter; The frame processing module is used to perform sliding window frame processing on the multi-channel acoustic emission signal to obtain the signal of each channel in each sliding window. The feature construction module is used to construct a fusion feature set for each channel signal within each sliding window, which is obtained by splicing CEEMDAN hierarchical IMF statistical features, dimensionless time-domain abrupt change features, and tool diameter condition coding features. The CEEMDAN hierarchical IMF statistical features are the time-domain statistical features of the high-frequency effective intrinsic mode components obtained by screening after CEEMDAN decomposition of each channel signal. The dimensionless time-domain abrupt change features include peak value ratio, which is the ratio of the peak value of the signal in the current sliding window to the mean of the steady-state baseline before the fault. The tool diameter condition coding features are generated based on the current tool diameter. The fault classification module is used to input the fused feature set into a pre-trained random forest classification model to obtain the fault probability of the current sliding window. The hierarchical decision module is used to determine the current sliding window as a suspected fault window when the fault probability is not less than the standard probability threshold and the peak multiplier is not less than the standard multiplier threshold. When the tool diameter is less than the diameter threshold, the current sliding window is also determined as a suspected fault window when the fault probability is not less than the combined probability threshold and the peak multiplier is not less than the micro-diameter multiplier threshold. The combined probability threshold is lower than the standard probability threshold and the micro-diameter multiplier threshold is higher than the standard multiplier threshold. The steady-state verification module is used to output the tool breakage fault detection result when the number of sliding windows that are continuously judged as suspected fault windows reaches the continuous hit threshold.

[0071] The specific implementation of each module in this embodiment can be referred to the description of the corresponding steps in Embodiment 1, and will not be repeated here. During system operation, the acoustic emission signal acquisition module continuously acquires multi-channel acoustic emission signals and tool diameter information; the frame processing module performs sliding window frame segmentation on the signal; the feature construction module constructs a fused feature set window by window; the fault classification module outputs the fault probability; the hierarchical decision module selects a decision rule based on the relationship between the tool diameter and the diameter threshold; and the steady-state verification module outputs the tool breakage fault detection result after continuous window verification. All modules of the system work together to complete the entire process of the method described in Embodiment 1, and correspondingly achieve the technical effects of the method described in Embodiment 1.

[0072] Example 3: This embodiment provides a tool breakage state detection device based on multi-dimensional feature fusion. The device includes an acoustic emission sensor, a preamplifier, a signal conditioning circuit, and an analog-to-digital conversion module connected in sequence, and also includes a processor, a memory, and an I / O interface module.

[0073] An acoustic emission sensor is mounted on the machine tool worktable to pick up acoustic emission signals excited during the cutting process. A preamplifier is connected to the acoustic emission sensor to amplify the weak signal output by the sensor, with its gain set to 40dB. A signal conditioning circuit is connected to the preamplifier to perform hardware bandpass filtering on the amplified signal to remove power frequency interference and low-frequency machine body vibration noise. This can be implemented using a conventional active filter circuit in the field. An analog-to-digital conversion module is connected to the signal conditioning circuit to convert the filtered analog signal into a digital signal at a sampling frequency of 1MHz, resulting in a multi-channel acoustic emission signal including an amp amplitude channel, a power channel, and an RMS effective value channel.

[0074] The processor is connected to both the analog-to-digital conversion module and the I / O interface module. The memory stores a computer program, which, when executed, implements the steps of the tool breakage detection method described in Embodiment 1. The processor can be an existing computing device with data processing capabilities, such as an industrial control computer or an embedded processor. The memory can be an existing storage medium, such as volatile memory or non-volatile memory. In addition to the computer program, the memory also stores a pre-trained random forest classification model, as well as parameters used during the detection process, such as standard probability thresholds, standard magnification thresholds, joint probability thresholds, micro-diameter magnification thresholds, diameter thresholds, and consecutive hit thresholds.

[0075] When the device is working, the acoustic emission signal is sequentially amplified, filtered by hardware bandpass, and converted from analog to digital before being sent to the processor as a multi-channel digital signal. The processor sequentially completes sliding window framing, feature set construction, fault probability calculation, hierarchical decision-making, and steady-state verification according to the method described in Example 1. When the number of sliding windows continuously judged as suspected fault windows reaches the continuous hit threshold, the broken tool fault detection result is output to the CNC system via the IO interface module, triggering feed hold or emergency stop control. This device integrates signal acquisition, signal conditioning, feature calculation, and decision control functions, and can be deployed on existing processing equipment as an independent monitoring instrument to achieve online real-time monitoring of the broken tool status during the processing.

[0076] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0077] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for detecting tool breakage state based on multi-dimensional feature fusion, characterized in that, Includes the following steps: Acquire multi-channel acoustic emission signals during the cutting process and obtain the current tool diameter; The multi-channel acoustic emission signal is subjected to sliding window framing processing to obtain the signal of each channel within each sliding window; A fusion feature set is constructed for each channel signal within each sliding window, consisting of CEEMDAN hierarchical IMF statistical features, dimensionless temporal abrupt change features, and tool diameter condition coding features. The CEEMDAN hierarchical IMF statistical features are the temporal statistical features of the high-frequency effective intrinsic mode components obtained by filtering after CEEMDAN decomposition of each channel signal. The dimensionless temporal abrupt change features include peak value multipliers, which are the ratio of the peak value of the signal within the current sliding window to the mean of the steady-state baseline before the fault. The tool diameter condition coding features are generated based on the current tool diameter. The fused feature set is input into a pre-trained random forest classification model to obtain the failure probability of the current sliding window; When the failure probability is not less than the standard probability threshold and the peak multiplier is not less than the standard multiplier threshold, the current sliding window is determined to be a suspected failure window; when the tool diameter is less than the diameter threshold, when the failure probability is not less than the combined probability threshold and the peak multiplier is not less than the micro-diameter multiplier threshold, the current sliding window is also determined to be a suspected failure window; wherein, the combined probability threshold is lower than the standard probability threshold, and the micro-diameter multiplier threshold is higher than the standard multiplier threshold; When the number of sliding windows that are consecutively identified as suspected fault windows reaches the consecutive hit threshold, the broken tool fault detection result is output.

2. The method according to claim 1, characterized in that, The high-frequency effective intrinsic mode components are obtained by screening in the following way: the effectiveness of each order of intrinsic mode components obtained by CEEMDAN decomposition is judged by kurtosis index, the noise-dominant components with weak impact information are eliminated, and the intrinsic mode components with the highest kurtosis ranking are selected as the high-frequency effective intrinsic mode components. The time-domain statistical features include standard deviation, peak value, RMS value, absolute mean, kurtosis, peak factor, impulse factor, and margin factor.

3. The method according to claim 1, characterized in that, The dimensionless time-domain abrupt change characteristics include peak scalar, maximum rise slope, pulse half-width, rise time, waveform sharpness, fall slope, and baseline drop before and after the peak. The maximum rising slope is an index obtained by normalizing the maximum amplitude increment per unit time within the rising edge range of the impact pulse to the mean of the pre-fault steady-state baseline. The pulse half-width is an index obtained by normalizing the time interval from when the amplitude of the impact pulse rises to half of the peak value to when it falls back to half of the peak value, using the window duration of the current sliding window. The rise time is an index obtained by normalizing the window duration of the current sliding window after the time it takes for the amplitude of the impact pulse to rise from 10% of the peak value to 90% of the peak value. The waveform sharpness is the ratio of the peak amplitude of the impact to the average amplitude of the signal within the impact pulse window; The fallback slope is an index obtained by normalizing the mean of the pre-fault steady-state baseline by the maximum amplitude attenuation per unit time within the falling edge range of the impact pulse. The baseline drop before and after the peak is the relative difference between the mean steady-state baseline after the end of the impact event and the mean steady-state baseline before the impact event.

4. The method according to claim 1, characterized in that, The tool diameter condition coding feature is generated in the following way: the tool is divided into four diameter ranges according to the tool diameter size, the four diameter ranges include micro diameter range, small diameter range, medium diameter range and large diameter range; the diameter range to which the current tool diameter belongs is determined, and the tool diameter condition coding feature corresponding to the diameter range is generated by one-hot coding method; wherein, the diameter threshold is consistent with the minimum diameter boundary of the large diameter range.

5. The method according to claim 1, characterized in that, The random forest classification model is obtained through the following training steps: Construct a training dataset, which includes normal cutting samples and tool breakage failure samples, wherein the normal cutting samples are truncated from the steady-state cutting interval; The fault sample is obtained by taking the time corresponding to the peak value of the power channel signal in the multi-channel acoustic emission signal as the fault center point, adaptively determining the expansion range of the fault window according to the tool diameter, and extracting the sample from the fault center point according to the expansion range. The extended range corresponding to a tool whose diameter is not less than the diameter threshold is greater than the extended range corresponding to a tool whose diameter is less than the diameter threshold.

6. The method according to claim 5, characterized in that, The training steps also include: using the recall rate of broken tool failures as the optimization objective, determining the hyperparameters of the random forest classification model through grid search combined with cross-validation; and introducing class weights during training to balance the class distribution of broken tool failure samples and normal cutting samples, with the class weight of each class being inversely proportional to the number of samples in that class.

7. The method according to claim 6, characterized in that, The training steps further include: after training, obtaining the feature importance weights of each feature output by the random forest classification model; summing the feature importance weights of each feature under the same feature category to obtain the feature importance weight sum of that feature category; and determining the ratio of the feature importance weight sum of that feature category to the total feature importance weight sum of all features as the contribution ratio of that feature category to the tool breakage fault judgment; wherein, the feature category includes the feature categories to which the CEEMDAN hierarchical IMF statistical features, the dimensionless temporal abrupt change features, and the tool diameter working condition coding features belong respectively.

8. The method according to claim 1, characterized in that, The method also includes: when the number of sliding windows that are consecutively identified as suspected faulty windows does not reach the consecutive hit threshold, it is determined to be random noise interference, and the corresponding judgment result is filtered out.

9. The method according to any one of claims 1 to 8, characterized in that, Also includes: The broken tool fault detection result is output to the CNC system in real time through the IO interface to trigger feed hold or emergency stop control.

10. A tool breakage state detection system based on multi-dimensional feature fusion, characterized in that, include: The acoustic emission signal acquisition module is used to acquire multi-channel acoustic emission signals during the cutting process and obtain the current tool diameter; The frame processing module is used to perform sliding window frame processing on the multi-channel acoustic emission signal to obtain the signal of each channel in each sliding window. The feature construction module is used to construct a fusion feature set for each channel signal within each sliding window, obtained by concatenating CEEMDAN hierarchical IMF statistical features, dimensionless temporal abrupt change features, and tool diameter condition coding features. The CEEMDAN hierarchical IMF statistical features are the temporal statistical features of the high-frequency effective intrinsic mode components obtained by filtering after CEEMDAN decomposition of each channel signal. The dimensionless temporal abrupt change features include a peak value ratio, which is the ratio of the peak value of the signal within the current sliding window to the mean of the steady-state baseline before the fault. The tool diameter condition coding features are generated based on the current tool diameter. The fault classification module is used to input the fused feature set into a pre-trained random forest classification model to obtain the fault probability of the current sliding window. The hierarchical decision module is used to determine the current sliding window as a suspected fault window when the fault probability is not less than a standard probability threshold and the peak multiplier is not less than a standard multiplier threshold; when the tool diameter is less than a diameter threshold, the current sliding window is also determined as a suspected fault window when the fault probability is not less than a combined probability threshold and the peak multiplier is not less than a micro-diameter multiplier threshold; the combined probability threshold is lower than the standard probability threshold, and the micro-diameter multiplier threshold is higher than the standard multiplier threshold. The steady-state verification module is used to output the tool breakage fault detection result when the number of sliding windows that are continuously judged as suspected fault windows reaches the continuous hit threshold.

Images