A numerical control machine tool machining process abnormal state recognition method and system

Abstract

This invention discloses a method and system for identifying abnormal states during CNC machine tool machining, relating to the field of machining anomaly identification technology. The method includes the following steps: acquiring machining signal data from the CNC machine tool and extracting the force establishment feature set of the machining path segment; performing feature matching on the force establishment feature set to obtain the force establishment offset, and constructing several machining state switching thresholds; identifying the state switching points of the machining path segment based on the several machining state switching thresholds and aligning the path segment boundaries to obtain a path alignment set; performing a first abnormal state identification on the path alignment set and determining the optimal machining state switching threshold based on the identification results; defining the optimal machining interval within the machining path segment based on the optimal machining state switching threshold, and performing a second abnormal state identification. This solves the problem that abnormal states are difficult to accurately identify during the transition phase of CNC machine tool machining conditions due to rapid fluctuations in load, vibration, or force distribution.

Description

Technical Field

[0001] This invention relates to the field of machining anomaly identification technology, and more specifically, to a method and system for identifying abnormal states during CNC machine tool machining. Background Technology

[0002] When CNC machine tools perform part machining, they typically control the cutting tool along a predetermined machining path according to a preset program to complete the cutting operation. During the machining process, process parameters such as spindle speed, feed rate, and depth of cut work together to generate continuously changing cutting forces, vibrations, and acoustic emission signals between the tool and the workpiece material. These physical quantities fluctuate dynamically with the machining stage, the degree of tool wear, and changes in the local properties of the material. When the machining state is stable, various monitoring signals usually show relatively stable or periodic variation characteristics. However, when there are situations such as accelerated tool wear, tool chipping, sudden changes in workpiece material hardness, loose clamping, or abnormal cutting parameters, it often causes phenomena such as a sudden increase in cutting force, changes in the vibration spectrum structure, or abnormal energy concentration, thereby affecting machining accuracy and surface quality.

[0003] However, during CNC machine tool machining, the stress state, vibration characteristics, and dynamic response of the machine tool often fluctuate rapidly and complexly during the transition from one machining condition to another, such as when the tool changes from no-load feed to cutting load, or from roughing to finishing. These fluctuations cause the machining signal to exhibit instantaneous deviations or abrupt changes, making it difficult for traditional anomaly monitoring methods based on overall machining data to accurately capture key anomaly features. This results in the difficulty in timely detection of abnormal states, increasing the risk of workpiece damage, tool wear, or decreased machining accuracy. To address these problems, this invention proposes a solution. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a method and system for identifying abnormal states during CNC machine tool machining. By first identifying state switching points in the machining path segment and constructing an optimal machining interval, and then performing anomaly analysis within this interval, the method addresses the problem that abnormal states are difficult to identify in a timely and accurate manner during the transition phase of CNC machine tool machining conditions when load, vibration, or force distribution fluctuates rapidly.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A method for identifying abnormal states during CNC machine tool machining includes the following steps: acquiring machining signal data of the CNC machine tool, and extracting the forces on the machining path segments based on the machining signal data to establish a feature set;

[0007] Feature matching is performed on the force establishment feature set to obtain the force establishment offset, and several processing state switching thresholds are constructed based on the force establishment offset. The state switching points of the processing path segments are identified based on the several processing state switching thresholds, and the path segment boundaries are aligned according to the state switching points to obtain the path alignment set. The first abnormal state identification is performed on the path alignment set, and the optimal processing state switching threshold is determined based on the identification results. The optimal processing interval is defined within the processing path segment based on the optimal processing state switching threshold, and the second abnormal state identification is performed within the optimal processing interval.

[0008] In a preferred embodiment, the step of extracting the force-based feature set of the machining path segment based on machining signal data specifically involves: parsing the tool movement trajectory based on a preset machining instruction code, and dividing the continuous machining process into several machining path segments based on the tool movement trajectory; performing time-frequency joint analysis on the machining signal data, and performing time-series correlation on each machining path segment based on the analysis results to obtain a time-series feature matrix for each machining path segment; performing sliding window segmentation on the time-series feature matrix along the time axis, and calculating the statistical distribution moments of the feature values ​​of each dimension within each time window; and concatenating the statistical distribution moments of all time windows according to a preset machining order to obtain the force-based feature set.

[0009] In a preferred embodiment, the step of performing feature matching on the force establishment feature set to obtain the force establishment offset, and constructing several processing state switching thresholds based on the force establishment offset, specifically involves: selecting a benchmark processing process that processes the same workpiece material and geometric features under standard working conditions, and extracting the benchmark force establishment feature matching template corresponding to the benchmark processing process; using a dynamic time warping algorithm, performing nonlinear sequence matching between each path segment feature subset of the force establishment feature set and the corresponding path segment in the benchmark force establishment feature matching template, and calculating the cumulative matching distance between the two in the feature space; defining the cumulative matching distance of each path segment as the force establishment offset of that path segment, and obtaining a force establishment offset sequence arranged along the processing time sequence; analyzing the statistical distribution characteristics of the force establishment offset sequence, and identifying abrupt change regions in the offset sequence based on the statistical distribution characteristics; calculating the boundary values ​​of the abrupt change regions and the distribution quantiles of the force establishment offset sequence, and adaptively generating several processing state switching thresholds based on the boundary values ​​and distribution quantiles.

[0010] In a preferred embodiment, the dynamic time warping algorithm is used to perform nonlinear sequence matching between each path segment feature subset of the force-established feature set and the corresponding path segment in the benchmark force-established feature matching template, and to calculate the cumulative matching distance between the two in the feature space. Specifically:

[0011] The force-induced feature subset of a certain path segment in the current processing is established and arranged into a first multidimensional feature sequence according to the preset signal acquisition time order. The corresponding reference path segment is retrieved from the reference force-induced feature matching template, and its path segment feature subset is arranged into a second multidimensional feature sequence. The Mahalanobis distance matrix between each feature point of the first and second multidimensional feature sequences is constructed. The dynamic time warping algorithm is used to find a warped path with the minimum cumulative distance from the starting point to the ending point in the Mahalanobis distance matrix and calculate the minimum cumulative distance of the warped path to obtain the cumulative matching distance.

[0012] In a preferred embodiment, the step of identifying the state switching points of processing path segments based on several processing state switching thresholds and aligning the path segment boundaries according to the state switching points to obtain a path alignment set specifically involves: for each processing state switching threshold, traversing the force establishment offset sequence, marking the point where the force establishment offset first exceeds the processing state switching threshold as the state switching start point, and marking the point where the offset falls back below the processing state switching threshold as the state switching end point; re-dividing the processing path segment based on the state switching start point and state switching end point to obtain several processing stage sequences; processing the several processing stage sequences and adjusting the start and end boundaries of each path segment to obtain a path alignment set.

[0013] In a preferred embodiment, the first abnormal state identification of the path alignment set and the determination of the optimal processing state switching threshold based on the identification results specifically involves: extracting the path segment feature vector of each path alignment subset within the path alignment set, and calculating the first standard deviation sequence of the path segment feature vector at each time point; performing spectral analysis on the first standard deviation sequence and extracting its dominant frequency component and corresponding harmonic energy distribution; constructing an evaluation function based on the dominant frequency component and corresponding harmonic energy distribution, and performing a consistency score on the path alignment set based on the evaluation function; and taking the processing state switching threshold corresponding to the path alignment subset with the lowest consistency score as the optimal processing state switching threshold.

[0014] In a preferred embodiment, the step of defining the optimal processing interval within the processing path segment based on the optimal processing state switching threshold specifically involves: identifying the first state switching point corresponding to the optimal processing state switching threshold within the processing path segment; extending a preset time window forward and backward with the first state switching point as the midpoint to obtain the extended time window; and limiting the range of the extended time window to the optimal processing interval.

[0015] In a preferred embodiment, the second abnormal state identification within the optimal processing interval specifically involves: extracting the multidimensional feature vector subsequence within the optimal processing interval of each processing path segment, and averaging them according to time points to obtain the first reference feature trajectory.

[0016] Calculate the point-by-point residuals between the multidimensional feature vector subsequence of each path segment and the first reference feature trajectory to obtain the residual sequence of each path segment; perform wavelet transform on the residual sequence of each path segment to obtain the time-frequency energy distribution map; based on the time-frequency energy distribution map, calculate the energy change value of the residual sequence of each path segment within a preset frequency band; based on the energy change value, calculate the time-frequency abnormal response vector of each processing path segment within the optimal processing interval; compare the time-frequency abnormal response vector with the preset statistical distribution model of historical normal processing samples within the same interval and calculate the anomaly confidence; when the anomaly confidence exceeds a preset threshold, it is determined that the corresponding processing path segment has an abnormal processing state within the optimal processing interval.

[0017] The technical effects and advantages of the present invention, a method and system for identifying abnormal states during CNC machine tool machining, are as follows:

[0018] This invention acquires machining signal data from CNC machine tools and extracts force features from machining path segments to establish a fine characterization of machining state changes. Further, it matches the force feature set to obtain force offsets and constructs multiple machining state switching thresholds based on these offsets, enabling precise location of machining condition transition stages within the machining path segment. By identifying state switching points based on thresholds and aligning path segment boundaries to form a path alignment set, the location of abnormal states becomes more accurate. The initial abnormal state identification can preliminarily screen potential abnormalities in the overall machining process and determine the optimal machining state switching threshold to limit the optimal machining range of the machining path segment. A second abnormal state identification is performed within the optimal machining range, further improving the sensitivity and accuracy of abnormalities in critical transition stages. Overall, this method effectively solves the problem of difficulty in timely and accurate identification of abnormal states during transition stages of CNC machine tool machining conditions due to rapid fluctuations in load, vibration, or force distribution. It achieves phased refinement of abnormal monitoring, path alignment, and threshold optimization, improving machining safety and accuracy while reducing the risk of tool wear and workpiece damage. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating a method for identifying abnormal states during CNC machine tool machining according to the present invention.

[0020] Figure 2 This is a schematic diagram of the structure of an abnormal state identification system for CNC machine tool machining process according to the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0022] Example 1, Figure 1 This invention provides a method for identifying abnormal states during CNC machine tool machining, comprising the following steps:

[0023] S1, acquire the machining signal data of the CNC machine tool, and extract the force of the machining path segment based on the machining signal data to establish a feature set;

[0024] In this embodiment, a feature set is established by extracting the force on the processing path segment based on the processing signal data, specifically as follows:

[0025] The tool movement trajectory is analyzed based on the preset machining instruction code, and the continuous machining process is divided into several machining path segments based on the tool movement trajectory;

[0026] Time-frequency joint analysis is performed on the processing signal data, and time-series correlation is performed on each processing path segment based on the analysis results to obtain the time-series feature matrix of each processing path segment;

[0027] The temporal feature matrix is ​​segmented along the time axis using a sliding window, and the statistical distribution moments of the feature values ​​of each dimension within each time window are calculated.

[0028] The statistical distribution moments of all time windows are spliced ​​together according to the preset processing order to obtain the force establishment feature set.

[0029] It should be noted that machining signal data refers to the physical quantity data related to the machining state that are collected in real time by the machine tool body, drive unit, and peripheral sensing devices during the machining process of a CNC machine tool. This mainly includes spindle current, spindle power, spindle speed, servo current of each feed axis, servo torque, feed rate, cutting force signal, vibration signal, acoustic emission signal, tool temperature, workpiece surface temperature, and coolant pressure. Taking milling as an example, at the moment the tool enters the workpiece, the spindle current will rise significantly, the X / Y / Z axis servo current will fluctuate due to the increased load, the vibration acceleration sensor will detect the impact component, and the acoustic emission sensor will show high-frequency abrupt changes. These multi-source data collectively reflect the force establishment process between the tool and the workpiece. In actual acquisition, these signals are usually recorded continuously at millisecond intervals, forming multi-channel time series data. Each time point corresponds to a multi-dimensional feature vector. For example, at a certain moment, the spindle current is recorded as 12A, the Z-axis servo current as 4.5A, the vibration acceleration as 0.8g, and the acoustic emission amplitude as 0.35V. These data collectively constitute the machining signal data.

[0030] Secondly, machining instruction codes are typically CNC program codes, such as G-codes and M-codes. G-codes describe the geometric motion trajectory and interpolation method; for example, G00 indicates rapid positioning, G01 indicates linear interpolation, and G02 / G03 indicates circular interpolation, along with X, Y, and Z coordinate values, as well as parameters such as feed rate (F) and spindle speed (S). M-codes control auxiliary functions, such as M03 for spindle forward rotation and M08 for turning on the coolant. The parsing process reads the machining instructions line by line according to the program execution order, extracting information such as target coordinates, motion type, feed rate, and spindle status from each instruction, and calculating the continuous positional changes of the tool in the machining space according to the interpolation rules. For example, a program like "G01 X50 Y0 F200" indicates that the tool moves along a straight line from its current position to a position of X=50, Y=0 at a feed rate of 200 mm per minute. This straight-line trajectory can be reconstructed by reading the current coordinates and the target coordinates. If it is "G02 X80 Y30 R20", the corresponding circular trajectory is generated based on the arc radius information. The tool motion trajectory refers to the spatial path of the tool tip or tool reference point in the workpiece coordinate system. It is a three-dimensional spatial point sequence arranged in time order. For example, the tool moves from (0,0,5) to (0,0,0), then moves along the workpiece surface to (50,0,0), and then turns to (50,50,0). These continuous spatial points constitute a complete motion trajectory.

[0031] Furthermore, after obtaining the complete tool movement trajectory, it can be segmented based on changes in the trajectory's geometric features, feed state, or machining process semantics. Specifically, it can be divided according to the following categories of boundary rules: The first category is points of geometric feature change, such as the transition from a straight line to an arc, or changes in the radius of an arc; the second category is points of motion state change, such as the transition from rapid positioning to cutting feed, the moment the tool enters the material, or the moment the tool retracts from the material; the third category is points of abrupt changes in spatial direction, such as when the angle of the motion direction exceeds a preset threshold. For example, in cavity milling, the tool's rapid movement from its idle stroke to above the workpiece constitutes one segment; the tool entering the material is the second segment; contour milling along the cavity boundary is the third segment; and finally, tool retraction is the fourth segment. By traversing the sequence of trajectory points and detecting changes in trajectory curvature, feed rate, or machining command type, the continuous machining process can be divided into multiple semantically clear machining path segments, each with relatively stable motion and force characteristics.

[0032] Joint time-frequency analysis refers to simultaneously examining the changes in a signal across both the time and frequency dimensions. For example, short-time Fourier analysis or wavelet decomposition can be used to decompose the original vibration or current signal into energy distributions of different frequency components that change over time. In practice, the spindle current and vibration signal can be segmented for analysis at fixed time intervals. Within each time interval, features such as low-frequency energy, mid-frequency energy, high-frequency energy, and amplitude variation trends can be extracted, resulting in multi-dimensional feature values ​​with time stamps. These features are then aligned with the previously defined machining path segments according to their timestamps. For example, if a certain time interval corresponds to the tool executing the second path segment, then the features within that time interval are assigned to the second path segment. Finally, for each machining path segment, a multi-dimensional feature set arranged chronologically can be obtained. For instance, in path segment A, 200 time points are collected, each containing multiple features such as the average spindle current, low-frequency vibration energy, high-frequency vibration energy, and acoustic emission intensity. These features, arranged chronologically, form the temporal feature matrix for that path segment.

[0033] Furthermore, the temporal feature matrix for each path segment can be divided along the time axis using a fixed-length sliding window, for example, every 50 milliseconds as a window, with partial overlap between windows. Each window contains multidimensional feature data at several consecutive time points. Within each window, statistical distribution characteristics, such as mean, range, variance, skewness trend, and peak level, are calculated for each dimension to characterize the stability and fluctuation characteristics of force changes within that time window. For example, during the stable cutting phase of a certain path segment, the mean spindle current may stabilize at around 11A with a small variance; however, when the tool enters the hard inclusion region, the current fluctuation amplitude within the window increases significantly, the high-frequency vibration components intensify, and its statistical distribution characteristics change significantly. After performing similar processing on each window, multiple statistical distribution moments corresponding to the windows are obtained, each representing the comprehensive distribution state of multidimensional features within that time window.

[0034] Finally, after completing the sliding window statistics for all path segments, the statistical results of each window can be concatenated according to the actual machining execution order. The preset machining order is usually consistent with the machining command execution order, for example, first performing rough machining of the outer contour, then finishing machining of the inner cavity, and finally chamfering. The statistical results of all time windows for the first path segment are arranged chronologically, and then the statistical results corresponding to the second and third path segments are sequentially concatenated to form an overall feature sequence unfolding along the machining timeline. Each element in this sequence corresponds to the statistical distribution state of a specific time window, reflecting the changes in the force establishment process of the tool at that stage. For example, if the entire machining process generates 120 time windows, the final result is a feature sequence of length 120, with each position containing multidimensional statistical distribution features. This overall set arranged according to the machining order is the force establishment feature set, used for subsequent matching analysis with the baseline machining process.

[0035] S2, perform feature matching on the force establishment feature set to obtain the force establishment offset, and construct several processing state switching thresholds based on the force establishment offset;

[0036] In this embodiment, feature matching is performed on the force establishment feature set to obtain the force establishment offset, and several processing state switching thresholds are constructed based on the force establishment offset, specifically:

[0037] Select a benchmark machining process that processes the same workpiece material and geometric features under standard working conditions, and extract the benchmark force corresponding to the benchmark machining process to establish a feature matching template;

[0038] The dynamic time warping algorithm is used to perform nonlinear sequence matching between each path segment feature subset of the force-established feature set and the corresponding path segment in the benchmark force-established feature matching template, and to calculate the cumulative matching distance between the two in the feature space.

[0039] The cumulative matching distance of each path segment is defined as the force establishment offset of that path segment, thus obtaining a sequence of force establishment offsets arranged along the processing time sequence.

[0040] The statistical distribution characteristics of the offset sequence established by the force are analyzed, and the abrupt change regions in the offset sequence are identified based on the statistical distribution characteristics.

[0041] Calculate the boundary values ​​of the mutation region and the distribution quantiles of the force offset sequence, and adaptively generate several processing state switching thresholds based on the boundary values ​​and distribution quantiles.

[0042] It should be noted that the so-called standard operating condition refers to a machining process completed under the following conditions: tool wear is in its initial or slight stage, machine tool rigidity is normal, cooling conditions are stable, machining parameters meet the recommended process values, and the machining results are confirmed to be normal by quality inspection. When selecting a benchmark machining process, it is essential to ensure that the machining materials are consistent (e.g., all are 45# steel or aluminum alloy 7075), the geometry is consistent (e.g., the same cavity depth and wall thickness), the tool type is consistent, and the spindle speed and feed parameters are consistent. Priority should be given to machining batches that meet the design requirements for surface roughness and dimensional accuracy. In actual operation, multiple qualified batches can be selected from the historical machining database, and the one with the smallest vibration fluctuation and the most stable spindle current change curve can be selected as the benchmark sample.

[0043] After obtaining the benchmark processing procedure, the force establishment feature set is extracted according to the method in claim 2, and further divided according to path segments. For each path segment, the multidimensional statistical features obtained by sliding window statistics are arranged in the order of processing time to form a standard force establishment feature trajectory. If there are multiple qualified samples in history, the average or interval statistics of multiple samples in the same path segment can be superimposed to obtain a more representative feature range. The so-called benchmark force establishment feature matching template is essentially a multidimensional statistical feature expression of the force establishment process of each path segment under standard working conditions, which includes both the time unfolding order and the normal fluctuation range of each feature dimension. For example, in a certain path segment, the average spindle current is stable between 10A and 11A under standard conditions, and the high-frequency vibration energy is at a low level. This multidimensional statistical trajectory and its stable range constitute the benchmark matching template of the path segment for subsequent comparative analysis.

[0044] In actual machining, the force establishment feature subset extracted from the current machining path segment is dynamically time-warped and matched with the corresponding path segment in the aforementioned benchmark matching template to obtain a cumulative matching distance. This cumulative matching distance reflects the overall degree of difference between the force establishment process of the current path segment and the standard working condition. The smaller the matching distance, the closer the current machining state is to the normal state; the larger the matching distance, the more the force establishment process deviates, such as tool wear or local hardness changes in the material.

[0045] The cumulative matching distance calculated for each path segment is directly defined as the force establishment offset of that path segment. For example, if a workpiece is processed into 8 path segments with matching results of 0.8, 1.0, 0.9, 1.2, 3.5, 3.8, 1.1, and 0.95, and the 5th and 6th path segments are significantly higher than the preceding and following segments, then these two path segments can be considered to have significant force offsets. These offsets are arranged sequentially according to the processing execution order, forming a force establishment offset sequence along the processing timeline, such as [0.8, 1.0, 0.9, 1.2, 3.5, 3.8, 1.1, 0.95]. This sequence directly reflects the trend of the entire workpiece's force deviating from the standard state during processing.

[0046] Furthermore, after obtaining the complete force establishment offset sequence, its overall distribution characteristics can be statistically analyzed, including the overall mean level, dispersion, skewness trend, and local fluctuation intensity. For example, in most normal processing sections, the offset is concentrated between 0.8 and 1.2, with small fluctuations; if a particular path segment suddenly rises above 3, it indicates that abnormal force establishment behavior may occur at that location. To identify abrupt change regions, local sliding statistics can be performed along the sequence to observe the changes in the difference between adjacent path segments. When the offset difference between two adjacent path segments significantly exceeds the overall average fluctuation range, and several consecutive segments maintain a high level, it can be identified as an abrupt change region. For example, if the first four segments remain around 1, the fifth segment suddenly rises to 3.5 and remains at 3.8 in the sixth segment, then the starting position of the fifth segment can be regarded as the abrupt change starting point; if it returns to 1.1 after the sixth segment ends, then the sixth segment is the abrupt change ending point. Through this analysis method based on overall distribution characteristics and local change amplitude, the offset sequence can be divided into stable and abrupt change regions.

[0047] Secondly, after identifying the mutation region, all offset values ​​within that region can be recorded, and their upper and lower boundary values ​​can be calculated. The boundary values ​​are typically the minimum and maximum values ​​of the mutation region. For example, in the above example, the mutation region contains 3.5 and 3.8, so its lower boundary value is 3.5, and its upper boundary value is 3.8. Simultaneously, the entire offset sequence is sorted to obtain an ascending order, for example, [0.8, 0.9, 0.95, 1.0, 1.1, 1.2, 3.5, 3.8]. Based on this, the quantiles corresponding to different positions can be determined. For example, values ​​in the middle position represent the median level, and values ​​in higher positions represent the high quantile level. If 3.5 and 3.8 are located in the highest interval of the sequence, they can be considered to be in the high quantile range, for example, in the quantile interval above the 90th percentile. In this way, the positional relationship between the normal fluctuation range and the abnormally high range within the overall distribution can be clearly defined.

[0048] Finally, based on the boundary values ​​of the mutation region and the overall distribution quantile information obtained above, multiple processing state switching thresholds can be adaptively generated according to the following steps:

[0049] The first step is to divide the entire offset sequence into a normal distribution interval and an abnormal high interval, record the upper limit of the normal interval (e.g., 1.2), and record the lower boundary of the abrupt change region (e.g., 3.5).

[0050] The second step is to determine several representative quantiles based on the overall distribution, such as the median level, the higher level, and the very higher level, and extract the corresponding values, such as 1.0, 1.2, and 3.5.

[0051] The third step is to divide the region into layers between the upper limit of the normal range and the lower boundary of the mutation region. For example, a transition threshold of 2.0 is set between 1.2 and 3.5 to distinguish between mild and severe shifts.

[0052] The fourth step is to use the normal upper limit value, the transition value, and the mutation lower boundary value as the switching thresholds for different levels of processing states, for example, threshold 1 is 1.2, threshold 2 is 2.0, and threshold 3 is 3.5.

[0053] Fifth, in subsequent processing, when the offset exceeds threshold 1 for the first time, the processing state is considered to have entered the mild fluctuation zone; when it exceeds threshold 2, it is considered to have entered the obvious offset zone; when it exceeds threshold 3, it is considered to have entered the significant abnormal zone.

[0054] Through the above steps, the threshold is not set manually, but is automatically generated based on the overall offset distribution characteristics of the current processing batch. It can be dynamically adjusted as the material batch changes or the tool status changes, thereby ensuring that the processing status switching judgment has adaptability and stability.

[0055] In this embodiment, a dynamic time warping algorithm is used to perform nonlinear sequence matching between each path segment feature subset of the force-established feature set and the corresponding path segment in the baseline force-established feature matching template, and to calculate the cumulative matching distance between the two in the feature space. Specifically:

[0056] The force on a certain path segment in the current processing is used to establish a feature subset, which is then arranged into the first multidimensional feature sequence according to the preset signal acquisition time order.

[0057] The corresponding baseline path segment is retrieved from the feature matching template established by the baseline force, and its path segment feature subset is arranged into a second multi-dimensional feature sequence.

[0058] Construct the Mahalanobis distance matrix between each feature point of the first multidimensional feature sequence and the second multidimensional feature sequence;

[0059] The dynamic time warping algorithm is used to find a warped path with the minimum cumulative distance from the starting point to the ending point in the Mahalanobis distance matrix and calculate the minimum cumulative distance of the warped path to obtain the cumulative matching distance.

[0060] It should be noted that during the current machining process, once a certain path segment has been divided according to the tool movement trajectory, all sliding window statistical features belonging to that path segment can be extracted from the force establishment feature set corresponding to that path segment. Since the machining signal already has a timestamp during acquisition, it can be sorted according to the timestamp from smallest to largest, and the multidimensional statistical features corresponding to all time windows within that path segment can be arranged sequentially in chronological order. For example, if 120 milliseconds of machining signal are acquired within a certain path segment, and statistics are performed in 20-millisecond windows, six time windows will be obtained. Each window contains several statistical features, such as the average spindle current, spindle current fluctuation amplitude, low-frequency vibration energy, high-frequency vibration energy, and acoustic emission intensity—a total of five dimensions. The five-dimensional features of the first window are used as the first feature point, and the five-dimensional features of the second window are used as the second feature point, arranged sequentially to form a time-expanded sequence, which is the first multidimensional feature sequence. Each element of this sequence is a multidimensional feature vector, reflecting the force establishment evolution trajectory of the current machining path segment from entry, stable cutting to exit.

[0061] Secondly, during the benchmark machining process, standard force establishment feature matching templates have been established for each path segment, and each path segment has a clear number or geometric identification information, such as "first segment of rough machining of outer contour" and "finishing segment of inner sidewall of cavity". After the current machining path segment is determined, path segments with the same geometric semantics and process attributes can be found in the benchmark template based on the path segment number, geometric position features, or tool movement direction features. For example, if the current path segment is a contour milling segment along the X direction, then path segments that are also contour milling segments in the same area will be searched in the benchmark template. After finding the corresponding path segment, the sliding window statistical features of the path segment are also arranged in chronological order to form a second multidimensional feature sequence. For example, if the benchmark path segment has a total of 8 time windows under standard working conditions, and each window contains statistical features of the same dimension, then they are arranged in chronological order to form a sequence of 8 feature points. This second multidimensional feature sequence represents the ideal change trajectory of the force establishment of the path segment under standard conditions.

[0062] Furthermore, after obtaining the first and second multidimensional feature sequences, pairwise comparisons need to be performed on each feature point in both sequences. The specific steps are as follows: First, ensure that the feature dimensions in the two sequences are consistent, for example, both should include the same dimensions such as the mean principal current, low-frequency vibration energy, and high-frequency vibration energy. Second, calculate the covariance statistical structure of the overall features of the reference sequence to reflect the correlation and scale differences between dimensions. Then, traverse each feature point in the first sequence and calculate the distance between it and each feature point in the second sequence. During the calculation, not only are the numerical differences between dimensions considered, but the correlation between dimensions is also taken into account to weight the differences, thus obtaining a distance value. If the length of the first sequence is 6 and the length of the second sequence is 8, a 6x8 distance matrix is ​​ultimately formed, where the element in the i-th row and j-th column of the matrix represents the comprehensive degree of difference between the i-th feature point in the first sequence and the j-th feature point in the second sequence. For example, if the average current in the third time window of the first sequence is 12A, while the average current in the fifth time window of the reference sequence is 10.5A, and the vibration energy also differs, then the distance value at the corresponding position will be relatively larger. After all pairwise comparisons are completed, a complete Mahalanobis distance matrix is ​​formed.

[0063] After obtaining the distance matrix, the top-left corner can be considered the starting point, and the bottom-right corner the ending point. The idea behind dynamic time warping is to allow two time series to be elastically stretched or compressed along the time axis to find a path from the starting point to the ending point that minimizes the sum of the distances along the way. In practice, starting from the starting point, each step only allows movement to the right, down, or lower right, corresponding to extension, compression, or synchronization in time alignment, respectively. During the traversal, the distances are accumulated, and at each step, the path direction with the smallest accumulated distance is selected to continue until the ending point is reached. For example, when the first sequence is shorter than the second sequence, the algorithm may move continuously to the right at certain positions, indicating that there are additional time points in the reference sequence that need to be merged and matched. The final complete path is the warped path, and the sum of all distances along this path is the minimum accumulated distance. This minimum accumulated distance is defined as the accumulated matching distance between the current path segment and the reference path segment, used to quantify the degree of deviation in force establishment.

[0064] Finally, the first multidimensional feature sequence originates from the current actual machining process, reflecting the force establishment behavior of a certain path segment under real-time machining conditions. It may be affected by factors such as tool wear, material hardness fluctuations, and changes in cooling conditions, thus exhibiting uncertainty and volatility. The second multidimensional feature sequence, on the other hand, originates from benchmark machining samples under standard operating conditions. It is an ideal force establishment trajectory extracted under conditions of good tool condition and acceptable machining quality, representing a reference pattern for normal machining behavior. Both sequences are structurally identical, being multidimensional feature vector sequences arranged in chronological order, but they may differ in numerical stability and fluctuation amplitude. For example, under normal conditions, the current change trends of both sequences are basically consistent; however, when tool wear is severe, the average current value in the first sequence is generally higher, and the vibration energy is enhanced, showing a significant deviation from the second sequence. By regularizing and matching these two sequences and calculating the cumulative distance, the degree of deviation between the current machining state and the standard state can be objectively assessed.

[0065] S3, based on several processing state switching thresholds, identify the state switching points of the processing path segments and align the path segment boundaries according to the state switching points to obtain the path alignment set;

[0066] In this embodiment, state switching points of processing path segments are identified based on several processing state switching thresholds, and the path segment boundaries are aligned according to the state switching points to obtain a path alignment set, specifically:

[0067] For each processing state switching threshold, traverse the force establishment offset sequence, mark the point where the force establishment offset first exceeds the processing state switching threshold as the state switching start point, and mark the point where the offset falls back below the processing state switching threshold as the state switching end point.

[0068] The processing path segment is re-divided based on the state transition start point and state transition end point to obtain several processing stage sequences.

[0069] By performing a series of processing stages and adjusting the start and end boundaries of each path segment, a path alignment set is obtained.

[0070] It should be noted that after obtaining the sequence of force establishment offsets arranged along the processing timeline, for example, a sequence obtained from a certain processing step is [0.9, 1.1, 1.0, 1.3, 2.8, 3.2, 3.0, 1.4, 1.1], if a certain processing state switching threshold is set to 2.5, the sequence can be traversed point by point in chronological order. During traversal, when it is found that the offset corresponding to a certain path segment first changes from below the threshold to above the threshold, for example, jumping from 1.3 to 2.8, the path segment number corresponding to that position is marked as the state switching start point. The state switching start point refers to the position of the first path segment where the processing state moves from the normal fluctuation zone to the abnormal offset zone, representing the time point when the force establishment behavior begins to significantly deviate from the baseline state. Continuing to traverse forward, when the offset falls from above the threshold back to below the threshold, for example, from 3.0 to 1.4, the path segment corresponding to that position is marked as the state switching end point. The state switching end point represents the time point when the processing state recovers from the abnormal offset zone to the normal fluctuation zone. If the offset remains above the threshold for multiple path segments, only the position where the offset first exceeds the threshold is considered the starting point, and the position where the offset first falls below the threshold is considered the ending point. In this way, one or more abnormal state intervals can be identified throughout the entire processing sequence.

[0071] Furthermore, after identifying the start and end points of the state transition, the original processing path segments divided by geometric trajectories can be reorganized into stages. For example, if the original division consists of 9 path segments, and the 5th segment is the start point of the state transition and the 8th segment is the end point, then path segments 1 to 4 can be divided into the first processing stage (normal stage), path segments 5 to 8 into the second processing stage (abnormal offset stage), and path segment 9 and thereafter into the third processing stage (recovery stage). If there are multiple start and end point pairs in the sequence, multiple stage intervals can be formed sequentially. Each processing stage sequence consists of several consecutive path segments, and the trend of offset change within the same stage is consistent, such as a sustained high or sustained low position. Through this re-division method based on offset state intervals, the division of processing stages no longer depends solely on geometric trajectories, but introduces the dynamic change characteristics of force establishment behavior, making the stage boundaries more consistent with the actual processing state changes.

[0072] Finally, after obtaining multiple processing stage sequences, the stage division results obtained under different thresholds can be compared and integrated. For example, when there are multiple processing state switching thresholds, different thresholds may produce different start and end intervals. These stage division results can be superimposed and compared, and the intervals identified as abnormal under most thresholds can be selected as the core abnormal intervals, and the path segment boundaries can be refined and adjusted accordingly. In specific operation, the start time of the path segment corresponding to the state switching start point can be used as the new stage boundary start point, and the end time of the end path segment can be used as the stage boundary end point. If a path segment is partially covered by multiple stage division results, the path segment can be re-divided according to the time ratio, so that part of it belongs to the normal stage and the other part belongs to the abnormal stage. For example, if the first half of the 5th path segment is still in normal fluctuation, and the second half of the 5th path segment enters the abnormal zone, the 5th path segment can be split into two sub-segments on the time axis and assigned to different stages respectively. After adjusting and unifying the boundaries of all path segments, a new path alignment set is finally obtained, in which the start and end boundaries of each path segment are consistent with the force establishment state change. This path alignment set achieves synchronous alignment between geometric trajectory and force state evolution, providing more accurate analysis units for subsequent anomaly identification.

[0073] S4, perform the first abnormal state identification on the path alignment set, and determine the optimal processing state switching threshold based on the identification results;

[0074] In this embodiment, the path alignment set undergoes an initial abnormal state identification, and the optimal processing state switching threshold is determined based on the identification results. Specifically:

[0075] Extract the path segment feature vectors of each path alignment subset within the path alignment set, and calculate the first standard deviation sequence of the path segment feature vectors at each time point;

[0076] Spectral analysis was performed on the first standard deviation sequence, and its dominant frequency component and corresponding harmonic energy distribution were extracted.

[0077] An evaluation function is constructed based on the dominant frequency component and the corresponding harmonic energy distribution, and the consistency score of the path alignment set is performed based on the evaluation function.

[0078] The processing state switching threshold corresponding to the path alignment subset with the lowest consistency score is taken as the optimal processing state switching threshold.

[0079] It should be noted that after path alignment is completed, each processing state switching threshold will form a corresponding path alignment subset, and each subset contains several path segments that have been boundary-corrected. For a certain path alignment subset, the multidimensional features of all corresponding path segments at the same relative time position can be aligned under a unified time reference framework. For example, in a certain subset, there are 6 aligned path segments, and each path segment is resampled into a time point sequence of the same length after alignment, for example, a unified 50 time points. At each time point, there is a corresponding multidimensional feature vector, which may contain multiple dimensions such as the mean spindle current, low-frequency vibration energy, high-frequency vibration energy, and acoustic emission intensity. This set of multidimensional vectors arranged in time within a single path segment is the path segment feature vector sequence. Subsequently, at the same time point, the dispersion of the feature values ​​corresponding to all path segments at that time point is statistically analyzed. For example, at the 10th time point, the mean spindle currents of the 6 path segments are 10.1, 10.3, 9.9, 10.2, 10.0, and 10.4, respectively, and the fluctuation level at that time point can be calculated. Performing this operation at each time point yields a dispersion curve unfolding over time; this curve is the first standard deviation sequence. The path segment feature vector reflects the force characteristics within a single path segment as they change over time, while the first standard deviation sequence reflects the degree of consistency between different path segments at the same time location.

[0080] After obtaining the first standard deviation sequence, it can be treated as a time signal describing uniform fluctuations and subjected to frequency analysis. Specifically, the sequence can be transformed into a frequency distribution by performing a frequency domain transformation. If the standard deviation sequence is generally stable and without significant periodicity, low-frequency components dominate in the frequency domain. If periodic fluctuations exist, such as a significant increase in the standard deviation at certain time points, a significant peak will form in the frequency domain. By scanning the spectrum, the frequency component with the highest amplitude can be identified and determined as the dominant frequency component. For example, in a path-aligned subset, if the standard deviation sequence shows a periodic increase every 5 time points, a significant peak will form at the corresponding frequency position. Further analysis of integer multiples of this dominant frequency reveals the harmonic energy distribution. If the harmonic energy distribution is concentrated and regular, it indicates that the uniform fluctuations between path segments are structural; if the spectral distribution is chaotic and lacks a clear dominant peak, it indicates that the uniform fluctuations lack a stable pattern.

[0081] Furthermore, when constructing the evaluation function, two aspects can be considered: the proportion of dominant frequency energy and the concentration of harmonic energy. First, the energy of the dominant frequency component in the spectrum is extracted, and its relative proportion in the overall spectrum energy is calculated to reflect whether the consistency fluctuations are concentrated in a specific rhythm. Second, the energy distribution at integer multiples of the dominant frequency is statistically analyzed to assess whether the harmonic energy shows a decreasing or concentrated distribution, which measures the regularity of the fluctuation structure. Third, the proportion of dominant frequency energy and the concentration of harmonic energy are weighted together to form a consistency index. For example, if the dominant frequency energy of a path-aligned subset accounts for 60% of the overall energy and the harmonic distribution is clear, its consistency index is high; if the dominant frequency accounts for only 20% and the spectrum shows a multi-peaked, scattered distribution, the consistency index is low. Fourth, the consistency index is calculated for each path-aligned subset formed under different thresholds and normalized to obtain a consistency score. The higher the score, the better the time alignment effect between path segments under that threshold division, and the stronger the consistency.

[0082] Finally, under multiple processing state switching thresholds, multiple path alignment subsets are formed, each corresponding to a consistency score. If a threshold is too coarse, it may mix normal and abnormal stages, leading to increased dispersion between path segments, chaotic spectral performance, and a low consistency score. If a threshold is too fine, it may over-segment, causing unstable alignment and also resulting in a lower score. By comparing the scores corresponding to all thresholds, the one with the most reasonable score can be selected as the final threshold. In this technical solution, the processing state switching threshold corresponding to the path alignment subset with the lowest consistency score is determined as the optimal processing state switching threshold. This means that under this threshold, the consistency fluctuation between path segments is most significant, revealing the true boundary of processing state changes to the greatest extent. This optimal processing state switching threshold is not set manually based on experience, but is automatically selected based on the force-induced offset distribution and consistency spectral characteristics. It is used to subsequently limit the optimal processing interval and perform fine-grained anomaly identification, thereby improving anomaly localization accuracy.

[0083] S5, based on the optimal processing state switching threshold, limits the optimal processing interval within the processing path segment, and performs a second abnormal state identification within the optimal processing interval.

[0084] In this embodiment, the optimal processing interval is defined within the processing path segment based on the optimal processing state switching threshold, specifically as follows:

[0085] Based on the optimal processing state switching threshold, backtrack to identify the first state switching point corresponding to it within the processing path segment;

[0086] Using the first state switching point as the midpoint, expand the preset time window forward and backward to obtain the expanded time window;

[0087] The extended time window is limited to the optimal processing interval.

[0088] It should be noted that after determining the optimal processing state switching threshold, the complete force establishment offset sequence can be re-traversed to find the position where the offset first exceeds the optimal threshold. During the traversal, segments are checked sequentially according to processing time. When the force establishment offset of a certain processing path segment changes from below the optimal threshold to above it, the corresponding time position of that path segment is marked as the first state switching point. For example, if the offset sequence of a batch of processing is [0.9, 1.1, 1.0, 1.4, 2.9, 3.2, 1.3], and the optimal threshold is 2.5, then the offset of the 5th path segment, 2.9, is the position where it first exceeds the threshold. The starting time of this path segment or the time point within it when it first exceeds the threshold is the first state switching point. This first state switching point represents the critical time node when the processing process transitions from the normal force establishment stage to the significant offset stage, and is the starting boundary of the entire abnormal evolution process. Unlike the aforementioned general state switching points, the first state switching point specifically refers to the most representative state transition position identified under the optimal threshold condition, used for subsequent local fine-grained analysis.

[0089] After determining the first state transition point, a symmetrical time window can be constructed centered on this point. Specifically, a preset time length is set based on experience or processing rhythm characteristics, such as extending it by 200 milliseconds before and after, or by several sliding window units on each side. For example, if the time corresponding to the first state transition point is 5.2 seconds after processing begins, then backtracking 200 milliseconds to 5.0 seconds and extending it by 200 milliseconds to 5.4 seconds forms an extended time window lasting 400 milliseconds. If the processing rhythm is fast, it can also be extended according to the number of sampling windows within the path segment, for example, extending by 5 statistical windows before and after. In actual processing scenarios, force anomalies often show gradual signs before entering the anomaly stage; therefore, extending forward can cover the anomaly initiation stage, and extending backward can cover the anomaly stabilization stage. The extended time window constructed in this way not only includes the instant of state change but also covers the transition process before and after, making subsequent analysis more complete.

[0090] Finally, after obtaining the expanded time window, its range needs to be limited based on its actual position within the entire processing path segment. If the expanded window falls entirely within the current processing path segment, this time range can be directly defined as the optimal processing interval. If the expanded window exceeds the path segment boundary, it needs to be trimmed to within the effective time range of the path segment. For example, if the total duration of a path segment is 4.8 to 5.6 seconds, and the expanded window is 5.0 to 5.4 seconds, then the window is fully retained; if the expanded window is 4.7 to 5.3 seconds, then the starting point needs to be limited to 4.8 seconds. The time range obtained after boundary correction is the optimal processing interval. The optimal processing interval refers to a local key analysis area constructed around the critical state inflection point identified by the optimal state switching threshold. This area centrally reflects the entire process of the processing force evolving from normal to abnormal. Compared to a unified analysis of the entire path segment, the optimal processing interval focuses more on the abnormal sensitive area, amplifying abnormal features within a smaller time range and improving the accuracy and response speed of subsequent second abnormal state identification.

[0091] In this embodiment, a second abnormal state identification is performed within the optimal processing range, specifically as follows:

[0092] Extract the multidimensional feature vector subsequence within the optimal processing interval of each processing path segment, and average them according to time points to obtain the first reference feature trajectory;

[0093] Calculate the point-by-point residuals between the multidimensional feature vector subsequence of each path segment and the first reference feature trajectory to obtain the residual sequence of each path segment;

[0094] Wavelet transform is performed on the residual sequence of each path segment to obtain the time-frequency energy distribution map;

[0095] Based on the time-frequency energy distribution map, calculate the energy change value of the residual sequence of each path segment within the preset frequency band;

[0096] Based on the energy change value, calculate the time-frequency anomaly response vector of each processing path segment within the optimal processing interval;

[0097] Compare the time-frequency anomaly response vector with the preset statistical distribution model of historical normal processing samples within the same interval and calculate the anomaly confidence level.

[0098] When the anomaly confidence level exceeds the preset threshold, it is determined that the corresponding processing path segment has an abnormal processing state within the optimal processing range.

[0099] It should be noted that after determining the optimal processing interval for each processing path segment, feature data within that time range can be extracted from the original multidimensional monitoring data. A multidimensional feature vector subsequence refers to a set of features corresponding to multiple time points arranged in chronological order within the optimal processing interval. For example, each time point may contain multiple dimensions such as spindle current, feed force, vibration amplitude, and acoustic emission intensity. Assuming an optimal processing interval length of 400 milliseconds and a sampling interval of 10 milliseconds, 40 time points can be obtained, each corresponding to a multidimensional feature vector. These 40 vectors, arranged chronologically, constitute the multidimensional feature vector subsequence for that path segment. After performing the same extraction on all processing path segments and aligning the time axis, the feature values ​​corresponding to all path segments are averaged at each relative time position. For example, at the 10th time point, the spindle current values ​​and vibration values ​​of all path segments are averaged, and so on, ultimately yielding an average feature trajectory that changes over time. This trajectory is the first reference feature trajectory. The first reference feature trajectory reflects the typical force change pattern within the optimal processing range and is a concentrated expression of the common features of multiple path segments.

[0100] After obtaining the first reference feature trajectory, the multidimensional feature vector subsequence of each path segment can be compared with the reference trajectory point-by-point over time. Specifically, at the same time index position, the actual feature value corresponding to the path segment is compared with the corresponding feature value in the reference trajectory, and the difference between the two is calculated. For example, at the 15th time point, the vibration amplitude of a certain path segment is 0.82, while the corresponding value in the reference trajectory is 0.75; the residual of the vibration dimension at that time point is the difference between the two. The same processing is performed for each time point and each feature dimension, and the multidimensional residuals can be combined into a single overall deviation index, ultimately resulting in a residual sequence arranged along time. This residual sequence reflects the degree of deviation of the current path segment from the typical processing mode within the optimal processing range; the larger the deviation, the higher the probability of abnormal fluctuations within that time period.

[0101] To further analyze the frequency structure of the residual over time, wavelet transform can be applied to the residual sequence. Specifically, the residual sequence is used as the input signal and decomposed into components at different frequency levels through multi-scale decomposition, thus obtaining the energy distribution at different time points and frequency ranges. Wavelet transform can simultaneously preserve both time and frequency information, allowing observation of high-frequency sudden fluctuations within a given time period. For example, in the optimal machining section of a certain path segment, if a slight tool chipping occurs in the latter half, the residual sequence will exhibit high-frequency oscillations at that time point. After wavelet decomposition, a local energy enhancement region will appear in the high-frequency band. Expanding the energy intensity at each time point and frequency band in a two-dimensional manner forms a time-frequency energy distribution map, where the horizontal axis represents time, the vertical axis represents the frequency band level, and color or numerical value represents energy intensity.

[0102] Secondly, after obtaining the time-frequency energy distribution map, a frequency band range sensitive to machining anomalies can be pre-defined, such as the mid-to-high frequency range, because tool anomalies usually cause higher frequency vibrations. Subsequently, within this frequency band, the energy values ​​within the time range corresponding to the optimal machining interval are statistically analyzed. For example, the energy of this frequency band at all time points is accumulated or averaged, and compared with the energy level of that path segment in the normal frequency band region, thus obtaining an energy change value. For instance, if the average energy of a certain path segment within the preset frequency band is twice the mean of the normal sample, it indicates an abnormal enhancement within that frequency range. This energy change value reflects the intensity level of the abnormal vibration component in the residual; the larger the value, the more significant the deviation from the typical machining trajectory, potentially indicating problems such as tool wear, material abnormalities, or cutting instability.

[0103] Furthermore, for each path segment, the corresponding energy change values ​​can be calculated within multiple preset frequency bands, such as the energy enhancement levels in the low-frequency, mid-frequency, and high-frequency ranges. Arranging these energy change values ​​from different frequency bands in a fixed order forms a multi-dimensional vector, which is the time-frequency anomaly response vector. For example, if a path segment has an energy change value of 0.8 in the low-frequency range, 1.5 in the mid-frequency range, and 2.3 in the high-frequency range, a three-dimensional response vector can be formed. This vector comprehensively reflects the distribution of anomaly intensity at different frequency levels of the path segment and is important input data for subsequent statistical comparisons.

[0104] Furthermore, statistical analysis can be performed on time-frequency anomaly response vectors within the same optimal processing interval in historical normal processing samples to obtain the distribution characteristics of energy change values ​​in each frequency band, such as the mean range, fluctuation range, and overall distribution pattern. These statistical characteristics constitute a statistical distribution model, used to describe the typical distribution range of time-frequency responses under normal processing conditions. When a new processing path segment obtains a time-frequency anomaly response vector, its position in the statistical distribution model can be determined, such as whether it falls outside the normal distribution range or how much it deviates from the center. Based on the degree of deviation, anomaly confidence can be calculated; a higher confidence indicates a greater probability that the current path segment belongs to an anomaly state. For example, if the energy change value of a path segment in the high-frequency interval far exceeds the upper limit of historical normal samples, its anomaly confidence will increase significantly. Anomaly confidence is essentially a probabilistic measure of whether the current state does not belong to the normal processing distribution.

[0105] In practical applications, an anomaly detection threshold can be set based on historical experience or verification data. For example, when the anomaly confidence level reaches a certain level, the path segment is considered to have an anomaly within the optimal processing range. When the anomaly confidence level calculated for a new path segment exceeds this threshold, the anomaly detection result can be output. For instance, if the threshold is set to 0.85, and the anomaly confidence level of a certain path segment is 0.92, then the path segment is determined to have an abnormal processing state within the optimal processing range. This detection method avoids relying solely on a single vibration peak or a single time point for judgment, but rather on the overall comparison results between the time-frequency energy distribution and historical normal statistical models, resulting in higher stability and reliability.

[0106] Finally, the first and second anomaly identification methods differ fundamentally in their identification targets, analysis scope, data processing depth, and judgment logic. The first anomaly identification focuses on determining the state transition structure across the entire path. Its core task is to establish the relationship between offset and different processing state transition thresholds through force analysis, identify potential abnormal stage boundaries during processing, and use a consistency spectrum evaluation mechanism to select the optimal processing state transition threshold, thereby determining the key state transition locations. The first identification leans more towards macro-level stage division and coarse anomaly boundary localization, resulting in finding the most representative state transition point and the optimal processing interval built around that point, essentially addressing the question of "at which stage might the anomaly occur." The second anomaly identification, however, takes into account the already determined optimal processing interval and performs refined modeling and deviation analysis on this local time period. This involves constructing a first reference feature trajectory, calculating point-by-point residuals, performing wavelet time-frequency decomposition, extracting frequency band energy change values, and comparing them with historical normal processing statistical distribution models to calculate the anomaly confidence and make a final judgment. The second identification belongs to the micro-level structured deviation quantification and probabilistic judgment, addressing whether an anomaly truly exists at this key stage and to what extent. In short, the first anomaly identification is a phased screening and key interval location based on the offset established by the force, which is a global coarse screening mechanism; the second anomaly identification is a local fine judgment mechanism based on residual time-frequency characteristics and statistical distribution modeling, which is an in-depth verification and precise quantification of the first location results. The two are progressive in level.

[0107] Example 2, Figure 2 This invention presents an abnormal state identification system for CNC machine tool machining processes, comprising a feature extraction module, a threshold construction module, a path alignment module, a threshold optimization module, and a state identification module.

[0108] The feature extraction module is used to acquire machining signal data of CNC machine tools and extract the force of machining path segments based on the machining signal data to establish a feature set;

[0109] The threshold construction module is used to perform feature matching on the force establishment feature set, obtain the force establishment offset, and construct several processing state switching thresholds based on the force establishment offset.

[0110] The path alignment module is used to identify the state switching points of the processing path segments based on several processing state switching thresholds and to align the boundaries of the path segments according to the state switching points to obtain a path alignment set.

[0111] The threshold optimization module is used to perform the first abnormal state identification on the path alignment set and determine the optimal processing state switching threshold based on the identification results.

[0112] The status recognition module is used to limit the optimal processing range within the processing path segment based on the optimal processing status switching threshold, and to perform a second abnormal status recognition within the optimal processing range.

[0113] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0114] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0115] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0116] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0117] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for identifying abnormal states during CNC machine tool machining, characterized in that, Includes the following steps: Acquire machining signal data from CNC machine tools, and extract the forces on the machining path segments based on the machining signal data to establish a feature set; Feature matching is performed on the force establishment feature set to obtain the force establishment offset, and several processing state switching thresholds are constructed based on the force establishment offset; The state switching points of the processing path segments are identified based on several processing state switching thresholds, and the path segment boundaries are aligned according to the state switching points to obtain the path alignment set; The first abnormal state identification is performed on the path alignment set, and the optimal processing state switching threshold is determined based on the identification results; Based on the optimal processing state switching threshold, the optimal processing interval is defined within the processing path segment, and a second abnormal state identification is performed within the optimal processing interval.

2. The method for identifying abnormal states during CNC machine tool machining according to claim 1, characterized in that, The process of extracting the force characteristics of the processing path segment based on processing signal data to establish a feature set is specifically as follows: The tool movement trajectory is analyzed based on the preset machining instruction code, and the continuous machining process is divided into several machining path segments based on the tool movement trajectory; Time-frequency joint analysis is performed on the processing signal data, and time-series correlation is performed on each processing path segment based on the analysis results to obtain the time-series feature matrix of each processing path segment; The temporal feature matrix is ​​segmented along the time axis using a sliding window, and the statistical distribution moments of the feature values ​​of each dimension within each time window are calculated. The statistical distribution moments of all time windows are spliced ​​together according to the preset processing order to obtain the force establishment feature set.

3. The method for identifying abnormal states during CNC machine tool machining according to claim 2, characterized in that, The process involves feature matching of the force establishment feature set to obtain the force establishment offset, and then constructing several processing state switching thresholds based on the force establishment offset. Specifically: Select a benchmark machining process that processes the same workpiece material and geometric features under standard working conditions, and extract the benchmark force corresponding to the benchmark machining process to establish a feature matching template; The dynamic time warping algorithm is used to perform nonlinear sequence matching between each path segment feature subset of the force-established feature set and the corresponding path segment in the benchmark force-established feature matching template, and to calculate the cumulative matching distance between the two in the feature space. The cumulative matching distance of each path segment is defined as the force establishment offset of that path segment, thus obtaining a sequence of force establishment offsets arranged along the processing time sequence. The statistical distribution characteristics of the offset sequence established by the force are analyzed, and the abrupt change regions in the offset sequence are identified based on the statistical distribution characteristics. Calculate the boundary values ​​of the mutation region and the distribution quantiles of the force offset sequence, and adaptively generate several processing state switching thresholds based on the boundary values ​​and distribution quantiles.

4. The method for identifying abnormal states during CNC machine tool machining according to claim 3, characterized in that, The dynamic time warping algorithm is employed to perform nonlinear sequence matching between each path segment feature subset of the force-established feature set and the corresponding path segment in the benchmark force-established feature matching template, and to calculate the cumulative matching distance between the two in the feature space. Specifically: The force on a certain path segment in the current processing is used to establish a feature subset, which is then arranged into the first multidimensional feature sequence according to the preset signal acquisition time order. The corresponding baseline path segment is retrieved from the feature matching template established by the baseline force, and its path segment feature subset is arranged into a second multi-dimensional feature sequence. Construct the Mahalanobis distance matrix between each feature point of the first multidimensional feature sequence and the second multidimensional feature sequence; The dynamic time warping algorithm is used to find a warped path with the minimum cumulative distance from the starting point to the ending point in the Mahalanobis distance matrix and calculate the minimum cumulative distance of the warped path to obtain the cumulative matching distance.

5. The method for identifying abnormal states during CNC machine tool machining according to claim 4, characterized in that, The process involves identifying state transition points of processing path segments based on several processing state transition thresholds and aligning the path segment boundaries according to these state transition points to obtain a path alignment set. For each processing state switching threshold, traverse the force establishment offset sequence, mark the point where the force establishment offset first exceeds the processing state switching threshold as the state switching start point, and mark the point where the offset falls back below the processing state switching threshold as the state switching end point. The processing path segment is re-divided based on the state transition start point and state transition end point to obtain several processing stage sequences. By performing a series of processing stages and adjusting the start and end boundaries of each path segment, a path alignment set is obtained.

6. The method for identifying abnormal states during CNC machine tool machining according to claim 5, characterized in that, The first abnormal state identification of the path alignment set, and the determination of the optimal processing state switching threshold based on the identification results, specifically involves: Extract the path segment feature vectors of each path alignment subset within the path alignment set, and calculate the first standard deviation sequence of the path segment feature vectors at each time point; Spectral analysis was performed on the first standard deviation sequence, and its dominant frequency component and corresponding harmonic energy distribution were extracted. An evaluation function is constructed based on the dominant frequency component and the corresponding harmonic energy distribution, and the consistency score of the path alignment set is performed based on the evaluation function. The processing state switching threshold corresponding to the path alignment subset with the lowest consistency score is taken as the optimal processing state switching threshold.

7. The method for identifying abnormal states during CNC machine tool machining according to claim 6, characterized in that, The method of defining the optimal processing interval within the processing path segment based on the optimal processing state switching threshold is as follows: Based on the optimal processing state switching threshold, backtrack to identify the first state switching point corresponding to it within the processing path segment; Using the first state switching point as the midpoint, expand the preset time window forward and backward to obtain the expanded time window; The extended time window is limited to the optimal processing interval.

8. The method for identifying abnormal states during CNC machine tool machining according to claim 7, characterized in that, The second abnormal state identification within the optimal processing range specifically involves: Extract the multidimensional feature vector subsequence within the optimal processing interval of each processing path segment, and average them according to time points to obtain the first reference feature trajectory; Calculate the point-by-point residuals between the multidimensional feature vector subsequence of each path segment and the first reference feature trajectory to obtain the residual sequence of each path segment; Wavelet transform is performed on the residual sequence of each path segment to obtain the time-frequency energy distribution map; Based on the time-frequency energy distribution map, calculate the energy change value of the residual sequence of each path segment within the preset frequency band; Based on the energy change value, calculate the time-frequency anomaly response vector of each processing path segment within the optimal processing interval; Compare the time-frequency anomaly response vector with the preset statistical distribution model of historical normal processing samples within the same interval and calculate the anomaly confidence level. When the anomaly confidence level exceeds the preset threshold, it is determined that the corresponding processing path segment has an abnormal processing state within the optimal processing range.

9. A CNC machine tool machining process abnormal state identification system, applied to the CNC machine tool machining process abnormal state identification method according to any one of claims 1-8, characterized in that, It includes a feature extraction module, a threshold construction module, a path alignment module, a threshold optimization module, and a state recognition module: The feature extraction module is used to acquire machining signal data of CNC machine tools and extract the force of machining path segments based on the machining signal data to establish a feature set; The threshold construction module is used to perform feature matching on the force establishment feature set, obtain the force establishment offset, and construct several processing state switching thresholds based on the force establishment offset. The path alignment module is used to identify the state switching points of the processing path segments based on several processing state switching thresholds and to align the boundaries of the path segments according to the state switching points to obtain a path alignment set. The threshold optimization module is used to perform the first abnormal state identification on the path alignment set and determine the optimal processing state switching threshold based on the identification results. The status recognition module is used to limit the optimal processing range within the processing path segment based on the optimal processing status switching threshold, and to perform a second abnormal status recognition within the optimal processing range.

Images