Data-driven full-automatic edge removal machine anomaly detection method and system

Through the data-driven fully automatic edge de-edge abnormal detection method, the timing current data is analyzed using SLIC and LOF algorithms, and the problems of artificial dependence and low detection accuracy in the existing technology are solved, achieving higher detection accuracy and robustness ensures product quality.

CN118521589BActive Publication Date: 2024-11-15JIANGSU LEIBO MICROELECTRONICS EQUIP CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410999537.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-07-24
Publication Date
2024-11-15
Estimated Expiration
2044-07-24

AI Technical Summary

Technical Problem

The existing fully automatic edge removal machine abnormal detection methods rely on manual experience, have low accuracy and are easily affected by human factors, and are difficult to detect in time when cutting blades wear, affecting product quality.

Method used

The data-driven fully automatic edge de-edge machine abnormal detection method is adopted. By obtaining the timing current data of the high-speed motor, clustering and abnormal detection are used using SLIC and LOF algorithms, the search window size is dynamically adjusted to improve detection accuracy.

Benefits of technology

It improves the accuracy and robustness of the abnormal detection of fully automatic edge removal machine, and can promptly detect abnormalities in cutting blades to ensure product quality.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118521589B_ABST
    Figure CN118521589B_ABST
Patent Text Reader

Abstract

The present invention relates to the field of semiconductor technology, and in particular to a data-driven fully automatic edge removal machine anomaly detection method and system, comprising: performing clustering operations on the time series current data sequences of high-speed motors of several fully automatic edge removal machines to obtain several seed points in each time series current data sequence; obtaining the degree of outliers of each seed point on each time series current data sequence to obtain the characteristic value of the update direction of each seed point; obtaining the adaptive weight of the search window of each seed point to obtain the adaptive search window size of each seed point on each time series current data sequence and update each seed point, clustering each time series current data sequence to obtain several update clusters; performing anomaly detection on all data in the update cluster to complete the anomaly detection of the fully automatic edge removal machine. The present invention improves the accuracy of anomaly detection of the fully automatic edge removal machine.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of semiconductor technology, and in particular to a data-driven full-automatic edge removal machine anomaly detection method and system. Background Art

[0002] Fully automatic deburring machines are often used to remove burrs from products such as wafers to ensure that the quality of the final product meets the standards. If an abnormality occurs in the deburring machine and is not detected and handled in time, it may cause defects in the product and affect the overall quality. Therefore, it is very important to perform abnormality detection on the fully automatic deburring machine.

[0003] The data-driven fully automatic de-margining machine anomaly detection method is a new type of de-margining machine anomaly detection method. It uses a large amount of data to perform de-margining machine anomaly detection and discovers anomalies through data analysis. Compared with the traditional de-margining machine anomaly detection method, this method has higher accuracy and robustness. The traditional de-margining machine anomaly detection method mainly relies on manual experience to judge anomalies. This method is easily affected by human factors and has low accuracy. The data-driven fully automatic de-margining machine anomaly detection method can discover anomalies through a large amount of data and improve the accuracy and robustness of detection.

[0004] During the operation of the fully automatic edge removal machine, the cutting blade will gradually wear out during use. If it is not replaced in time, it will cause loss of the wafer surface or uneven edge removal. The high-speed motor is used to drive the cutting blade to work, so it is usually necessary to measure the current signal of the high-speed motor during the operation of the fully automatic edge removal machine. The abnormal situation of the cutting blade during the operation of the fully automatic edge removal machine is detected by analyzing the abnormal data of the current signal. Among them, CBLOF is a commonly used abnormal method. The algorithm first uses the existing clustering method to cluster the data set and then performs subsequent operations. SLIC is a commonly used clustering method. The algorithm usually sets several seed points and then updates the position of the seed points. The search window size of each seed point is the same during the update process. Since the data in the present invention is time-series current data, and there will be a certain degree of noise and abnormal data therein. If the initial seed points are distributed in the abnormal data area, if the search window is too large, the seed points are very likely to be updated to a more normal data area during the update process, thereby affecting the subsequent clustering results and reducing the accuracy of abnormal detection of the fully automatic edge removal machine. Summary of the invention

[0005] The present invention provides a data-driven fully automatic edge removal machine anomaly detection method and system to solve the existing problems.

[0006] The data-driven fully automatic edge removal machine abnormality detection method and system of the present invention adopts the following technical solutions:

[0007] An embodiment of the present invention provides a data-driven fully automatic edge removal machine abnormality detection method, the method comprising the following steps:

[0008] Obtain the time series current data of the high-speed motors of several fully automatic edge removal machines;

[0009] Perform clustering operations on all time series current data sequences respectively to obtain a number of seed points in each time series current data sequence; segment each time series current data sequence to determine the working stage of each seed point in each time series current data sequence;

[0010] Perform anomaly detection on each time series current data sequence to obtain the degree of abnormality of each data in each time series current data sequence and a number of abnormal data screened out from all data; obtain the degree of outlier of each seed point on each time series current data sequence according to the degree of abnormality of the data at the working stage where each seed point on each time series current data sequence is located;

[0011] According to the outlier degree of each seed point on each time series current data sequence and the number of abnormal data around the seed point, the characteristic value of the update direction of each seed point on each time series current data sequence is obtained;

[0012] According to the outlier degree of each seed point on each time series current data sequence and the characteristic value of the update direction, an adaptive weight of the search window of each seed point on each time series current data sequence is obtained;

[0013] According to the number of all data on each time series current data sequence and the adaptive weight of the search window of each seed point, the adaptive search window size of each seed point on each time series current data sequence is obtained;

[0014] Each seed point is updated according to the adaptive search window size of each seed point on each time series current data sequence, and each time series current data sequence is clustered to obtain several updated clusters; anomaly detection is performed on all data in the updated cluster to complete anomaly detection of the fully automatic edge removal machine.

[0015] Furthermore, the step of segmenting each time series current data sequence and determining the working stage of each seed point in each time series current data sequence includes the following specific steps:

[0016] Using the least square method to perform curve fitting on each time series current data sequence, to obtain a fitting curve for each time series current data sequence;

[0017] Using Newton's method to calculate the fitting curve of each time series current data sequence, a number of extreme value points on the fitting curve of each time series current data sequence are obtained;

[0018] According to the extreme value points on the fitting curve of each time series current data sequence, a plurality of target fitting curve segments on the fitting curve of each time series current data sequence are obtained;

[0019] The working stage of each seed point in each time-series current data sequence is determined according to the target fitting curve segment on the fitting curve of each time-series current data sequence.

[0020] Furthermore, the step of obtaining a plurality of target fitting curve segments on the fitting curve of each time series current data sequence according to the extreme value points on the fitting curve of each time series current data sequence includes the following specific steps:

[0021] On the fitting curve of the mth time series current data sequence, the curve segment between two adjacent extreme value points is recorded as a sub-fitting curve segment; on any sub-fitting curve segment, the slope of the line connecting the first data point and the last data point is recorded as the slope of any sub-fitting curve segment;

[0022] According to the slopes of all sub-fitting curve segments on the fitting curve of the mth time-series current data sequence, the K-means algorithm is used to cluster all sub-fitting curve segments into two clusters, and the mean of the absolute values ​​of the slopes of all sub-fitting curve segments in each cluster is calculated. The cluster with the largest mean of the absolute values ​​of the slopes of all sub-fitting curve segments in the cluster is recorded as a feature cluster, and each sub-fitting curve segment in the feature cluster is recorded as a target fitting curve segment.

[0023] Furthermore, the step of determining the working stage of each seed point in each time series current data sequence according to the target fitting curve segment on the fitting curve of each time series current data sequence includes the following specific steps:

[0024] On the fitting curve of the mth time-series current data sequence, the first target fitting curve segment corresponds to the sequence segment in the time-series current data sequence, which is recorded as the first idling segment, and the curve segment between the first and second target fitting curve segments corresponds to the sequence segment in the time-series current data sequence, which is recorded as the first cutting segment; the third target fitting curve segment corresponds to the sequence segment in the time-series current data sequence, which is recorded as the first cooling segment;

[0025] The fourth target fitting curve segment corresponds to the sequence segment in the time series current data sequence, which is recorded as the second idling segment; the curve segment between the fourth and fifth target fitting curve segments corresponds to the sequence segment in the time series current data sequence, which is recorded as the second cutting segment; the fifth target fitting curve segment corresponds to the sequence segment in the time series current data sequence, which is recorded as the second cooling segment;

[0026] Similarly, the mth time series current data sequence is divided into three working stages: a plurality of idling stages, a cutting stage and a cooling stage.

[0027] Furthermore, the outlier degree of each seed point on each time series current data sequence is obtained according to the abnormal degree of data at the working stage where each seed point on each time series current data sequence is located. The specific calculation formula is:

[0028] ;

[0029] Among them, θ j(m) Indicates the degree of outlier of the jth seed point on the mth time series current data sequence; LOF j(m) Indicates the mean value of the abnormality of all data in the working stage where the jth seed point in the mth time series current data sequence is located; LOF j(p) represents the mean value of the abnormality degree of all data in the working stage where the jth seed point on the pth time series current data sequence is located except the mth time series current data sequence; M represents the number of time series current data sequences.

[0030] Furthermore, the method of obtaining the characteristic value of the update direction of each seed point on each time series current data sequence according to the outlier degree of each seed point on each time series current data sequence and the number of abnormal data around the seed point includes the following specific steps:

[0031] In each time series current data sequence, the left neighborhood of each seed point is formed by I data adjacent to the left of each seed point, and the right neighborhood of each seed point is formed by I data adjacent to the right of each seed point, where I represents a preset quantity threshold;

[0032] According to the outlier degree of each seed point on each time series current data sequence, the abnormal degree of data in the left neighborhood of each seed point on each time series current data sequence, and the number of abnormal data, the characteristic value of the left neighborhood of each seed point on each time series current data sequence is obtained;

[0033] According to the method of obtaining the characteristic value of the left neighborhood of each seed point on each time series current data sequence, the characteristic value of the right neighborhood of each seed point on each time series current data sequence can be calculated;

[0034] Among the eigenvalues ​​of the right neighborhood and the eigenvalues ​​of the left neighborhood of each seed point on each time series current data sequence, the side corresponding to the largest eigenvalue is recorded as the update direction of each seed point on each time series current data sequence, and the largest eigenvalue is recorded as the eigenvalue of the update direction of each seed point on each time series current data sequence.

[0035] Furthermore, according to the outlier degree of each seed point on each time series current data sequence, the abnormal degree of data in the left neighborhood of each seed point on each time series current data sequence, and the number of abnormal data, the characteristic value of the left neighborhood of each seed point on each time series current data sequence is obtained. The specific calculation formula is:

[0036] ;

[0037] Among them, ρ j(m) represents the characteristic value of the left neighborhood of the jth seed point on the mth time series current data sequence; B represents the number of data on the left neighborhood of the jth seed point on the mth time series current data sequence; θ j(m) represents the outlier degree of the jth seed point on the mth time series current data sequence; θ j(m,a) Indicates the abnormality of the ath data in the left neighborhood of the jth seed point on the mth time series current data sequence; represents the number of abnormal data in all data in the left neighborhood of the jth seed point on the mth time series current data sequence; exp( ) represents an exponential function with a natural constant as the base; || is the absolute value function.

[0038] Furthermore, according to the outlier degree of each seed point on each time series current data sequence and the eigenvalue of the update direction, the adaptive weight of the search window of each seed point on each time series current data sequence is obtained, and the specific calculation formula is:

[0039] ;

[0040] Among them, λ j(m) represents the adaptive weight of the search window of the jth seed point on the mth time series current data sequence; θ j(m) represents the outlier degree of the jth seed point on the mth time series current data sequence; σ j(m) The characteristic value representing the update direction of the jth seed point on the mth time series current data sequence; σ v(m) is the eigenvalue of the update direction of the vth seed point on the mth time-series current data sequence, which has the same update direction as the jth seed point; V represents the number of seed points with the same update direction as the jth seed point among all seed points on the mth time-series current data sequence; norm() is a linear normalization function; K is the number of seed points on the mth time-series current data sequence.

[0041] Furthermore, the adaptive search window size of each seed point on each time series current data sequence is obtained according to the number of all data on each time series current data sequence and the adaptive weight of the search window of each seed point. The specific calculation formula is:

[0042] ;

[0043] Among them, D j(m) represents the adaptive search window size of the jth seed point on the mth time series current data sequence; j(m) represents the adaptive weight of the search window of the jth seed point on the mth time series current data sequence; Q m represents the number of all data in the mth time series current data sequence; K is the number of seed points in the mth time series current data sequence; floor() represents the floor rounding function.

[0044] The present invention also proposes a data-driven fully automatic de-edging machine anomaly detection system, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program stored in the memory to implement the steps of the aforementioned data-driven fully automatic de-edging machine anomaly detection method.

[0045] The beneficial effects of the technical solution of the present invention are: obtaining the time-series current data sequences of the high-speed motors of several fully automatic de-edging machines, performing clustering operations on all the time-series current data sequences respectively, obtaining several seed points in each time-series current data sequence, segmenting each time-series current data sequence, and determining the working stage of each seed point in each time-series current data sequence, which is helpful to understand the changes and abnormalities in the working state of the fully automatic de-edging machine; performing abnormality detection on each time-series current data sequence to obtain the abnormality degree of each data in each time-series current data sequence and several abnormal data screened out from all the data, which is helpful to discover abnormal situations during the operation of the fully automatic de-edging machine and to handle and repair them in time; according to the abnormality degree of the data at the working stage of each seed point on each time-series current data sequence, the outlier degree of each seed point on each time-series current data sequence is obtained, and according to the outlier degree of each seed point on each time-series current data sequence and the number of abnormal data around the seed point, the outlier degree of each seed point on each time-series current data sequence is obtained. The characteristic value of the update direction of each seed point on each time series current data sequence is obtained according to the outlier degree of each seed point on each time series current data sequence and the characteristic value of the update direction, and the adaptive weight of the search window of each seed point on each time series current data sequence is obtained, which can more accurately describe the abnormal situation of the data and help to determine the importance and influence range of the abnormal data point; according to the number of all data on each time series current data sequence and the adaptive weight of the search window of each seed point, the adaptive search window size of each seed point on each time series current data sequence is obtained, and the search window size is dynamically adjusted, which helps to improve the accuracy and efficiency of abnormal detection; according to the adaptive search window size of each seed point on each time series current data sequence, each seed point is updated, and each time series current data sequence is clustered to obtain a number of updated clusters; all data in the updated cluster are detected for abnormality, and the abnormal detection of the fully automatic edge removal machine is completed. The present invention improves the accuracy of abnormal detection of the fully automatic edge removal machine. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0047] Figure 1 The present invention is a flowchart of the steps of the data-driven fully automatic edge removal machine anomaly detection method. DETAILED DESCRIPTION

[0048] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the specific implementation method, structure, features and effects of the data-driven fully automatic edge removal machine abnormality detection method and system proposed by the present invention are described in detail below in combination with the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures or characteristics in one or more embodiments may be combined in any suitable form.

[0049] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0050] The specific scheme of the data-driven full-automatic edge removal machine abnormality detection method and system provided by the present invention is described in detail below with reference to the accompanying drawings.

[0051] See also Figure 1 , which shows a flowchart of a method for detecting abnormalities of a fully automatic edge removal machine based on data drive according to an embodiment of the present invention, the method comprising the following steps:

[0052] Step S001: Obtaining the time series current data sequences of the high-speed motors of several fully automatic edge removing machines.

[0053] First, the time series current data of the high-speed motor of the fully automatic edge trimming machine in the database is obtained, where the current data is collected by the current sensor installed at the high-speed motor. Since the change of the current data of the high-speed motor of a fully automatic edge trimming machine may be occasional, it will affect the subsequent abnormal detection results.

[0054] Therefore, the present invention obtains the time-series current data sequence of the high-speed motors of several fully automatic edge removing machines with the same working time, wherein the various parameters and data collection frequency during the working process of each fully automatic edge removing machine are the same.

[0055] It should be noted that the fully automatic edge removal machine is often used to remove the burrs of wafers. The wafer is one of the key materials in the semiconductor manufacturing process. It is a very thin round silicon wafer that is usually used as the basis for manufacturing integrated circuits.

[0056] Step S002: performing clustering operations on all time series current data sequences respectively to obtain a number of seed points in each time series current data sequence; segmenting each time series current data sequence to determine the working stage of each seed point in each time series current data sequence.

[0057] Among the time series current data sequences of all fully automatic edge removal machines, taking the mth time series current data sequence as an example, the SLIC algorithm is used to perform a clustering operation on the mth time series current data sequence to obtain a number of seed points in the mth time series current data sequence, wherein the SLIC algorithm is a well-known technology, and the specific method is not introduced here.

[0058] According to the above method, a number of seed points in each time series current data sequence are obtained.

[0059] It should be noted that: in this embodiment, the number of seed points set for the SLIC algorithm is 50, and all seed points are evenly distributed in the time series current data sequence.

[0060] For each seed point on the mth time series current data sequence, since the blade of the fully automatic edge trimming machine will experience three working stages during the wafer edge trimming process, namely, the idling stage, the cutting stage, and the cooling stage, these three working stages constitute a working cycle, and the acquired time series current data sequence of the fully automatic edge trimming machine will experience multiple working cycles, the working stage of each seed point is first determined, and then the degree of abnormality of each seed point in the working stage is calculated to determine the degree of outlier of each seed point.

[0061] Taking the mth time series current data sequence as an example, the least squares method is used to perform curve fitting on the mth time series current data sequence to obtain a fitting curve of the mth time series current data sequence, wherein the least squares method is a well-known technology and the specific method is not introduced here.

[0062] According to the above method, the fitting curves of all time series current data sequences are obtained.

[0063] During the idling stage, the current of the blade will quickly increase to the predetermined current and fluctuate steadily. After entering the cutting stage, the current will fluctuate unstably as the cutting progresses. During the cooling stage, as the cutting stops, the current will gradually decrease until the next working cycle. Therefore, the timing current data sequence of the fully automatic de-edging machine has a periodic fluctuation characteristic.

[0064] The fitting curves of all time-series current data sequences are obtained by the least squares method. For the fitting curve of the mth time-series current data sequence, all extreme points on the fitting curve are calculated using the Newton method, and the first and last data points on the fitting curve are set as extreme points. Among them, the least squares method and the Newton method are well-known technologies, and the specific methods are not introduced here.

[0065] On the fitting curve of the mth time series current data sequence, the curve segment between two adjacent extreme value points is recorded as a sub-fitting curve segment. On any sub-fitting curve segment, the slope of the line connecting the first data point and the last data point is recorded as the slope of any sub-fitting curve segment.

[0066] According to the slope of the sub-fitting curve segment, the K-means algorithm is used to cluster all the sub-fitting curve segments into two clusters, and the mean of the absolute values ​​of the slopes of all the sub-fitting curve segments in each cluster is calculated. The cluster with the largest mean is recorded as the feature cluster, and each sub-fitting curve segment in the feature cluster is recorded as the target fitting curve segment. Among them, the K-means algorithm is a well-known technology, and the specific method is not introduced here.

[0067] On the fitting curve of the mth time-series current data sequence, the first target fitting curve segment corresponds to the sequence segment in the time-series current data sequence, which is recorded as the first idling segment, and the curve segment between the first and second target fitting curve segments corresponds to the sequence segment in the time-series current data sequence, which is recorded as the first cutting segment; the third target fitting curve segment corresponds to the sequence segment in the time-series current data sequence, which is recorded as the first cooling segment;

[0068] The fourth target fitting curve segment corresponds to the sequence segment in the time series current data sequence, which is recorded as the second idling segment; the curve segment between the fourth and fifth target fitting curve segments corresponds to the sequence segment in the time series current data sequence, which is recorded as the second cutting segment; the fifth target fitting curve segment corresponds to the sequence segment in the time series current data sequence, which is recorded as the second cooling segment;

[0069] Similarly, the mth time series current data sequence is divided into three working stages: a plurality of idling stages, a cutting stage and a cooling stage.

[0070] According to the above method, each timing current sequence is divided into three working stages: idling stage, cutting stage and cooling stage.

[0071] Step S003: Perform anomaly detection on each time-series current data sequence to obtain the degree of abnormality of each data in each time-series current data sequence and a number of abnormal data screened out from all the data; obtain the degree of outlier of each seed point on each time-series current data sequence according to the degree of abnormality of the data in the working stage where each seed point on each time-series current data sequence is located.

[0072] Taking the mth time series current data sequence as an example, the LOF algorithm is used to perform anomaly detection on the mth time series current data sequence to obtain the abnormality degree of each data in the mth time series current data sequence and several abnormal data screened out from all the data. Among them, the LOF algorithm is a well-known technology and the specific method is not introduced here.

[0073] For the jth seed point on the mth time-series current data sequence, the degree of outlier of the jth seed point is determined according to the degree of abnormality of the working stage in which the jth seed point on the mth time-series current data sequence is located and the degree of abnormality of the corresponding working stages on other time-series current data sequences.

[0074] The calculation formula for the outlier degree of the jth seed point on the mth time series current data sequence is:

[0075] ;

[0076] Among them, θ j(m) Indicates the degree of outlier of the jth seed point on the mth time series current data sequence; LOF j(m) Indicates the mean value of the abnormality of all data in the working stage where the jth seed point in the mth time series current data sequence is located; LOF j(p) represents the mean value of the abnormality degree of all data in the working stage where the jth seed point on the pth time series current data sequence is located except the mth time series current data sequence; M represents the number of time series current data sequences.

[0077] Through the above method, the outlier degree of each seed point on each time series current data sequence is obtained.

[0078] Step S004: according to the outlier degree of each seed point on each time series current data sequence and the number of abnormal data around the seed point, obtain the characteristic value of the update direction of each seed point on each time series current data sequence.

[0079] In the mth time-series current data sequence, with the jth seed point as the center, the I data adjacent to the left side of the jth seed point constitute the left neighborhood of the jth seed point, and the I data adjacent to the right side of the jth seed point constitute the right neighborhood of the jth seed point.

[0080] The preset quantity threshold value I=10 in this embodiment is used as an example for description. Other values ​​may be set in other implementation modes, which are not limited in this embodiment.

[0081] It should be noted that: if the number of data adjacent to the left and right sides of the j-th seed point is less than I, the existing data constitute the left neighborhood and the right neighborhood; in the m-th time series current data sequence, the first and last data are not used as seed points.

[0082] Through the above steps, the abnormality degree of each data in the mth time series current data sequence is obtained, and a number of abnormal data are screened out from all the data in the mth time series current data sequence.

[0083] Calculate the degree of deviation between each seed point and the data in its left and right neighborhoods. When the seed point position is subsequently updated, the higher the degree of deviation of the data in the left neighborhood, the more the update direction will be biased to the left during the subsequent update process; otherwise, the update direction will be biased to the right.

[0084] Then the calculation formula for the eigenvalue of the left neighborhood of the jth seed point on the mth time series current data sequence is:

[0085] ;

[0086] Among them, ρ j(m) represents the characteristic value of the left neighborhood of the jth seed point on the mth time series current data sequence; B represents the number of data on the left neighborhood of the jth seed point on the mth time series current data sequence; θ j(m) represents the outlier degree of the jth seed point on the mth time series current data sequence; θ j(m,a) Indicates the abnormality of the ath data in the left neighborhood of the jth seed point on the mth time series current data sequence; represents the number of abnormal data in all data in the left neighborhood of the jth seed point on the mth time series current data sequence; exp( ) represents an exponential function with a natural constant as the base; || is the absolute value function.

[0087] |θ j(m)- θ j(m,a) The smaller the value of | is, the smaller the difference in the degree of outliers between the jth seed point on the mth time series current data sequence and the ath data in its left neighborhood is, and the greater the deviation of the ath data in the left neighborhood of the jth seed point on the mth time series current data sequence is. The larger the value of is, the greater the deviation of the left neighborhood of the jth seed point on the mth time series current data sequence is. The smaller the value of , the more similar the data of the left neighborhood of the jth seed point on the mth time series current data sequence is, and the larger the eigenvalue of the left neighborhood of the jth seed point on the mth time series current data sequence is.

[0088] According to the method of obtaining the characteristic value of the left neighborhood of the jth seed point on the mth time series current data sequence, the characteristic value of the right neighborhood of the jth seed point on the mth time series current data sequence can be calculated, which is recorded as .

[0089] Among the eigenvalues ​​of the neighborhoods on both sides of the j-th seed point on the m-th time series current data sequence, the neighborhood on the side where the maximum eigenvalue is located is selected as the update direction of the j-th seed point;

[0090] ;

[0091] Among them, σ j(m) The characteristic value representing the update direction of the jth seed point on the mth time series current data sequence; ρ j(m) represents the characteristic value of the jth seed point in the left neighborhood of the mth time series current data sequence; represents the characteristic value of the jth seed point on the mth time series current data sequence in the right neighborhood; max( ) is the maximum value function.

[0092] Through the above method, the characteristic value of the update direction of each seed point on each time series current data sequence is obtained.

[0093] Step S005: Obtaining an adaptive weight of a search window for each seed point on each time series current data sequence according to the outlier degree of each seed point on each time series current data sequence and the eigenvalue of the update direction.

[0094] After obtaining the outlier degree of each seed point on each time series current data sequence and the characteristic value of the update direction according to the above steps, set the size of the adaptive search window in the update direction according to the outlier degree of each seed point. The larger the outlier degree of the seed point, the smaller the corresponding adaptive window, and include as little data as possible to avoid updating the seed point to data with a smaller outlier degree, thereby affecting the subsequent clustering results;

[0095] ;

[0096] Among them, λ j(m) represents the adaptive weight of the search window of the jth seed point on the mth time series current data sequence; θ j(m) represents the outlier degree of the jth seed point on the mth time series current data sequence; σ j(m) The characteristic value representing the update direction of the jth seed point on the mth time series current data sequence; σ v(m) is the characteristic value of the update direction of the vth seed point on the mth time series current data sequence, which has the same update direction as the jth seed point; V represents the number of seed points on the mth time series current data sequence, which have the same update direction as the jth seed point; norm() is a linear normalization function, which normalizes the data value to the interval [0,1]; K is the number of seed points on the mth time series current data sequence;

[0097] The larger the value of is, the greater the degree of outlier of the jth seed point in the mth time series current data sequence among all seed points. The smaller the value is, the smaller the characteristic value of the update direction of the seed point in the same update direction as the jth seed point on the mth time-series current data sequence is, the smaller the corresponding deviation in the update direction is, and the smaller the adaptive weight of the search window of the jth seed point on the mth time-series current data sequence is.

[0098] By the above method, the adaptive weight of the search window of each seed point on each time series current data sequence is obtained.

[0099] Step S006: obtaining the adaptive search window size of each seed point on each time series current data sequence according to the number of all data on each time series current data sequence and the adaptive weight of the search window of each seed point;

[0100] ;

[0101] Among them, D j(m) represents the adaptive search window size of the jth seed point on the mth time series current data sequence; j(m) represents the adaptive weight of the search window of the jth seed point on the mth time series current data sequence; Q m represents the number of all data in the mth time series current data sequence; K is the number of seed points in the mth time series current data sequence; Indicates the maximum empirical value of the adaptive search window; Represents the minimum empirical value of the adaptive search window; floor() represents the rounding down function.

[0102] According to the above method, the adaptive search window size of each seed point on each time series current data sequence can be calculated.

[0103] Step S007: Update each seed point according to the adaptive search window size of each seed point on each time series current data sequence, cluster each time series current data sequence to obtain several updated clusters; perform anomaly detection on all data in the updated cluster to complete the anomaly detection of the fully automatic edge removal machine.

[0104] Each seed point is updated according to the adaptive search window size of each seed point on each time series current data sequence, and each time series current data sequence is clustered using the DBSCAN clustering algorithm to obtain several update clusters corresponding to each time series current data sequence; the CFLOF algorithm is used to perform anomaly detection on all data in any update cluster to obtain the final abnormal data in any update cluster, wherein the DBSCAN clustering algorithm and the CFLOF clustering algorithm are well-known technologies, and the specific methods are not introduced here.

[0105] Thus, a number of final abnormal data corresponding to each time series current data sequence are obtained, thereby completing the abnormality detection of the fully automatic edge removal machine.

[0106] So far, the present invention is completed.

[0107] To sum up, in an embodiment of the present invention, the present invention also provides a data-driven fully-automatic de-edging machine anomaly detection system, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program stored in the memory to implement the steps of the aforementioned data-driven fully-automatic de-edging machine anomaly detection method.

[0108] 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 principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A data-driven method for detecting abnormalities in a fully automatic edge removal machine, characterized in that: The method comprises the following steps: Obtain the time series current data of the high-speed motors of several fully automatic edge removal machines; Perform clustering operations on all time series current data sequences respectively to obtain a number of seed points in each time series current data sequence; segment each time series current data sequence to determine the working stage of each seed point in each time series current data sequence; Perform anomaly detection on each time series current data sequence to obtain the degree of abnormality of each data in each time series current data sequence and a number of abnormal data screened out from all data; obtain the degree of outlier of each seed point on each time series current data sequence according to the degree of abnormality of the data at the working stage where each seed point on each time series current data sequence is located; According to the outlier degree of each seed point on each time series current data sequence and the number of abnormal data around the seed point, the characteristic value of the update direction of each seed point on each time series current data sequence is obtained; According to the outlier degree of each seed point on each time series current data sequence and the characteristic value of the update direction, an adaptive weight of the search window of each seed point on each time series current data sequence is obtained; According to the number of all data on each time series current data sequence and the adaptive weight of the search window of each seed point, the adaptive search window size of each seed point on each time series current data sequence is obtained; Each seed point is updated according to the adaptive search window size of each seed point on each time series current data sequence, and each time series current data sequence is clustered to obtain several updated clusters; all data in the updated cluster are detected for anomalies to complete the anomaly detection of the fully automatic edge removal machine; The specific steps of segmenting each time series current data sequence and determining the working stage of each seed point in each time series current data sequence are as follows: Using the least square method to perform curve fitting on each time series current data sequence, to obtain a fitting curve for each time series current data sequence; Using Newton's method to calculate the fitting curve of each time series current data sequence, a number of extreme value points on the fitting curve of each time series current data sequence are obtained; According to the extreme value points on the fitting curve of each time series current data sequence, a plurality of target fitting curve segments on the fitting curve of each time series current data sequence are obtained; Determine the working stage of each seed point in each time series current data sequence according to the target fitting curve segment on the fitting curve of each time series current data sequence; The specific steps of obtaining a plurality of target fitting curve segments on the fitting curve of each time series current data sequence according to the extreme value points on the fitting curve of each time series current data sequence are as follows: On the fitting curve of the mth time series current data sequence, the curve segment between two adjacent extreme value points is recorded as a sub-fitting curve segment; on any sub-fitting curve segment, the slope of the line connecting the first data point and the last data point is recorded as the slope of any sub-fitting curve segment; According to the slopes of all sub-fitting curve segments on the fitting curve of the mth time series current data sequence, all sub-fitting curve segments are clustered into two clusters using the K-means algorithm, the mean of the absolute values ​​of the slopes of all sub-fitting curve segments in each cluster is calculated, and the cluster with the largest mean of the absolute values ​​of the slopes of all sub-fitting curve segments in the cluster is recorded as a feature cluster, and each sub-fitting curve segment in the feature cluster is recorded as a target fitting curve segment; The specific steps of determining the working stage of each seed point in each time series current data sequence according to the target fitting curve segment on the fitting curve of each time series current data sequence are as follows: On the fitting curve of the mth time-series current data sequence, the first target fitting curve segment corresponds to the sequence segment in the time-series current data sequence, which is recorded as the first idling segment, and the curve segment between the first and second target fitting curve segments corresponds to the sequence segment in the time-series current data sequence, which is recorded as the first cutting segment; the third target fitting curve segment corresponds to the sequence segment in the time-series current data sequence, which is recorded as the first cooling segment; The fourth target fitting curve segment corresponds to the sequence segment in the time series current data sequence, which is recorded as the second idling segment; the curve segment between the fourth and fifth target fitting curve segments corresponds to the sequence segment in the time series current data sequence, which is recorded as the second cutting segment; the fifth target fitting curve segment corresponds to the sequence segment in the time series current data sequence, which is recorded as the second cooling segment; By analogy, the mth time series current data sequence is divided into three working stages: a plurality of idling stages, a cutting stage and a cooling stage; According to the abnormal degree of data at the working stage where each seed point on each time series current data sequence is located, the outlier degree of each seed point on each time series current data sequence is obtained, and the specific calculation formula is: ; Among them, θ j(m) Indicates the outlier degree of the jth seed point on the mth time series current data sequence; LOF j(m) Indicates the mean value of the abnormality of all data in the working stage where the jth seed point in the mth time series current data sequence is located; LOF j(p) represents the mean value of the abnormality of all data at the working stage where the jth seed point on the pth time series current data sequence is located except the mth time series current data sequence; M represents the number of time series current data sequences; The specific steps of obtaining the characteristic value of the update direction of each seed point on each time series current data sequence according to the outlier degree of each seed point on each time series current data sequence and the number of abnormal data around the seed point include the following: In each time series current data sequence, the left neighborhood of each seed point is formed by I data adjacent to the left of each seed point, and the right neighborhood of each seed point is formed by I data adjacent to the right of each seed point, where I represents a preset quantity threshold; According to the outlier degree of each seed point on each time series current data sequence, the abnormal degree of data in the left neighborhood of each seed point on each time series current data sequence, and the number of abnormal data, the characteristic value of the left neighborhood of each seed point on each time series current data sequence is obtained; According to the method of obtaining the characteristic value of the left neighborhood of each seed point on each time series current data sequence, the characteristic value of the right neighborhood of each seed point on each time series current data sequence can be calculated; Among the eigenvalues ​​of the right neighborhood and the eigenvalues ​​of the left neighborhood of each seed point on each time series current data sequence, the side corresponding to the largest eigenvalue is recorded as the update direction of each seed point on each time series current data sequence, and the largest eigenvalue is recorded as the eigenvalue of the update direction of each seed point on each time series current data sequence; Anomaly detection is performed on each time series current data sequence to obtain the abnormality degree of each data in each time series current data sequence, including: For the mth time series current data sequence, the LOF algorithm is used to perform anomaly detection on the mth time series current data sequence to obtain the abnormality degree of each data in the mth time series current data sequence and a number of abnormal data screened out from all the data.

2. The data-driven fully automatic edge removal machine abnormality detection method according to claim 1 is characterized in that: According to the outlier degree of each seed point on each time series current data sequence, the abnormal degree of data in the left neighborhood of each seed point on each time series current data sequence, and the number of abnormal data, the characteristic value of the left neighborhood of each seed point on each time series current data sequence is obtained. The specific calculation formula is: ; Among them, ρ j(m) represents the characteristic value of the left neighborhood of the jth seed point on the mth time series current data sequence; B represents the number of data on the left neighborhood of the jth seed point on the mth time series current data sequence; θ j(m) represents the outlier degree of the jth seed point on the mth time series current data sequence; θ j(m,a) Indicates the abnormality of the ath data in the left neighborhood of the jth seed point on the mth time series current data sequence; represents the number of abnormal data in all data in the left neighborhood of the jth seed point on the mth time series current data sequence; exp( ) represents an exponential function with a natural constant as the base; || is the absolute value function.

3. The data-driven fully automatic edge removal machine abnormality detection method according to claim 1 is characterized in that: According to the outlier degree of each seed point on each time series current data sequence and the characteristic value of the update direction, the adaptive weight of the search window of each seed point on each time series current data sequence is obtained. The specific calculation formula is: ; Among them, λ j(m) represents the adaptive weight of the search window of the jth seed point on the mth time series current data sequence; θ j(m) represents the outlier degree of the jth seed point on the mth time series current data sequence; σ j(m) The characteristic value representing the update direction of the jth seed point on the mth time series current data sequence; σ v(m) is the eigenvalue of the update direction of the vth seed point on the mth time-series current data sequence, which has the same update direction as the jth seed point; V represents the number of seed points with the same update direction as the jth seed point among all seed points on the mth time-series current data sequence; norm() is a linear normalization function; K is the number of seed points on the mth time-series current data sequence.

4. The data-driven fully automatic edge removal machine abnormality detection method according to claim 1 is characterized in that: According to the number of all data on each time series current data sequence and the adaptive weight of the search window of each seed point, the adaptive search window size of each seed point on each time series current data sequence is obtained. The specific calculation formula is: ; Among them, D j(m) represents the adaptive search window size of the jth seed point on the mth time series current data sequence; j(m) represents the adaptive weight of the search window of the jth seed point on the mth time series current data sequence; Q m represents the number of all data in the mth time series current data sequence; K is the number of seed points in the mth time series current data sequence; floor() represents the floor rounding function.

5. A data-driven fully automatic edge removal machine anomaly detection system, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the computer program is executed by a processor, the steps of the data-driven fully automatic edge removal machine anomaly detection method as described in any one of claims 1 to 4 are implemented.