CN111027606B - Multi-mode time series anomaly detection method, storage medium and equipment - Google Patents

Multi-mode time series anomaly detection method, storage medium and equipment Download PDF

Info

Publication number
CN111027606B
CN111027606B CN201911199574.5A CN201911199574A CN111027606B CN 111027606 B CN111027606 B CN 111027606B CN 201911199574 A CN201911199574 A CN 201911199574A CN 111027606 B CN111027606 B CN 111027606B
Authority
CN
China
Prior art keywords
clustering
sub
time sequence
mode
pattern
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN201911199574.5A
Other languages
Chinese (zh)
Other versions
CN111027606A (en
Inventor
段江永
郭丽丽
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Technology and Engineering Center for Space Utilization of CAS
Original Assignee
Technology and Engineering Center for Space Utilization of CAS
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Technology and Engineering Center for Space Utilization of CAS filed Critical Technology and Engineering Center for Space Utilization of CAS
Priority to CN201911199574.5A priority Critical patent/CN111027606B/en
Publication of CN111027606A publication Critical patent/CN111027606A/en
Application granted granted Critical
Publication of CN111027606B publication Critical patent/CN111027606B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/23—Clustering techniques
    • G06F18/232—Non-hierarchical techniques
    • G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
    • G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/243—Classification techniques relating to the number of classes
    • G06F18/2433—Single-class perspective, e.g. one-against-all classification; Novelty detection; Outlier detection

Landscapes

  • Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Artificial Intelligence (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Probability & Statistics with Applications (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

The invention relates to a multimode time sequence anomaly detection method which comprises the steps of constructing a multimode time sequence clustering model based on time sequence segmentation and a time sequence clustering mode according to a historical time sequence; iteratively optimizing a multi-mode sub-time sequence clustering model based on a sub-time sequence clustering mode initializing, splitting, combining and removing method, and solving to obtain each sub-time sequence clustering mode; and extracting each sub-time sequence from the time sequence to be detected based on the sub-time sequence clustering mode to perform anomaly detection. The invention can adaptively extract various sub-time sequence modes with different lengths from the time sequence, avoids the condition of insufficient mode expression or interference of mode expression caused by extracting the sub-time sequence with fixed length at present, and has higher accuracy of time sequence abnormality detection on the basis. The invention also relates to a storage medium and a device.

Description

Multi-mode time series anomaly detection method, storage medium and equipment
Technical Field
The invention relates to the technical field of data analysis, in particular to a multi-mode time series abnormity detection method, a storage medium and equipment.
Background
The application range of time series abnormality detection is very wide, for example, time series data acquired by a sensor in industry is subjected to abnormality detection so as to evaluate the performance of equipment, heartbeat time series is subjected to abnormality detection in the medical field to judge the disease condition of a patient, and stock market time series is subjected to abnormality detection in the economic field to determine the stock market condition.
One class of time series anomaly detection methods employs a prediction-based method that determines whether a predicted value is anomalous or not from the difference between the predicted value and the true value, and the prediction models typically employ differential auto-regressive moving average models (ARIMA models), Kalman filtering, online support vector machines, and the like. For the method, the model is established on the basis of certain assumptions, for example, the Kalman filtering method needs the time sequence to meet the random process assumption, and the ARIMA model needs the time sequence to meet the difference stationary process assumption, so that the application range of the ARIMA model is limited.
Another type of time series anomaly detection method employs a clustering-based method, defining sub-time series that do not belong to each cluster as anomalies. Some methods firstly extract fixed-length sub-time sequences, then perform clustering by using a traditional clustering method, such as a k-means, k-medoids, one-class SVM method and the like, and perform anomaly detection based on clustering, but the extracted sub-time sequences cannot reflect inherent patterns in the original time sequences.
Therefore, how to extract various sub-sequence patterns with unknown lengths from the time sequence, so as to perform anomaly detection based on the sub-sequence patterns, is a problem to be solved at present.
Disclosure of Invention
The invention provides a multimode time series abnormity detection method, a storage medium and equipment aiming at the defects of the prior art.
The technical scheme for solving the technical problems is as follows:
a multi-mode time series anomaly detection method comprises the following steps:
s1, constructing a multi-mode sub-time sequence clustering model target function based on time sequence segmentation and sub-time sequence clustering modes according to the historical time sequence;
s2, iteratively optimizing the multi-mode sub-time sequence clustering model based on the method of sub-time sequence clustering mode initialization, splitting, combining and removing, and solving to obtain each sub-time sequence clustering mode;
s3, extracting each time subsequence from the time sequences to be detected, and determining whether the time subsequences are abnormal or not according to the distance between each time subsequence and the center of each clustering pattern in the time subsequences clustering model.
The invention has the beneficial effects that: the method can adaptively extract various subsequence modes with different lengths from a time sequence, avoids the condition of insufficient pattern expression or interference of pattern expression caused by extracting the subsequence with fixed length at present, and has higher time sequence abnormity detection precision on the basis; in addition, the invention can perform sub-time sequence clustering on the basis of only assuming the length range of the sub-time sequence without adding other model assumptions to the time sequence, so that the application range of the anomaly detection performed on the basis of the self-adaptive sub-time sequence clustering is wider.
On the basis of the technical scheme, the invention can be further improved as follows.
Further, the step S1 specifically includes:
time series X of length T1:TLength range l based on a predetermined sub-time series patternminAnd lmaxAnd the number k of the sub-time series modes, the multi-mode sub-time series clustering model is realized by minimizing an objective function related to the sub-time series clustering mode C and the sub-time series segmentation set Z under the following conditions:
Figure BDA0002295533850000021
s.t.∪x＝X1:T，lmin≤|μi|≤lmax，
wherein X is the historical sub-time series, k is the number of clustering patterns, CiFor the ith cluster, μiFor the ith cluster CiI μmiL is muiLength of (l)minAnd lmaxFor each muiThe time sequence length constraint must be satisfied, and ═ X1:TAnd (3) representing a time series coverage constraint that each point in X must be at least in a certain class of sub-time series, and Z is the sub-time series segmentation set, namely the sub-time series extracted from the time series X and corresponding to each sub-time series clustering pattern C.
The beneficial effect of adopting the above further scheme is: the length of each cluster is unknown and variable, so that the model is different from other methods, and the length of a cluster mode is not required to be given in advance; meanwhile, in the sub-time sequence clustering, accurate segmentation Z is very important for clustering, and the model can realize automatic extraction of various sub-time sequence patterns by simultaneously optimizing a sub-time sequence clustering pattern C and a sub-time sequence segmentation Z; and when extracting the sub-time sequence, the time sequence is required to cover and restrict, so that the problem of meaningless clustering caused by repeated extraction of adjacent fields when extracting the sub-time sequence through a sliding window by a plurality of methods is avoided.
Further, the step S2 specifically includes:
s21, initializing a sub-time sequence clustering mode set based on the sub-time sequence clustering model and an inner sleeve optimization algorithm according to the historical time sequence, and specifically comprising the following steps:
setting a length value of each clustering pattern in the historical sub-time sequence clustering pattern set, randomly initializing each clustering pattern, and sequentially setting the length value to be a constraint value l from the minimum preset lengthminTo said maximum predetermined length constraint value lmaxEach value within the range;
obtaining the sub-time sequence clustering mode and the time sequence segmentation set corresponding to each length value through the inner sleeve optimization algorithm;
aggregating all of the cluster patterns and the segmented set as an initial solution to minimize an objective function.
The step S21 has the advantages that: by traversing each fixed length in the sub-sequence length range, clustering patterns with different lengths can be obtained, and the clustering patterns with different lengths can be used as better initial solutions of the sub-time sequence clustering model.
S22, based on the current sub-time sequence clustering mode set, according to the sub-time sequence clustering model and the inner sleeve optimization algorithm, splitting the sub-time sequence clustering mode and optimizing the minimized objective function, the specific steps are as follows:
giving a current sub-time sequence clustering pattern set, and calculating the variance value of each clustering pattern in the sub-time sequence clustering pattern set;
taking the clustering pattern with the maximum variance value as a clustering pattern to be split;
splitting the clustering pattern to be split at each position meeting the minimum preset length constraint value and the maximum preset length constraint value, and calculating the value of the minimized objective function based on the inner sleeve optimization algorithm to obtain the minimized objective function value corresponding to each position;
splitting the clustering mode to be split at the position corresponding to the minimum value in the minimized objective function value to obtain a result after splitting of a time sequence clustering mode set;
and optimizing the minimized objective function according to the sub-time sequence clustering mode set and the inner sleeve optimization algorithm.
The beneficial effects of adopting the step S22 are as follows: in the cluster pattern set, some clusters may be formed by combining a plurality of patterns, so that we propose that a sub-time series pattern splitting algorithm can split a plurality of corresponding patterns from the cluster patterns.
S23, based on the current sub-time sequence clustering mode set, according to the sub-time sequence clustering model and the inner sleeve optimization algorithm, combining sub-time sequence clustering modes and optimizing the minimized objective function, and according to the optimized clustering result, comparing the clustering result with the clustering result before the execution of S22, judging the next optimization mode of sub-time sequence clustering, and the specific steps are as follows:
giving a current sub-time sequence clustering pattern set, and calculating adjacency values among clustering patterns in the historical sub-time sequence clustering pattern set, wherein the adjacency values are obtained by the following method:
clustering pattern CjAnd CjHas a neighborhood value of P (C)i|Cj)＝max{P1(Ci|Cj),P2(Ci|Cj) And (c) the step of (c) in which,
Figure BDA0002295533850000041
as a cluster pattern CiThe cluster pattern of the immediately adjacent position is CjThe probability of (a) of (b) being,
Figure BDA0002295533850000051
as a cluster pattern CjThe clustering pattern of the immediately preceding neighboring position is CiProbability of (n)iIndicates belongings to clustering Pattern C in ZiNumber of sub-time series of (2), nijRepresenting belongings to clustering Pattern C in ZiIs followed by a sub-time sequence belonging to cluster pattern CjThe number of sub-time series of (2).
Taking the clustering pattern pair with the maximum adjacency value as a clustering pattern pair to be combined, and if P (C)i|Cj) Source of maximum value ofIn P1(Ci|Cj) Then combine CiAnd CjTo replace original CjIf P (C)i|Cj) Is derived from P2(Ci|Cj) Then combine CiAnd CjTo replace original Ci。
Combining the clustering modes to be combined at each position meeting the minimum preset length constraint value and the maximum preset length constraint value, and calculating a minimized objective function value corresponding to each position based on the inner sleeve optimization algorithm;
combining the clustering modes to be combined with the positions corresponding to the minimum values in the minimized objective function values to obtain the result after the clustering mode sets of the sub-time sequences are combined;
and optimizing the minimized objective function according to the sub-time sequence clustering mode set and the inner sleeve optimization algorithm.
And comparing the clustering result after the clustering mode combination optimization with the clustering result before the S22 execution, and judging the next optimization mode:
comparing the current number of the clustering modes obtained after the clustering mode combination optimization in the iteration step with the number of the clustering modes obtained before the S22 is executed;
if the number of the current clustering modes obtained after the clustering mode combination optimization is greater than the number of the clustering modes obtained before S22 is executed, executing S24;
otherwise, judging whether the current minimized objective function value obtained after the cluster mode combination optimization in the iteration step is larger than the minimized objective function value obtained before the execution of S22;
if not, go to S22;
if yes, judging whether the number of the clustering modes obtained after the clustering mode combination optimization in the iteration step reaches the preset target number of the clustering modes;
if not, go to step S24;
if so, the sub-time sequence clustering mode obtained after the clustering mode combination optimization in the iteration step is the final sub-time sequence clustering mode to be solved, and the multi-mode sub-time sequence clustering model optimization is finished;
the step S23 has the advantages that: in the cluster pattern set, some cluster patterns may only contain a part of the original sub-time series pattern, so we propose that the sub-time series cluster pattern combines multiple cluster patterns to obtain a complete sub-time series pattern.
S24, based on the current sub-time sequence clustering mode set, according to the minimized objective function and the inner sleeve optimization algorithm, removing the clustering mode corresponding to the minimum value of the minimized objective function, and optimizing the minimum objective function, the specific steps are as follows:
giving a current sub-time sequence clustering mode set, traversing all modes in the current sub-time sequence clustering mode set based on the minimized objective function and an internal optimization algorithm, and calculating the value of the minimized objective function when each clustering mode in the sub-time sequence clustering mode set is deleted in sequence;
and removing the deleted clustering mode corresponding to the minimum value in the values of the minimized objective function, and optimizing the minimized objective function based on the inner sleeve optimization algorithm.
The beneficial effects of adopting the step S24 are as follows: because the initialized clustering patterns are large in number, and when the sub-time clustering patterns are split, the clustering patterns are also split into two, and the number of the clustering patterns is also increased, a sub-time clustering pattern removing algorithm is provided to remove some patterns under a certain condition, so that the number of the clustering patterns is gradually reduced to achieve the preset number of the clustering patterns.
S25, executing S22 based on the current sub-time sequence clustering mode set and the minimum objective function value obtained by optimizing the S24;
the beneficial effects of adopting the scheme S21-S25 in the S2 are as follows:
for the executed sub-time series clustering model in S1, there is no analytical solution because the pattern length, pattern series, and sub-time series segmentation of each sub-time cluster need to be optimized simultaneously. Therefore, an iterative optimization model based on a sub-time sequence clustering mode initialization, splitting, combining and removing method is adopted, and in the optimization process, the optimization algorithm automatically adjusts the length of each sub-time sequence clustering mode, the mode sequence and the sub-time sequence segmentation, so that the required solution is obtained.
Further, the inner sleeve optimization algorithm in the steps S21, S22, S23 and S24 specifically includes the following steps:
based on the multi-mode sub-time series clustering model, given the randomly initialized clustering mode set C in S21, the current sub-time series clustering mode set C in S22 or S23 or S24, optimizing a sub-time series segmentation set Z and a sub-time series clustering mode C by using a dynamic programming-based inner casing optimization algorithm under the condition of fixing the number of clustering modes and the length of each clustering mode:
d1, given C-optimized Z: the cluster pattern CiHas a length of li，l0Is the cluster pattern CiThe minimum value of the length of (a),
Figure BDA0002295533850000071
is to cover said time series
Figure BDA0002295533850000072
Is the optimal segmentation set of
Figure BDA0002295533850000073
Figure BDA0002295533850000074
Is to cover said time series
Figure BDA0002295533850000075
Is minimized by an objective function, i.e.
Figure BDA0002295533850000076
ZpIs to coverThe time series X1:pWherein p has a value ranging from 1 to t-1, t>2 and is an integer value, DpIs covering said time series X1:pThe Z is obtained by traversing all the lower minimum objective function values meeting the time series coverage constrainttAnd said DtI.e. by
Figure BDA0002295533850000077
Wherein the content of the first and second substances,
Figure BDA0002295533850000078
when T is T, the above algorithm can obtain the time series X given C1:TOptimal coverage Z;
d2, given Z-optimized C:
calculating the center mu of each time subsequence and each cluster pattern in the time subsequence segmentation set ZiA distance value of (d);
classifying each of said sub-time series into a cluster pattern center μiCluster pattern with the smallest distance value;
updating the cluster pattern by averaging all the sub-time series belonging to one of the cluster patterns,
Figure BDA0002295533850000081
and alternately executing the steps D1 and D2 until the sub time sequence segmentation set Z is not changed any more, converging the inner optimization algorithm, wherein the clustering mode C at the convergence is the final solution after optimization.
The beneficial effect of adopting the further scheme is that: in the provided inner sleeve optimization algorithm, when the sub-time sequence segmentation set Z and the sub-time sequence clustering mode C are iteratively optimized, the algorithm can ensure to obtain a global optimal solution, so that the inner sleeve optimization algorithm can ensure that the algorithm is converged under the condition of fixing the number of clustering modes and the length of each clustering mode.
The invention provides a computer-readable storage medium, which comprises instructions that, when executed on a computer, cause the computer to execute the multi-mode time series anomaly detection method according to any one of the above technical solutions.
In addition, the present invention also provides a computer device, which includes a memory, a processor, and a computer program stored in the memory and capable of running on the processor, wherein the processor implements the multi-mode time series abnormality detection method according to any one of the above technical solutions when executing the program.
Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention.
Drawings
In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings needed to be used in the embodiments of the present invention or in the description of the prior art will be briefly described below, and it is obvious that the drawings described below are only some embodiments of the present invention, and it is obvious for those skilled in the art that other drawings can be obtained according to the drawings without creative efforts.
FIG. 1 is a flowchart illustrating a multi-mode time series anomaly detection method according to an embodiment of the present invention;
fig. 2 is a flowchart of an optimization algorithm of a multi-mode sub-time series clustering model according to an embodiment of the present invention.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are some, not all, embodiments of the present invention. All other embodiments, which can be obtained by a person skilled in the art without any inventive step based on the embodiments of the present invention, shall fall within the scope of protection of the present invention.
As shown in fig. 1, a schematic flowchart of a sub-multimode time-series anomaly detection method according to an embodiment of the present invention, the multimode time-series anomaly detection method includes the following steps:
and S1, constructing a multi-mode sub-time sequence clustering model objective function based on time sequence segmentation and sub-time sequence clustering modes according to the historical time sequence.
And S2, iteratively optimizing the multi-mode sub-time sequence clustering model based on the sub-time sequence clustering mode initializing, splitting, combining and removing method, and solving to obtain each sub-time sequence clustering mode.
S3, extracting each sub-time sequence from the time sequences to be detected, and determining whether the sub-time sequences are abnormal or not according to the distance between each sub-time sequence and the center of each clustering pattern in the sub-time sequence clustering model.
The invention has the beneficial effects that: the method can adaptively extract various subsequence modes with different lengths from a time sequence, avoids the condition of insufficient pattern expression or interference of pattern expression caused by extracting the subsequence with fixed length at present, and has higher time sequence abnormity detection precision on the basis; in addition, the invention can perform sub-time sequence clustering on the basis of only assuming the length range of the sub-time sequence without adding other model assumptions to the time sequence, so that the application range of the anomaly detection performed on the basis of the self-adaptive sub-time sequence clustering is wider.
Based on the foregoing embodiment, further, the step S1 specifically includes:
time series X of length T1:TLength range l based on a predetermined sub-time series patternminAnd lmaxAnd the number k of the sub-time series modes, the multi-mode sub-time series clustering model is realized by minimizing an objective function related to the sub-time series clustering mode C and the sub-time series segmentation set Z under the following conditions:
Figure BDA0002295533850000101
s.t.∪x＝X1:T，lmin≤|μi|≤lmax，
wherein X is the historical sub-time series, k is the number of clustering patterns, CiFor the ith cluster, μiFor the ith cluster CiI μmiL is muiLength of (l)minAnd lmaxFor each muiThe time series length constraint must be satisfied, and ═ X1:TAnd representing a time series coverage constraint that each point in X must be at least in a sub-time series of a certain class, and Z is the sub-time series segmentation set, namely a sub-time series extracted from the time series X and corresponding to each sub-time series clustering pattern C.
It should be understood that since a general time series contains a plurality of sub-time patterns with different lengths, in order to realize the automatic extraction of the variable-length patterns, a sub-time series clustering model is provided.
Based on the embodiment, the length of each cluster is unknown and variable, so that the model is different from other methods, and the length of the cluster does not need to be given in advance; in the sub-time sequence clustering, accurate segmentation Z is very important for clustering, and the model can realize automatic extraction of various sub-time sequence patterns by simultaneously optimizing a sub-time sequence clustering pattern C and a sub-time sequence segmentation Z; when extracting the sub-time sequence, the time sequence is required to be covered and restricted, so that the problem of meaningless clustering caused by repeated extraction of adjacent fields when extracting the sub-time sequence through a sliding window by a plurality of methods is avoided.
Further, as shown in fig. 2, which is a flowchart of an optimization algorithm of the multi-mode sub-time series clustering model according to the embodiment of the present invention, step S2 specifically includes:
and S21, performing sub time series clustering pattern initialization.
Initializing a sub-time sequence clustering mode set based on the sub-time sequence clustering model and an inner sleeve optimization algorithm according to a historical time sequence, and specifically comprising the following steps:
setting a length value of each clustering pattern in the historical sub-time sequence clustering pattern set, randomly initializing each clustering pattern, and sequentially setting the length value to be a constraint value l from the minimum preset lengthminTo said maximum predetermined length constraint value lmaxEach value within the range.
And obtaining the sub time sequence clustering mode and the time sequence segmentation set corresponding to each length value through the inner sleeve optimization algorithm.
Aggregating all of the cluster patterns and the segmented set as an initial solution to minimize an objective function.
It should be understood that in the sub-time series clustering pattern model, since the length of each sub-time cluster is unknown, the initial solution of the sub-time series clustering pattern model with respect to the clustering pattern and the segmentation set cannot be directly given. To this end, traverse from lminTo lmaxAnd (4) initializing a segmentation set randomly for each fixed length, and then performing iterative optimization by using an internal algorithm to obtain a clustering mode and a segmentation set under each fixed length. And finally aggregating all optimized clustering patterns and segmentation sets under the fixed length as initial solutions of the sub-time sequence clustering pattern model.
Based on the embodiment, based on all fixed length values set by traversal, clustering is initialized according to an inner sleeve optimization algorithm, so that a better target function initial solution can be obtained.
S22, performing sub-time sequence clustering pattern splitting: based on the current sub-time sequence clustering mode set, splitting the sub-time sequence clustering mode and optimizing the minimized objective function according to the sub-time sequence clustering model and an inner sleeve optimization algorithm, and the specific steps are as follows:
and giving a current sub-time sequence clustering pattern set, and calculating the variance value of each clustering pattern in the sub-time sequence clustering pattern set.
And taking the clustering pattern with the maximum variance value as a clustering pattern to be split.
Splitting the clustering mode to be split at each position meeting the minimum preset length constraint value and the maximum preset length constraint value, and calculating the value of the minimized objective function based on the inner sleeve optimization algorithm to obtain the minimized objective function value corresponding to each position.
And splitting the clustering mode to be split at the position corresponding to the minimum value in the minimized objective function value to obtain the split result of the time sequence clustering mode set.
And optimizing the minimized objective function according to the sub-time sequence clustering mode set and the inner sleeve optimization algorithm.
It should be understood that some of the initialized cluster patterns may be formed by combining multiple patterns, and thus, we propose a sub-time series pattern splitting algorithm to split the corresponding multiple patterns from the cluster. Two problems are involved here: 1) which mode is selected for splitting. If some clusters contain a plurality of patterns, the variance of the corresponding sub-time sequences may be larger, so that the clustering pattern with the largest variance in the current clustering patterns is split; 2) the clustering pattern to be split is divided into two at which position. A traversal method is adopted to split each feasible position meeting the clustering mode length constraint, and an inner algorithm is adopted to optimize after splitting to obtain an optimized objective function. We select those splitting points of the best minimum objective function to split.
The embodiment optimizes the sub-sequence clustering pattern splitting process, so that the pattern formed by combining a plurality of patterns can be split into a plurality of corresponding patterns, thereby being beneficial to better obtaining each sub-time sequence pattern inherent in the time sequence.
S23, performing sub-time series clustering pattern combination: based on the current sub-time sequence clustering mode set, combining sub-time sequence clustering modes and optimizing the minimized objective function according to the sub-time sequence clustering model and the inner sleeve optimization algorithm, comparing the optimized clustering result with the clustering result before the execution of S22, and judging the optimization mode of the next sub-time sequence clustering, wherein the specific steps are as follows:
giving a current sub-time sequence clustering pattern set, and calculating adjacency values among clustering patterns in the historical sub-time sequence clustering pattern set, wherein the adjacency values are obtained by the following method:
clustering pattern CjAnd CjHas a neighborhood value of P (C)i|Cj)＝max{P1(Ci|Cj),P2(Ci|Cj) And (c) the step of (c) in which,
Figure BDA0002295533850000121
as a cluster pattern CiThe cluster pattern of the immediately adjacent position is CjThe probability of (a) of (b) being,
Figure BDA0002295533850000122
as a cluster pattern CjThe clustering pattern of the immediately preceding neighboring position is CiProbability of (n)iIndicates belongings to clustering Pattern C in ZiNumber of sub-time series of (2), nijIndicates belongings to clustering Pattern C in ZiIs followed by a sub-time sequence belonging to cluster pattern CjThe number of sub-time series of (2).
Taking the clustering pattern pair with the maximum adjacency value as a clustering pattern pair to be combined, and if P (C)i|Cj) Is derived from P1(Ci|Cj) Then combine CiAnd CjTo replace original CjIf P (C)i|Cj) Is derived from P2(Ci|Cj) Then combine CiAnd CjTo replace original Ci。
And combining the clustering modes to be combined at each position meeting the minimum preset length constraint value and the maximum preset length constraint value, and calculating a minimum objective function value corresponding to each position based on the inner sleeve optimization algorithm.
And combining the clustering modes to be combined with the position corresponding to the minimum value in the minimized objective function value to obtain the result after the clustering mode sets of the sub-time sequence are combined.
And optimizing the minimized objective function according to the sub-time sequence clustering mode set and the inner sleeve optimization algorithm.
And comparing the clustering result after the clustering mode combination optimization with the clustering result before the S22 execution, and judging the next optimization mode:
and S231, comparing the number of the current clustering modes obtained after the clustering mode combination optimization in the iteration step with the number of the clustering modes obtained before the execution of the step S22.
And if the current clustering mode number obtained after the clustering mode combination optimization is greater than the current clustering mode number obtained before the S22 is executed, executing the step S24.
Otherwise, step S232 is executed.
And S232, judging whether the current minimized objective function value obtained after the cluster mode combination optimization in the iteration step is larger than the minimized objective function value obtained before the execution of S22.
If not, S22 is executed.
If yes, go to step S233.
And S233, judging whether the number of the clustering patterns obtained after the clustering pattern combination optimization in the iteration step reaches the preset target number of the clustering patterns.
If not, step S24 is executed.
If yes, go to step S234.
And S234, the sub-time sequence clustering mode obtained after the clustering mode combination optimization in the iteration step is the final sub-time sequence clustering mode to be solved, and the multi-mode sub-time sequence clustering model optimization is finished.
It should be understood that, in all the sub-time series clustering patterns, some clustering patterns may only contain a part of the original time series patterns, and therefore, the sub-time series clustering combination algorithm needs to be combined for the sub-time series clustering patterns. There are two problems to be solved here: 1) which two time series are combined. If two adjacent parts of a time series pattern appear in two cluster patterns, the sub-time series corresponding to the two cluster patterns are also adjacent. Thus, two cluster patterns with the greatest adjacency are selected for combination; 2) and combining the two selected clustering patterns at which position. Similar to sub-time clustering pattern splitting, clustering pattern combination is carried out on each position meeting the clustering pattern length constraint, and a new clustering pattern generated by combination is obtained by averaging clustering values of the two positions. And optimizing by using a new clustering mode and an inner sleeve optimization algorithm to obtain an optimized objective function, and selecting the combination position of the optimal minimum objective function to perform clustering mode combination.
The above-described embodiment optimizes the sub-time-series pattern combination so that a plurality of portions of a certain pattern existing in a plurality of cluster patterns can be combined, thereby contributing to better obtaining each sub-time-series pattern inherent in the time series.
S24, based on the current sub-time sequence clustering mode set, according to the minimized objective function and the inner sleeve optimization algorithm, removing the clustering mode corresponding to the minimum value of the minimized objective function, and optimizing the minimum objective function, the specific steps are as follows:
and giving a current sub-time sequence clustering mode set, traversing all modes in the current sub-time sequence clustering mode set based on the minimized objective function and an internal optimization algorithm, and calculating the value of the minimized objective function when each clustering mode in the sub-time sequence clustering mode set is deleted in sequence.
And removing the deleted clustering mode corresponding to the minimum value in the values of the minimized objective function, and optimizing the minimized objective function based on the inner sleeve optimization algorithm.
It should be understood that, because the number of initialized clustering patterns is large, and when a sub-time clustering pattern is split, the clustering pattern is also split into two from one clustering pattern, and the number of clustering patterns is increased, a sub-time clustering pattern removing algorithm is proposed to remove some patterns under certain conditions, so that the number of clustering patterns is gradually reduced until the number of clustering patterns required by the sub-time series clustering model is met. When the subsequence clustering modes are removed, each clustering mode is traversed, the value of the objective function obtained by adopting an inner sleeve optimization algorithm after a certain clustering mode is removed is calculated, and the clustering model which enables the value of the objective function to be minimum is selected for removal.
The above embodiment will gradually reduce the number of clustering patterns until the number of clusters meeting the requirements of the sub-time series clustering model is obtained.
S25, executing sub-time sequence clustering iteration: executing S22 based on the current sub-time sequence clustering mode set obtained through optimization in S24, the minimum objective function value and the number of modes in the clustering mode set;
the beneficial effects of adopting the scheme S21-S25 in the S2 are as follows:
for the executed sub-time series clustering model in S1, there is no analytical solution because the pattern length, pattern series, and sub-time series segmentation of each sub-time cluster need to be optimized simultaneously. Therefore, an iterative optimization model based on a sub-time sequence clustering mode initialization, splitting, combining and removing method is adopted, and in the optimization process, the optimization algorithm automatically adjusts the length of each sub-time sequence clustering mode, the mode sequence and the sub-time sequence segmentation, so that the required solution is obtained.
Further, the inner sleeve optimization algorithm in the steps S21, S22, S23 and S24 specifically includes the following steps:
based on the multi-mode sub-time series clustering model, given the randomly initialized clustering mode set C in S21, the current sub-time series clustering mode set C in S22 or S23 or S24, optimizing a sub-time series segmentation set Z and a sub-time series clustering mode C by using a dynamic programming-based inner casing optimization algorithm under the condition of fixing the number of clustering modes and the length of each clustering mode:
d1, given C-optimized Z: the cluster pattern CiHas a length of li，l0Is the cluster pattern CiThe minimum value of the length of (a),
Figure BDA0002295533850000151
is to cover said time series
Figure BDA0002295533850000152
Is the optimal segmentation set of
Figure BDA0002295533850000153
Figure BDA0002295533850000165
Is to cover said time series
Figure BDA0002295533850000164
Is minimized by an objective function, i.e.
Figure BDA0002295533850000166
ZpIs covering said time series X1:pWherein p has a value ranging from 1 to t-1, t>2 and is an integer value, DpIs covering said time series X1:pThe Z is obtained by traversing all the lower minimum objective function values meeting the time series coverage constrainttAnd said DtI.e. by
Figure BDA0002295533850000161
Wherein the content of the first and second substances,
Figure BDA0002295533850000162
when T is T, the above algorithm can obtain the time series X given C1:TThe optimal coverage Z.
D2, given Z-optimized C:
calculating each sub-time sequence and each clustering pattern in the sub-time sequence segmentation set ZHeart muiThe distance value of (2).
Assigning each of the sub-time series to the cluster pattern center μiCluster pattern with the smallest distance value.
Updating the cluster pattern by averaging all the sub-time series belonging to one of the cluster patterns,
Figure BDA0002295533850000163
and alternately executing the steps D1 and D2 until the sub time sequence segmentation set Z is not changed any more, converging the inner optimization algorithm, wherein the clustering mode C at the convergence is the final solution after optimization.
In the above embodiment, in the inner optimization algorithm, when the sub-time sequence segmentation set Z and the sub-time sequence clustering pattern C are iteratively optimized, the algorithm can ensure that a globally optimal solution is obtained, and therefore, the inner optimization algorithm can ensure that the algorithm converges under the condition that the number of clustering patterns and the lengths of the clustering patterns are fixed for the sub-time sequence clustering pattern.
It should be understood that, in the step S2, an internal optimization algorithm and an optimization algorithm based on the sub-time series clustering model thereof are proposed. The optimization process comprises the steps of firstly giving an initialization algorithm of a clustering mode and a segmentation set in a sub-time sequence clustering model, then optimizing the clustering mode and the segmentation set of the model from the large number of sub-time clustering modes through an algorithm of sub-time clustering mode splitting, sub-time clustering mode combination and sub-time sequence clustering mode removing, and gradually reducing the number of the clustering modes until the number of clusters required by the sub-time sequence clustering mode model is reached, thereby realizing the optimization of the sub-time sequence clustering model.
Further, the step S3 specifically includes:
based on the sub-time sequence clustering model and the model optimization result, extracting each sub-time sequence from the time sequence to be detected, extracting each sub-time sequence x from the time sequence, calculating whether the time sequence belongs to a certain sub-time sequence clustering mode, and if so, determining whether the time sequence belongs to the sub-time sequence clustering modeIf the cluster pattern does not belong to any cluster pattern, the cluster pattern is judged to be abnormal. Calculating whether x belongs to cluster CiThen calculate x and μiD (x, mu) betweeni) If d (x, μ)i) If the threshold value is greater than a certain threshold value thr, the user is judged not to belong to the class. The threshold thr may be determined by the cluster CiIs determined.
In the embodiment, due to the fact that the sub-time sequence clustering model is adopted, various sub-time sequence modes with different lengths can be extracted from the time sequence in a self-adaptive mode, the situation that mode expression is insufficient or interference exists in the mode expression caused by the fact that the sub-time sequence is extracted by adopting a fixed length at present is avoided, and therefore the time sequence abnormity detection accuracy is higher on the basis.
The invention provides a computer-readable storage medium comprising instructions which, when run on a computer, cause the computer to perform the steps of the multi-mode time series anomaly detection method according to any one of the above-mentioned aspects.
The invention also relates to a computer device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor executes the program to implement the steps of the multi-mode time series anomaly detection method according to any one of the above embodiments.
While the invention has been described with reference to specific embodiments, it will be understood by those skilled in the art that various changes in form and details may be made therein without departing from the spirit and scope of the invention as defined by the appended claims. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims (6)

1. A multi-mode time series abnormality detection method is characterized by comprising the following steps:
s1, constructing a multi-mode sub-time sequence clustering model target function based on time sequence segmentation and sub-time sequence clustering modes according to the historical time sequence;
s2, iteratively optimizing the multi-mode sub-time sequence clustering model based on the method of sub-time sequence clustering mode initialization, splitting, combining and removing, and solving to obtain each sub-time sequence clustering mode;
s3, extracting each sub-time sequence from the time sequences to be detected, and determining whether the sub-time sequences are abnormal or not according to the distance between each sub-time sequence and the center of each clustering pattern in the sub-time sequence clustering model;
the step S1 specifically includes:
time series X of length T1：TLength range l based on a predetermined sub-time series patternminAnd lmaxAnd the number k of the sub-time series modes, the multi-mode sub-time series clustering model is realized by minimizing an objective function related to the sub-time series clustering mode C and the sub-time series segmentation set Z under the following conditions:
Figure FDA0003591903100000011
s.t.∪x＝X1：T，lmin≤|μi|≤lmax，
wherein the content of the first and second substances,
Figure FDA0003591903100000012
to minimize the objective function,/minFor a minimum predetermined length constraint value,/maxFor maximum predetermined length constraint value, X is the historical sub-time series, k is the number of clustering patterns, CiFor the ith cluster, μiFor the ith cluster CiI μmiL is muiLength of (a) < l >minAnd lmaxFor each muiThe time series length constraint must be satisfied, and ═ X1：TRepresenting a time series coverage constraint, i.e. each point in X must be in at least one class of sub-time series, Z is a set of partitions of said sub-time series, i.e. the cluster modulus extracted from time series X corresponding to each sub-time seriesA sub-time sequence of formula C;
the step S2 specifically includes: s21, initializing a sub-time sequence clustering mode set based on the sub-time sequence clustering model and an inner sleeve optimization algorithm according to the historical time sequence;
s22, based on the current sub-time sequence clustering mode set, splitting the sub-time sequence clustering mode and optimizing the minimized objective function according to the sub-time sequence clustering model and an inner sleeve optimization algorithm;
s23, based on the current sub-time sequence clustering mode set, according to the sub-time sequence clustering model and the inner sleeve optimization algorithm, combining sub-time sequence clustering modes and optimizing the minimized objective function, and according to the optimized clustering result, comparing the clustering result with the clustering result before the execution of S22, and judging the optimization mode of the next sub-time sequence clustering;
s24, based on the current sub-time sequence clustering mode set, according to the minimized objective function and the inner sleeve optimization algorithm, removing the clustering mode corresponding to the minimum value of the minimized objective function, and optimizing the minimum objective function;
s25, executing S22 based on the current sub-time sequence clustering mode set and the minimum objective function value obtained by optimizing the S24;
the step S21 specifically includes:
the specific steps for executing the initialization of the sub-time sequence clustering pattern are as follows:
setting a length value of each clustering pattern in the historical sub-time sequence clustering pattern set, randomly initializing each clustering pattern, and sequentially setting the length value to be a constraint value l from the minimum preset lengthminTo said maximum predetermined length constraint value lmaxEach value within the range;
obtaining the sub-time sequence clustering mode and the time sequence segmentation set corresponding to each length value through the inner sleeve optimization algorithm;
aggregating all of the cluster patterns and the segmented set as an initial solution to a minimization objective function;
the inner sleeve optimization algorithm specifically comprises the following steps:
based on the multi-mode sub-time series clustering model, given the randomly initialized clustering mode set C in S21, the current sub-time series clustering mode set C in S22 or S23 or S24, optimizing a sub-time series segmentation set Z and a sub-time series clustering mode C by using a dynamic programming-based inner casing optimization algorithm under the condition of fixing the number of clustering modes and the length of each clustering mode:
d1, given C-optimized Z: the cluster pattern CiHas a length of li，l0Is the cluster pattern CiThe minimum value of the length of (a),
Figure FDA0003591903100000031
is to cover said time series
Figure FDA0003591903100000032
Is the optimal segmentation set of
Figure FDA0003591903100000033
Figure FDA0003591903100000034
Is to cover said time series
Figure FDA0003591903100000035
Is minimized by an objective function, i.e.
Figure FDA0003591903100000036
ZpIs covering said time series X1：pWherein p has a value ranging from 1 to t-1, t>2 and is an integer value, DpIs to cover said time series X1：pThe Z is obtained by traversing all the minimum objective function values under the time series coverage constrainttAnd said DtI.e. by
Figure FDA0003591903100000037
Wherein the content of the first and second substances,
Figure FDA0003591903100000038
when T is T, the above algorithm can obtain the time series X given C1：TOptimal coverage Z;
d2, given Z-optimized C:
calculating the center mu of each time subsequence and each cluster pattern in the time subsequence segmentation set ZiA distance value of (d);
assigning each of the sub-time series to the cluster pattern center μiCluster pattern with the smallest distance value;
updating the cluster pattern by averaging all the sub-time series belonging to one of the cluster patterns,
Figure FDA0003591903100000039
where x denotes each sub-time sequence, d (x, μ)i) Is x and muiThe distance between them; and D1 and D2 are executed alternately until the time sequence sub-segmentation set Z is not changed any more, the inner optimization algorithm converges, and the clustering mode set C at the time of convergence is the solution of the clustering model.
2. The method for detecting the abnormality in the multi-pattern time series according to claim 1, wherein the step S22 specifically includes:
the specific steps for performing the splitting of the clustering pattern of the sub-time series are as follows:
giving a current sub-time sequence clustering pattern set, and calculating a variance value of each clustering pattern in the sub-time sequence clustering pattern set;
taking the clustering pattern with the maximum variance value as a clustering pattern to be split;
splitting the clustering pattern to be split at each position meeting the minimum preset length constraint value and the maximum preset length constraint value, and calculating the value of the minimized objective function based on the inner sleeve optimization algorithm to obtain the minimized objective function value corresponding to each position;
splitting the clustering mode to be split at the position corresponding to the minimum value in the minimized objective function value to obtain a result after splitting of a time sequence clustering mode set;
and optimizing the minimized objective function according to the sub-time sequence clustering mode set and the inner sleeve optimization algorithm.
3. The method for detecting the abnormality in the multi-pattern time series according to claim 1, wherein the step S23 specifically includes:
the specific steps for performing the sub-time series clustering pattern combination are as follows:
giving a current sub-time sequence clustering pattern set, and calculating the adjacency value among all clustering patterns in the historical sub-time sequence clustering pattern set;
the adjacency value is obtained by the following method:
clustering pattern CjAnd CjHas a neighborhood value of P (C)i|Cj)＝max{P1(Ci|Cj)，P2(Ci|Cj) -means for, among other things,
Figure FDA0003591903100000041
as a cluster pattern CiThe cluster pattern of the immediately adjacent position is CjThe probability of (a) of (b) being,
Figure FDA0003591903100000042
as a cluster pattern CjThe clustering pattern of the immediately preceding neighboring position is CiProbability of (n)iIndicates belongings to clustering Pattern C in ZiOf a sub-time series ofNumber, nijIndicates belongings to clustering Pattern C in ZiIs immediately followed by a sub-temporal sequence belonging to clustering pattern CjThe number of sub-time series of (a);
taking the clustering pattern pair with the maximum adjacency value as a clustering pattern pair to be combined if P (C)i|Cj) Is derived from P1(Ci|Cj) Then combine CiAnd CjTo replace original CjIf P (C)i|Cj) Is derived from P2(Ci|Cj) Then the combination CiAnd CjTo replace original Ci；
Combining the clustering modes to be combined at each position meeting the minimum preset length constraint value and the maximum preset length constraint value, and calculating a minimized objective function value corresponding to each position based on the inner sleeve optimization algorithm;
combining the clustering modes to be combined with the positions corresponding to the minimum values in the minimized objective function values to obtain the result after the clustering mode sets of the sub-time sequences are combined;
optimizing the minimized objective function according to the sub-time sequence clustering mode set and the inner sleeve optimization algorithm;
and comparing the clustering result after the clustering mode combination optimization with the clustering result before the S22 execution, and judging the next optimization mode:
comparing the number of the current clustering modes obtained after the clustering mode combination optimization in the iteration step with the number of the clustering modes obtained before the S22 is executed;
if the number of the current clustering modes obtained after the clustering mode combination optimization is greater than the number of the clustering modes obtained before the S22 is executed, executing S24;
otherwise, judging whether the current minimized objective function value obtained after the cluster mode combination optimization in the iteration step is larger than the minimized objective function value obtained before the execution of S22;
if not, go to S22;
if yes, judging whether the number of the clustering modes obtained after the clustering mode combination optimization in the iteration step reaches the preset clustering mode target number or not;
if not, go to step S24;
if so, the sub-time sequence clustering mode obtained after the clustering mode combination optimization in the iteration step is the final sub-time sequence clustering mode to be solved, and the multi-mode sub-time sequence clustering model optimization is finished.
4. The method according to claim 1, wherein the step S24 specifically includes the steps of:
the specific steps for removing the clustering pattern of the sub-time series are as follows:
giving a current sub-time sequence clustering mode set, traversing all modes in the current sub-time sequence clustering mode set based on the minimized objective function and an inner sleeve optimization algorithm, and calculating the value of the minimized objective function when deleting each clustering mode in the sub-time sequence clustering mode set in sequence;
and removing the deleted clustering mode corresponding to the minimum value in the values of the minimized objective function, and optimizing the minimized objective function based on the inner sleeve optimization algorithm.
5. A computer-readable storage medium comprising instructions that, when executed on a computer, cause the computer to perform the steps of the multi-modal time series anomaly detection method according to any one of claims 1-4.
6. A computer device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, characterized in that the processor implements the steps of the multi-modal time series anomaly detection method according to any one of claims 1-4 when executing the program.
CN201911199574.5A 2019-11-29 2019-11-29 Multi-mode time series anomaly detection method, storage medium and equipment Active CN111027606B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201911199574.5A CN111027606B (en) 2019-11-29 2019-11-29 Multi-mode time series anomaly detection method, storage medium and equipment

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201911199574.5A CN111027606B (en) 2019-11-29 2019-11-29 Multi-mode time series anomaly detection method, storage medium and equipment

Publications (2)

Publication Number Publication Date
CN111027606A CN111027606A (en) 2020-04-17
CN111027606B true CN111027606B (en) 2022-05-31

Family

ID=70207290

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201911199574.5A Active CN111027606B (en) 2019-11-29 2019-11-29 Multi-mode time series anomaly detection method, storage medium and equipment

Country Status (1)

Country Link
CN (1) CN111027606B (en)

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113420070B (en) * 2021-06-24 2023-06-30 平安国际智慧城市科技股份有限公司 Pollution discharge monitoring data processing method and device, electronic equipment and storage medium
CN116756595B (en) * 2023-08-23 2023-12-01 深圳市森瑞普电子有限公司 Conductive slip ring fault data acquisition and monitoring method

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102880621A (en) * 2011-07-14 2013-01-16 富士通株式会社 Method and device for extracting similar sub time sequences
CN103020643A (en) * 2012-11-30 2013-04-03 武汉大学 Classification method based on kernel feature extraction early prediction multivariate time series category

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102930294A (en) * 2012-10-18 2013-02-13 上海交通大学 Chaotic characteristic parameter-based motion mode video segmentation and traffic condition identification method
CN103034122A (en) * 2012-11-28 2013-04-10 上海交通大学 Multi-model self-adaptive controller and control method based on time series
CN105678409A (en) * 2015-12-31 2016-06-15 哈尔滨工业大学 Adaptive and distribution-free time series abnormal point detection method

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102880621A (en) * 2011-07-14 2013-01-16 富士通株式会社 Method and device for extracting similar sub time sequences
CN103020643A (en) * 2012-11-30 2013-04-03 武汉大学 Classification method based on kernel feature extraction early prediction multivariate time series category

Also Published As

Publication number Publication date
CN111027606A (en) 2020-04-17

Similar Documents

Publication Publication Date Title
CN110245721B (en) Training method and device for neural network model and electronic equipment
CN112640380B (en) Apparatus and method for anomaly detection of an input stream of events
CN109118500B (en) Image-based three-dimensional laser scanning point cloud data segmentation method
KR101889722B1 (en) Method and Apparatus for Diagnosing Malignant Tumor
CN111027606B (en) Multi-mode time series anomaly detection method, storage medium and equipment
CN109241126B (en) Spatio-temporal trajectory aggregation mode mining algorithm based on R-tree index
WO2017161632A1 (en) Cleaning robot optimal target path planning method based on model learning
CN110493221B (en) Network anomaly detection method based on clustering contour
JP2009515585A (en) Method, system and computer program for volumetric alignment
CN111562787B (en) Method, device, medium and equipment for dividing planning area of full-coverage path of robot
WO2018036547A1 (en) Data processing method and device thereof
JP2015191261A (en) state estimation device, program, and integrated circuit
KR101889723B1 (en) Method and Apparatus for Diagnosing Malignant Tumor
CN111460234A (en) Graph query method and device, electronic equipment and computer readable storage medium
JP4375308B2 (en) Ridge direction extraction device, ridge direction extraction method, ridge direction extraction program
CN110473224B (en) Automatic RSF level set image segmentation method based on KL entropy
CN111160459A (en) Device and method for optimizing hyper-parameters
KR101708042B1 (en) Cylinder estimation device and method by ransac algorithm in point cloud
Karlsson et al. Forests of randomized shapelet trees
CN108898611B (en) Fuzzy region active contour segmentation model based on significant perception prior
JP6643368B2 (en) Image processing apparatus and image processing method
CN112102178A (en) Point cloud feature-preserving denoising method and device, electronic equipment and storage medium
CN110032585B (en) Time sequence double-layer symbolization method and device
CN110275895B (en) Filling equipment, device and method for missing traffic data
KR101919698B1 (en) Group search optimization data clustering method and system using silhouette

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant