CN112529678A - Financial index time sequence abnormity detection method based on self-supervision discriminant network - Google Patents

Financial index time sequence abnormity detection method based on self-supervision discriminant network Download PDF

Info

Publication number
CN112529678A
CN112529678A CN202011543431.4A CN202011543431A CN112529678A CN 112529678 A CN112529678 A CN 112529678A CN 202011543431 A CN202011543431 A CN 202011543431A CN 112529678 A CN112529678 A CN 112529678A
Authority
CN
China
Prior art keywords
financial index
network
time sequence
index time
supervision
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.)
Granted
Application number
CN202011543431.4A
Other languages
Chinese (zh)
Other versions
CN112529678B (en
Inventor
马千里
黄德森
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.)
South China University of Technology SCUT
Original Assignee
South China University of Technology SCUT
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 South China University of Technology SCUT filed Critical South China University of Technology SCUT
Priority to CN202011543431.4A priority Critical patent/CN112529678B/en
Publication of CN112529678A publication Critical patent/CN112529678A/en
Application granted granted Critical
Publication of CN112529678B publication Critical patent/CN112529678B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q40/00—Finance; Insurance; Tax strategies; Processing of corporate or income taxes
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/06—Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
    • G06Q10/063—Operations research, analysis or management
    • G06Q10/0639—Performance analysis of employees; Performance analysis of enterprise or organisation operations
    • G06Q10/06393—Score-carding, benchmarking or key performance indicator [KPI] analysis

Landscapes

  • Engineering & Computer Science (AREA)
  • Business, Economics & Management (AREA)
  • Theoretical Computer Science (AREA)
  • Human Resources & Organizations (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Strategic Management (AREA)
  • Economics (AREA)
  • Development Economics (AREA)
  • Marketing (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • General Business, Economics & Management (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Educational Administration (AREA)
  • Molecular Biology (AREA)
  • Biophysics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Evolutionary Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • General Health & Medical Sciences (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Computational Linguistics (AREA)
  • Game Theory and Decision Science (AREA)
  • Computing Systems (AREA)
  • Biomedical Technology (AREA)
  • Operations Research (AREA)
  • Quality & Reliability (AREA)
  • Tourism & Hospitality (AREA)
  • Health & Medical Sciences (AREA)
  • Accounting & Taxation (AREA)
  • Finance (AREA)
  • Technology Law (AREA)
  • Data Exchanges In Wide-Area Networks (AREA)
  • Financial Or Insurance-Related Operations Such As Payment And Settlement (AREA)

Abstract

The invention discloses a financial index time sequence abnormity detection method based on an automatic supervision discriminant network, which is realized based on two components of an automatic supervision downsampling module and an agent supervision network module. The self-monitoring down-sampling module performs down-sampling on the time sequence information and gives each time sequence subsequence a mark related to down-sampling scale information. The input time sequence generates a sample set with scale information marks after passing through an auto-supervision down-sampling module, and then the classification of different scales in the sample set is realized through an agent supervision network module, so that the characteristic modeling of the input time sequence is realized. And finally, taking the loss function value of the agent supervision network module as an abnormality detection index to perform abnormality detection on the financial index time sequence sample. The anomaly detection method disclosed by the invention has the characteristic of high training efficiency, and simultaneously achieves the effect of high-precision financial index time sequence anomaly detection.

Description

Financial index time sequence abnormity detection method based on self-supervision discriminant network
Technical Field
The invention relates to the technical field of anomaly detection research of financial index time series, in particular to a financial index time series anomaly detection method based on an automatic supervision discriminant network.
Background
The financial index time series has been the research focus of people for a long time, and whether the fluctuation of the financial index time series is normal or not brings huge challenges to investment and economic operation. How to effectively and quickly analyze the financial index time data and dig out an abnormal financial index time sequence has great significance to the field of financial indexes.
Most existing model-based anomaly detection methods are based on constructing a summary of normal instances, and then identifying instances that do not fit the normal summary as anomalous samples. Most of the existing classical anomaly detection methods realize anomaly detection based on similarity search and density clustering, and are essentially based on the anomaly detection of the distance between samples. Although anomaly detection based on the distance between samples is well interpretable and has been experimentally and theoretically verified as the most advanced anomaly detection shelf, it requires re-traversing the entire training set to find the nearest neighbor sample during the testing phase, which is very time consuming. The classical methods based on classification, such as OCSVM, implement detection of abnormal samples by establishing a boundary, which is often not robust enough in the task of abnormal detection. Furthermore, classical methods tend to suffer from temporal and spatial limitations when applied to large data sets.
Compared with the traditional classical anomaly detection algorithm, deep learning has made great progress in the anomaly detection field in recent years. The mainstream framework is based on a generative model, and the core idea is to reduce the dimension of data in an original space into a low-dimensional space and then reconstruct the data. The latent assumption of this generative framework is that the features of the anomaly are eliminated in the low dimensional space, and therefore anomaly detection can be performed using the reconstruction error as an indicator. The sample with large reconstruction error corresponds to an abnormal sample. In a generative framework of the field of temporal anomaly detection, work based on autoencoders includes the BeatGAN model extracted in 2019 by Bin Zhou et al. Furthermore, the network model based on predictive type reconstruction includes LSTM predictive model, sequence-to-sequence model. The generative model plays a major role in time sequence anomaly detection, however, the generative model has poor modeling capability on normal samples, and the extracted features are difficult to distinguish from the extracted features corresponding to the anomaly samples.
As a way of unsupervised feature extraction, the self-supervision technology has recently exerted excellent performance in the field of image anomaly detection, such as the self-supervision anomaly detection framework of the image anomaly detection field proposed by Siqi Wang in NIPS 2019. Before the application of the self-supervision technology to image anomaly detection, the most advanced image anomaly detection framework is mainly based on generation modules such as a noise reduction self-encoder, a variation self-encoder and a generation countermeasure network. This is very similar to the mainstream framework of the current time sequence anomaly detection, however, the self-supervision technology has not been integrated into the field of financial index time sequence anomaly detection. The main reason is that the self-supervision information in the time series field is not yet clear, and how to mine the self-supervision information in the time series is an unexplored problem.
Disclosure of Invention
The invention aims to solve the defects in the existing time sequence anomaly detection technology and provides a financial index time sequence anomaly detection method based on an automatic supervision discriminant network. Extracting subsequence samples representing time sequence different scale information from training samples through downsampling, and giving class marks of the scale information to the subsequence samples; then, the agent supervision network module learns the characteristic information of different scales of the samples, so that the abnormal samples can be better distinguished from the normal samples in the testing stage.
The purpose of the invention can be achieved by adopting the following technical scheme:
a financial index time sequence abnormity detection method based on an auto-supervision discriminant network is applied to the field of financial index time sequence abnormity detection, and high-precision financial index time sequence abnormity detection is realized through the auto-supervision discriminant network, and the method comprises the following steps:
s1, generating a time sequence subsequence through a down-sampling mechanism, and obtaining subsequence samples of the original time sequence under different scale information through sliding sampling according to different window sizes f belonging to [1,2, …, K ] of input financial index time sequence samples, wherein K is the upper limit of the size of a sampling window;
s2, marking the subsequence samples under the financial index time sequence different scale information obtained in the step S1 through the scale information to obtain a time sequence sample set with marking information;
s3, extracting first-layer convolution characteristics, namely, taking the time sequence sample set with the marked information obtained in the step S2 as a training set, taking the training set as input information of a network, and accessing a first-layer convolution layer to obtain convolution characteristics;
s4, pooling the first-layer convolution characteristics, and performing pooling operation on the convolution characteristics output in the step S3 to obtain pooled characteristics;
s5, performing second-layer convolution feature extraction, namely performing convolution feature extraction on the pooled features output in the step S4;
s6, pooling the second layer of convolution characteristics, and performing pooling operation on the convolution characteristics output in the step S5 to obtain pooled characteristics;
s7, Softmax classification, connecting the pooled features output in the step S6 to a Softmax network layer, and obtaining classification of each sample in the time series sample set with the mark information generated by the financial index time series;
s8, training the agent supervision network established in the steps S3-S7, and optimizing the network parameter weights in the steps S3-S7 based on a back propagation algorithm to further obtain internal parameters of the high-precision financial index time sequence abnormity detection network model;
s9, calculating abnormal indexes, namely calculating the classified cross entropy loss of each financial index time series sample generated by downsampling and provided with scale information mark samples based on the optimized agent supervision network, and taking the classified cross entropy loss as the abnormal indexes of the financial index time series samples;
s10, inputting a test sample, repeating the step S1 to construct a plurality of subsequence samples with different scale information after the test sample is sampled, then repeating the step S2 to mark the subsequence samples through the scale information, then repeating the steps S3 to S7 to obtain the classification result of the subsequence samples on the network by a trained agent monitoring network, repeating the step S9 to calculate the classification cross entropy loss of the agent monitoring network under the test sample, and taking the classification cross entropy loss as the index of financial index time sequence abnormity detection.
Further, in step S2, the financial index time series signal sample set is used
Figure BDA0002853150870000041
Obtaining a time series sample set through an auto-supervision down-sampling module,
representing an input signal as financial index time series data of length T dimension M
Figure BDA0002853150870000042
Wherein, Xi(t)∈RM×1Identifying the characteristic value, x, of the ith sample at time tm,tThe characteristics of the sample at the t-th time of the mth dimension are represented as the time sequence after the downsampling
Xi,f＝(Xi(1),Xi(f+1),…,Xi(l*f+1))
Where f denotes the sliding sampling window size,
Figure BDA0002853150870000043
the ith downsampling sliding window is identified.
Further, in steps S3 to S7, feature extraction of the input financial index time series is implemented by an agent monitoring network module, specifically: features are extracted by convolution and pooling, and discriminant classification of financial index time sequences is realized by combining a Softmax network layer.
Further, in step S8, the proxy monitoring network module is optimized by using the down-sampled scale information as the classification label and using the cross-entropy loss function as the classification loss function, where the expression is as follows:
Figure BDA0002853150870000051
wherein y represents the size of the sampling window, K is the upper limit of the size of the sampling window, theta is a network parameter,
Figure BDA0002853150870000052
representative sample XiA time sequence subsequence generated after a down-sampling operation with a sliding window size of y,
Figure BDA0002853150870000053
corresponding agent supervision network to sequence sub-sequence under parameter theta
Figure BDA0002853150870000054
The classification predicts the probability of correctness, log is a log function with base 2, Loss (X)iTheta) is used as a sample X of the agent monitoring network module under the parameter thetaiLoss of self-supervised classification.
Compared with the prior art, the invention has the following advantages and effects:
the invention establishes a self-monitoring discriminant network with the function of capturing multi-scale information of financial index time series by introducing a self-monitoring down-sampling mechanism. After the financial index time sequence is acted by the self-supervision down-sampling module and a series of time sequence sample sets with scale information supervision are generated, classification of different scale information of the financial index time sequence is achieved through an agent supervision network module. The whole agent supervision network module is optimized based on a back propagation algorithm. Different from the existing time sequence abnormity detection method based on the generative framework, the invention utilizes the self-supervision technology to capture the information of the normal financial index time sequence samples from various scales. A large number of experiments show that the financial index time sequence abnormity detection method based on the self-supervision discriminant network has excellent performance in a financial index time sequence abnormity detection task and has the advantage of high training efficiency.
Drawings
FIG. 1 is a network architecture diagram of a financial index timing anomaly detection method based on an auto-supervised discriminant network according to the present invention;
FIG. 2 is a schematic diagram of a financial index time series abnormality detection method based on an unsupervised discriminant network according to the present disclosure.
Detailed Description
In order to make the objects, technical solutions and advantages of the embodiments of the present invention clearer, 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, but not all, embodiments of the present invention. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
Examples
As shown in fig. 1, the present embodiment discloses a financial index time series abnormality detection method based on an unsupervised discriminant network. After the financial index time sequence is acted by the self-supervision down-sampling module and a series of time sequence sample sets with scale information marks are generated, classification of different scale information of the financial index time sequence is achieved through an agent supervision network module. The whole agent supervision network module is optimized based on a back propagation algorithm. Compared with the traditional generating-based framework, the generating-based framework has the advantages of better capturing characteristic information of multiple scales of original financial index time series samples, and specifically comprises the following steps:
s1, obtaining subsequence samples of the financial index time series samples under different scale information (K is the upper limit of the size of a sampling window) by sliding and sampling windows with the size f belonging to [1,2, …, K ] aiming at the input financial index time series samples, taking K as 50 in an experiment, and constructing a time series sample with different scale information;
the input signal is financial index time sequence data with the length of T-200 and the dimension of M-10
Figure BDA0002853150870000061
Wherein, Xi(t)∈R10×1Identifying the characteristic value, x, of the ith sample at time tm,tThe characteristic of the sample at the tth moment of the mth dimension is represented by a time sequence after downsampling
Xi,f＝(Xi(1),Xi(f+1),…,Xi(l*f+1))
Wherein f represents the sliding sampling window size, each window size corresponds to the scale information of one kind of sampling,
Figure BDA0002853150870000071
identifying an l-th downsampling sliding window;
s2, labeling the sub-sequence samples with different scale information obtained after the step S1 by the scale information to obtain a time sequence sample set with labeled information, as shown in FIG. 2, the sub-sequence obtained by sliding different down-sampling window sizes corresponds to information (gray time stamp corresponds to the sampled sub-sequence sample) of multiple scales of the time sequence sample, the sub-sequence obtained by down-sampling the specific window size corresponds to a scale information label as a supervision signal, the multiple signals correspond to the characteristic information of the time sequence sample in different scales, and the multiple signals are the reaction of the multi-scale characteristic information of the time sequence sample;
and S3, extracting the first layer convolution characteristics. And taking a time sequence sample set with scale information marks collected after down sampling as a training set, taking the training set as input information of an agent monitoring network, and accessing the training set to the first layer of convolutional layer to obtain primary feature extraction. In practical application, a one-dimensional convolution operation is adopted, three types of lengths in the time direction of a convolution kernel are set to be 1, 3 and 5 and are selected as parameter candidate sets, the dimension M which corresponds to the input channel number and is a time sequence variable is set to be 10, the number of convolution channels is set to be 16, 32 and 64 and are selected as parameter candidate sets, and the optimal hyper-parameter of the proxy monitoring network on a financial index time sequence abnormity detection task is found in a grid search mode;
and S4, pooling the first-layer convolution characteristics. And performing a pooling operation on the convolution characteristics output after the action of the step S3 to improve the robustness and generalization characteristics of the characteristic output, thereby further obtaining pooled characteristics. In the specific implementation, the length of the pooling is 2, and the characteristic sliding step length is also set to be 2;
and S5, extracting the second layer of convolution characteristics. And performing further convolution feature extraction on the pooled features output in the step S4 to obtain deeper sample feature output. In practical application, one-dimensional convolution operation is adopted, three types of lengths in the time direction of a convolution kernel, namely 1, 3 and 5, are set as parameter candidate sets, the number of convolution channels is set to be one of 16, 32 and 64, and the optimal network parameters are found in a grid search mode;
s6, pooling of the second layer convolution characteristics. And performing further pooling operation on the convolution characteristics output in the step S5, and enhancing the robustness and generalization performance of the characteristics to obtain the further pooled characteristics. In the specific implementation, the length of the pooling is 2, and the characteristic sliding step length is also set to be 2;
s7, Softmax classification, connecting the characteristics output in the step S6 to a Softmax network layer, and obtaining the classification of each sample in the time series sample set with the mark information generated by the financial index time series;
s8, training the agent supervision network established in the steps S3-S7, and optimizing the network parameter weights in the steps S3-S7 based on a back propagation algorithm to further obtain internal parameters of the high-precision financial index time sequence abnormity detection network model; specifically, the agent monitoring network uses the size of a downsampling window as a scale information label, and uses a cross entropy loss function as a classification loss function, and the expression is as follows:
Figure BDA0002853150870000081
wherein y represents the size of the sampling window and corresponds to the scale information marking signal of the self-supervision downsampling, K is 50 which is the upper limit of the size of the sampling window, theta is a network parameter,
Figure BDA0002853150870000082
representative sample XiA time sequence subsequence generated after a down-sampling operation with a sliding window size of y,
Figure BDA0002853150870000083
corresponding agent supervision network to sequence sub-sequence under parameter theta
Figure BDA0002853150870000084
The classification predicts the probability of correctness, log is a log function with base 2, Loss (X)iTheta) is sample X of agent supervision network under parameter thetai(iii) a loss of self-supervised classification;
s9, calculating abnormal indexes, namely calculating the classification cross entropy loss of each financial index time series sample generated by downsampling and provided with scale information mark samples based on the optimized agent supervision network, and taking the classification cross entropy loss as the abnormal indexes of the financial index time series samples, wherein the larger the index is, the more likely the sample is to be an abnormal sample sequence;
s10, inputting a test sample, repeating the step S1 to construct a plurality of subsequence samples with different scale information after the test sample is sampled, repeating the step S2 to mark the subsequence samples through the scale information, repeating the steps S3 to S7 to obtain a classification result of the subsequence samples on the network by a trained agent monitoring network, repeating the step S9 to calculate the classification cross entropy loss of the agent monitoring network under the test sample, taking the classification cross entropy loss as an index of financial index time sequence abnormity detection, wherein the sample with a larger index value represents that the test sample is more likely to be abnormal because the discriminant characteristics extracted by the test sample have larger difference with the normal training sample characteristics.
As shown in fig. 2, the financial index time series abnormality detection method based on the self-supervised discriminant network disclosed in this embodiment has an advantage of capturing multi-scale information of time series samples compared with a generative framework. The generative-based framework performs only single reconstruction modeling on time-series samples from an input sample space, omits multi-scale information of the time-series samples, and is difficult to capture rich information in the time-series samples. The invention realizes the extraction of multi-scale information in time sequence samples based on the self-monitoring down-sampling module, and realizes the feature capture of the multi-scale information of the time sequence samples by utilizing the agent monitoring network module. The financial index time sequence abnormity detection method based on the self-supervision discriminant network has superior performance in a financial index time sequence abnormity detection task, and has the advantage of high calculation efficiency compared with a traditional recurrent neural network-based generating abnormity detection framework.
The above embodiments are preferred embodiments of the present invention, but the present invention is not limited to the above embodiments, and any other changes, modifications, substitutions, combinations, and simplifications which do not depart from the spirit and principle of the present invention should be construed as equivalents thereof, and all such changes, modifications, substitutions, combinations, and simplifications are intended to be included in the scope of the present invention.

Claims (5)

1. A financial index time sequence abnormity detection method based on an automatic supervision discriminant network is characterized by comprising the following steps:
s1, obtaining subsequence samples of the financial index time sequence under different scale information through sliding sampling by aiming at the input financial index time sequence samples through a window with the size of f;
s2, marking the subsequence samples under the financial index time sequence different scale information obtained in the step S1 through the scale information to obtain a time sequence sample set with marking information;
s3, extracting first-layer convolution characteristics, namely, taking the time sequence sample set with the marked information obtained in the step S2 as a training set, taking the training set as input information of a network, and accessing a first-layer convolution layer to obtain convolution characteristics;
s4, pooling the first-layer convolution characteristics, and performing pooling operation on the convolution characteristics output in the step S3 to obtain pooled characteristics;
s5, performing second-layer convolution feature extraction, namely performing convolution feature extraction on the pooled features output in the step S4;
s6, pooling the second layer of convolution characteristics, and performing pooling operation on the convolution characteristics output in the step S5 to obtain pooled characteristics;
s7, Softmax classification, connecting the pooled features output in the step S6 to a Softmax network layer, and obtaining classification of each sample in the time series sample set with the mark information generated by the financial index time series;
s8, training the agent supervision network established in the steps S3-S7, and optimizing the network parameter weights in the steps S3-S7 based on a back propagation algorithm to further obtain internal parameters of the high-precision financial index time sequence abnormity detection network model;
s9, calculating abnormal indexes, namely calculating the classified cross entropy loss of each financial index time series sample generated by downsampling and provided with scale information mark samples based on the optimized agent supervision network, and taking the classified cross entropy loss as the abnormal indexes of the financial index time series samples;
s10, inputting a test sample, repeating the step S1 to construct a plurality of subsequence samples with different scale information after the test sample is sampled, then repeating the step S2 to mark the subsequence samples through the scale information, then repeating the steps S3 to S7 to obtain the classification result of the subsequence samples on the network by a trained agent monitoring network, repeating the step S9 to calculate the classification cross entropy loss of the agent monitoring network under the test sample, and taking the classification cross entropy loss as the index of financial index time sequence abnormity detection.
2. The method according to claim 1, wherein in step S1, for the inputted financial index time series samples, sliding the samples through a window with a size f to obtain subsequence samples of the financial index time series with different scale information, and each sampling window has a size corresponding to one of the scale information:
f∈{1,2,…,K-1,K}
wherein, K corresponds to the upper limit of the size of the down-sampling window and also corresponds to the total number of the scale information in the self-supervision down-sampling module.
3. The method as claimed in claim 1, wherein the financial index time series abnormality detection method based on the self-supervised discriminant network comprises the step of collecting the financial index time series signal samples in step S2
Figure FDA0002853150860000021
Obtaining a time series sample set through an auto-supervision down-sampling module,
representing an input signal as financial index time series data of length T dimension M
Figure FDA0002853150860000022
Wherein, Xi(t)∈RM×1Identifying the characteristic value, x, of the ith sample at time tm,tThe characteristics of the sample at the t-th time of the mth dimension are represented as the time sequence after the downsampling
Xi,f＝(Xi(1),Xi(f+1),…,Xi(l*f+1))
Where f denotes the sliding sampling window size,
Figure FDA0002853150860000035
the ith downsampling sliding window is identified.
4. The method as claimed in claim 1, wherein in steps S3 to S7, the method for detecting financial index time series abnormality based on an unsupervised discriminant network realizes feature extraction of the input financial index time series through a proxy supervised network module, specifically: features are extracted by convolution and pooling, and discriminant classification of financial index time sequences is realized by combining a Softmax network layer.
5. The financial index time series abnormality detection method based on the self-supervision discriminant network as claimed in claim 1, wherein in the step S8, the proxy supervision network module is optimized by using the down-sampling scale information as the classification label and using the cross entropy loss function as the classification loss function, and the expression is as follows:
Figure FDA0002853150860000031
wherein y represents the size of the sampling window, K is the upper limit of the size of the sampling window, theta is a network parameter,
Figure FDA0002853150860000032
representative sample XiA time sequence subsequence generated after a down-sampling operation with a sliding window size of y,
Figure FDA0002853150860000033
corresponding agent supervision network to sequence sub-sequence under parameter theta
Figure FDA0002853150860000034
The classification predicts the probability of correctness, log is a log function with base 2, Loss (X)iTheta) is used as a sample X of the agent monitoring network module under the parameter thetaiLoss of self-supervised classification.
CN202011543431.4A 2020-12-23 2020-12-23 Financial index time sequence anomaly detection method based on self-supervision discriminant network Active CN112529678B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202011543431.4A CN112529678B (en) 2020-12-23 2020-12-23 Financial index time sequence anomaly detection method based on self-supervision discriminant network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202011543431.4A CN112529678B (en) 2020-12-23 2020-12-23 Financial index time sequence anomaly detection method based on self-supervision discriminant network

Publications (2)

Publication Number Publication Date
CN112529678A true CN112529678A (en) 2021-03-19
CN112529678B CN112529678B (en) 2023-06-23

Family

ID=74976092

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202011543431.4A Active CN112529678B (en) 2020-12-23 2020-12-23 Financial index time sequence anomaly detection method based on self-supervision discriminant network

Country Status (1)

Country Link
CN (1) CN112529678B (en)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113240011A (en) * 2021-05-14 2021-08-10 烟台海颐软件股份有限公司 Deep learning driven abnormity identification and repair method and intelligent system
CN114090396A (en) * 2022-01-24 2022-02-25 华南理工大学 Cloud environment multi-index unsupervised anomaly detection and root cause analysis method
CN115099321A (en) * 2022-06-17 2022-09-23 杭州电子科技大学 Bidirectional autoregression unsupervised pre-training fine-tuning type abnormal pollution discharge monitoring method and application

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106411597A (en) * 2016-10-14 2017-02-15 广东工业大学 Network traffic abnormality detection method and system
WO2018121157A1 (en) * 2016-12-29 2018-07-05 北京神州绿盟信息安全科技股份有限公司 Network traffic anomaly detection method and apparatus
CN111814685A (en) * 2020-07-09 2020-10-23 西安电子科技大学 Hyperspectral image classification method based on double-branch convolution self-encoder
US20200393812A1 (en) * 2019-06-14 2020-12-17 Renesas Electronics Corporation Abnormality detection system, abnormality detection apparatus, and abnormality detection method

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106411597A (en) * 2016-10-14 2017-02-15 广东工业大学 Network traffic abnormality detection method and system
WO2018121157A1 (en) * 2016-12-29 2018-07-05 北京神州绿盟信息安全科技股份有限公司 Network traffic anomaly detection method and apparatus
US20200393812A1 (en) * 2019-06-14 2020-12-17 Renesas Electronics Corporation Abnormality detection system, abnormality detection apparatus, and abnormality detection method
CN111814685A (en) * 2020-07-09 2020-10-23 西安电子科技大学 Hyperspectral image classification method based on double-branch convolution self-encoder

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
连鸿飞;张浩;郭文忠;: "一种数据增强与混合神经网络的异常流量检测", 小型微型计算机系统 *

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113240011A (en) * 2021-05-14 2021-08-10 烟台海颐软件股份有限公司 Deep learning driven abnormity identification and repair method and intelligent system
CN113240011B (en) * 2021-05-14 2023-04-07 烟台海颐软件股份有限公司 Deep learning driven abnormity identification and repair method and intelligent system
CN114090396A (en) * 2022-01-24 2022-02-25 华南理工大学 Cloud environment multi-index unsupervised anomaly detection and root cause analysis method
CN114090396B (en) * 2022-01-24 2022-06-14 华南理工大学 Cloud environment multi-index unsupervised anomaly detection and root cause analysis method
CN115099321A (en) * 2022-06-17 2022-09-23 杭州电子科技大学 Bidirectional autoregression unsupervised pre-training fine-tuning type abnormal pollution discharge monitoring method and application
CN115099321B (en) * 2022-06-17 2023-08-04 杭州电子科技大学 Bidirectional autoregressive non-supervision pretraining fine-tuning type pollution discharge abnormality monitoring method and application

Also Published As

Publication number Publication date
CN112529678B (en) 2023-06-23

Similar Documents

Publication Publication Date Title
CN112529678B (en) Financial index time sequence anomaly detection method based on self-supervision discriminant network
CN109639739B (en) Abnormal flow detection method based on automatic encoder network
CN111914873A (en) Two-stage cloud server unsupervised anomaly prediction method
CN112561165A (en) Multidimensional time series data prediction method based on combined model
CN111243674B (en) Base sequence identification method, device and storage medium
CN111273623A (en) Fault diagnosis method based on Stacked LSTM
CN113673346A (en) Motor vibration data processing and state recognition method based on multi-scale SE-Resnet
CN112329536A (en) Single-sample face recognition method based on alternative pair anti-migration learning
CN111835738B (en) Network abnormal flow automatic detection method based on time series mining
CN115165366A (en) Variable working condition fault diagnosis method and system for rotary machine
CN116167370A (en) Log space-time characteristic analysis-based distributed system anomaly detection method
CN117596191A (en) Power Internet of things abnormality detection method, device and storage medium
CN116383747A (en) Anomaly detection method for generating countermeasure network based on multi-time scale depth convolution
CN116992757A (en) Wellhead pressure prediction method and device based on deep learning and rolling optimization
CN115810127A (en) Small sample image classification method based on supervision and self-supervision combined contrast learning
CN113326371B (en) Event extraction method integrating pre-training language model and anti-noise interference remote supervision information
Oh et al. Multivariate Time Series Open-Set Recognition Using Multi-Feature Extraction and Reconstruction
CN117725491B (en) SCINet-based power system fault state detection and classification method
CN116361673B (en) Quasi-periodic time sequence unsupervised anomaly detection method, system and terminal
CN114021663B (en) Industrial process off-line data segmentation method based on sequence local discrimination information mining network
CN114255373B (en) Sequence anomaly detection method, device, electronic equipment and readable medium
CN117942038A (en) Sleep stage identification system integrating cross prediction and discrimination tasks
Wang et al. Multi-Scale Time Series Ensemble Learning for Information System Anomaly Detection
CN117611901A (en) Small sample image classification method based on global and local contrast learning
CN116861214A (en) Health state identification method and system based on convolution long short-time memory network

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