CN107256333A - A kind of Argo profile anomaly detection methods based on prediction and dynamic threshold - Google Patents
A kind of Argo profile anomaly detection methods based on prediction and dynamic threshold Download PDFInfo
- Publication number
- CN107256333A CN107256333A CN201710382710.9A CN201710382710A CN107256333A CN 107256333 A CN107256333 A CN 107256333A CN 201710382710 A CN201710382710 A CN 201710382710A CN 107256333 A CN107256333 A CN 107256333A
- Authority
- CN
- China
- Prior art keywords
- mrow
- msub
- section
- section point
- argo
- 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.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
- G16Z—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS, NOT OTHERWISE PROVIDED FOR
- G16Z99/00—Subject matter not provided for in other main groups of this subclass
Landscapes
- Testing Or Calibration Of Command Recording Devices (AREA)
Abstract
The invention discloses a kind of Argo profile anomaly detection methods based on prediction and dynamic threshold, for given Argo section time serieses PN={ p1,p2,…,pn, its profiling observation value sequence is ON={ o1,o2,…,on, (1≤i≤n).Define section point p to be measurediK neighbour's section point sequences ζi, ARMA forecast models are set up, by ζiT is obtained as input parameteriMoment corresponding section predicted valueT is calculated using central-limit theoremiMoment correspondence threshold value thi.By judging tiMoment corresponding profiling observation value oiWhether t judged in confidential intervaliMoment correspondence section point p to be measurediIt is whether abnormal.If oiIn threshold range thiIt is interior, then piFor normal cross section point, flag=1 is made；If oiIn threshold range thiOutside, then piFor abnormal profile point, flag=0 is made.Sliding window one is moved afterwards, said process is repeated, until having detected all section points to be measured.This method can accurately judge normal cross section point or abnormal profile point, and the reliability of abnormality detection is high.
Description
Technical field
The present invention relates to oceanographic data abnormality detection technical field, specifically a kind of Argo based on prediction and dynamic threshold
Profile anomaly detection method.
Background technology
(ARRAY for REAL-TIME GEOSTROPHIC OCEANOGRAPHY, ground turns oceanography real-time monitored to Argo
Battle array) it is first real-time, high-resolution three-dimensional ocean observation net in the whole world, because of its activity and scale, it may be said that be ocean
The revolution observed in history.Since implementing from 2000, temperature, the salt profile data volume obtained was received than 100 years in the past
More than the total amount of collection.But it is due to the randomness and disposability of Argo buoys itself, it is difficult to be corrected for sensor, so
Probably there is data error or mistake.And in order to further serve scientific research and other social applications, it is necessary to for data
Mass accuracy have certain guarantee, so the abnormality detection of Argo cross-sectional datas is just particularly important.
Abnormality detection (Anomaly Detection) is exactly to identify the mistake for having marked difference with history or expectancy model
Journey, abnormal data may be produced because sensor gathers the reasons such as mistake or data transfer.For abnormality detection, learn both at home and abroad
Person has done substantial amounts of research.Network traffics are divided into high frequency flow and low frequency flow by Haiyan Wang using wavelet decomposition,
RVM (Relevance Vector Machine, Method Using Relevance Vector Machine) and ARMA (Autoregressive is utilized respectively again
Moving average, ARMA model) model realization for data prediction, finally using prediction data obtain
Threshold value realizes the abnormality detection for flow.Agrawal S etc. are utilized respectively the pre- of a variety of forecast model computing environment sensors
Measured value obtains confidential interval, realizes the abnormality detection that data are observed for sensor.Yu Yufeng etc. utilizes sliding window and LN
Forecast model obtains the predicted value of hydrographic data, and by comparative observation value and predicted value, realizes for the different of hydrographic data
Often detection.In summary, the relation between comparison prediction value and observation can effectively realize the exception for time series
Detection.
ARMA is a kind of typical, most widely used time series models.Wang K etc. realize estimation using arma modeling
The purpose of observation noise variance.The exception that Li Lei etc. can effectively detect the appearance of earthquake ionosphere by arma modeling is disturbed
It is dynamic, and with higher precision.The method that Shen Y etc. are combined using least square method (LS) with arma modeling, is realized
For the prediction of polarity parameters important in satellite navigation positioning and spacecraft tracking.Chen Chunyan etc. is predicted using arma modeling
The popularity of line TV play, the accuracy that predicts the outcome is high, and with higher reference value.Therefore, can using arma modeling
To effectively realize time series forecasting, and it is a kind of preferable forecast model with higher prediction accuracy.
With developing rapidly for ocean observation technology, the growth of oceanographic data exponentially.On Argo cross-sectional datas
Abnormality detection, also has substantial amounts of scholar to be studied work.Wong A etc. are proposed one kind and picked using conventional threshold values determination methods
Except abnormal data, required section calculation formula is：Detected value=| V2- (V3+V1)/2 |-| (V3-V1)/2 |, wherein V2
Numerical value to be measured is represented, V1, V3 represent the up and down two section point data adjacent with V2.When test data is salinity, work as pressure<
During 500dbar, detected value is more than 0.9PSU, or as pressure >=500dbar, and detected value then treats measured value mark more than 0.3PSU
It is designated as exception.As can be seen that this method carries out abnormality detection just for single cross-sectional data, and sampled point inspection is not accounted for
Survey the change of depth interval.Wang Huizan etc. judges that the method being combined is realized to Argo using " three times standard deviation " with conventional threshold values
Cross-sectional data abnormality detection, this method combines neighbouring cross-sectional data, compensate for the deficiency of conventional threshold values decision method, but should
Method uses global static threshold, there is the erroneous judgement or leakage that some abnormal profile points are caused when threshold value sets improper
Sentence.History thermohaline for many years is observed data (Nan Senzhan and CTD data) and obtains certain by least square fitting by Ji Fengying etc.
The T-S relation model in area, and quality control is carried out to Argo cross-sectional datas using the model.But it is due to thermohaline data
Nonlinear feature, data characteristics, therefore the T-S relation model essence obtained can not be extracted well using least square method
Exactness is not high, causes the reliability of abnormality detection also not high.
The content of the invention
It is an object of the invention to provide a kind of Argo profile anomaly detection methods based on prediction and dynamic threshold, this method
Normal cross section point or abnormal profile point can accurately be judged, the reliability of abnormality detection is high.
Realizing the technical scheme of the object of the invention is：
A kind of Argo profile anomaly detection methods based on prediction and dynamic threshold, specifically include following steps：
1) Argo files to be measured are chosen, the data attribute in Argo files is pre-processed, data attribute is extracted, and
The section time series for setting the Argo files of section point to be measured is Pn={ p1,p2,…pi,…,pn, k- neighbour's section point sequences
Width k and confidence level p, the profiling observation value sequence of section point to be measured is On={ o1,o2,…,oi,…on, (1≤i≤n)；
2) section point p to be measured is definediK- neighbour's section point sequences ζi={ pi-2k,pi-2k+1,…,pi-1}；
3) ARMA forecast models are set up, by ζiT is obtained as input parameteriMoment corresponding section predicted value
4) t is calculated using center line limit theoremiMoment corresponding threshold value thi, obtain confidential interval
5) in order to travel through all section subsequences, from sliding window technique, judgment step 1) in profiling observation value
oiWhether in step 4) in obtained confidential interval, if oiIn threshold value thiIn the range of, then piFor normal cross section point, flag is made
=1, a sliding window is moved afterwards, step 5 is performed)；If oiIn threshold range thiOutside, then piFor abnormal profile point, flag=is made
0, a sliding window is moved afterwards, with step 3) obtained section predicted valueReplace oi, continue executing with step 4), until having detected
After all section points to be measured, step 5 is performed)；
5) judge by step 4) after obtained normal section point piWhether i is met<N, if meeting, i=i+1, is jumped
To step 3)；If it is not satisfied, then performing step 6)；
6) to PnAbnormality detection terminate, export
By above-mentioned steps, complete to detect the profile anomaly of Argo files.
Step 4) in, described confidential interval determines that it is calculated as follows according to central-limit theorem：
In above-mentioned formula (1),For tiMoment corresponding predicted value, ta/2(2k-1) is to obey the t that the free degree is 2k-1 to divide
The cloth value corresponding when confidence level is p, siFor the standard deviation of the residual error of ARMA forecast models, k is the width of sliding window；Cause
This, threshold range is：
In above-mentioned formula (2), threshold value lower bound isThe threshold value upper bound is
Step 5) in, described sliding window is set up, and choose unilateral window as k- neighbours based on k- nearest neighbouring rules
Sliding window.
Described unilateral window, the length for referring to the section point left side to be measured is 2k section subsequence, i.e., included in it
All elements be all data after abnormality detection.
Beneficial effect：The method for detecting abnormality specificity that the present invention is provided is high, is maintained at more than 99%, illustrates this method
It can be very good to detect that normal cross section point is normal.But, as confidence level p >=95%, susceptibility only has 60% or so, says
It is bright when fiducial interval range set it is excessive when, the effect that can really judge abnormal profile point is not fine, and works as confidence level
It is arranged on p ∈ [80%,When in the range of 90%], susceptibility can maintain more than 80%, and with sliding window width k's
Increase, susceptibility is in rising trend.In addition, the accuracy of this method maintains more than 99% always, illustrate that this method can be accurate
True judges normal cross section point or abnormal profile point, and other indexs also maintain higher level.As sliding window width k
∈[10,20], confidence level p ∈ [80%,When 90%], susceptibility can reach more than 85%, and specificity can be maintained
99%, the degree of accuracy is more than 99%, and illustrating the method for detecting abnormality of the present invention has higher reliability.
Brief description of the drawings
Fig. 1 is a kind of Argo profile anomaly detection method flow charts based on prediction and dynamic threshold；
Fig. 2 is abnormality detection mitigation strategy schematic diagram；
Fig. 3 is the Argo section files of embodiment；
Fig. 4 is 15 ° of N~18 ° N, 138 ° of W~141 ° W salt profile sequences；
Fig. 5 predicts the outcome for salt profile sequence；
Fig. 6 is the abnormality detection result under the conditions of k=10, p=95%；
Fig. 7 is the ROC curve of distinct methods.
Embodiment
The present invention is further elaborated with reference to the accompanying drawings and examples, but is not limitation of the invention.
As shown in figure 1, a kind of Argo profile anomaly detection methods based on prediction and dynamic threshold, specifically include following step
Suddenly：
1) Argo files to be measured are chosen, the data attribute in Argo files is pre-processed, data attribute is extracted, and
The section time series for setting the Argo files of section point to be measured is Pn={ p1,p2,…pi,…,pn, k- neighbour's section point sequences
Width k and confidence level p, the profiling observation value sequence of section point to be measured is On={ o1,o2,…,oi,…on, (1≤i≤n)；
2) section point p to be measured is definediK- neighbour's section point sequences ζi={ pi-2k,pi-2k+1,…,pi-1}；
3) ARMA forecast models are set up, by ζiT is obtained as input parameteriMoment corresponding section predicted value
4) t is calculated using center line limit theoremiMoment corresponding threshold value thi, obtain confidential interval
5) in order to travel through all section subsequences, from sliding window technique, judgment step 1) in profiling observation value
oiWhether in step 4) in obtained confidential interval, if oiIn threshold value thiIn the range of, then piFor normal cross section point, flag is made
=1, a sliding window is moved afterwards, uses piSubstitute ζi={ pi-2k,pi-2k+1,…,pi-1In pi-1, obtain new k- arest neighbors windows
Mouth ζi+1, perform step 5)；If oiIn threshold range thiOutside, then piFor abnormal profile point, flag=0 is made, a sliding window is moved afterwards
Mouthful, use piSubstitute ζi={ pi-2k,pi-2k+1,…,pi-1In pi-1, obtain new k- arest neighbors windows ζi+1, with step 3) obtain
Section predicted valueReplace original observation oi, i.e.,As shown in Fig. 2 continuing executing with step 4), until detection
After complete all section points to be measured, step 5 is performed)；
5) judge by step 4) after obtained normal section point piWhether i is met<N, if meeting, i=i+1, is jumped
To step 3)；If it is not satisfied, then performing step 6)；
6) to PnAbnormality detection terminate, export
By above-mentioned steps, complete to detect the profile anomaly of Argo files.
Step 4) in, described confidential interval determines that it is calculated as follows according to central-limit theorem：
In above-mentioned formula (1),For tiMoment corresponding predicted value, ta/2(2k-1) is to obey the t that the free degree is 2k-1 to divide
The cloth value corresponding when confidence level is p, siFor the standard deviation of the residual error of ARMA forecast models, k is the width of sliding window；Cause
This, threshold range is：
In above-mentioned formula (2), threshold value lower bound isThe threshold value upper bound is
Step 5) in, described sliding window is set up, and choose unilateral window as k- neighbours based on k- nearest neighbouring rules
Sliding window.
Described unilateral window, the length for referring to the section point left side to be measured is 2k section subsequence, i.e., included in it
All elements be all data after abnormality detection.
Embodiment：
Tested from the whole world Argo buoy section data in 2016 obtained from Chinese Argo real time datas center, it is real
Data are tested with .dat representation of file, as shown in figure 3, the experimental situation that the present embodiment is used is MyEclipse2016.
From figure 3, it can be seen that multiple attributes are included in original Argo section files, mainly including temperature, pressure, salt
Degree.The inventive method is can not directly to detect original Argo sections file, it is necessary to be pre-processed for data, needed for extracting
The attribute wanted.This experiment is verified using salinity attribute, but the variation tendency of salinity can become with the change of longitude and latitude
Change, so this experiment is divided Salinity Data for 3 ° × 3 ° according to longitude and latitude grid, then the data chosen in a certain grid are entered
Row experimental verification.
As shown in figure 4, be that longitude and latitude is salt profile data and curves in the range of 15 ° of N~18 ° N, 138 ° of W~141 ° W,
2011 section points are had, abscissa is section point number, and ordinate is salinity observation, and as shown in Figure 4, salt profile has
Periodically, data are integrally steady, but there is also some obvious suspicious points.
The salinity Argo section sequences P inputted in the range of 15 ° of N~18 ° N in 2016,138 ° of W~141 ° Wn={ p1,
p2,…,pi,…,pn, the width k, confidence level p of k- neighbour's section point sequences.
Export Argo section flags sequence
Comprise the concrete steps that：
Step 1 defines salt profile point pi to be measured k- arest neighbors windows ζi={ pi-2k,pi-2k+1,…,pi-1}。
Step 2 sets up ARMA forecast models, inputs ζiAs input parameter, the predicted value for obtaining salinity is calculated
Step 3 is according to the salinity predicted value obtained by formula (2) and step 2Calculate the corresponding salinity upper boundAnd salinity
Lower bound
Step 4 judges salt profile point p to be measurediIt is whether abnormal.If salinity observation oiIn the boundary that step 3 is obtained,
Then point piFor normal point, flag=1 is made, step 5 is jumped to；Otherwise point piFor abnormity point, flag=0 is made, step 6 is jumped to.
Step 5 uses p by one sliding window of rear shiftingiSubstitute ζi={ pi-2k,pi-2k+1,…,pi-1In pi-1, obtain
New k- arest neighbors windows ζi+1, jump to step 7).
Step 6 uses p by one sliding window of rear shiftingiSubstitute ζi={ pi-2k,pi-2k+1,…,pi-1In pi-1, obtain
New k- arest neighbors windows ζi+1, and with predicted valueReplace original observation oi, i.e.,
If step 7 i<N, then i=i+1, skips to step 2；Otherwise, to PNAbnormality detection terminate, export
A, prediction
In the method, prediction cross-sectional data is core procedure, therefore, and in order to detect prediction effect, this experiment is using equal
Square error (Root Mean Square Error, RMSE) and relative root-mean-square error (Relative Root Mean
Square Error, RRMSE) come for the carry out quantitative evaluation that predicts the outcome.
Root-mean-square error is expressed as：
It is expressed as with respect to root-mean-square error：
Wherein, xiFor profiling observation value,For section predicted value, N is sample number.
It is as shown in table 1 below for ARMA forecast models and LN (Single-layer linear network predictor,
The linear Network Prediction Model of individual layer) forecast model salt profile predicated error under conditions of sliding window width k=10, from table
1 as can be seen that the RMSE and RRMSE of ARMA forecast models are respectively less than LN forecast models, with higher accuracy.
The salt profile predicated error of table 1
As shown in table 2 below is salt profile predicated error of the ARMA forecast models under different sliding window width K,
From table 2 it can be seen that with sliding window width K continuous increase, RMSE and RRMSE are steadily decreasing, because with input
Prediction Parameters increase, predicting the outcome can be more accurate.
The salt profile predicated error of the different sliding window width of table 2
Fig. 5 show predicting the outcome for salt profile sequence, from fig. 5, it can be seen that prediction data mostly with original observation
Data are sufficiently close to, and only partial data and observation data has larger deviation, illustrates that the precision of prediction of this forecast model is higher,
In the abnormality detection for being effectively applied to profiling observation data.
B. abnormality detection
This method is can be seen that from predicting the outcome and uses ARMA forecast models, can preferably be predicted the outcome, because
This, predicts the outcome according to this, it is possible to achieve for the abnormality detection of salt profile sequence.Sliding window width k=10, confidence level
Abnormality detection result under the conditions of p=95% is as shown in Figure 6.
The result of abnormality detection is probably different under different sliding window width and confidential interval.In order to effective
Evaluation this method, abnormality detection result is divided into 4 classes herein, as shown in table 3.Wherein, TN and TP are desirable to the result occurred,
And FN and FP are to judge the result that mistake occurs occur.
The testing result of table 3 is classified
Susceptibility (Sensitivity, SEN) is to describe the probability that abnormality detection finds real abnormal profile point, and it is public
Formula is described as follows：
Specificity (Specificity, SPE) is to describe the probability that normal cross section point is properly separated out, and its formula is retouched
State as follows：
Positive predictive value (Positive Predictive Value, PPV) is that the abnormal profile point that description is detected is true
The probability of positive abnormal profile point, its formula is described as follows：
Negative predictive value (Negative Predictive Value, NPV) is that the normal cross section point that description is detected is true
The probability of positive normal cross section point, its formula is described as follows：
Accuracy (Accuracy, ACC) is describe that abnormal profile point and normal cross section point be detected correctly general
Rate, its formula is described as follows：
Comparative result of the different parameters of table 4 to abnormality detection
Classified according to the testing result of table 3 and formula (5)~(9), the abnormality detection result such as institute of table 4 under different parameters
Show.As known from Table 4, under the conditions of selection sliding window width k=10, confidence level p=95%, method for detecting abnormality of the invention
Abnormal profile point 10 (TP=10) can be correctly detected, the section point that normal cross section point is appropriately determined out there are 1994
(TN=1994), but have 2 normal cross section points by mistake judgement be abnormal profile point (FP=2), finally, also 5 are different
Normal section point is not detected among out (FN=5).
The specificity that the method for detecting abnormality of the present invention is can be seen that by the assessment result of contrast table 4 is very high, protects always
Hold more than 99%, illustrate that this method can be very good to detect that normal cross section point is normal.But, when confidence level p >=95%
When, susceptibility only has 60% or so, illustrates, when fiducial interval range sets excessive, can really judge abnormal profile point
Effect is not fine, and when confidence level is arranged in the range of p ∈ [80%, 90%], susceptibility can maintain more than 80%,
And with sliding window width k increase, susceptibility is in rising trend.In addition, the accuracy of this method is maintained always
More than 99%, illustrate that this method can accurately judge normal cross section point or abnormal profile point, other indexs are also maintained
Higher level.As sliding window width k ∈ [10,20], confidence level p ∈ [80%, 90%], susceptibility can reach 85%
More than, and specificity can maintain 99%, the degree of accuracy more than 99%, illustrate the present invention method for detecting abnormality have compared with
High reliability.
In order to preferably assess the method for detecting abnormality of the present invention, by the inventive method and other abnormality detection sides
Method passes through on same " Receiver Operating Characteristics " (Receiver Operating Characteristic, ROC) curve.
In ROC curve, abscissa is " false positive rate " (False Positive Rate, FPR), and ordinate is " real ratio " (True
Positive Rate, TPR), both formula are respectively：
When carry out method compares, when a method ROC curve by the ROC curve of another method completely " encasing ",
The performance that the latter then can be explained is better than the former.From formula (11)~(12) as can be seen that FPR is exactly " 1- specificities ", and TPR is just
It is " susceptibility ".In abnormality detection, it is generally desirable to obtain high TPR, low FPR, it is meant that curve is closer to the upper left of reference axis
Angle, the accuracy of method is higher, and performance is better.
Shown in Fig. 7 is method for detecting abnormality of the method for detecting abnormality with T-S relation model of the present invention, based on " LN "
The method for detecting abnormality of forecast model and the ROC curve comparison diagram of k nearest neighbor method.Fig. 7 shows, the abnormality detection of the inventive method
Effect is better than other method for detecting abnormality.T-S relation model method uses least square fitting history thermohaline data and obtained
Bound realizes abnormality detection, but be due to that the not high effect for causing abnormality detection of accuracy of the model is least preferable
's.Although the effect of k nearest neighbor method is better than T-S relation model method, compared to other two methods, Detection results one
As.It is closer to based on " LN " method predicted with the inventive method Detection results, but is due to the prediction of the selection of this method
Model is concentrated applied to Argo cross-sectional datas, and predicated error is larger, so its Detection results is compared to the inventive method slightly inferior one
Raise.And the ROC curve of the inventive method is always positioned at the top, other three kinds of methods " are encased " completely, so Detection results
Preferably, accuracy is high.
Claims (4)
1. a kind of Argo profile anomaly detection methods based on prediction and dynamic threshold, it is characterised in that specifically include following step
Suddenly：
1) Argo files to be measured are chosen, the data attribute in Argo files is pre-processed, data attribute are extracted, and set
The section time series of the Argo files of section point to be measured is Pn={ p1,p2,…pi,…,pn, the width of k- neighbour's section point sequences
K and confidence level p is spent, the profiling observation value sequence of section point to be measured is On={ o1,o2,…,oi,…on, (1≤i≤n)；
2) section point p to be measured is definediK- neighbour's section point sequences ζi={ pi-2k,pi-2k+1,…,pi-1}；
3) ARMA forecast models are set up, by ζiT is obtained as input parameteriMoment corresponding section predicted value
4) t is calculated using center line limit theoremiMoment corresponding threshold value thi, obtain confidential interval
5) in order to travel through all section subsequences, from sliding window technique, judgment step 1) in profiling observation value oiWhether
In step 4) in obtained confidential interval, if oiIn threshold value thiIn the range of, then piFor normal cross section point, flag=1 is made, after
A sliding window is moved, step 5 is performed)；If oiIn threshold range thiOutside, then piFor abnormal profile point, flag=0 is made, it is rear to move
One sliding window, with step 3) obtained section predicted valueReplace oi, continue executing with step 4), until having detected all
After section point to be measured, step 5 is performed)；
5) judge by step 4) after obtained normal section point piWhether i is met<N, if meeting, i=i+1 skips to step
3)；If it is not satisfied, then performing step 6)；
6) to PnAbnormality detection terminate, export
By above-mentioned steps, complete to detect the profile anomaly of Argo files.
2. according to the method described in claim 1, it is characterised in that step 4), described confidential interval is according to center pole
Reason is limited come what is determined, it is calculated as follows：
<mrow>
<mo>&lsqb;</mo>
<msub>
<mover>
<mi>o</mi>
<mo>~</mo>
</mover>
<mi>i</mi>
</msub>
<mo>-</mo>
<msub>
<mi>t</mi>
<mrow>
<mi>&alpha;</mi>
<mo>/</mo>
<mn>2</mn>
</mrow>
</msub>
<mrow>
<mo>(</mo>
<mn>2</mn>
<mi>k</mi>
<mo>-</mo>
<mn>1</mn>
<mo>)</mo>
</mrow>
<mo>&times;</mo>
<msub>
<mi>s</mi>
<mi>i</mi>
</msub>
<msqrt>
<mrow>
<mn>1</mn>
<mo>+</mo>
<mfrac>
<mn>1</mn>
<mrow>
<mn>2</mn>
<mi>k</mi>
</mrow>
</mfrac>
</mrow>
</msqrt>
<mo>,</mo>
<msub>
<mover>
<mi>o</mi>
<mo>~</mo>
</mover>
<mi>i</mi>
</msub>
<mo>+</mo>
<msub>
<mi>t</mi>
<mrow>
<mi>&alpha;</mi>
<mo>/</mo>
<mn>2</mn>
</mrow>
</msub>
<mrow>
<mo>(</mo>
<mn>2</mn>
<mi>k</mi>
<mo>-</mo>
<mn>1</mn>
<mo>)</mo>
</mrow>
<mo>&times;</mo>
<msub>
<mi>s</mi>
<mi>i</mi>
</msub>
<msqrt>
<mrow>
<mn>1</mn>
<mo>+</mo>
<mfrac>
<mn>1</mn>
<mrow>
<mn>2</mn>
<mi>k</mi>
</mrow>
</mfrac>
</mrow>
</msqrt>
<mo>&rsqb;</mo>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>1</mn>
<mo>)</mo>
</mrow>
</mrow>
In above-mentioned formula (1),For tiMoment corresponding predicted value, ta/2(2k-1) is to obey the t that the free degree is 2k-1 to be distributed in
Confidence level value corresponding when being p, siFor the standard deviation of the residual error of ARMA forecast models, k is the width of sliding window；
Therefore, threshold range is：
<mrow>
<msub>
<mi>th</mi>
<mi>i</mi>
</msub>
<mo>=</mo>
<mo>&lsqb;</mo>
<msub>
<mover>
<mi>o</mi>
<mo>~</mo>
</mover>
<mi>i</mi>
</msub>
<mo>-</mo>
<mn>3</mn>
<mo>&times;</mo>
<msub>
<mi>t</mi>
<mrow>
<mi>&alpha;</mi>
<mo>/</mo>
<mn>2</mn>
</mrow>
</msub>
<mrow>
<mo>(</mo>
<mn>2</mn>
<mi>k</mi>
<mo>-</mo>
<mn>1</mn>
<mo>)</mo>
</mrow>
<mo>&times;</mo>
<msub>
<mi>s</mi>
<mi>i</mi>
</msub>
<msqrt>
<mrow>
<mn>1</mn>
<mo>+</mo>
<mfrac>
<mn>1</mn>
<mrow>
<mn>2</mn>
<mi>k</mi>
</mrow>
</mfrac>
</mrow>
</msqrt>
<mo>,</mo>
<msub>
<mover>
<mi>o</mi>
<mo>~</mo>
</mover>
<mi>i</mi>
</msub>
<mo>+</mo>
<mn>3</mn>
<mo>&times;</mo>
<msub>
<mi>t</mi>
<mrow>
<mi>&alpha;</mi>
<mo>/</mo>
<mn>2</mn>
</mrow>
</msub>
<mrow>
<mo>(</mo>
<mn>2</mn>
<mi>k</mi>
<mo>-</mo>
<mn>1</mn>
<mo>)</mo>
</mrow>
<mo>&times;</mo>
<msub>
<mi>s</mi>
<mi>i</mi>
</msub>
<msqrt>
<mrow>
<mn>1</mn>
<mo>+</mo>
<mfrac>
<mn>1</mn>
<mrow>
<mn>2</mn>
<mi>k</mi>
</mrow>
</mfrac>
</mrow>
</msqrt>
<mo>&rsqb;</mo>
<mo>-</mo>
<mo>-</mo>
<mo>-</mo>
<mrow>
<mo>(</mo>
<mn>2</mn>
<mo>)</mo>
</mrow>
</mrow>
In above-mentioned formula (2), threshold value lower bound isThe threshold value upper bound is
3. according to the method described in claim 1, it is characterised in that step 5), described sliding window, based on k- arest neighbors
Principle is set up, and chooses unilateral window as k- neighbour's sliding windows.
4. method according to claim 3, it is characterised in that described unilateral window, refers to the section point left side to be measured
Length is 2k section subsequence, i.e. all elements included in it are all the data after abnormality detection.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710382710.9A CN107256333A (en) | 2017-05-26 | 2017-05-26 | A kind of Argo profile anomaly detection methods based on prediction and dynamic threshold |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710382710.9A CN107256333A (en) | 2017-05-26 | 2017-05-26 | A kind of Argo profile anomaly detection methods based on prediction and dynamic threshold |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN107256333A true CN107256333A (en) | 2017-10-17 |
Family
ID=60027740
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710382710.9A Pending CN107256333A (en) | 2017-05-26 | 2017-05-26 | A kind of Argo profile anomaly detection methods based on prediction and dynamic threshold |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107256333A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109141380A (en) * | 2018-09-19 | 2019-01-04 | 天津大学 | A kind of hydrographic data processing method based on CTD instrument |
| CN110569912A (en) * | 2019-09-09 | 2019-12-13 | 自然资源部第一海洋研究所 | Method for removing singular values of observation data of sea water profile |
| CN110648030A (en) * | 2019-10-31 | 2020-01-03 | 吉林大学 | Method and device for predicting seawater temperature |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101651568A (en) * | 2009-07-01 | 2010-02-17 | 青岛农业大学 | Method for predicting network flow and detecting abnormality |
| CN104994539A (en) * | 2015-06-30 | 2015-10-21 | 电子科技大学 | Wireless sensor network traffic abnormality detection method based on ARIMA model |
-
2017
- 2017-05-26 CN CN201710382710.9A patent/CN107256333A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101651568A (en) * | 2009-07-01 | 2010-02-17 | 青岛农业大学 | Method for predicting network flow and detecting abnormality |
| CN104994539A (en) * | 2015-06-30 | 2015-10-21 | 电子科技大学 | Wireless sensor network traffic abnormality detection method based on ARIMA model |
Non-Patent Citations (3)
| Title |
|---|
| HAIYAN WANG: "Anomaly Detection of Network Traffic Based on Prediction and Self-Adaptive Threshold", 《INTERNATIONAL JOURNAL OF FUTURE GENERATION COMMUNICATION AND NETWORKING》 * |
| 余宇峰等: "基于滑动窗口预测的水文时间序列异常检测", 《计算机应用》 * |
| 刘仁山等: "含自适应阈值的ARMA网络流量异常检测算法", 《信阳师范学院学报：自然科学版》 * |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109141380A (en) * | 2018-09-19 | 2019-01-04 | 天津大学 | A kind of hydrographic data processing method based on CTD instrument |
| CN110569912A (en) * | 2019-09-09 | 2019-12-13 | 自然资源部第一海洋研究所 | Method for removing singular values of observation data of sea water profile |
| CN110569912B (en) * | 2019-09-09 | 2022-02-01 | 自然资源部第一海洋研究所 | Method for removing singular values of observation data of sea water profile |
| CN110648030A (en) * | 2019-10-31 | 2020-01-03 | 吉林大学 | Method and device for predicting seawater temperature |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN103974311B (en) | Based on the Condition Monitoring Data throat floater detection method for improving Gaussian process regression model | |
| CN103136539B (en) | Ground net corrosion speed grade Forecasting Methodology | |
| CN104156615A (en) | Sensor test data point anomaly detection method based on LS-SVM | |
| CN112033463B (en) | Nuclear power equipment state evaluation and prediction integrated method and system | |
| CN111767517B (en) | BiGRU multi-step prediction method, system and storage medium applied to flood prediction | |
| CN109543356A (en) | Consider the ocean interior temperature-salinity structure remote sensing inversion method of Space atmosphere | |
| CN109447152B (en) | Fault prediction method based on Monte Carlo tree search and neural network | |
| CN102855634A (en) | Image detection method and image detection device | |
| CN111191191A (en) | Construction method of combined model for accurately predicting deformation effect of concrete dam | |
| CN107256333A (en) | A kind of Argo profile anomaly detection methods based on prediction and dynamic threshold | |
| CN112558185A (en) | Bidirectional GRU typhoon track intelligent prediction and forecast system based on attention mechanism, computer equipment and storage medium | |
| CN105005822A (en) | Optimal step length and dynamic model selection based ultrahigh arch dam response prediction method | |
| Zhao et al. | On-line least squares support vector machine algorithm in gas prediction | |
| CN105844217A (en) | Multi-target tracking method based on measure-driven target birth intensity PHD (MDTBI-PHD) | |
| CN116341901B (en) | Integrated evaluation method for landslide surface domain-monomer hazard early warning | |
| CN116415481A (en) | Regional landslide hazard risk prediction method and device, computer equipment and storage medium | |
| CN115100819B (en) | Landslide hazard early warning method and device based on big data analysis and electronic equipment | |
| CN112529053A (en) | Short-term prediction method and system for time sequence data in server | |
| CN115185937A (en) | SA-GAN architecture-based time sequence anomaly detection method | |
| Ye et al. | A deep learning-based method for automatic abnormal data detection: Case study for bridge structural health monitoring | |
| CN117521512A (en) | Bearing residual service life prediction method based on multi-scale Bayesian convolution transducer model | |
| CN116522594A (en) | Time self-adaptive transient stability prediction method and device based on convolutional neural network | |
| CN113887119A (en) | River water quality prediction method based on SARIMA-LSTM | |
| Mu et al. | Bayesian Learning–Based Data Analysis of Uniaxial Compressive Strength of Rock: Relevance Feature Selection and Prediction Reliability Assessment | |
| CN115618173A (en) | Geological drilling process anomaly detection method based on interval-augmented Mahalanobis distance |
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 | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20171017 |