WO2016155241A1 - 基于Kalman滤波器的容量预测方法、系统和计算机设备 - Google Patents

基于Kalman滤波器的容量预测方法、系统和计算机设备 Download PDF

Info

Publication number
WO2016155241A1
WO2016155241A1 PCT/CN2015/089025 CN2015089025W WO2016155241A1 WO 2016155241 A1 WO2016155241 A1 WO 2016155241A1 CN 2015089025 W CN2015089025 W CN 2015089025W WO 2016155241 A1 WO2016155241 A1 WO 2016155241A1
Authority
WO
WIPO (PCT)
Prior art keywords
capacity
time series
segmentation
point
acceleration
Prior art date
Application number
PCT/CN2015/089025
Other languages
English (en)
French (fr)
Inventor
苗贝贝
陈宇
金学波
曲显平
陶仕敏
臧志
王博
Original Assignee
百度在线网络技术（北京）有限公司
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 百度在线网络技术（北京）有限公司 filed Critical 百度在线网络技术（北京）有限公司
Priority to EP15849818.8A priority Critical patent/EP3279819B1/en
Priority to JP2016526273A priority patent/JP6343001B2/ja
Priority to KR1020167011244A priority patent/KR101829560B1/ko
Priority to US15/033,604 priority patent/US10437942B2/en
Publication of WO2016155241A1 publication Critical patent/WO2016155241A1/zh

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
    • G06Q10/00—Administration; Management
    • G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/20—Design optimisation, verification or simulation
    • G—PHYSICS
    • G05—CONTROLLING; REGULATING
    • G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
    • G05B23/00—Testing or monitoring of control systems or parts thereof
    • G05B23/02—Electric testing or monitoring
    • G05B23/0205—Electric testing or monitoring by means of a monitoring system capable of detecting and responding to faults
    • G05B23/0218—Electric testing or monitoring by means of a monitoring system capable of detecting and responding to faults characterised by the fault detection method dealing with either existing or incipient faults
    • G05B23/0243—Electric testing or monitoring by means of a monitoring system capable of detecting and responding to faults characterised by the fault detection method dealing with either existing or incipient faults model based detection method, e.g. first-principles knowledge model
    • G05B23/0254—Electric testing or monitoring by means of a monitoring system capable of detecting and responding to faults characterised by the fault detection method dealing with either existing or incipient faults model based detection method, e.g. first-principles knowledge model based on a quantitative model, e.g. mathematical relationships between inputs and outputs; functions: observer, Kalman filter, residual calculation, Neural Networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
    • G06F16/23—Updating
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
    • G06F17/10—Complex mathematical operations
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
    • G06F17/10—Complex mathematical operations
    • G06F17/16—Matrix or vector computation, e.g. matrix-matrix or matrix-vector multiplication, matrix factorization
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
    • G06F17/10—Complex mathematical operations
    • G06F17/18—Complex mathematical operations for evaluating statistical data, e.g. average values, frequency distributions, probability functions, regression analysis
    • 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
    • G06Q10/0631—Resource planning, allocation, distributing or scheduling for enterprises or organisations
    • 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
    • H—ELECTRICITY
    • H03—ELECTRONIC CIRCUITRY
    • H03H—IMPEDANCE NETWORKS, e.g. RESONANT CIRCUITS; RESONATORS
    • H03H17/00—Networks using digital techniques
    • H03H17/02—Frequency selective networks
    • H03H17/0248—Filters characterised by a particular frequency response or filtering method
    • H03H17/0255—Filters based on statistics
    • H03H17/0257—KALMAN filters
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2111/00—Details relating to CAD techniques
    • G06F2111/10—Numerical modelling

Definitions

  • the present invention relates to the field of communications technologies, and in particular, to a capacity prediction method, system, and computer device based on a Kalman filter.
  • the method of predicting the future trend of capacity through the historical data of the capacity time series has the following disadvantages: (1) Human observation method, although experienced professionals can make more accurate predictions on the trend of the capacity time series data. Judgment, but its prediction error is still large; (2) Threshold alarm method, this method can prompt the relevant operation and maintenance personnel to make purchases in time, but can not give reasonable advice on the purchase amount, which will easily lead to excessive expansion and waste.
  • Embodiments of the present invention provide a Kalman filter-based capacity prediction method, system, and computer device to accurately predict capacity growth, thereby facilitating operation and maintenance personnel to formulate a reasonable expansion plan.
  • embodiments of the present invention provide a Kalman-based filter. Capacity prediction method, the method includes:
  • the capacity of the future time is predicted based on the at least one segmentation point determined in the capacity time series.
  • Embodiments of the present invention also provide a capacity prediction system based on a Kalman filter, the system comprising:
  • An obtaining module configured to acquire a time series of capacity of the object to be predicted
  • Establishing an extraction module configured to establish a dynamic model for the capacity time series, and extract state transition parameters and process noise parameters of the dynamic model
  • An estimation generating module configured to perform Kalman filter estimation on the capacity time series by using the state transition parameter and the process noise parameter to generate at least one state feature signal
  • a segmentation module configured to segment the capacity time series according to the at least one state feature signal, and determine a corresponding at least one segmentation point
  • a prediction module configured to predict a capacity of a future time according to the at least one segment point determined in the capacity time series.
  • Embodiments of the present invention provide a computer device comprising: one or more processors; a memory; one or more programs, the one or more programs being stored in the memory, and configured to be Executing, by the one or more processors, instructions for performing a Kalman filter-based capacity prediction method by the one or more processors: acquiring a capacity time series of the object to be predicted; establishing a dynamics model for the capacity time series, And extracting a state transition parameter and a process noise parameter of the dynamic model; performing a Kalman filter estimation on the capacity time series by using the state transition parameter and the process noise parameter to generate at least one state feature signal; according to the at least one a state feature signal segments the capacity time series and determines a corresponding at least one segment point; according to the time sequence of the capacity The at least one segmentation point determined in the column predicts the capacity of the future time.
  • the Kalman filter-based capacity prediction method, system and computer device provided by the embodiments of the present invention construct a dynamic model by acquiring the time series of the capacity, and then use the Kalman filter to filter and estimate the capacity time series under the dynamic model. At least one state feature signal; segmenting the generated state feature signal, and predicting the capacity of the future time according to the segmentation point, thereby improving the accuracy of the prediction.
  • FIG. 1 is a flowchart of a method for an embodiment of a Kalman filter-based capacity prediction method according to the present invention
  • FIG. 2 is a schematic diagram of an example of a capacity time series provided by the present invention.
  • FIG. 3 is a schematic diagram of state characteristic signals of a capacity time series filtered by a Kalman filter according to the present invention
  • FIG. 4 is a flowchart of a method for another embodiment of a Kalman filter-based capacity prediction method according to the present invention.
  • FIG. 5 is a schematic diagram of signals obtained by performing median filtering and moving average filtering on a capacity time series according to the present invention
  • FIG. 6 is a flowchart of a method for determining a segmentation point according to a state feature signal according to the present invention
  • FIG. 7 is a diagram showing a sample of a segmentation suspect point determined by an analysis of variance of a speed time series of speeds provided by the present invention
  • FIG. 8 is a diagram showing a sample of a segmentation suspect point determined by performing an analysis of variance on the acceleration of a capacity time series according to the present invention
  • FIG. 9 is a diagram showing a sample of a segmentation suspect point determined by performing an analysis of variance on an acceleration derivative of a capacity time series according to the present invention.
  • FIG. 10 is a sample diagram of four stages of a time series in a capacity time series obtained by performing an analysis of variance on the acceleration signal of FIG. 8;
  • FIG. 11 is a schematic structural diagram of an embodiment of a Kalman filter-based capacity prediction system according to the present invention.
  • FIG. 12 is a schematic structural diagram of another embodiment of a Kalman filter-based capacity prediction system according to the present invention.
  • Figure 13 is a schematic structural view of the segmentation module of Figure 12;
  • FIG. 14 is a schematic structural diagram of a prediction module in FIG. 12;
  • FIG. 15 is a schematic structural diagram of still another embodiment of a Kalman filter-based capacity prediction system according to the present invention.
  • FIG. 16 is a schematic structural diagram of an embodiment of a computer device according to the present invention.
  • the technical solution provided by the invention is to filter the capacity time series based on the Kalman filter to generate a state feature signal; segment the state feature signal, determine the segmentation point and the growth trend of the signals in each segment, and then use the determined The segmentation point predicts the capacity of the future time.
  • the technical solution of the embodiment of the present invention can be applied to various capacity prediction systems based on Kalman filters.
  • FIG. 1 is a flowchart of a method for an embodiment of a Kalman filter-based capacity prediction method according to the present invention.
  • step S110 a capacity time series of an object to be predicted is acquired.
  • the object to be predicted may be the growing information data in the information industry or the Internet.
  • a sequence formed by changing the capacity size of information over time is referred to as a capacity time series.
  • each time point corresponds to a capacity value of an information capacity.
  • the above capacity time series can be formed by acquiring historical capacity data of the object to be predicted online or offline.
  • step S120 a dynamic model is established for the capacity time series, and the state transition parameters and process noise parameters of the dynamic model are extracted.
  • the dynamic model of maneuvering targets has a long history and good results.
  • the kinetic models that have achieved good results and are widely used include Constant Velocity (CV) model, Constant acceleration (CA) model, Singer model, "current” statistical model, and Jerk model.
  • Various different forms of maneuvering target dynamics models are formed by assuming that the target velocity and acceleration high-order state characteristic signals satisfy different statistical characteristics, which is a mathematical relationship.
  • the selected dynamic model is consistent with the state characteristics of the actual target motion, it indicates the tracking performance of the target state estimation using this dynamic model. The better you can (the smaller the global variance can be considered).
  • the research on the existing dynamic model is applied to the analysis of the growth trend of the above capacity time series, thereby establishing a dynamic model for the capacity time series.
  • X(k) is the system state quantity corresponding to the kth state of the system
  • U(k) is the control quantity of the kth state
  • A is the state transition matrix corresponding to the system state equation, that is, corresponding to the above state Transfer parameters.
  • W(k) is the process noise term corresponding to the state equation of the system
  • Q can be the process noise parameter under the Jerk model, and Q is expressed as follows:
  • T 0 is the sampling time interval
  • ⁇ is the maneuvering frequency
  • Kalman filter estimation is performed on the capacity time series using state transition parameters and process noise parameters to generate at least one state feature signal.
  • Kalman filter is an optimized autoregressive data processing algorithm, which can estimate the state of dynamic system from a series of measurement data that does not completely contain noise.
  • Modeling try to build a dynamic model that is consistent with the changing laws of the system process. Then, for the constructed dynamic model, two parameters in the corresponding system state equation are extracted, which respectively correspond to the above state transition parameters and process noise parameters. Using these two parameters and the original system process data, the Kalman filter can more accurately estimate the trend of the original system process data over time.
  • the Kalman filter is used to estimate the capacity time series in the dynamic model, and at least one state feature signal including the position, velocity, acceleration, and acceleration derivative corresponding to each time point in the time series can be obtained online.
  • the estimated values are then used to analyze the variation of the time series of the acquired capacity.
  • step S140 the capacity time series is segmented according to the at least one state feature signal, and the corresponding at least one segment point is determined.
  • Z t represents the observed time series of capacity
  • ⁇ t is the basic trend
  • N t is an independent noise process.
  • the Z t has the following characteristics: there is noise, there is a “mutation point”, and the basic trend of the data Z t undergoes a phase change, forming a small stage of a plurality of linear changes.
  • the Kalman filter estimation can obtain the optimal estimation result in the linear mode with relatively stable changes.
  • the capacity time series does not conform to the characteristics of Gaussian white noise, and the model estimation value is inconsistent with the actual data, which results in the difference between the predicted value and the correction value of the Kalman filter.
  • the Kalman filter automatically adjusts the derivative of velocity, acceleration, and acceleration to slowly perform tracking estimation learning, resulting in "fluctuations" in this interval.
  • the state characteristic signal pattern of the derivative of the velocity, acceleration, and acceleration of the capacity time series is obtained by filtering the capacity time series by the Kalman filter.
  • the time sequence of the capacity to be studied can be determined by the position and number of occurrences of the "fluctuation" signal segment of at least one of the derivatives of velocity, acceleration, and acceleration in FIG.
  • the column is segmented (the number of segments that can be patterned in this embodiment), and the corresponding at least one segmentation point is determined, and each segmentation point can be located on a "fluctuation" signal segment.
  • the capacity of the future time is predicted based on at least one segmentation point determined in the capacity time series.
  • the variation rule of the capacity time series in the adjacent segmentation points can be grasped, and then the variation law of the entire capacity time series can be grasped, and the capacity of the future time is performed. prediction. Segmentation prediction can accurately grasp the trend of capacity time series growth change more than single non-segment prediction, making the prediction result more accurate.
  • the pattern of all patterns or partial pattern content time series can be selected according to actual needs for capacity prediction.
  • the Kalman filter-based capacity prediction method constructs a dynamic model by using the acquired capacity time series, and then uses a Kalman filter to filter and estimate the capacity time series under the dynamic model to generate at least one state feature signal.
  • the segmentation of the generated state feature signal is performed, and the capacity of the future time is predicted according to the segmentation point, thereby improving the accuracy of the prediction.
  • FIG. 4 is a flowchart of a method for another method for predicting a capacity of a Kalman filter based on the present invention, which may be regarded as a specific implementation manner of the embodiment shown in FIG. 1.
  • the Jerk model is used as the dynamic model established for the capacity time series, and the state transition matrix A and the process noise variance matrix Q under the Jerk model are respectively used as the state transition parameters and process noise parameters for capacity time.
  • step S410 a capacity time series of an object to be predicted is acquired.
  • the process of step S410 is the same as the process of the aforementioned step S110.
  • step S430 a dynamics model is established for the capacity time series, and the state transition parameters and process noise parameters of the dynamic model are extracted.
  • the processing of step S430 is similar to the processing of the foregoing step S120.
  • the present embodiment uses a system state equation based on the dynamics of the Jerk model as a dynamic model established for the time series of the studied capacity, and extracts a state transition matrix A and a process noise variance matrix Q corresponding to the system state equation, respectively. State transfer parameter And the process noise parameters, the Kalman filter estimation of the capacity time series.
  • the specific acquisition formulas of A and Q are shown in step S120, and are not described herein.
  • the capacity time series data of the acquired object to be tested generally has the following characteristics: a noise point, a set of abrupt points that deviate far from the basic trend, and a basic trend in a phase change. In order to achieve accurate prediction, these noise and mutation points need to be removed. Therefore, the embodiment may further perform step S420 to perform data cleaning on the capacity time series before step S430.
  • step S420 the capacity time series is filtered by median filtering and/or moving average filtering to generate a filtered capacity time series.
  • the median filtering can effectively remove the abrupt points in the data sequence, and the moving average filtering can effectively remove the noise in the data sequence.
  • the median filtering may not be performed, and only the moving average filtering method is used for filtering processing, and the processing is simple.
  • median filtering and moving average filtering are used to sequentially filter the capacity time series to obtain a basic trend of the capacity time series.
  • FIG. 5 it is a schematic diagram of signals after performing median filtering and moving average filtering on the capacity time series.
  • step S440 Kalman filter estimation is performed on the capacity time series using the state transition parameter and the process noise parameter to generate at least one state feature signal; the process of step S440 is similar to the process of step S130 described above.
  • the system state equation under the Jerk model is used as the dynamic model established by the predicted capacity time series. Therefore, the state transition matrix A corresponding to the system state equation and the process noise variance matrix Q can be directly used as The state transition parameters and the process noise parameters are Kalman filter estimates for the capacity time series to generate at least one of the following state characteristic signals: velocity, acceleration, and acceleration derivative.
  • the filtering process of the Kalman filter by taking the system state equation of the capacity time series in the Jerk model as an example.
  • the filtering process of the Kalman filter mainly includes two steps of state prediction and correction.
  • the next state of the capacity particle is predicted using the state transition matrix A corresponding to the capacity time series in the Jerk model.
  • the current system state is k, according to the state transition matrix A, Predict the current state based on the previous state of the system.
  • k-1) is the result of prediction using the previous state
  • k-1) is the optimal result of the previous state
  • U(k) is the current state.
  • a T Represents the transposed matrix of A, where Q is the system's process noise variance matrix.
  • H the measurement matrix
  • Kg Kalman Gain
  • Kg P(k
  • the entire Kalman filter estimation can be completed, and state characteristic signals such as the position, velocity, acceleration, and derivative of the acceleration time series are extracted.
  • state characteristic signals are relatively stable signals and are more able to reflect the main features of the trend of the time series of capacity.
  • step S450 the capacity time series is segmented according to the at least one state feature signal, and the corresponding at least one segment point is determined.
  • the processing of step S450 is similar to the processing of the foregoing step S140.
  • the system state equation under the Jerk model is used as the dynamic model established by the predicted capacity time series. Therefore, the state signal of the generated capacity state signal can be estimated by the Kalman filter under the capacity time series of the Jerk model: A signal including a velocity, an acceleration, and an acceleration derivative, segmenting the time series of capacity, and determining a corresponding at least one segmentation point.
  • the present embodiment provides a specific segmentation method, including the method steps shown in FIG. 6.
  • FIG. 6 is a flowchart of a method for determining a segmentation point according to a state feature signal according to the present invention
  • At step S610 at least one phase change signal region in which the presence of the velocity, acceleration, and acceleration derivative signals obtained by the Kalman filter is estimated is determined.
  • the estimated value of the Kalman filter does not match the capacity time series data due to the transition phase between adjacent two patterns in the capacity time series, ie near the segmentation point.
  • the Kalman filter realizes the matching estimation by adjusting its own state parameters, which causes the vicinity of the segmentation point to be significantly different from the state characteristic signal of the system state in the adjacent pattern.
  • the state characteristic signals have distinct signal numbers.
  • a phase change point or a segmentation suspect point an area composed of these phase change points or segmentation suspect points is called a phase change signal area.
  • the signal in the phase change signal region has distinct characteristics (such as "fluctuation” characteristics) from the signals in the adjacent pattern, the signal can be mathematically analyzed to determine at least one phase change signal region.
  • This embodiment provides a statistical analysis method for performing a significant analysis of high-order state characteristic signals such as velocity, acceleration, and acceleration derivative of a system state extracted by a Kalman filter, and detecting a segmentation suspect point from a probabilistic angle. And then using the obtained segmentation suspect points to determine the phase change signal region.
  • high-order state characteristic signals such as velocity, acceleration, and acceleration derivative of a system state extracted by a Kalman filter
  • the method of statistical analysis includes the following steps:
  • Step 1 for at least one of the signals of the speed, acceleration, and acceleration derivatives obtained by the Kalman filter after the capacity time series is calculated, using two consecutive sliding windows of the same length to calculate each sliding window in time point order.
  • the variance or mean of the internal signal is
  • the width of the two sliding windows to be 24, then the first pair of slidings calculated in chronological order
  • the windows are 1 to 24 and 25 to 48, respectively, and the variance or mean of all the capacities in the corresponding time series of the two windows are calculated.
  • the two sliding windows are moved backward in the order of time, that is, the second pair of sliding windows are respectively 2 to 25, 26 to 49, ..., and sequentially move to the last time point, and each of the two sliding windows is calculated.
  • the variance or mean of the internal signal is calculated.
  • Step 2 If the probability that the variance or the mean value of the signals in the two sliding windows are different is greater than the preset corresponding probability threshold, mark the middle point of the two consecutive sliding windows as a segmentation suspect point.
  • the principle of determining a segmentation suspect point is determined according to the degree of difference between two variances or mean values in each pair of windows obtained.
  • the test for different variances often uses the F test for the mean value.
  • Different tests often use T test or displacement test. The following tests are performed separately for the two parameters.
  • the probability of Y 1 and Y 2 vector variance can be obtained by mathematical analysis such as vartest2 function in Matlab. Find the probability that the variances of the two samples of Y 1 and Y 2 are different. When the probability of the two variances is greater than the preset corresponding probability threshold, such as 0.99, the variance of the data in the two sliding windows is considered to be different.
  • the middle point of two consecutive sliding windows is a segmentation suspect point.
  • the intermediate point of the first pair of sliding windows for a sliding window length of 24 may be any of 24, 25.
  • the probability that the mean values of Y 1 and Y 2 vectors are the same is obtained, and the probability that the mean values of the two samples of Y 1 and Y 2 are different can be obtained.
  • the corresponding probability threshold such as 0.99, considers the mean values of the data in the two sliding windows to be different, and determines that the intermediate point of the two consecutive sliding windows is a segmentation suspected point.
  • the intermediate point of the first pair of sliding windows for a sliding window length of 24 may be any of 24, 25.
  • At least one segmentation suspect point in the capacity time series can be determined by the above-described different test of the difference and the test with different mean values. As shown in Fig. 7, Fig. 8, and Fig. 9, the variance analysis of the velocity, acceleration, and acceleration derivatives of the capacity time series is shown. The identified segmentation suspect points, as can be seen from the figure, the segmentation suspect points are unevenly distributed throughout the capacity time series, but concentrated in the "fluctuation" signal region of the transition segment.
  • step 3 a signal region composed of a plurality of consecutive segmentation suspect points is determined as a phase change signal region of the capacity time series.
  • the signal in which the phase change actually occurs may be spread throughout the time series of the capacity, especially during the transition phase between the two adjacent patterns. Therefore, it is more reasonable to select a plurality of consecutive segmentation suspect points obtained by the detection, and the signal regions composed of the segmentation suspect points are used as a phase change signal region of the capacity time series. Specifically, the determined position of the phase change signal region can also be locked by setting the number of consecutive segmentation suspect points required to determine a phase change signal region and the spacing condition between the two consecutive segments of the suspect segment.
  • the capacity time series is segmented according to the determined at least one phase change signal region, and the corresponding at least one segment point is determined.
  • any point in the region can be selected as the segmentation point.
  • the segmentation suspect point corresponding to the maximum value or the maximum value of the difference between the variances of the variances of the signals in the corresponding two sliding windows in each phase change signal region may be determined as one point.
  • the change of the signal may be "fluctuated" by default; or the intermediate point in each phase change signal region is determined as a segmentation point.
  • the rules determined by the segmentation points are not limited.
  • the determination of the segmentation point in the capacity time series is completed.
  • the object characterized by the segmentation point is the state characteristic signal of the capacity time series, including: velocity, acceleration and acceleration derivative signals, then which signal is determined Is the segmentation point more accurate?
  • the present embodiment introduces the concept of precision and recall. Take the phase transition point (segmented suspect point) obtained by analysis of variance as an example, which is defined as follows:
  • the total number of true phase transition points detected by the variance analysis is the total number of segmentation points of the final determined capacity time series; the total number of points detected by the variance analysis is the total number of segmentation suspect points; It is expected that the total number of phase change points detected is artificially specified by observing the basic trend of the sample and is constant. The higher the precision, the more the actual phase change points detected, the greater the recall rate, indicating that the number of phase change points that can be detected by variance analysis is more.
  • the acceleration derivative signal can be used for the variance analysis when the segmentation point determination of the capacity time series is actually performed.
  • the results are segmented to ensure that the capacity prediction model generated later is more accurate.
  • the derivative signals for velocity, acceleration and acceleration can be detected by analysis of variance, the detection results of different signals have advantages and disadvantages. After testing a large number of data samples, it is found that the acceleration is generally When the derivative signal of acceleration is analyzed by variance, the detection result is better. Therefore, in the present embodiment, it is preferable to perform capacity prediction on the basis of segmentation of the derivative signal of acceleration or acceleration.
  • step S480 the capacity of the future time is predicted based on at least one segment point determined in the capacity time series.
  • the processing of step S480 is similar to the processing of the aforementioned step S150.
  • the embodiment shows a specific implementation manner for predicting the capacity of the future time, that is, including steps S481 to S482:
  • Step S481 using the last segment point in the capacity time series as a starting point, linearly or nonlinearly fitting the data of the subsequent capacity time series to generate a capacity prediction model; step S482: predicting the future time according to the capacity prediction model Capacity is predicted.
  • the capacity time series of the corresponding position after filtering by the Kalman filter may also be fitted to generate a capacity prediction model.
  • steps S491 to S492 may be further performed to determine the reasonableness and accuracy of the generated capacity prediction model.
  • step S491 a goodness of fit evaluation is performed on the generated capacity prediction model.
  • the goodness of fit is often used for evaluation.
  • the variance analysis of the acceleration signal in Fig. 8 is performed by linear regression method using R 2 as the evaluation standard.
  • the four stages of the time series obtained by the segmentation are fitted, and the goodness of fit and the serial number of the corresponding stage are shown in Table 2.
  • step S492 if the evaluation value obtained by the goodness of fit evaluation is greater than the preset goodness-of-fit threshold, the processing operation of predicting the capacity of the future time according to the capacity prediction model is triggered.
  • the preset goodness of fit threshold may be 0.99.
  • the last stage (stage 4) capacity time series generation capacity prediction model is described in the present embodiment.
  • the default is that in the near future period, the capacity changes with a greater probability of increasing the phase of the last stable change. Therefore, the last segmentation point in the capacity time series is selected as the starting point, and the curve or straight line fitting of the subsequent data can make more reasonable and accurate prediction of the capacity data in the future.
  • Table 3 shows the statistical results of the predicted value obtained by predicting the data in the last stage from the capacity prediction model generated by the last stage data in the capacity time series shown in Fig. 10 and the true value of the stage.
  • Prediction point actual value Predictive value error Relative error(%) 1501 19405616 19626584 220967.71 1.14 1502 19402632 19646480 243847.92 1.26 1503 19401049 19666376 265327.13 1.37 1504 19399736 19686272 286536.34 1.48 1505 19396443 19706169 309725.55 1.60 1506 19394328 19726065 331736.76 1.71
  • x is the position of the predicted point (corresponding to the time node)
  • y is the capacity value.
  • the error is the difference between the predicted value and the true value
  • the relative error error / true value.
  • the Kalman filter-based capacity prediction method extracts the corresponding state transition matrix A and the process noise variance matrix by taking the system state equation under the Jerk model as an example on the basis of the method embodiment shown in FIG. Q to realize the Kalman filter estimation of the capacity time series; the high-order state characteristic signals estimated by the Kalman filter, such as the derivative of velocity, acceleration and acceleration, segment the capacity time series; finally, the last segment The point is the starting point, and the subsequent capacity time series is fitted to generate a capacity prediction model to predict the capacity in the future time period. Improve the accuracy of the forecast.
  • FIG. 11 is a schematic structural diagram of an embodiment of a Kalman filter-based capacity prediction system according to the present invention, which can be used to perform the method steps of the embodiment shown in FIG. 1.
  • the Kalman filter-based capacity prediction system specifically includes an acquisition module 111, an establishment extraction module 112, an estimation generation module 113, a segmentation module 114, and a prediction module.
  • Block 115 wherein:
  • An obtaining module 111 configured to acquire a capacity time series of the object to be predicted
  • An extraction module 112 is configured to establish a dynamic model for the capacity time series, and extract state transition parameters and process noise parameters of the dynamic model;
  • the estimation generating module 113 is configured to perform Kalman filter estimation on the capacity time series by using the state transition parameter and the process noise parameter to generate at least one state feature signal;
  • the segmentation module 114 is configured to segment the capacity time series according to the at least one state feature signal, and determine a corresponding at least one segment point;
  • the prediction module 115 is configured to predict the capacity of the future time according to the at least one segment point determined in the capacity time series.
  • the Kalman filter-based capacity prediction system constructs a dynamic model by acquiring the time series of the capacity, and then uses a Kalman filter to filter and estimate the capacity time series under the dynamic model to generate at least one state feature signal.
  • the segmentation of the generated state feature signal is performed, and the capacity of the future time is predicted according to the segmentation point, thereby improving the accuracy of the prediction.
  • FIG. 12 is a schematic structural diagram of another embodiment of a Kalman filter-based capacity prediction system according to the present invention, which may be regarded as a specific implementation structure of the embodiment shown in FIG. Method steps.
  • the Kalman filter-based capacity prediction system includes an acquisition module 111, an establishment extraction module 112, an estimation generation module 113, a segmentation module 114, and a prediction module 115, and is the same as the corresponding module in FIG.
  • the Kalman filter-based capacity prediction system shown in FIG. 12 may further include a pre-processing module 116 for filtering the capacity time series by using median filtering and/or moving average filtering to generate a filtered Capacity time series.
  • the above dynamic model may include at least one of a constant velocity (CV) model, a constant acceleration (CA) model, a Singer model, a “current” statistical model, and a Jerk model.
  • CV constant velocity
  • CA constant acceleration
  • Singer model a Singer model
  • current statistical model a “current” statistical model
  • Jerk model a Jerk model
  • the foregoing establishment extraction module 112 can also be used to:
  • the foregoing estimation generating module 113 is further configured to:
  • the state transition matrix A and the process noise variance matrix Q are used to perform Kalman filter estimation on the capacity time series, and generate at least one state feature signal as follows: velocity, acceleration, and acceleration derivative.
  • segmentation module 114 is further configured to:
  • At least one of the velocity, acceleration, and acceleration derivatives obtained by estimating the Kalman filter according to the capacity time sequence segments the capacity time series and determines a corresponding at least one segmentation point.
  • the segmentation module 114 may include:
  • a determining unit 1141 configured to determine at least one phase change signal region of the presence of the speed, acceleration, and acceleration derivative signals obtained by the Kalman filter after the capacity time series is determined;
  • the segmentation unit 1142 is configured to segment the capacity time series according to the determined at least one phase change signal region, and determine a corresponding at least one segmentation point.
  • determining unit 1141 is further configured to:
  • a signal region composed of a plurality of consecutive segmentation suspect points is determined as a phase change signal region of the capacity time series.
  • segmentation unit 1142 is further configured to:
  • segmentation suspect points corresponding to the maximum value or the maximum value of the differences among the variances of the variances of the signals in the corresponding two sliding windows in each phase change signal region are determined as one segmentation point.
  • segmentation unit 1142 is further configured to:
  • the intermediate point in each phase change signal region is determined as a segmentation point.
  • the prediction module 115 may further include:
  • the prediction model generating unit 1151 is configured to perform linear or nonlinear fitting on the data of the subsequent capacity time series by using the last segmentation point in the capacity time series as a starting point to generate a capacity prediction model;
  • the prediction unit 1152 is configured to predict the capacity of the future time according to the capacity prediction model.
  • the method further includes:
  • a goodness-of-fit evaluation module 117 is configured to perform a goodness-of-fit evaluation on the generated capacity prediction model
  • the triggering module 118 is configured to: if the evaluation value obtained by the goodness of fit evaluation is greater than a preset goodness-of-fit threshold, trigger the prediction unit to perform a processing operation of predicting the capacity of the future time according to the capacity prediction model.
  • FIG. 4, FIG. 5 and FIG. 6 can be performed by the corresponding functional modules in FIG. 11 to FIG. 15, and the principle of the steps is not described herein.
  • the Kalman filter-based capacity prediction system extracts the corresponding state transition matrix A and the process noise variance matrix by taking the system state equation under the Jerk model as an example on the basis of the system embodiment shown in FIG. Q to realize the Kalman filter estimation of the capacity time series; the high-order state characteristic signals estimated by the Kalman filter, such as the derivative of velocity, acceleration and acceleration, segment the capacity time series; finally, the last segment The point is the starting point, and the subsequent capacity time series is fitted to generate a capacity prediction model to predict the capacity in the future time period. Improve the accuracy of the forecast.
  • FIG. 16 is a schematic structural diagram of an embodiment of a computer device according to the present invention.
  • a computer device can be used to implement the Kalman filter-based capacity prediction method provided in the above embodiments. Specifically:
  • Computer devices may vary considerably depending on configuration or performance, and may include one or more processors (such as Central Processing Units, CPU) 710 and memory 720.
  • the memory 720 can be short-term storage or persistent storage.
  • One or more programs may be stored in the memory 720, and each program may include a pair of computer devices Column instruction operation.
  • the processor 710 can communicate with the memory 720 to perform a series of instruction operations in the memory 720 on the computer device.
  • the memory 720 also stores data of one or more operating systems, such as Windows ServerTM, Mac OS XTM, UnixTM, LinuxTM, FreeBSDTM, and the like.
  • the computer device may also include one or more power sources 730, one or more wired or wireless network interfaces 740, one or more input and output interfaces 750, and the like.
  • the computer device includes one or more processors 710, memory 720, and one or more programs, one or more programs stored in memory 720, and configured to be executed by one or more processors. 710 executing instructions included in one or more programs for performing a Kalman filter-based capacity prediction method: acquiring a capacity time series of the object to be predicted; establishing a dynamic model for the capacity time series, and extracting the dynamic model a state transition parameter and a process noise parameter; performing a Kalman filter estimation on the capacity time series using the state transition parameter and the process noise parameter to generate at least one state feature signal; and the capacity according to the at least one state feature signal.
  • the time series is segmented and the corresponding at least one segmentation point is determined; the capacity of the future time is predicted based on the at least one segmentation point determined in the capacity time series.
  • the memory 720 further includes an instruction to perform filtering processing on the capacity time series by median filtering and/or moving average filtering to generate the filtered capacity time series.
  • the dynamic model includes at least one of a constant velocity (CV) model, a constant acceleration (CA) model, a Singer model, a “current” statistical model, and a Jerk model.
  • CV constant velocity
  • CA constant acceleration
  • Singer model a Singer model
  • current statistical model a “current” statistical model
  • Jerk model a Jerk model
  • the memory 720 further includes an instruction to: construct a system state equation of the capacity time series under the Jerk model, and extract a state transition matrix A and a process noise variance matrix Q corresponding to the system state equation.
  • the memory 720 further includes an instruction to perform Kalman filter estimation on the capacity time series by using the state transition matrix A and the process noise variance matrix Q, and generate at least one state feature signal as follows : Speed, acceleration and acceleration derivatives.
  • the memory 720 further includes an instruction to: process the speed, acceleration, and addition obtained by the Kalman filter according to the capacity time series. At least one of the velocity characteristics signals segments the capacity time series and determines a corresponding at least one segmentation point.
  • the memory 720 further includes an instruction to: determine at least one of the signals of the speed, acceleration, and acceleration derivatives obtained by the Kalman filter after the capacity time series is determined to be at least one of its existence a phase change signal region; segmenting the capacity time series according to the determined at least one phase change signal region, and determining a corresponding at least one segment point.
  • the memory 720 further includes an instruction to: use at least one of the signals of the speed, acceleration, and acceleration derivatives obtained by the Kalman filter after the capacity time series, using two consecutive and The sliding windows of the same length calculate the variance or mean of the signals in each of the two sliding windows in time-point order; if the probability of the variance or the mean of the signals in the two sliding windows is different than the preset corresponding probability threshold, Then marking an intermediate point of the two consecutive sliding windows as a segmentation suspect point; determining a signal region composed of a plurality of consecutive segmentation suspect points as one phase change signal region of the capacity time series.
  • the memory 720 further includes an instruction to: perform, in each of the phase change signal regions, a difference between a maximum value or a mean value of a difference between variances of signals in the corresponding two sliding windows The segmentation suspect point corresponding to the maximum value is determined as a segmentation point.
  • the memory 720 further includes an instruction to: determine an intermediate point in each of the phase change signal regions as a segmentation point.
  • the memory 720 further includes an instruction to: perform a linear or nonlinear fitting on the data of the subsequent capacity time series by using the last segmentation point in the capacity time series as a starting point a capacity prediction model; predicting the capacity of the future time based on the capacity prediction model.
  • the memory 720 further includes an instruction to: perform a goodness-of-fit evaluation on the generated capacity prediction model; if the evaluation value obtained by the goodness of fit evaluation is greater than a preset goodness-of-fit threshold And triggering the processing operation of predicting the capacity of the future time according to the capacity prediction model.
  • the computer device constructs dynamics by acquiring the time series of capacity
  • the model uses the Kalman filter to filter the capacity time series under the dynamic model to generate at least one state feature signal; segment the generated state feature signal, and predict the capacity of the future time according to the segmentation point, and improve The accuracy of the prediction.
  • the present invention extracts the corresponding state transition matrix A and the process noise variance matrix Q by using the system state equation in the Jerk model as an example to realize Kalman filter estimation for the capacity time series; the high order estimated by the Kalman filter is obtained.
  • State characteristic signals such as velocity, acceleration, and acceleration derivatives, segment the time series of capacity; finally, starting with the last segmentation point, fitting the subsequent capacity time series to generate a capacity prediction model for the future The capacity within the time period is predicted. Improve the accuracy of the forecast.
  • the above method and apparatus according to the present invention may be implemented in hardware, firmware, or as software or computer code that may be stored in a recording medium such as a CD ROM, RAM, floppy disk, hard disk, or magneto-optical disk, or implemented.
  • the network downloads computer code originally stored in a remote recording medium or non-transitory machine readable medium and stored in a local recording medium so that the methods described herein can be stored using a general purpose computer, a dedicated processor or programmable
  • Such software processing on a recording medium of dedicated hardware such as an ASIC or an FPGA.
  • a computer, processor, microprocessor controller or programmable hardware includes storage components (eg, RAM, ROM, flash memory, etc.) that can store or receive software or computer code, when the software or computer code is The processing methods described herein are implemented when the processor or hardware is accessed and executed. Moreover, when a general purpose computer accesses code for implementing the processing shown herein, the execution of the code converts the general purpose computer into a special purpose computer for performing the processing shown herein.

Abstract

本发明实施例提供的基于Kalman滤波器的容量预测方法、系统和计算机设备，方法包括：获取待预测对象的容量时间序列；为容量时间序列建立动力学模型，并提取动力学模型的状态转移参量和过程噪声参量；利用状态转移参量和过程噪声参量对容量时间序列进行Kalman滤波器估计，生成至少一个状态特征信号；根据至少一个状态特征信号对容量时间序列进行分段，并确定相应的至少一个分段点；根据在容量时间序列中确定的至少一个分段点对未来时间的容量进行预测。本发明的技术方案实现对容量增长的准确预测，进而便于运维人员制定合理的扩容方案。

Description

基于Kalman滤波器的容量预测方法、系统和计算机设备
本申请要求于2015年04月03日提交中国专利局、申请号为201510158699.9、发明名称为“基于Kalman滤波器的容量预测方法和系统”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本发明涉及通信技术领域，尤其涉及一种基于Kalman滤波器的容量预测方法、系统和计算机设备。
背景技术
随着互联网和信息行业的快速发展，大量的信息数据随之产生。运维人员需要及时采购存储设备，以有效存储不断增长的数据。但是，由于采购预算不准确，常常出现设备因采购过量而闲置，给公司和企业造成人力浪费、财力损失。基于此，业内技术人员积极开展数据分析和数据可视化研究项目，以期望通过容量时间序列的历史数据，分析其基本变化趋势，准确预测容量的未来走势，实现智能化定制采购计划的目标，为公司节省采购成本，避免浪费。
现有技术中，通过容量时间序列的历史数据，预测容量的未来走势的方法具有以下缺点：(1)人为观测法，尽管经验丰富的专业人士可以对容量时间序列数据的走势做出较准确的判断，但其预测误差仍较大；(2)阈值报警法，这种方法可以及时提醒相关运维人员进行采购，但不能对采购量给出合理建议，容易导致扩容过量，造成浪费。
发明内容
本发明的实施例提供一种基于Kalman滤波器的容量预测方法、系统和计算机设备，以实现对容量增长的准确预测，进而便于运维人员制定合理的扩容方案。
为达到上述目的，本发明的实施例提供了一种基于Kalman滤波器 的容量预测方法，所述方法包括：
获取待预测对象的容量时间序列；
为所述容量时间序列建立动力学模型，并提取所述动力学模型的状态转移参量和过程噪声参量；
利用所述状态转移参量和过程噪声参量对所述容量时间序列进行Kalman滤波器估计，生成至少一个状态特征信号；
根据所述至少一个状态特征信号对所述容量时间序列进行分段，并确定相应的至少一个分段点；
根据在所述容量时间序列中确定的所述至少一个分段点对未来时间的容量进行预测。
本发明的实施例还提供了一种基于Kalman滤波器的容量预测系统，所述系统包括：
获取模块，用于获取待预测对象的容量时间序列；
建立提取模块，用于为所述容量时间序列建立动力学模型，并提取所述动力学模型的状态转移参量和过程噪声参量；
估计生成模块，用于利用所述状态转移参量和过程噪声参量对所述容量时间序列进行Kalman滤波器估计，生成至少一个状态特征信号；
分段模块，用于根据所述至少一个状态特征信号对所述容量时间序列进行分段，并确定相应的至少一个分段点；
预测模块，用于根据在所述容量时间序列中确定的所述至少一个分段点对未来时间的容量进行预测。
本发明的实施例提供了一种计算机设备，包括：一个或多个处理器；存储器；一个或多个程序，所述一个或多个程序存储在所述存储器中，且经配置以由所述一个或者多个处理器执行所述一个或者多个程序包含的用于执行基于Kalman滤波器的容量预测方法的指令：获取待预测对象的容量时间序列；为所述容量时间序列建立动力学模型，并提取所述动力学模型的状态转移参量和过程噪声参量；利用所述状态转移参量和过程噪声参量对所述容量时间序列进行Kalman滤波器估计，生成至少一个状态特征信号；根据所述至少一个状态特征信号对所述容量时间序列进行分段，并确定相应的至少一个分段点；根据在所述容量时间序 列中确定的所述至少一个分段点对未来时间的容量进行预测。
本发明实施例提供的基于Kalman滤波器的容量预测方法、系统和计算机设备，通过对获取的容量时间序列搭建动力学模型，然后利用Kalman滤波器对动力学模型下的容量时间序列进行滤波估计生成至少一个状态特征信号；对生成的状态特征信号进行分段，并依据分段点对未来时间的容量进行预测，提高了预测的准确性。
附图说明
图1为本发明提供的基于Kalman滤波器的容量预测方法一个实施例的方法流程图；
图2为本发明提供的容量时间序列的样例示意图；
图3为本发明提供的基于Kalman滤波器滤波后的容量时间序列的状态特征信号示意图；
图4为本发明提供的基于Kalman滤波器的容量预测方法另一个实施例的方法流程图；
图5为本发明提供的对容量时间序列进行中值滤波和移动平均滤波处理后的信号示意图；
图6为本发明提供的根据状态特征信号确定分段点的方法流程图；
图7为本发明提供的对容量时间序列的速度进行方差分析检验确定的分段疑似点样例图；
图8为本发明提供的对容量时间序列的加速度进行方差分析检验确定的分段疑似点样例图；
图9为本发明提供的对容量时间序列的加速度导数进行方差分析检验确定的分段疑似点样例图；
图10为对图8中加速度信号进行方差分析分段得到的容量时间序列中4个阶段信号样例图；
图11为本发明提供的基于Kalman滤波器的容量预测系统一个实施例的结构示意图；
图12为本发明提供的基于Kalman滤波器的容量预测系统另一个实施例的结构示意图；
图13为图12中分段模块的结构示意图；
图14为图12中预测模块的结构示意图；
图15为本发明提供的基于Kalman滤波器的容量预测系统又一个实施例的结构示意图。
图16为本发明提供的计算机设备一个实施例的结构示意图。
具体实施方式
本发明提供的技术方案，是基于Kalman滤波器对容量时间序列进行滤波，生成状态特征信号；对状态特征信号进行分段，确定分段点以及各段内的信号的增长趋势，然后利用确定的分段点对未来时间的容量进行预测。本发明实施例的技术方案可以适用于各种基于Kalman滤波器的容量预测系统。
实施例一
图1为本发明提供的基于Kalman滤波器的容量预测方法一个实施例的方法流程图。
参照图1，在步骤S110，获取待预测对象的容量时间序列。
所述的待预测对象可以为信息行业或互联网中不断增长的信息数据。通常将信息的容量大小随时间变化所形成的序列称为容量时间序列。在容量时间序列中，每一时间点对应着一个信息容量的容量值。
通过在线或离线的获取待预测对象的历史容量数据可以形成上述容量时间序列。
在步骤S120，为容量时间序列建立动力学模型，并提取动力学模型的状态转移参量和过程噪声参量。
机动目标的动力学模型研究历史悠久且效果良好。目前，已经取得良好效果并广泛使用的动力学模型主要有常速度(Constant velocity，CV)模型、常加速度(Constant acceleration，CA)模型、Singer模型、“当前”统计模型、Jerk模型等。各种不同形式的机动目标动力学模型是通过假设目标速度及加速度高阶的状态特征信号满足不同统计特性形成的，是一种数学关系。当选取的动力学模型与实际目标运动的状态特征一致程度越高时，表明采用这种动力学模型进行的目标状态估计得到的跟踪性 能越好(可以认为全局方差越小)。
本实施例中，将对现有的动力学模型的研究运用到上述容量时间序列的增长趋势的分析中，从而建立针对容量时间序列的动力学模型。
建立容量时间序列的动力学模型，例如，对于线性信号，我们假设其理想加速度值为0，可以建立CA模型，实现对速度的估计；对于二次函数的非线性信号，其理想加速度不为0，但其加速度的导数为0，可以建立Jerk模型，实现对加速度的估计；对于周期信号，可以建立三角函数趋势模型。基于这种思想，我们可以根据容量时间序列的变化情况，通过数学运算，选择建立对应的动力学模型。
通常，一个固定的动力学模型中，存在两个描述目标状态特征的参量，分别为目标从当前状态变换到下一状态对应的状态转移参量和过程噪声参量。例如，对于Jerk模型，其对应的系统状态方程可表示为：
X(k)＝AX(k-1)+BU(k)+W(k)          (1)
其中，X(k)为系统在第k个状态对应的系统状态量，U(k)为第k个状态的控制量，A为该系统状态方程对应的状态转移矩阵，即对应为上述的状态转移参量。A表式如下：
Figure PCTCN2015089025-appb-000001
W(k)为该系统状态方程对应的过程噪声项，其对应的过程噪声方差矩阵Q可以为Jerk模型下的过程噪声参量，Q表示如下：
Figure PCTCN2015089025-appb-000002
其中：T0为采样时间间隔、α为机动频率、
Figure PCTCN2015089025-appb-000003
为Jerk模型的瞬时方差；
Figure PCTCN2015089025-appb-000004
Figure PCTCN2015089025-appb-000005
参照Jerk模型下的状态转移参量和过程噪声参量的选取，类似的我们还可以提取出其他动力学模型中相应的参量。基于针对上述容量时间序列在包括但不限于Jerk模型下的状态转移参量和过程噪声参量，可以初步获悉当前被测容量时间序列的基本增长趋势特征。
在步骤S130，利用状态转移参量和过程噪声参量对容量时间序列进行Kalman滤波器估计，生成至少一个状态特征信号。
其中，Kalman滤波器是一种最优化自回归数据处理算法，它能够从一系列不完全包含噪声的测量数据中，估计动态系统的状态。通常，在Kalman滤波器对一个系统过程进行估计之前，需要预先对该系统过程进 行建模，尽量构建与系统过程的变化规律相一致的动力学模型。然后针对构建的动力学模型，提取对应的系统状态方程中的两个参量，这两个参量即分别对应于上述的状态转移参量和过程噪声参量。Kalman滤波器利用这两个参量和原始的系统过程数据可以较为准确的估计出原始的系统过程数据随时间的变化趋势。
本实施例中，利用Kalman滤波器对动力学模型下的容量时间序列进行估计，可以在线得到容量在时间序列中各时间点对应的包括如位置、速度、加速度以及加速度导数的至少一个状态特征信号的估计值，然后，通过这些估计值可分析获取容量时间序列的变化规律。
在步骤S140，根据至少一个状态特征信号对容量时间序列进行分段，并确定相应的至少一个分段点。
通常，在针对容量时间序列建立动力学模型之前，假设待研究的数据序列为基本趋势和噪声分量的和，即假定该观测得到的容量时间序列为：
Zt＝μt+Nt              (6)
其中，Zt表示观测得到的容量时间序列，μt是基本趋势，Nt是一个独立噪声过程。如图2所示，该Zt具有以下特征：有噪声、有“突变点”、数据的基本趋势Zt会发生相变，形成多个线性变化的小阶段。
当上述容量时间序列与建立的动力学模型假设一致时，在变化比较稳定的线性阶段(pattern)内，采用Kalman滤波器估计可以得到最优估计结果。但是，在相邻的两个不同pattern之间发生改变的过渡段内，容量时间序列不符合高斯白噪声的特点，模型估计值与实际数据不一致，这导致Kalman滤波器的预测值和校正值相差很大。这种情况下，Kalman滤波器会自动调整速度、加速度及加速度的导数来慢慢的进行跟踪估计学习，从而导致这个区间呈现出“波动”。如图3所示，为容量时间序列经过Kalman滤波器滤波后得到容量时间序列的速度、加速度及加速度的导数的状态特征信号图形。
从图3中获悉，当前被研究的容量时间序列存在三个pattern，且“波动”信号分别发生在时间序列中的400以及1200附近的位置。
通过对图3中速度、加速度、加速度的导数中的至少一个状态特征信号的“波动”信号段发生的位置以及数量可以将被研究的容量时间序 列进行分段(本实施例中可以pattern出现的个数划分段的个数)，并确定相应的至少一个分段点，每个分段点可分别位于一个“波动”信号段上。
在步骤S150，根据在容量时间序列中确定的至少一个分段点对未来时间的容量进行预测。
根据在容量时间序列中确定的至少一个分段点，可以掌握相邻分段点内(各pattern段)容量时间序列的变化规律，进而掌握整个容量时间序列的变化规律，对未来时间的容量进行预测。分段预测较单一的不分段预测更能准确把握容量时间序列增长变化的趋势，使得预测结果更加准确。在预测过程中，可根据实际需求选取所有pattern或部分pattern内容量时间序列的变化规律进行容量预测。
本发明实施例提供的基于Kalman滤波器的容量预测方法，通过对获取的容量时间序列搭建动力学模型，然后利用Kalman滤波器对动力学模型下的容量时间序列进行滤波估计生成至少一个状态特征信号；对生成的状态特征信号进行分段，并依据分段点对未来时间的容量进行预测，提高了预测的准确性。
实施例二
图4为本发明提供的基于Kalman滤波器的容量预测方法另一个实施例的方法流程图，可视为图1所示实施例的一种具体实现方式。本实施例中，以Jerk模型作为针对容量时间序列所建立的动力学模型，并以Jerk模型下的状态转移矩阵A和过程噪声方差矩阵Q分别作为前述的状态转移参量和过程噪声参量进行容量时间序列的Kalman滤波器估计。
参照图4，在步骤S410，获取待预测对象的容量时间序列。步骤S410的处理与前述步骤S110的处理相同。
在步骤S430，为容量时间序列建立动力学模型，并提取动力学模型的状态转移参量和过程噪声参量。步骤S430的处理与前述步骤S120的处理类似。
具体地，本实施例以基于Jerk模型下的动力学的系统状态方程作为针对被研究容量时间序列所建立的动力学模型，并提取系统状态方程对应的状态转移矩阵A和过程噪声方差矩阵Q分别作为前述状态转移参量 和过程噪声参量，进行容量时间序列的Kalman滤波器估计。其中，A和Q的具体获取公式由步骤S120中示出，在此不做赘述。
在步骤S430之前，针对获取的待测对象的容量时间序列数据，如图2所示，通常具有以下特点：含有噪声、存在远远偏离基本趋势的突变点集、基本趋势呈阶段变化。为实现准确预测，需要去除这些噪声和突变点。因此，本实施例在步骤S430之前还可执行步骤S420对容量时间序列进行数据清洗。
在步骤S420，采用中值滤波和/或移动平均滤波对容量时间序列进行滤波处理，生成滤波后的容量时间序列。
具体的，中值滤波可有效去除数据序列中的突变点，移动平均滤波可有效去除数据序列中的噪声。当容量时间序列无明显突变点时，可以不进行中值滤波，仅采用移动平均滤波的方法进行滤波处理，简便处理过程。
优选的，本实施例采用中值滤波、移动平均滤波依次对容量时间序列进行滤波处理，得到容量时间序列的基本趋势。如图5所示，为对容量时间序列进行中值滤波和移动平均滤波处理后的信号示意图。
在步骤S440，利用状态转移参量和过程噪声参量对容量时间序列进行Kalman滤波器估计，生成至少一个状态特征信号；步骤S440的处理与前述步骤S130的处理类似。
具体地，基于本实施例是采用Jerk模型下的系统状态方程作为预测容量时间序列所建立的动力学模型，因此，可直接采用系统状态方程对应的状态转移矩阵A和过程噪声方差矩阵Q分别作为状态转移参量和过程噪声参量对容量时间序列进行Kalman滤波器估计，生成至少一个如下的状态特征信号：速度、加速度和加速度导数。
下面就容量时间序列在Jerk模型下的系统状态方程为例对Kalman滤波器的滤波过程进行说明。通常，Kalman滤波器的滤波过程主要包括状态预测和校正两步。
(1)状态预测
利用容量时间序列在Jerk模型下对应的状态转移矩阵A，来预测容量质点的下一状态。假设现在的系统状态是k，根据状态转移矩阵A，可 以基于系统的上一状态预测现在状态。
X(k|k-1)＝AX(k-1|k-1)+BU(k)            (7)
式(7)中，X(k|k-1)是利用上一状态预测的结果，X(k-1|k-1)是上一状态的最优结果，U(k)为现在状态的控制量，容量序列中没有控制量，因此U(k)＝0。
(7)中已将X(k-1|k-1)的结果进行了更新，可是，对应于X(k|k-1)的方差还没有更新，我们用P表示方差，则有：
P(k|k-1)＝AP(k-1|k-1)AT+Q               (8)
其中，P(k|k-1)是X(k|k-1)对应的方差，P(k-1|k-1)是X(k-1|k-1)对应的方差，AT表示A的转置矩阵，Q为系统的过程噪声方差矩阵。利用式子(6)、(7)可以完成对系统的状态预测。
(2)校正
现在有了现在状态的预测结果，再根据现在状态的测量序列Z(k)，就可以得到现在状态k的最优化估算值X(k|k)：
X(k|k)＝X(k|k-1)+Kg(k)(Z(k)-HX(k|k-1))            (9)
其中H为测量矩阵，我们设定为H＝[1 0 0 0]，Kg为卡尔曼增益(Kalman Gain)：
Kg＝P(k|k-1)HT/(HP(k|k-1)HT+R)       (10)
为了令卡尔曼滤波器不断的运行下去直到系统过程结束，需要更新k状态下X(k|k)的方差。
P(k|k)＝(I-Kg(k)H)P(k|k-1)           (11)
其中I为单位矩阵，当系统进入k+1状态时，P(k|k)就是式子(7)中的P(k-1|k-1)。这样，整个算法就可以自回归的运算下去。
通过式子(7)-(11)，可以完成整个Kalman滤波器估计，并估计提取出容量时间序列的位置、速度、加速度以及加速度的导数等状态特征信号。这些状态特征信号是比较平稳的信号，更能反映容量时间序列变化趋势的主要特征。
在步骤S450，根据至少一个状态特征信号对容量时间序列进行分段，并确定相应的至少一个分段点。步骤S450的处理与前述步骤S140的处理类似。
具体地，基于本实施例是采用Jerk模型下的系统状态方程作为预测容量时间序列所建立的动力学模型，因此，可以Jerk模型下的容量时间序列经Kalman滤波器估计，生成的状态特征信号：包括速度、加速度和加速度导数中的一种信号，对容量时间序列进行分段，并确定相应的至少一个分段点。
在根据状态特征信号对容量时间序列进行分段的过程中，本实施例给出了一种具体的分段方法，包括如图6所示的方法步骤。
参照图6，为本发明提供的根据状态特征信号确定分段点的方法流程图；
在步骤S610，针对容量时间序列经Kalman滤波器估计后得到的速度、加速度、加速度导数的信号中的至少一种信号，确定其存在的至少一个相变信号区域。
由于在容量时间序列中相邻两个pattern之间的过渡阶段，即分段点附近，Kalman滤波器的估计值与容量时间序列数据不匹配。此时Kalman滤波器通过调整自身的状态参数实现匹配估计，这导致分段点附近区域与相邻pattern内的系统状态的状态特征信号有明显不同，这里将这些状态特征信号有明显不同的信号称为相变点或分段疑似点，由这些相变点或分段疑似点构成的区域称为相变信号区域。
由于相变信号区域内的信号与相邻pattern内的信号有着明显的不同特征(如“波动”特征)，因此，可以通对信号进行数学分析，确定至少一个相变信号区域。
本实施例提供了一种采用统计学分析的方法，对经Kalman滤波器提取的系统状态的速度、加速度以及加速度导数等高阶的状态特征信号进行显著性分析，从概率角度检测分段疑似点，然后利用得到的分段疑似点确定相变信号区域。
具体地，该统计学分析的方法包括步骤如下：
步骤1，针对容量时间序列经Kalman滤波器估计后得到的速度、加速度、加速度导数的信号中的至少一种信号，采用两个连续且长度相同的滑动窗口按时间点顺序计算每两个滑动窗口内信号的方差或均值。
设定两个滑动窗口宽度均为24，则按时间点顺序计算的第一对滑动 窗口分别为1～24、25～48，计算这两个窗口内对应的容量时间序列中所有容量的方差或均值。计算完成后，将两个滑动窗口按时间点顺序向后移动，即第二对滑动窗口分别为2～25、26～49，……，依次移动至最后一个时间点，计算每两个滑动窗口内信号的方差或均值。
步骤2，若两个滑动窗口内信号的方差或均值不相同的概率大于预置的相应的概率阈值，则标记两个连续滑动窗口的中间点为一个分段疑似点。
这里，确定一个分段疑似点的原理是依据得到的每对窗口内两个方差或均值之间的不相同的程度来确定，在统计学里，针对方差不同的检验常采用F检验，针对均值不同的检验常采用T检验或置换检验。以下针对两种参数的检验分别进行说明。
1.针对方差不同的检验
假设每次选取的一对窗口Y1、Y2均服从正态分布，且具有相同方差的样本，通过数学分析如Matlab中vartest2函数可以求得Y1、Y2向量方差相同的概率，进而可以求得Y1、Y2两个样本方差不同的概率，当两个方差不同概率大于预置的相应的概率阈值，如0.99时，则认为这两个滑动窗口内的数据的方差不同，并确定两个连续滑动窗口的中间点为一个分段疑似点。例如，针对滑动窗口长度为24的第一对滑动窗口的中间点可以为24、25中任一点。
2.针对均值不同的检验
与针对方差不同的检验类似，首先，设定两个相同长度的滑动窗口，每个窗口Y1、Y2均服从正态分布，且具有相同均值的样本。然后通过统计学中的T检验或置换检验求得Y1、Y2向量均值相同的概率，进而可以求得Y1、Y2两个样本均值不同的概率，当两个均值不同概率大于预置的相应的概率阈值，如0.99时，则认为这两个滑动窗口内的数据的均值不同，并确定两个连续滑动窗口的中间点为一个分段疑似点。例如，针对滑动窗口长度为24的第一对滑动窗口的中间点可以为24、25中任一点。
通过上述对方差不同的检验以及均值不同的检验都可以确定容量时间序列上的至少一个分段疑似点。如图7、图8和图9所示，分别示出了对容量时间序列的速度、加速度以及加速度导数进行方差分析检验 确定的分段疑似点，从图中可以看出，分段疑似点在整个容量时间序列中分布不均，但集中在过渡段的“波动”信号区域。
步骤3，将由多个连续的分段疑似点构成的信号区域确定为容量时间序列的一个相变信号区域。
由于实际发生相变的信号可能遍布于整个容量时间序列中，特别是在相邻的两个pattern之间的过渡阶段内尤为集中。因此，选取检测得到的多个连续的分段疑似点，由这些分段疑似点构成的信号区域作为容量时间序列的一个相变信号区域更为合理。具体地，还可以通过设置确定一个相变信号区域所需的连续分段疑似点的个数以及两段连续分段疑似点区域之间的间距条件来锁定相变信号区域的确定位置。
在步骤S620，根据确定的至少一个相变信号区域对容量时间序列进行分段，确定相应的至少一个分段点。
在确定了容量时间序列中的相变信号区域后，可以选取该区域中的任一点作为分段点。
具体地，可将各相变信号区域中，对应的前述两个滑动窗口内信号的方差的差值中的最大值或均值的差值中的最大值所对应的分段疑似点确定为一个分段点。在这些分段点上，可默认为信号的改变“波动”最大；或者，将各相变信号区域中的中间点确定为一个分段点。这里对分段点确定的规则不做限定。
至此，完成了容量时间序列中分段点的确定。但是，上述分段点的确定过程还存在一个问题，即：被研究确定分段点的对象为容量时间序列的状态特征信号，包括：速度、加速度以及加速度导数信号，那么哪一种信号确定的分段点更为准确呢？为此，本实施例引入了查准率与查全率的概念。以方差分析得到的相变点(分段疑似点)为例，其定义如下：
Figure PCTCN2015089025-appb-000006
Figure PCTCN2015089025-appb-000007
其中，方差分析检测到的真实相变点总数即为最终确定的容量时间序列的分段点总数；方差分析检测到总点数即为总的分段疑似点数；期 望检测到的相变点总数是通过观察样本的基本趋势人为规定的，为常数。查准率越大，表明检测出的真实相变点越多，查全率越大，表明采用方差分析可以检测到的相变点数越多。
通过对如图7、图8和图9所示出的对容量时间序列的速度、加速度以及加速度导数进行方差分析检验确定的分段疑似点进行查准率与查全率计算分析，得到如表1所示的结果：
表1 查准率与查全率计算分析结果
信号 查准率 查全率
速度 71.4 100
加速度 71.4 100
加速度的导数 100 100
从表1中可以看出，由于加速度的导数信号具有很高的查准率和查全率，所以，在实际对容量时间序列进行分段点确定时，可利用加速度导数信号进行方差分析的检测结果进行分段，进而保证后期生成的容量预测模型更为准确。这里说明：尽管对于速度、加速度以及加速度的导数信号都可以通过方差分析来检测分段点，但是不同信号的检测结果却有优劣之分，经过对大量数据样本进行试验，发现一般对加速度或者加速度的导数信号进行方差分析时，检测结果比较好。所以，在本实施例中，优先考虑在加速度或者加速度的导数信号进行分段的基础上进行容量预测。
在此之后，在步骤S480，根据在容量时间序列中确定的至少一个分段点对未来时间的容量进行预测。步骤S480的处理与前述步骤S150的处理类似。
在其基础上，本实施例示出了一种对未来时间的容量进行预测的具体实现方式，即包括步骤S481～步骤S482：
步骤S481：以容量时间序列中最后一个分段点作为起始点，对其后的容量时间序列的数据进行线性或非线性拟合，生成容量预测模型；步骤S482：根据容量预测模型对未来时间的容量进行预测。在实际预测时，也可以Kalman滤波器滤波后的相应位置的容量时间序列进行拟合，生成容量预测模型。
为了使得生成的容量预测模型在满足贴近容量时间序列实际的变化规律的情况下，再对未来时间的容量进行预测。本实施例在步骤S482之前，还可以执行步骤S491～步骤S492，以对生成的容量预测模型的合理和准确性进行判定。
在步骤S491，对生成的容量预测模型进行拟合优度评估。
通常，为了评估拟合模型的优劣，常采用拟合优度进行评估，如图10所示，为采用R2作为评估标准示出的采用线性回归方法对前述图8中加速度信号进行方差分析分段得到的容量时间序列中4个阶段进行拟合，拟合优度及对应阶段序号如表2所示。
表2 容量预测模型的拟合优度评估结果
阶段序号 拟合优度
1 R2＝0.999293
2 R2＝0.997479
3 R2＝0.991105
4 R2＝0.997525
从表2中可以看出，本实施例选用的容量预测模型其四个阶段的拟合优度均大于0.9，甚至大于0.99，这也间接地说明了本实施例采用的容量时间序列的分段方法比较合理。
在步骤S492，若经过拟合优度评估得到的评估值大于预置的拟合优度阈值，则触发根据容量预测模型对未来时间的容量进行预测的处理操作。
其中，预置的拟合优度阈值可以为0.99。结合图10，对本实施例采用最后一个阶段(阶段4)容量时间序列生成容量预测模型进行说明。在容量时间的变化过程中，默认为在不远的未来时间段内，容量以更大概率的接近最后一个稳定变化的阶段的增长趋势进行变化。因此，选用容量时间序列中最后一个分段点为起始点，对其以后的数据进行曲线或直线拟合可以对未来不远处容量数据进行更合理准确的预测。
表3示出了根据图10所示的容量时间序列中最后一个阶段数据生成的容量预测模型对最后一个阶段内的数据进行预测得到的预测值与该阶段真实值的统计结果。
表3 预测值与真实值的结果统计
预测点 真实值 预测值 误差 相对误差(％)
1501 19405616 19626584 220967.71 1.14
1502 19402632 19646480 243847.92 1.26
1503 19401049 19666376 265327.13 1.37
1504 19399736 19686272 286536.34 1.48
1505 19396443 19706169 309725.55 1.60
1506 19394328 19726065 331736.76 1.71
其中，该容量预测模型为y＝19896x-10237629.1031976。x为预测点的位置(与时间节点相对应)，y为容量值。误差为预测值与真实值之差，相对误差＝误差/真实值。从表3中可以看出，所有相对误差均在2％之下，而即使工作经验丰富的工程师仅仅通过人为的方法进行预测的时候，其相对误差也要在5％左右，因此，本实施例生成的容量预测模型具有很好的预测效果。
本发明实施例提供的基于Kalman滤波器的容量预测方法，在图1所示方法实施例的基础上，以Jerk模型下的系统状态方程为例提取出相应的状态转移矩阵A和过程噪声方差矩阵Q来实现对容量时间序列的Kalman滤波器估计；依据Kalman滤波器估计得到的高阶的状态特征信号，如速度、加速度以及加速度的导数对容量时间序列进行分段；最后，以最后一个分段点为起始点，对其以后的容量时间序列进行拟合生成容量预测模型，对未来时间段内的容量进行预测。提高了预测的准确性。
以上所述，仅为本发明的具体实施方式，但本发明的保护范围并不局限于此，任何熟悉本技术领域的技术人员在本发明揭露的技术范围内，可轻易想到变化或替换，都应涵盖在本发明的保护范围之内。因此，本发明的保护范围应以所述权利要求的保护范围为准。
实施例三
图11为本发明提供的基于Kalman滤波器的容量预测系统一个实施例的结构示意图，可用于执行如图1所示实施例的方法步骤。
参照图11，该基于Kalman滤波器的容量预测系统具体包括获取模块111、建立提取模块112、估计生成模块113、分段模块114和预测模 块115；其中：
获取模块111，用于获取待预测对象的容量时间序列；
建立提取模块112，用于为容量时间序列建立动力学模型，并提取动力学模型的状态转移参量和过程噪声参量；
估计生成模块113，用于利用状态转移参量和过程噪声参量对容量时间序列进行Kalman滤波器估计，生成至少一个状态特征信号；
分段模块114，用于根据至少一个状态特征信号对容量时间序列进行分段，并确定相应的至少一个分段点；
预测模块115，用于根据在容量时间序列中确定的至少一个分段点对未来时间的容量进行预测。
本发明实施例提供的基于Kalman滤波器的容量预测系统，通过对获取的容量时间序列搭建动力学模型，然后利用Kalman滤波器对动力学模型下的容量时间序列进行滤波估计生成至少一个状态特征信号；对生成的状态特征信号进行分段，并依据分段点对未来时间的容量进行预测，提高了预测的准确性。
实施例四
图12为本发明提供的基于Kalman滤波器的容量预测系统另一个实施例的结构示意图，可视为图11所示实施例的一种具体实现结构，用于执行如图4所示实施例的方法步骤。
参照图12，该基于Kalman滤波器的容量预测系统包括获取模块111、建立提取模块112、估计生成模块113、分段模块114和预测模块115，且与图11中的相应模块相同。
在此基础上，图12所示的基于Kalman滤波器的容量预测系统还可以包括预处理模块116，用于采用中值滤波和/或移动平均滤波对容量时间序列进行滤波处理，生成滤波后的容量时间序列。
进一步的，上述的动力学模型可以包括：常速度(CV)模型、常加速度(CA)模型、Singer模型、“当前”统计模型、Jerk模型中的至少一种。
进一步的，上述建立提取模块112还可以用于：
构建容量时间序列在Jerk模型下的系统状态方程，并提取系统状态 方程对应的状态转移矩阵A和过程噪声方差矩阵Q。
进一步的，上述的估计生成模块113还用于：
利用状态转移矩阵A和过程噪声方差矩阵Q对容量时间序列进行Kalman滤波器估计，并生成至少一个如下的状态特征信号：速度、加速度和加速度导数。
进一步的，上述的分段模块114还用于：
根据容量时间序列经Kalman滤波器估计后得到的所述速度、加速度和加速度导数中的至少一个状态特征信号对容量时间序列进行分段，并确定相应的至少一个分段点。
进一步的，如图13所示，为上述的分段模块114的具体结构示意图，该分段模块114可包括：
确定单元1141，用于针对容量时间序列经Kalman滤波器估计后得到的速度、加速度、加速度导数的信号中的至少一种信号，确定其存在的至少一个相变信号区域；
分段单元1142，用于根据确定的至少一个相变信号区域对容量时间序列进行分段，确定相应的至少一个分段点。
进一步的，上述的确定单元1141还用于：
针对容量时间序列经Kalman滤波器估计后得到的速度、加速度、加速度导数的信号中的至少一种信号，采用两个连续且长度相同的滑动窗口按时间点顺序计算每两个滑动窗口内信号的方差或均值；
若两个滑动窗口内信号的方差或均值不相同的概率大于预置的相应的概率阈值，则标记两个连续滑动窗口的中间点为一个分段疑似点；
将由多个连续的分段疑似点构成的信号区域确定为容量时间序列的一个相变信号区域。
进一步的，上述分段单元1142还用于：
将各相变信号区域中，对应的两个滑动窗口内信号的方差的差值中的最大值或均值的差值中的最大值所对应的分段疑似点确定为一个分段点。
进一步的，上述分段单元1142还用于：
将各相变信号区域中的中间点确定为一个分段点。
进一步的，如图14所示，为上述的预测模块115的具体结构示意图，如图14所示，该预测模块115还可包括：
预测模型生成单元1151，用于以容量时间序列中最后一个分段点作为起始点，对其后的容量时间序列的数据进行线性或非线性拟合，生成容量预测模型；
预测单元1152，用于根据容量预测模型对未来时间的容量进行预测。
进一步的，如图15所示，在如图12所示的基于Kalman滤波器的容量预测系统的基础上，还可以包括：
拟合优度评估模块117，用于对生成的容量预测模型进行拟合优度评估；
触发模块118，用于若经过拟合优度评估得到的评估值大于预置的拟合优度阈值，则触发预测单元执行根据容量预测模型对未来时间的容量进行预测的处理操作。
上述图4、图5和图6所示的方法实施例的方法步骤，可通过图11～图15中相应的功能模块执行完成，在此对其步骤原理不做赘述。
本发明实施例提供的基于Kalman滤波器的容量预测系统，在图11所示系统实施例的基础上，以Jerk模型下的系统状态方程为例提取出相应的状态转移矩阵A和过程噪声方差矩阵Q来实现对容量时间序列的Kalman滤波器估计；依据Kalman滤波器估计得到的高阶的状态特征信号，如速度、加速度以及加速度的导数对容量时间序列进行分段；最后，以最后一个分段点为起始点，对其以后的容量时间序列进行拟合生成容量预测模型，对未来时间段内的容量进行预测。提高了预测的准确性。
图16为本发明提供的计算机设备一个实施例的结构示意图。
参照图16，计算机设备可用于实施上述实施例中提供的基于Kalman滤波器的容量预测方法。具体来讲：
计算机设备可因配置或性能不同而产生比较大的差异，可以包括一个或一个以上处理器(如Central Processing Units，CPU)710和存储器720。其中，存储器720可以是短暂存储或持久存储。存储器720中可存储有一个或一个以上的程序，每个程序可包括对计算机设备中的一系 列指令操作。更进一步地，处理器710可与存储器720通信，在计算机设备上执行存储器720中的一系列指令操作。特别地，存储器720中还存储有一个或一个以上的操作系统的数据，例如Windows ServerTM，Mac OS XTM，UnixTM，LinuxTM，FreeBSDTM等等。计算机设备还可包括一个或一个以上电源730，一个或一个以上有线或无线网络接口740，一个或一个以上输入输出接口750等。
具体在本实施例中，计算机设备包括一个或者多个处理器710、存储器720，以及一个或者多个程序，一个或者多个程序存储于存储器720中，且经配置以由一个或者多个处理器710执行一个或者多个程序包含的用于执行基于Kalman滤波器的容量预测方法的指令：获取待预测对象的容量时间序列；为所述容量时间序列建立动力学模型，并提取所述动力学模型的状态转移参量和过程噪声参量；利用所述状态转移参量和过程噪声参量对所述容量时间序列进行Kalman滤波器估计，生成至少一个状态特征信号；根据所述至少一个状态特征信号对所述容量时间序列进行分段，并确定相应的至少一个分段点；根据在所述容量时间序列中确定的所述至少一个分段点对未来时间的容量进行预测。
进一步地，存储器720中还包括执行以下处理的指令：采用中值滤波和/或移动平均滤波对所述容量时间序列进行滤波处理，生成滤波后的所述容量时间序列。
可选地，所述动力学模型包括：常速度(CV)模型、常加速度(CA)模型、Singer模型、“当前”统计模型、Jerk模型中的至少一种。
进一步地，存储器720中还包括执行以下处理的指令：构建所述容量时间序列在Jerk模型下的系统状态方程，并提取所述系统状态方程对应的状态转移矩阵A和过程噪声方差矩阵Q。
进一步地，存储器720中还包括执行以下处理的指令：利用所述状态转移矩阵A和所述过程噪声方差矩阵Q对所述容量时间序列进行Kalman滤波器估计，并生成至少一个如下的状态特征信号：速度、加速度和加速度导数。
进一步地，存储器720中还包括执行以下处理的指令：根据所述容量时间序列经所述Kalman滤波器估计后得到的所述速度、加速度和加 速度导数中的至少一个状态特征信号对所述容量时间序列进行分段，并确定相应的至少一个分段点。
优选地，存储器720中还包括执行以下处理的指令：针对所述容量时间序列经所述Kalman滤波器估计后得到的速度、加速度、加速度导数的信号中的至少一种信号，确定其存在的至少一个相变信号区域；根据确定的所述至少一个相变信号区域对所述容量时间序列进行分段，确定相应的至少一个分段点。
进一步地，存储器720中还包括执行以下处理的指令：针对所述容量时间序列经所述Kalman滤波器估计后得到的速度、加速度、加速度导数的信号中的至少一种信号，采用两个连续且长度相同的滑动窗口按时间点顺序计算每所述两个滑动窗口内信号的方差或均值；若所述两个滑动窗口内信号的方差或均值不相同的概率大于预置的相应的概率阈值，则标记所述两个连续滑动窗口的中间点为一个分段疑似点；将由多个连续的所述分段疑似点构成的信号区域确定为所述容量时间序列的一个所述相变信号区域。
进一步地，存储器720中还包括执行以下处理的指令：将各所述相变信号区域中，对应的所述两个滑动窗口内信号的方差的差值中的最大值或均值的差值中的最大值所对应的所述分段疑似点确定为一个分段点。
优选地，存储器720中还包括执行以下处理的指令：将各所述相变信号区域中的中间点确定为一个分段点。
进一步地，存储器720中还包括执行以下处理的指令：以所述容量时间序列中最后一个所述分段点作为起始点，对其后的容量时间序列的数据进行线性或非线性拟合，生成容量预测模型；根据所述容量预测模型对未来时间的容量进行预测。
此外，存储器720中还包括执行以下处理的指令：对生成的所述容量预测模型进行拟合优度评估；若经过所述拟合优度评估得到的评估值大于预置的拟合优度阈值，则触发所述根据所述容量预测模型对未来时间的容量进行预测的处理操作。
本发明提供的计算机设备，通过对获取的容量时间序列搭建动力学 模型，然后利用Kalman滤波器对动力学模型下的容量时间序列进行滤波估计生成至少一个状态特征信号；对生成的状态特征信号进行分段，并依据分段点对未来时间的容量进行预测，提高了预测的准确性。
进一步地，本发明以Jerk模型下的系统状态方程为例提取出相应的状态转移矩阵A和过程噪声方差矩阵Q来实现对容量时间序列的Kalman滤波器估计；依据Kalman滤波器估计得到的高阶的状态特征信号，如速度、加速度以及加速度的导数对容量时间序列进行分段；最后，以最后一个分段点为起始点，对其以后的容量时间序列进行拟合生成容量预测模型，对未来时间段内的容量进行预测。提高了预测的准确性。
上述根据本发明的方法和装置可在硬件、固件中实现，或者被实现为可存储在记录介质(诸如CD ROM、RAM、软盘、硬盘或磁光盘)中的软件或计算机代码，或者被实现通过网络下载的原始存储在远程记录介质或非暂时机器可读介质中并将被存储在本地记录介质中的计算机代码，从而在此描述的方法可被存储在使用通用计算机、专用处理器或者可编程或专用硬件(诸如ASIC或FPGA)的记录介质上的这样的软件处理。可以理解，计算机、处理器、微处理器控制器或可编程硬件包括可存储或接收软件或计算机代码的存储组件(例如，RAM、ROM、闪存等)，当所述软件或计算机代码被计算机、处理器或硬件访问且执行时，实现在此描述的处理方法。此外，当通用计算机访问用于实现在此示出的处理的代码时，代码的执行将通用计算机转换为用于执行在此示出的处理的专用计算机。
以上所述，仅为本发明的具体实施方式，但本发明的保护范围并不局限于此，任何熟悉本技术领域的技术人员在本发明揭露的技术范围内，可轻易想到变化或替换，都应涵盖在本发明的保护范围之内。因此，本发明的保护范围应以所述权利要求的保护范围为准。

Claims (25)

  1. 一种基于Kalman滤波器的容量预测方法，其特征在于，包括：
    获取待预测对象的容量时间序列；
    为所述容量时间序列建立动力学模型，并提取所述动力学模型的状态转移参量和过程噪声参量；
    利用所述状态转移参量和过程噪声参量对所述容量时间序列进行Kalman滤波器估计，生成至少一个状态特征信号；
    根据所述至少一个状态特征信号对所述容量时间序列进行分段，并确定相应的至少一个分段点；
    根据在所述容量时间序列中确定的所述至少一个分段点对未来时间的容量进行预测。
  2. 根据权利要求1所述的方法，其特征在于，所述方法还包括：
    采用中值滤波和/或移动平均滤波对所述容量时间序列进行滤波处理，生成滤波后的所述容量时间序列。
  3. 根据权利要求1所述的方法，其特征在于，所述动力学模型包括：常速度(CV)模型、常加速度(CA)模型、Singer模型、“当前”统计模型、Jerk模型中的至少一种。
  4. 根据权利要求3所述的方法，其特征在于，所述为所述容量时间序列建立动力学模型，并提取所述动力学模型的状态转移参量和过程噪声参量的处理包括：
    构建所述容量时间序列在Jerk模型下的系统状态方程，并提取所述系统状态方程对应的状态转移矩阵A和过程噪声方差矩阵Q。
  5. 根据权利要求4所述的方法，其特征在于，所述利用所述状态转移参量和过程噪声参量对所述容量时间序列进行Kalman滤波器估计，生成至少一个状态特征信号的处理包括：
    利用所述状态转移矩阵A和所述过程噪声方差矩阵Q对所述容量时间序列进行Kalman滤波器估计，并生成至少一个如下的状态特征信号：速度、加速度和加速度导数。
  6. 根据权利要求5所述的方法，其特征在于，所述根据所述至少一个状态特征信号对所述容量时间序列进行分段，并确定相应的至少一 个分段点的处理包括：
    根据所述容量时间序列经所述Kalman滤波器估计后得到的所述速度、加速度和加速度导数中的至少一个状态特征信号对所述容量时间序列进行分段，并确定相应的至少一个分段点。
  7. 根据权利要求6所述的方法，其特征在于，所述根据所述容量时间序列经所述Kalman滤波器估计后得到的所述速度、加速度和加速度导数中的至少一个状态特征信号对所述容量时间序列进行分段，并确定相应的至少一个分段点的处理包括：
    针对所述容量时间序列经所述Kalman滤波器估计后得到的速度、加速度、加速度导数的信号中的至少一种信号，确定其存在的至少一个相变信号区域；
    根据确定的所述至少一个相变信号区域对所述容量时间序列进行分段，确定相应的至少一个分段点。
  8. 根据权利要求7所述的方法，其特征在于，所述针对所述容量时间序列经所述Kalman滤波器估计后得到的速度、加速度、加速度导数的信号中的至少一种信号，确定其存在的至少一个相变信号区域的处理包括：
    针对所述容量时间序列经所述Kalman滤波器估计后得到的速度、加速度、加速度导数的信号中的至少一种信号，采用两个连续且长度相同的滑动窗口按时间点顺序计算每所述两个滑动窗口内信号的方差或均值；
    若所述两个滑动窗口内信号的方差或均值不相同的概率大于预置的相应的概率阈值，则标记所述两个连续滑动窗口的中间点为一个分段疑似点；
    将由多个连续的所述分段疑似点构成的信号区域确定为所述容量时间序列的一个所述相变信号区域。
  9. 根据权利要求8所述的方法，其特征在于，所述根据确定的所述至少一个相变信号区域对所述容量时间序列进行分段，确定相应的至少一个分段点的处理包括：
    将各所述相变信号区域中，对应的所述两个滑动窗口内信号的方差 的差值中的最大值或均值的差值中的最大值所对应的所述分段疑似点确定为一个分段点。
  10. 根据权利要求8所述的方法，其特征在于，所述根据确定的所述至少一个相变信号区域对所述容量时间序列进行分段，确定相应的至少一个分段点的处理包括：
    将各所述相变信号区域中的中间点确定为一个分段点。
  11. 根据权利要求9或10所述的方法，其特征在于，所述根据在所述容量时间序列中确定的所述至少一个分段点对未来时间的容量进行预测的处理包括：
    以所述容量时间序列中最后一个所述分段点作为起始点，对其后的容量时间序列的数据进行线性或非线性拟合，生成容量预测模型；
    根据所述容量预测模型对未来时间的容量进行预测。
  12. 根据权利要求11所述的方法，其特征在于，所述方法还包括：
    对生成的所述容量预测模型进行拟合优度评估；
    若经过所述拟合优度评估得到的评估值大于预置的拟合优度阈值，则触发所述根据所述容量预测模型对未来时间的容量进行预测的处理操作。
  13. 一种基于Kalman滤波器的容量预测系统，其特征在于，所述系统包括：
    获取模块，用于获取待预测对象的容量时间序列；
    建立提取模块，用于为所述容量时间序列建立动力学模型，并提取所述动力学模型的状态转移参量和过程噪声参量；
    估计生成模块，用于利用所述状态转移参量和过程噪声参量对所述容量时间序列进行Kalman滤波器估计，生成至少一个状态特征信号；
    分段模块，用于根据所述至少一个状态特征信号对所述容量时间序列进行分段，并确定相应的至少一个分段点；
    预测模块，用于根据在所述容量时间序列中确定的所述至少一个分段点对未来时间的容量进行预测。
  14. 根据权利要求13所述的系统，其特征在于，所述系统还包括：
    预处理模块，用于采用中值滤波和/或移动平均滤波对所述容量时间 序列进行滤波处理，生成滤波后的所述容量时间序列。
  15. 根据权利要求13所述的系统，其特征在于，所述动力学模型包括：常速度(CV)模型、常加速度(CA)模型、Singer模型、“当前”统计模型、Jerk模型中的至少一种。
  16. 根据权利要求15所述的系统，其特征在于，所述建立提取模块用于：
    构建所述容量时间序列在Jerk模型下的系统状态方程，并提取所述系统状态方程对应的状态转移矩阵A和过程噪声方差矩阵Q。
  17. 根据权利要求16所述的系统，其特征在于，所述估计生成模块用于：
    利用所述状态转移矩阵A和所述过程噪声方差矩阵Q对所述容量时间序列进行Kalman滤波器估计，并生成至少一个如下的状态特征信号：速度、加速度和加速度导数。
  18. 根据权利要求17所述的系统，其特征在于，所述分段模块用于：
    根据所述容量时间序列经所述Kalman滤波器估计后得到的所述速度、加速度和加速度导数中的至少一个状态特征信号对所述容量时间序列进行分段，并确定相应的至少一个分段点。
  19. 根据权利要求18所述的系统，其特征在于，所述分段模块包括：
    确定单元，用于针对所述容量时间序列经所述Kalman滤波器估计后得到的速度、加速度、加速度导数的信号中的至少一种信号，确定其存在的至少一个相变信号区域；
    分段单元，用于根据确定的所述至少一个相变信号区域对所述容量时间序列进行分段，确定相应的至少一个分段点。
  20. 根据权利要求19所述的系统，其特征在于，所述确定单元还用于：
    针对所述容量时间序列经所述Kalman滤波器估计后得到的速度、加速度、加速度导数的信号中的至少一种信号，采用两个连续且长度相同的滑动窗口按时间点顺序计算每所述两个滑动窗口内信号的方差或 均值；
    若所述两个滑动窗口内信号的方差或均值不相同的概率大于预置的相应的概率阈值，则标记所述两个连续滑动窗口的中间点为一个分段疑似点；
    将由多个连续的所述分段疑似点构成的信号区域确定为所述容量时间序列的一个所述相变信号区域。
  21. 根据权利要求20所述的系统，其特征在于，所述分段单元还用于：
    将各所述相变信号区域中，对应的所述两个滑动窗口内信号的方差的差值中的最大值或均值的差值中的最大值所对应的所述分段疑似点确定为一个分段点。
  22. 根据权利要求20所述的系统，其特征在于，所述分段单元还用于：
    将各所述相变信号区域中的中间点确定为一个分段点。
  23. 根据权利要求21或22所述的系统，其特征在于，所述预测模块包括：
    预测模型生成单元，用于以所述容量时间序列中最后一个所述分段点作为起始点，对其后的容量时间序列的数据进行线性或非线性拟合，生成容量预测模型；
    预测单元，用于根据所述容量预测模型对未来时间的容量进行预测。
  24. 根据权利要求23所述的系统，其特征在于，所述系统还包括：
    拟合优度评估模块，用于对生成的所述容量预测模型进行拟合优度评估；
    触发模块，用于若经过所述拟合优度评估得到的评估值大于预置的拟合优度阈值，则触发所述预测单元执行根据所述容量预测模型对未来时间的容量进行预测的处理操作。
  25. 一种计算机设备，其特征在于，包括：
    一个或多个处理器；
    存储器；
    一个或多个程序，所述一个或多个程序存储在所述存储器中，且经配置以由所述一个或者多个处理器执行所述一个或者多个程序包含的用于执行基于Kalman滤波器的容量预测方法的指令：
    获取待预测对象的容量时间序列；
    为所述容量时间序列建立动力学模型，并提取所述动力学模型的状态转移参量和过程噪声参量；
    利用所述状态转移参量和过程噪声参量对所述容量时间序列进行Kalman滤波器估计，生成至少一个状态特征信号；
    根据所述至少一个状态特征信号对所述容量时间序列进行分段，并确定相应的至少一个分段点；
    根据在所述容量时间序列中确定的所述至少一个分段点对未来时间的容量进行预测。
PCT/CN2015/089025 2015-04-03 2015-09-07 基于Kalman滤波器的容量预测方法、系统和计算机设备 WO2016155241A1 (zh)

Priority Applications (4)

Application Number Priority Date Filing Date Title
EP15849818.8A EP3279819B1 (en) 2015-04-03 2015-09-07 Method, system and computer device for capacity prediction based on kalman filter
JP2016526273A JP6343001B2 (ja) 2015-04-03 2015-09-07 カルマンフィルタに基づく容量予測方法、システム及びコンピュータ機器
KR1020167011244A KR101829560B1 (ko) 2015-04-03 2015-09-07 칼만 필터를 기반으로 하는 용량 예측 방법, 시스템 및 컴퓨터 장치
US15/033,604 US10437942B2 (en) 2015-04-03 2015-09-07 Kalman filter based capacity forecasting method, system and computer equipment

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201510158699.9A CN104809333B (zh) 2015-04-03 2015-04-03 基于Kalman滤波器的容量预测方法和系统
CN2015101586999 2015-04-03

Publications (1)

Publication Number Publication Date
WO2016155241A1 true WO2016155241A1 (zh) 2016-10-06

Family

ID=53694149

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2015/089025 WO2016155241A1 (zh) 2015-04-03 2015-09-07 基于Kalman滤波器的容量预测方法、系统和计算机设备

Country Status (6)

Country Link
US (1) US10437942B2 (zh)
EP (1) EP3279819B1 (zh)
JP (1) JP6343001B2 (zh)
KR (1) KR101829560B1 (zh)
CN (1) CN104809333B (zh)
WO (1) WO2016155241A1 (zh)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110649911A (zh) * 2019-07-17 2020-01-03 电子科技大学 一种基于α散度的分布式非线性卡尔曼滤波方法
CN110727923A (zh) * 2019-11-01 2020-01-24 江苏省有色金属华东地质勘查局地球化学勘查与海洋地质调查研究院 一种短期潮位预报方法
CN110895727A (zh) * 2019-11-14 2020-03-20 佛山科学技术学院 基于卡尔曼滤波模型的鱼塘水质预测方法及系统

Families Citing this family (19)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104809333B (zh) * 2015-04-03 2017-08-29 百度在线网络技术（北京）有限公司 基于Kalman滤波器的容量预测方法和系统
CN105094708B (zh) 2015-08-25 2018-06-12 北京百度网讯科技有限公司 一种磁盘容量的预测方法及装置
CN108615017B (zh) * 2018-04-28 2019-04-02 山东科技大学 滑动窗分段直线表示法的总拟合误差阈值估计方法与系统
CN109067381B (zh) * 2018-07-05 2023-05-02 无锡北微传感科技有限公司 一种mems陀螺仪随机噪声的实时滤波系统及方法
CN109545113B (zh) * 2018-09-30 2022-09-06 合肥鑫晟光电科技有限公司 一种异常数据处理方法及装置
CN109408545B (zh) * 2018-09-30 2019-10-25 中科驭数（北京）科技有限公司 目标聚合操作处理方法和装置
WO2021000278A1 (zh) * 2019-07-03 2021-01-07 深圳市速腾聚创科技有限公司 一种多传感器状态估计方法、装置及终端设备
CN110493072A (zh) * 2019-07-11 2019-11-22 网宿科技股份有限公司 基于深度学习的带宽滤波方法、装置、服务器及存储介质
US11005689B2 (en) 2019-07-11 2021-05-11 Wangsu Science & Technology Co., Ltd. Method and apparatus for bandwidth filtering based on deep learning, server and storage medium
CN110414442B (zh) * 2019-07-31 2022-04-01 广东省智能机器人研究院 一种压力时序数据分段特征值预测方法
CN111002858B (zh) * 2019-12-18 2023-06-20 中兴新能源汽车有限责任公司 一种无线充电引导定位系统及方法、车载设备
CN111044907B (zh) * 2019-12-24 2022-06-14 苏州正力新能源科技有限公司 一种基于微片段数据和电压滤波的soh统计方法
CN111152795B (zh) * 2020-01-08 2022-12-13 东南大学 一种基于模型和参数动态调整的自适应车辆状态预测系统及预测方法
CN111313985B (zh) * 2020-03-05 2022-05-13 北京振中电子技术有限公司 宽带电力线载波通信模拟噪声生成方法、装置及电子设备
CN111625196A (zh) * 2020-05-26 2020-09-04 北京海益同展信息科技有限公司 区块链节点扩容方法、装置、计算机设备及存储介质
CN112958840B (zh) * 2021-02-10 2022-06-14 西南电子技术研究所（中国电子科技集团公司第十研究所） 精密零件加工削力信号自动分段方法
CN113859236A (zh) * 2021-09-07 2021-12-31 中汽创智科技有限公司 一种跟车控制系统、车辆、方法、装置、设备及存储介质
CN114660638A (zh) * 2022-03-31 2022-06-24 大连海事大学 一种锁频环辅助的矢量相位锁定环路系统
CN115097526B (zh) * 2022-08-22 2022-11-11 江苏益捷思信息科技有限公司 地震采集资料质量评价方法

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2011019979A2 (en) * 2009-08-13 2011-02-17 Siemens Corporation System and method for resilient control over wireless networks
CN102608631A (zh) * 2011-10-28 2012-07-25 北京航空航天大学 基于模糊逻辑的自适应强跟踪ukf定位滤波算法
CN103399496A (zh) * 2013-08-20 2013-11-20 中国能源建设集团广东省电力设计研究院 智能电网海量实时数据负载仿真测试云平台及其测试方法
CN104809333A (zh) * 2015-04-03 2015-07-29 百度在线网络技术（北京）有限公司 基于Kalman滤波器的容量预测方法和系统

Family Cites Families (18)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US6993458B1 (en) * 2000-11-07 2006-01-31 International Business Machines Corporation Method and apparatus for preprocessing technique for forecasting in capacity management, software rejuvenation and dynamic resource allocation applications
JP4083997B2 (ja) * 2001-04-16 2008-04-30 株式会社日立製作所 レンタルストレージのサービス方法、および、レンタルストレージシステム
US7043899B2 (en) * 2004-05-26 2006-05-16 Daimlerchrysler Corporation On-line catalyst monitoring using a Kalman filter
JP2007087121A (ja) * 2005-09-22 2007-04-05 Nec Corp 全体管理装置、データベースシステム、データベース管理方法、および、データベース管理プログラム
DE102005050563A1 (de) 2005-10-21 2007-04-26 Robert Bosch Gmbh Verfahren zur Vorhersage der Leistungsfähigkeit elektrischer Energiespeicher
US20070198328A1 (en) * 2006-02-09 2007-08-23 Fuller William T Storage Capacity Planning
JP2009064354A (ja) * 2007-09-07 2009-03-26 Hitachi Information Systems Ltd データベースの容量監視システム及びデータベースの容量予測方法
US20090228225A1 (en) 2008-03-04 2009-09-10 Eaton Corporation Battery Service Life Estimation Methods, Apparatus and Computer Program Products Using State Estimation Techniques Initialized Using a Regression Model
CN101629992B (zh) * 2009-05-27 2011-11-16 重庆大学 磷酸铁锂动力电池剩余容量的估算方法
JP2011005039A (ja) * 2009-06-26 2011-01-13 Toshiba Corp 医用画像診断装置
CN101598769B (zh) * 2009-06-29 2011-04-20 杭州电子科技大学 一种基于采样点卡尔曼滤波的电池剩余电量估计方法
ES2672222T3 (es) * 2010-02-15 2018-06-13 Carrier Corporation Sistema y método basados en modelos para estimar parámetros y estados en espacios de temperatura controlada
US8751463B1 (en) * 2011-06-30 2014-06-10 Emc Corporation Capacity forecasting for a deduplicating storage system
KR101271694B1 (ko) * 2012-01-09 2013-06-05 전남대학교산학협력단 시계열데이터의 연속적 결측값 대체 시스템 및 그 방법
US9261863B2 (en) 2012-01-23 2016-02-16 Earth Networks, Inc. Optimizing and controlling the energy consumption of a building
US9633359B2 (en) * 2012-08-10 2017-04-25 Itron, Inc. Near-term data filtering, smoothing and load forecasting
KR101430258B1 (ko) * 2013-02-27 2014-08-14 부경대학교 산학협력단 디지털 감시 카메라 시스템에서의 자동 장면 전환 검출 방법
KR20140147456A (ko) 2013-06-20 2014-12-30 엘에스산전 주식회사 상수관망의 일일 수요량 예측방법

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2011019979A2 (en) * 2009-08-13 2011-02-17 Siemens Corporation System and method for resilient control over wireless networks
CN102608631A (zh) * 2011-10-28 2012-07-25 北京航空航天大学 基于模糊逻辑的自适应强跟踪ukf定位滤波算法
CN103399496A (zh) * 2013-08-20 2013-11-20 中国能源建设集团广东省电力设计研究院 智能电网海量实时数据负载仿真测试云平台及其测试方法
CN104809333A (zh) * 2015-04-03 2015-07-29 百度在线网络技术（北京）有限公司 基于Kalman滤波器的容量预测方法和系统

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110649911A (zh) * 2019-07-17 2020-01-03 电子科技大学 一种基于α散度的分布式非线性卡尔曼滤波方法
CN110649911B (zh) * 2019-07-17 2023-10-27 电子科技大学 一种基于α散度的分布式非线性卡尔曼滤波方法
CN110727923A (zh) * 2019-11-01 2020-01-24 江苏省有色金属华东地质勘查局地球化学勘查与海洋地质调查研究院 一种短期潮位预报方法
CN110727923B (zh) * 2019-11-01 2023-02-24 江苏省有色金属华东地质勘查局地球化学勘查与海洋地质调查研究院 一种短期潮位预报方法
CN110895727A (zh) * 2019-11-14 2020-03-20 佛山科学技术学院 基于卡尔曼滤波模型的鱼塘水质预测方法及系统

Also Published As

Publication number Publication date
CN104809333B (zh) 2017-08-29
EP3279819A4 (en) 2018-12-26
EP3279819A1 (en) 2018-02-07
JP2017520029A (ja) 2017-07-20
EP3279819B1 (en) 2024-04-24
KR20160130209A (ko) 2016-11-10
KR101829560B1 (ko) 2018-02-14
US20180196891A1 (en) 2018-07-12
JP6343001B2 (ja) 2018-06-13
US10437942B2 (en) 2019-10-08
CN104809333A (zh) 2015-07-29

Similar Documents

Publication Publication Date Title
WO2016155241A1 (zh) 基于Kalman滤波器的容量预测方法、系统和计算机设备
CN107765347B (zh) 一种高斯过程回归和粒子滤波的短期风速预测方法
CN107145645B (zh) 带不确定冲击的非平稳退化过程剩余寿命预测方法
US9779361B2 (en) Method for learning exemplars for anomaly detection
US20190285517A1 (en) Method for evaluating health status of mechanical equipment
Huang et al. Remaining useful life prediction for a nonlinear heterogeneous Wiener process model with an adaptive drift
Gašperin et al. Model-based prognostics of gear health using stochastic dynamical models
JP2015011027A (ja) 時系列データにおける異常を検出する方法
CN108960329B (zh) 一种包含缺失数据的化工过程故障检测方法
JP2019112049A (ja) 陸上車両の運転者の運転スタイルを認識するための方法及び対応する装置
CN103678869A (zh) 一种飞行参数缺失数据的预测估计方法
WO2019161589A1 (zh) 一种结构模态参数实时追踪方法
CN110375736B (zh) 智能设备的路径规划方法、系统、设备及可读存储介质
CN112070272A (zh) 一种输电线路的覆冰厚度预测方法及装置
WO2021068589A1 (zh) 一种图像中物体及其关键点的确定方法和装置
US20240038324A1 (en) Knowledge reuse-based method and system for predicting cell concentration in fermentation process
CN113449254A (zh) 任意网型变形监测稳定性分析方法及监控点位置确定方法
CN109065176B (zh) 一种血糖预测方法、装置、终端和存储介质
WO2016098250A1 (ja) 波形推定装置及び波形推定方法
CN113986700A (zh) 数据采集频率的优化方法、系统、装置及存储介质
CN110083804B (zh) 基于条件分布回归的风电场scada数据缺失的智能修复方法
JP6116398B2 (ja) 波形推定装置及び波形推定方法
CN117131465B (zh) 单管塔损伤识别方法、装置、电子设备及可读存储介质
Kosorus et al. On the identification of frequencies and damping ratios for structural health monitoring using autoregressive models
CN117723782B (zh) 用于桥梁结构健康监测的传感器故障识别定位方法及系统

Legal Events

Date Code Title Description
REEP Request for entry into the european phase

Ref document number: 2015849818

Country of ref document: EP

ENP Entry into the national phase

Ref document number: 2016526273

Country of ref document: JP

Kind code of ref document: A

ENP Entry into the national phase

Ref document number: 20167011244

Country of ref document: KR

Kind code of ref document: A

121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 15849818

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

WWE Wipo information: entry into national phase

Ref document number: 2015849818

Country of ref document: EP