US20180053111A1 - Anomaly Detection for Complex Physical System via Nonlinear Relationship Modeling - Google Patents
Anomaly Detection for Complex Physical System via Nonlinear Relationship Modeling Download PDFInfo
- Publication number
- US20180053111A1 US20180053111A1 US15/680,831 US201715680831A US2018053111A1 US 20180053111 A1 US20180053111 A1 US 20180053111A1 US 201715680831 A US201715680831 A US 201715680831A US 2018053111 A1 US2018053111 A1 US 2018053111A1
- Authority
- US
- United States
- Prior art keywords
- time series
- determining
- predictive model
- periodic
- pair
- 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.)
- Abandoned
Links
Images
Classifications
-
- 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
- G05B17/00—Systems involving the use of models or simulators of said systems
- G05B17/02—Systems involving the use of models or simulators of said systems electric
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N7/00—Computing arrangements based on specific mathematical models
- G06N7/08—Computing arrangements based on specific mathematical models using chaos models or non-linear system models
-
- 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
- G06F17/14—Fourier, Walsh or analogous domain transformations, e.g. Laplace, Hilbert, Karhunen-Loeve, transforms
- G06F17/141—Discrete Fourier transforms
- G06F17/142—Fast Fourier transforms, e.g. using a Cooley-Tukey type algorithm
-
- 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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
- G06F18/2134—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on separation criteria, e.g. independent component analysis
- G06F18/21345—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on separation criteria, e.g. independent component analysis enforcing sparsity or involving a domain transformation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
- G06F2218/12—Classification; Matching
- G06F2218/16—Classification; Matching by matching signal segments
- G06F2218/20—Classification; Matching by matching signal segments by applying autoregressive analysis
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02B—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO BUILDINGS, e.g. HOUSING, HOUSE APPLIANCES OR RELATED END-USER APPLICATIONS
- Y02B90/00—Enabling technologies or technologies with a potential or indirect contribution to GHG emissions mitigation
- Y02B90/20—Smart grids as enabling technology in buildings sector
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y04—INFORMATION OR COMMUNICATION TECHNOLOGIES HAVING AN IMPACT ON OTHER TECHNOLOGY AREAS
- Y04S—SYSTEMS INTEGRATING TECHNOLOGIES RELATED TO POWER NETWORK OPERATION, COMMUNICATION OR INFORMATION TECHNOLOGIES FOR IMPROVING THE ELECTRICAL POWER GENERATION, TRANSMISSION, DISTRIBUTION, MANAGEMENT OR USAGE, i.e. SMART GRIDS
- Y04S20/00—Management or operation of end-user stationary applications or the last stages of power distribution; Controlling, monitoring or operating thereof
Definitions
- the present invention relates to anomaly detection and, more particularly, to modeling non-linear relationships between time series to test for abnormal behavior.
- Complex physical systems may include thousands of individual components, each having a different set of relationships with the other components. These relationships may be modeled as being, e.g., linear, quadratic, trigonometric, etc. In such a system, the failure of one or more components may lead to a breakdown of the functioning of the entire system.
- sensors may be added to monitor different components of the system with a very fine granularity.
- the readings from the sensors may be modeled as a time series of measured values.
- Current analysis techniques use kernel regression to model nonlinear relationships in time series. However, these techniques do not respect the underlying physical laws of the system and cannot express the modeled relationship in the form of formulas for diagnosis.
- a method for detecting anomalies includes determining a predictive model for each pair of a set of time series, each time series being associated with a component of a system. New values of each pair of time series are compared to values predicted by the respective predictive model to determine if the respective predictive model is broken. A number of broken predictive models is determined. An anomaly alert is generated if the number of broken predictive models exceeds a threshold.
- a method for detecting anomalies includes determining a predictive model for each pair of a set of time series, each time series being associated with a component of a system, by determining a periodic relationship between the time series and a non-periodic relationship between the time series. New values of pairs of time series having predictive models that have a fitness score higher than a threshold are compared to values predicted by the respective predictive model to determine if the respective predictive model is broken. A number of broken predictive models is determined. An anomaly alert is generated if the number of broken predictive models exceeds a threshold.
- a system for detecting anomalies includes a fitting module configured to determine a predictive model for each pair of a set of time series, each time series being associated with a component of a system.
- An anomaly detection module is configured to compare new values of each pair of time series to values predicted by the respective predictive model to determine if the respective predictive model is broken and to determine a number of broken predictive models.
- An alert module is configured to generate an anomaly alert if the number of broken predictive models exceeds a threshold.
- FIG. 1 is a block diagram of a monitoring system for anomaly detection and response in accordance with an embodiment of the present invention
- FIG. 2 is a block/flow diagram of anomaly detection and response in accordance with an embodiment of the present invention
- FIG. 3 is a block/flow diagram of training predictive models for pairs of time series in accordance with an embodiment of the present invention
- FIG. 4 is a block diagram of a processing system in accordance with an embodiment of the present invention.
- FIG. 5 is a block diagram of a power plant with sensors for physical systems and a monitoring system in accordance with an embodiment of the present invention.
- Embodiments of the present invention provide modeling and anomaly detection processes and systems that model the nonlinear, complex relationship between time series generated by sensors.
- the present embodiments use such models to detect anomalous behavior based on the nature of the physical systems being monitored and on observations from production workload patterns. Not only can the nonlinear relationships between sensors be so identified, but those relationships can be expressed as close-form formulas, greatly assisting operators in the system operators who wish to incorporate their domain knowledge for diagnosis.
- a regression function is used to model a non-linear relationship pairs of time series.
- One time series is fit using the function of another time series.
- This function is designed respecting the physical laws and workload patterns inside the physical system, providing a template of bases that include, for example, trigonometric and polynomial relationships.
- the regression is fitted by selecting correct bases out of the template.
- the relationship between each pair is then expressed using the estimated regression function.
- the fitting problem may be formulated as a Sparse Group Lasso with Overlapping problem.
- the formulation is further improved by decomposing the coefficient matrix into two matrices that are solved separately, greatly improving the solution speed.
- a function is formulated for every pair of the time series. A set of relationships is selected out of the pairs by choosing those with a high fitness score.
- the trained models are applied to newly acquired time series. For each pair of time series it is determined whether the pair matches the training model. If so, for each time point in the two time series, a fitted value is generated and compared with the measured value. The trained model is considered broken if the difference between the fitted value and the measured value is high (e.g., if the difference exceeds a threshold). The number of broken models is counted and time points that have a high number of broken models are considered anomalies.
- the present embodiments thus accurately identify anomalies, providing early warning to operators to avoid system failure.
- Explicit sensor relationships are provided to the operators to help them diagnose the potential problems.
- Large real-world datasets can be processed quickly in this manner.
- the monitoring system includes a hardware processor 102 and memory 104 . It should be understood that the monitoring system 100 may be distinct from the monitored system, or may be integrated therewith.
- the monitoring system 100 therefore may include a number of physical components 106 , or these physical components 106 may be external to the monitoring system 100 .
- These physical components 106 may include any include any type of device in a complex system such as, e.g., a power generator.
- the physical components 106 may include, for example, servers in a rack enclosure with their attendant power, cooling, and communications components.
- the physical components 106 are each fitted with one or more sensors 108 that measure operational parameters of the physical components 106 .
- These operational parameters may include, for example, temperature, power consumption, data throughput, vibrational amplitude, luminosity, physical location, humidity, contaminant concentration, and any other measurable quantity.
- the sensors 108 store their measurements in memory 104 as time series, with each measurement being stored with a respective timestamp.
- the sensors 108 keep a common clock, such that measurements stored at a particular time by multiple sensors 108 will share the same, or a very similar, timestamp.
- the monitoring system 100 also includes functional modules that may, in some embodiments, be implemented as software that is stored in memory 104 and executed by hardware processor 102 .
- the functional modules may be implemented as one or more discrete hardware components in the form of, e.g., application specific integrated chips or field programmable gate arrays.
- a fitting module 110 analyzes a set of time series stored in the memory 104 and generates models that represent the relationship between each respective pair of time series. The models are thus based on historical measurements from the sensors 108 and characterize normal operation of the physical components 106 . Anomaly detection module 112 then compares new measurements of from the sensors 108 to the values predicted by the models and, if a sufficient number of discrepancies (referred to herein as “broken models”) are found, anomaly detection module 112 flags a particular timestamp as being anomalous.
- alert module 114 takes one of a number of actions.
- the alert module simply generates an alert to a human operator, including the anomalous timestamp and information regarding the relationships between sensors that are characterized by the broken models.
- the report may indicate which physical components 106 are at issue, what the expected behavior was, and what the measured behavior is.
- the report may also indicate a measure of the magnitude of the anomaly (e.g., the number of broken models).
- the alert module 114 may be configured to take automatic actions to address anomalies.
- the alert module 114 may automatically issue instructions to those components 106 to slow or stop their processes.
- Block 202 trains a model for each pair of time series stored in the memory 104 .
- each time series corresponds to a respective sensor 108 that measures a physical component 106 .
- the model training will be described in greater detail below.
- Block 204 then trims out the models that have a fitness score below a threshold value. The fitness score may be based on, e.g., the r-squared error.
- Block 206 tests pairs of current measurements from the sensors 108 against the respective models. Block 210 then counts the number of broken models at each timestamp (i.e., the number of pairs for which the measured values differ from values predicted by the model by more than a threshold amount).
- Block 212 determines whether a number of broken models exceeds an anomaly threshold. If not, processing returns to block 206 and the current measurements from the sensors 108 at a new timestamp are tested. If the number of broken models does exceed the anomaly threshold, then block 214 generates an alert by, e.g., sending a report to a human operator or taking an automated administrative action.
- Block 306 performs coefficient correction for the discovered functions.
- Periodic signals can be cleanly separated out from polynomial signals using a fast Fourier transform (FFT), which turns a time-domain series into a frequency-domain series, showing high amplitudes for periodic components.
- FFT fast Fourier transform
- Non-periodic components are separated out from both the periodic components and from the noise.
- Periodic patterns may result from repeated workload patterns. Periodic components are shown as peaks in higher-frequency portions of the FFT output, while nonperiodic components are shown as a peak in the low-frequency portion of the output.
- block 304 can operate on only the nonperiodic portion.
- Block 302 further determines the time shift of the periodic component, with values for x at different points in the time series being used in the FFT.
- the results using incorrect time shifts will show only noise, while strong signals will be evident at particular frequencies when the time shift is correct.
- the fitting error (also known as the residue) is used as an input for the polynomial relationship discovery in block 304 .
- x is considered the independent signal and y is considered the response signal.
- Each power base includes the current sensor value and ⁇ previous sensor values.
- a ( ⁇ +1) ⁇ d(d+2) kernel matrix is defined as:
- K ( x , y ; t ) [ k ( x ; t ) T k ( x ; t - 1 ) T ... k ( x ; t - ⁇ ) T 0 k ( x , y ; t - 1 ) T ... k ( x , y ; t - ⁇ ) T ]
- Each matrix k (x,y;t) in K (x,y;t) is set to zero because the values for y t in k (x,y;t) cannot be modeled on y t itself.
- the index t of K (x,y;t) starts from ⁇ +1 to have enough delays in the first training example. If K (x,y;t) is a feature matrix, then a model with ⁇ delays can be expressed as:
- W is a model parameter matrix
- block 304 treats each base with the same orders as a group and all the bases at each delay as a group.
- each row of K (x,y;t) is a structured group, where all the bases are at a same delay
- each column is a structured group, with the same base at all delays.
- Regularizers are applied to the model parameter matrix W to select relevant groups of bases and delays and to add sparsity within the groups to control the model capacity.
- Block 204 therefore forms an optimization problem as a sparse group lasso with overlapping structures:
- W has overlapping structures
- the present embodiments provide a special case because the parameters naturally form a matrix, where row groups only overlap with column groups, but never with themselves. Similarly, column groups only overlap with row groups, but never with themselves.
- W can be decomposed into a summation of two components, P and Q, with only row groups being selected in P and column groups being selected in Q. Both rows and columns are selected in W after summation. This decomposition degenerates the regularization part of the original selection to two easier, sparse group Lasso penalties on both P and Q.
- the 1,2+1 matrix norm can be used. More particularly, the 1,2+1 norm is applied row-wise to P and the optimal solution of P will include mostly rows with all-zero entries and a few rows with non-zero entries. Similarly, the 1,2+1 norm of Q T encourages selection among columns of Q.
- W in the least loss is replaced by P+Q.
- the parameter ⁇ plays a role of balancing the weight between the 1,2 norm and the 1 norm.
- the regularization is decomposed into two sparse group Lasso penalties, the least squares loss part is not.
- P and Q are estimated together, because they both contribute to the optimal solution of W. Therefore existing sparse group Lasso solvers are inadequate to address this problem.
- the optimization problem can be iteratively solved using fast iterative shrinkage-thresholding algorithm.
- the optimization problem may be reformulated as:
- R(P, Q) is the non-smooth part that constitutes the regularizers, is an iteration number, is the continuously differentiable part
- ⁇ P and ⁇ Q are the partial derivatives of at ( ⁇ tilde over (P) ⁇ , ⁇ tilde over (Q) ⁇ ), ⁇ tilde over (P) ⁇ and ⁇ tilde over (Q) ⁇ are defined as:
- t is a scalar that is updated at each iteration as:
- the optimization problem is solved to obtain P and Q .
- the value of t +1 , ⁇ tilde over (P) ⁇ +1 , ⁇ tilde over (Q) ⁇ +1 are calculated as described above.
- the iteration number is incremented and these steps are repeated. Iteration continues until the values for P and Q converge.
- the optimization converges to an optimal solution with a convergence rate of
- the sum of the final values of P and Q is the parameter matrix W.
- a template of bases is generated using different orders of polynomial relationships with different time delays. Bases are generated with at most degree-d and ⁇ delays to model y as a function of x. Thus the function is expressed as:
- the x terms are bases, while the w terms are learning coefficients of the parameter matrix W.
- the problem is formulated as Sparse Group Lasso with Overlapping and can be solved efficiently.
- the output of block 304 is an estimated function for each pair of the time series.
- Block 306 corrects the coefficients of the functions generated by block 304 . After solving the Group Lasso problem, the feature coefficients are set to zero and non-zero values (such that particular coefficients are selected or deselected). The non-zero valued features may not represent the true coefficients. Block 306 corrects these coefficients using linear regression, regressing each estimated function to the time series to minimize the regression error. For example, if the coefficients in the estimated function are too large, a non-normalized kernel K S can be used to fit the regression, where:
- K S [k d 1 , ⁇ 1 (1: T ), . . . , k d S , ⁇ S (1: T )]
- Embodiments described herein may be entirely hardware, entirely software or including both hardware and software elements.
- the present invention is implemented in software, which includes but is not limited to firmware, resident software, microcode, etc.
- Embodiments may include a computer program product accessible from a computer-usable or computer-readable medium providing program code for use by or in connection with a computer or any instruction execution system.
- a computer-usable or computer readable medium may include any apparatus that stores, communicates, propagates, or transports the program for use by or in connection with the instruction execution system, apparatus, or device.
- the medium can be magnetic, optical, electronic, electromagnetic, infrared, or semiconductor system (or apparatus or device) or a propagation medium.
- the medium may include a computer-readable storage medium such as a semiconductor or solid state memory, magnetic tape, a removable computer diskette, a random access memory (RAM), a read-only memory (ROM), a rigid magnetic disk and an optical disk, etc.
- Each computer program may be tangibly stored in a machine-readable storage media or device (e.g., program memory or magnetic disk) readable by a general or special purpose programmable computer, for configuring and controlling operation of a computer when the storage media or device is read by the computer to perform the procedures described herein.
- the inventive system may also be considered to be embodied in a computer-readable storage medium, configured with a computer program, where the storage medium so configured causes a computer to operate in a specific and predefined manner to perform the functions described herein.
- a data processing system suitable for storing and/or executing program code may include at least one processor coupled directly or indirectly to memory elements through a system bus.
- the memory elements can include local memory employed during actual execution of the program code, bulk storage, and cache memories which provide temporary storage of at least some program code to reduce the number of times code is retrieved from bulk storage during execution.
- I/O devices including but not limited to keyboards, displays, pointing devices, etc. may be coupled to the system either directly or through intervening I/O controllers.
- Network adapters may also be coupled to the system to enable the data processing system to become coupled to other data processing systems or remote printers or storage devices through intervening private or public networks.
- Modems, cable modem and Ethernet cards are just a few of the currently available types of network adapters.
- the processing system 400 includes at least one processor (CPU) 404 operatively coupled to other components via a system bus 402 .
- a cache 406 operatively coupled to the system bus 402 .
- ROM Read Only Memory
- RAM Random Access Memory
- I/O input/output
- sound adapter 430 operatively coupled to the system bus 402 .
- network adapter 440 operatively coupled to the system bus 402 .
- user interface adapter 450 operatively coupled to the system bus 402 .
- display adapter 460 operatively coupled to the system bus 402 .
- a first storage device 422 and a second storage device 424 are operatively coupled to system bus 402 by the I/O adapter 420 .
- the storage devices 422 and 424 can be any of a disk storage device (e.g., a magnetic or optical disk storage device), a solid state magnetic device, and so forth.
- the storage devices 422 and 424 can be the same type of storage device or different types of storage devices.
- a speaker 432 is operatively coupled to system bus 402 by the sound adapter 430 .
- a transceiver 442 is operatively coupled to system bus 402 by network adapter 440 .
- a display device 462 is operatively coupled to system bus 402 by display adapter 460 .
- a first user input device 452 , a second user input device 454 , and a third user input device 456 are operatively coupled to system bus 402 by user interface adapter 450 .
- the user input devices 452 , 454 , and 456 can be any of a keyboard, a mouse, a keypad, an image capture device, a motion sensing device, a microphone, a device incorporating the functionality of at least two of the preceding devices, and so forth. Of course, other types of input devices can also be used, while maintaining the spirit of the present principles.
- the user input devices 452 , 454 , and 456 can be the same type of user input device or different types of user input devices.
- the user input devices 452 , 454 , and 456 are used to input and output information to and from system 400 .
- processing system 400 may also include other elements (not shown), as readily contemplated by one of skill in the art, as well as omit certain elements.
- various other input devices and/or output devices can be included in processing system 400 , depending upon the particular implementation of the same, as readily understood by one of ordinary skill in the art.
- various types of wireless and/or wired input and/or output devices can be used.
- additional processors, controllers, memories, and so forth, in various configurations can also be utilized as readily appreciated by one of ordinary skill in the art.
- a power plant with a monitoring system is shown.
- a power generator 12 may represent any kind of power generation mechanism including, e.g., solar power, wind power, hydroelectric power, geothermal power, fossil fuel power, nuclear power, etc.
- the power generator 12 will have a variety of different physical mechanisms and systems, each of which may operate within designed parameters or may malfunction.
- a plurality of sensors 13 are therefore used to monitor respective physical systems. These sensors 13 may monitor physical properties of the system, including for example temperature, humidity, position, speed, etc., and may also monitor logical properties of the system including, for example, recording logs from software applications and monitoring the status of workloads.
- the sensors 13 provide their sensed information to monitoring system 14 which performs analyses on the sensor information. It is particularly contemplated that these analyses identify correlations between different sensor time series and further identify anomalous behavior.
- the monitoring system 14 reports the results of these analyses to a user 16 and to a control system 18 .
- the control system 18 receives instructions directly from the monitoring system 14 and/or from the user 16 to perform control actions on the power generator 12 .
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Mathematical Physics (AREA)
- Pure & Applied Mathematics (AREA)
- Mathematical Analysis (AREA)
- Mathematical Optimization (AREA)
- Computational Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Software Systems (AREA)
- Algebra (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Databases & Information Systems (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computing Systems (AREA)
- Automation & Control Theory (AREA)
- Operations Research (AREA)
- Probability & Statistics with Applications (AREA)
- Discrete Mathematics (AREA)
- Medical Informatics (AREA)
- Nonlinear Science (AREA)
- Testing And Monitoring For Control Systems (AREA)
Abstract
Methods and systems for detecting anomalies include determining a predictive model for each pair of a set of time series, each time series being associated with a component of a system. New values of each pair of time series are compared to values predicted by the respective predictive model to determine if the respective predictive model is broken. A number of broken predictive models is determined. An anomaly alert is generated if the number of broken predictive models exceeds a threshold.
Description
- This application claims priority to U.S. Application Ser. No. 62/377,728, filed on Aug. 22, 2016, incorporated herein by reference in its entirety. This application is related to an application entitled “POWER GENERATION SYSTEMS WITH MONITORING FOR ANOMALY DETECTION VIA NONLINEAR RELATIONSHIP MODELING,” having attorney docket number 16022A, and which is incorporated by reference herein in its entirety.
- The present invention relates to anomaly detection and, more particularly, to modeling non-linear relationships between time series to test for abnormal behavior.
- Complex physical systems may include thousands of individual components, each having a different set of relationships with the other components. These relationships may be modeled as being, e.g., linear, quadratic, trigonometric, etc. In such a system, the failure of one or more components may lead to a breakdown of the functioning of the entire system.
- To address this problem, sensors may be added to monitor different components of the system with a very fine granularity. The readings from the sensors may be modeled as a time series of measured values. Current analysis techniques use kernel regression to model nonlinear relationships in time series. However, these techniques do not respect the underlying physical laws of the system and cannot express the modeled relationship in the form of formulas for diagnosis.
- A method for detecting anomalies includes determining a predictive model for each pair of a set of time series, each time series being associated with a component of a system. New values of each pair of time series are compared to values predicted by the respective predictive model to determine if the respective predictive model is broken. A number of broken predictive models is determined. An anomaly alert is generated if the number of broken predictive models exceeds a threshold.
- A method for detecting anomalies includes determining a predictive model for each pair of a set of time series, each time series being associated with a component of a system, by determining a periodic relationship between the time series and a non-periodic relationship between the time series. New values of pairs of time series having predictive models that have a fitness score higher than a threshold are compared to values predicted by the respective predictive model to determine if the respective predictive model is broken. A number of broken predictive models is determined. An anomaly alert is generated if the number of broken predictive models exceeds a threshold.
- A system for detecting anomalies includes a fitting module configured to determine a predictive model for each pair of a set of time series, each time series being associated with a component of a system. An anomaly detection module is configured to compare new values of each pair of time series to values predicted by the respective predictive model to determine if the respective predictive model is broken and to determine a number of broken predictive models. An alert module is configured to generate an anomaly alert if the number of broken predictive models exceeds a threshold.
- These and other features and advantages will become apparent from the following detailed description of illustrative embodiments thereof, which is to be read in connection with the accompanying drawings.
- The disclosure will provide details in the following description of preferred embodiments with reference to the following figures wherein:
-
FIG. 1 is a block diagram of a monitoring system for anomaly detection and response in accordance with an embodiment of the present invention; -
FIG. 2 is a block/flow diagram of anomaly detection and response in accordance with an embodiment of the present invention; -
FIG. 3 is a block/flow diagram of training predictive models for pairs of time series in accordance with an embodiment of the present invention; -
FIG. 4 is a block diagram of a processing system in accordance with an embodiment of the present invention; and -
FIG. 5 is a block diagram of a power plant with sensors for physical systems and a monitoring system in accordance with an embodiment of the present invention. - Embodiments of the present invention provide modeling and anomaly detection processes and systems that model the nonlinear, complex relationship between time series generated by sensors. The present embodiments use such models to detect anomalous behavior based on the nature of the physical systems being monitored and on observations from production workload patterns. Not only can the nonlinear relationships between sensors be so identified, but those relationships can be expressed as close-form formulas, greatly assisting operators in the system operators who wish to incorporate their domain knowledge for diagnosis.
- Toward that end, a regression function is used to model a non-linear relationship pairs of time series. One time series is fit using the function of another time series. This function is designed respecting the physical laws and workload patterns inside the physical system, providing a template of bases that include, for example, trigonometric and polynomial relationships. For each pair of time series, the regression is fitted by selecting correct bases out of the template. The relationship between each pair is then expressed using the estimated regression function. The fitting problem may be formulated as a Sparse Group Lasso with Overlapping problem. The formulation is further improved by decomposing the coefficient matrix into two matrices that are solved separately, greatly improving the solution speed. After modeling, a function is formulated for every pair of the time series. A set of relationships is selected out of the pairs by choosing those with a high fitness score.
- During anomaly detection, the trained models are applied to newly acquired time series. For each pair of time series it is determined whether the pair matches the training model. If so, for each time point in the two time series, a fitted value is generated and compared with the measured value. The trained model is considered broken if the difference between the fitted value and the measured value is high (e.g., if the difference exceeds a threshold). The number of broken models is counted and time points that have a high number of broken models are considered anomalies.
- The present embodiments thus accurately identify anomalies, providing early warning to operators to avoid system failure. Explicit sensor relationships are provided to the operators to help them diagnose the potential problems. Large real-world datasets can be processed quickly in this manner.
- Referring now to
FIG. 1 , amonitoring system 100 is shown. The monitoring system includes ahardware processor 102 andmemory 104. It should be understood that themonitoring system 100 may be distinct from the monitored system, or may be integrated therewith. Themonitoring system 100 therefore may include a number ofphysical components 106, or thesephysical components 106 may be external to themonitoring system 100. Thesephysical components 106 may include any include any type of device in a complex system such as, e.g., a power generator. In one specific embodiment thephysical components 106 may include, for example, servers in a rack enclosure with their attendant power, cooling, and communications components. Thephysical components 106 are each fitted with one ormore sensors 108 that measure operational parameters of thephysical components 106. These operational parameters may include, for example, temperature, power consumption, data throughput, vibrational amplitude, luminosity, physical location, humidity, contaminant concentration, and any other measurable quantity. Thesensors 108 store their measurements inmemory 104 as time series, with each measurement being stored with a respective timestamp. Thesensors 108 keep a common clock, such that measurements stored at a particular time bymultiple sensors 108 will share the same, or a very similar, timestamp. - The
monitoring system 100 also includes functional modules that may, in some embodiments, be implemented as software that is stored inmemory 104 and executed byhardware processor 102. In alternative embodiments, the functional modules may be implemented as one or more discrete hardware components in the form of, e.g., application specific integrated chips or field programmable gate arrays. - A
fitting module 110 analyzes a set of time series stored in thememory 104 and generates models that represent the relationship between each respective pair of time series. The models are thus based on historical measurements from thesensors 108 and characterize normal operation of thephysical components 106.Anomaly detection module 112 then compares new measurements of from thesensors 108 to the values predicted by the models and, if a sufficient number of discrepancies (referred to herein as “broken models”) are found,anomaly detection module 112 flags a particular timestamp as being anomalous. - When an anomaly is found,
alert module 114 takes one of a number of actions. In one embodiment, the alert module simply generates an alert to a human operator, including the anomalous timestamp and information regarding the relationships between sensors that are characterized by the broken models. Thus, for example, if a pair ofsensors 108 are expected to have a particular trigonometric relationship (e.g., where a first sensor's value is a function of the sine of a second sensor's value) and thephysical components 106 in question have deviated from the expected behavior, the report may indicate whichphysical components 106 are at issue, what the expected behavior was, and what the measured behavior is. The report may also indicate a measure of the magnitude of the anomaly (e.g., the number of broken models). - In another embodiment, the
alert module 114 may be configured to take automatic actions to address anomalies. Thus, for example, if the anomaly implicatesphysical components 106 in a particularly sensitive or dangerous sub-system, thealert module 114 may automatically issue instructions to thosecomponents 106 to slow or stop their processes. - Referring now to
FIG. 2 , a method for detecting anomalies is shown. Block 202 trains a model for each pair of time series stored in thememory 104. As noted above, each time series corresponds to arespective sensor 108 that measures aphysical component 106. The model training will be described in greater detail below.Block 204 then trims out the models that have a fitness score below a threshold value. The fitness score may be based on, e.g., the r-squared error. -
Block 206 tests pairs of current measurements from thesensors 108 against the respective models.Block 210 then counts the number of broken models at each timestamp (i.e., the number of pairs for which the measured values differ from values predicted by the model by more than a threshold amount). -
Block 212 determines whether a number of broken models exceeds an anomaly threshold. If not, processing returns to block 206 and the current measurements from thesensors 108 at a new timestamp are tested. If the number of broken models does exceed the anomaly threshold, then block 214 generates an alert by, e.g., sending a report to a human operator or taking an automated administrative action. - Referring now to
FIG. 3 , additional detail is shown onmodel training 202. Given two time series, x and y, trigonometric andperiodic recovery 302 determines whether there is a functional relationship between the time series that can be expressed as y(x)=g(x)+m.Polynomial recovery 304 determines whether there is a functional relationship between the time series that can be expressed as y=g(x)+h(x).Block 306 performs coefficient correction for the discovered functions. - Periodic signals can be cleanly separated out from polynomial signals using a fast Fourier transform (FFT), which turns a time-domain series into a frequency-domain series, showing high amplitudes for periodic components. Non-periodic components are separated out from both the periodic components and from the noise. The relationships between the time series can therefore be expressed as y(x)=fperiodic(x)+fnonperiodic(x)+ε, where fperiodic(•) represents the periodic component with a mean of zero and fnonperiodic(x) represents the polynomial component. Periodic patterns may result from repeated workload patterns. Periodic components are shown as peaks in higher-frequency portions of the FFT output, while nonperiodic components are shown as a peak in the low-frequency portion of the output. Thus, after
block 302 finds the periodic components, block 304 can operate on only the nonperiodic portion. -
Block 302 further determines the time shift of the periodic component, with values for x at different points in the time series being used in the FFT. The results using incorrect time shifts will show only noise, while strong signals will be evident at particular frequencies when the time shift is correct. - The fitting error (also known as the residue) is used as an input for the polynomial relationship discovery in
block 304. For the two time series x and y, x is considered the independent signal and y is considered the response signal. The time series x is extended to d different power bases as k(x;i)=[xi,xi 2,xi 3, . . . ,xi d]. The interactions between the two signals are extended to an autoregressive kernel as k(x,y;i )=[yi, . . . , yi d, xi,yixy,yixi 2 , . . . yi dxi d−1 ,yi dxi d]. Each power base includes the current sensor value and τ previous sensor values. A (τ+1)×d(d+2) kernel matrix is defined as: -
- Each matrix k(x,y;t) in K(x,y;t) is set to zero because the values for yt in k(x,y;t) cannot be modeled on yt itself. Given N total measurements for each signal, the index t of K(x,y;t) starts from τ+1 to have enough delays in the first training example. If K(x,y;t) is a feature matrix, then a model with τ delays can be expressed as:
-
y t =tr(K (x,y;t) T W) - where W is a model parameter matrix.
- The underlying physical laws that determine the relationship models sparsely rely on only a few bases. Furthermore, physical objects and events need time to propagate through a system. As a result, sensors at different locations may monitor the same event at different times, making their relationship sparsely rely on a time lag that is determined by their physical distance. This shows that the sensor relationships should sparsely appear in certain bases and delays. To select among bases and discover the certain delay time points in the kernels, block 304 treats each base with the same orders as a group and all the bases at each delay as a group. In other words, each row of K(x,y;t) is a structured group, where all the bases are at a same delay, and each column is a structured group, with the same base at all delays. Regularizers are applied to the model parameter matrix W to select relevant groups of bases and delays and to add sparsity within the groups to control the model capacity.
-
Block 204 therefore forms an optimization problem as a sparse group lasso with overlapping structures: -
- where (W)=Σt=1 N∥yt−tr(K(x,y;t) TW∥2 2, tr(•) represents the trace of a matrix, g is a particular group from the set of groups G, and α, λ1, and λ2 are tuning parameters. Ωg(•) refers to a structured-sparsity-inducing penalty and plays a role in encouraging closely related inputs to be selected jointly as relevant to the output by setting the corresponding regression coefficients to non-zero values based on the 1/ 2 mixed-norm.
- Although W has overlapping structures, the present embodiments provide a special case because the parameters naturally form a matrix, where row groups only overlap with column groups, but never with themselves. Similarly, column groups only overlap with row groups, but never with themselves. As a result, W can be decomposed into a summation of two components, P and Q, with only row groups being selected in P and column groups being selected in Q. Both rows and columns are selected in W after summation. This decomposition degenerates the regularization part of the original selection to two easier, sparse group Lasso penalties on both P and Q.
- To model the regularization part, the 1,2+1 matrix norm can be used. More particularly, the 1,2+1 norm is applied row-wise to P and the optimal solution of P will include mostly rows with all-zero entries and a few rows with non-zero entries. Similarly, the 1,2+1 norm of QT encourages selection among columns of Q.
- The regularization is therefore rewritten in terms of P and Q as:
-
- where W in the least loss is replaced by P+Q. The parameter α plays a role of balancing the weight between the 1,2 norm and the 1 norm. Although the regularization is decomposed into two sparse group Lasso penalties, the least squares loss part is not. In addition, P and Q are estimated together, because they both contribute to the optimal solution of W. Therefore existing sparse group Lasso solvers are inadequate to address this problem.
- The optimization problem can be iteratively solved using fast iterative shrinkage-thresholding algorithm. To fit the optimization algorithm, the optimization problem may be reformulated as:
-
- where •,• denotes an inner product, R(P, Q) is the non-smooth part that constitutes the regularizers, is an iteration number, is the continuously differentiable part, ∇P and ∇Q are the partial derivatives of at ({tilde over (P)}, {tilde over (Q)}), {tilde over (P)} and {tilde over (Q)} are defined as:
-
-
- and where L is an upper bound of the Lipschitz modulus.
- The iteration number is initialized to one, t1 is initialized to one, and P0={tilde over (P)}1=Q0={tilde over (Q)}1=0. The optimization problem is solved to obtain P and Q. The value of t +1, {tilde over (P)} +1, {tilde over (Q)} +1 are calculated as described above. The iteration number is incremented and these steps are repeated. Iteration continues until the values for P and Q converge. The optimization converges to an optimal solution with a convergence rate of
-
- As noted above, the sum of the final values of P and Q is the parameter matrix W.
- A template of bases is generated using different orders of polynomial relationships with different time delays. Bases are generated with at most degree-d and τ delays to model y as a function of x. Thus the function is expressed as:
-
- The x terms are bases, while the w terms are learning coefficients of the parameter matrix W. The problem is formulated as Sparse Group Lasso with Overlapping and can be solved efficiently. The output of
block 304 is an estimated function for each pair of the time series. -
Block 306 corrects the coefficients of the functions generated byblock 304. After solving the Group Lasso problem, the feature coefficients are set to zero and non-zero values (such that particular coefficients are selected or deselected). The non-zero valued features may not represent the true coefficients.Block 306 corrects these coefficients using linear regression, regressing each estimated function to the time series to minimize the regression error. For example, if the coefficients in the estimated function are too large, a non-normalized kernel KS can be used to fit the regression, where: -
K S =[k d1 ,τ1 (1:T), . . . , k dS ,τS (1:T)] -
- Embodiments described herein may be entirely hardware, entirely software or including both hardware and software elements. In a preferred embodiment, the present invention is implemented in software, which includes but is not limited to firmware, resident software, microcode, etc.
- Embodiments may include a computer program product accessible from a computer-usable or computer-readable medium providing program code for use by or in connection with a computer or any instruction execution system. A computer-usable or computer readable medium may include any apparatus that stores, communicates, propagates, or transports the program for use by or in connection with the instruction execution system, apparatus, or device. The medium can be magnetic, optical, electronic, electromagnetic, infrared, or semiconductor system (or apparatus or device) or a propagation medium. The medium may include a computer-readable storage medium such as a semiconductor or solid state memory, magnetic tape, a removable computer diskette, a random access memory (RAM), a read-only memory (ROM), a rigid magnetic disk and an optical disk, etc.
- Each computer program may be tangibly stored in a machine-readable storage media or device (e.g., program memory or magnetic disk) readable by a general or special purpose programmable computer, for configuring and controlling operation of a computer when the storage media or device is read by the computer to perform the procedures described herein. The inventive system may also be considered to be embodied in a computer-readable storage medium, configured with a computer program, where the storage medium so configured causes a computer to operate in a specific and predefined manner to perform the functions described herein.
- A data processing system suitable for storing and/or executing program code may include at least one processor coupled directly or indirectly to memory elements through a system bus. The memory elements can include local memory employed during actual execution of the program code, bulk storage, and cache memories which provide temporary storage of at least some program code to reduce the number of times code is retrieved from bulk storage during execution. Input/output or I/O devices (including but not limited to keyboards, displays, pointing devices, etc.) may be coupled to the system either directly or through intervening I/O controllers.
- Network adapters may also be coupled to the system to enable the data processing system to become coupled to other data processing systems or remote printers or storage devices through intervening private or public networks. Modems, cable modem and Ethernet cards are just a few of the currently available types of network adapters.
- Referring now to
FIG. 4 , anexemplary processing system 400 is shown which may represent themonitoring system 100. Theprocessing system 400 includes at least one processor (CPU) 404 operatively coupled to other components via a system bus 402. Acache 406, a Read Only Memory (ROM) 408, a Random Access Memory (RAM) 410, an input/output (I/O)adapter 420, asound adapter 430, anetwork adapter 440, auser interface adapter 450, and adisplay adapter 460, are operatively coupled to the system bus 402. - A
first storage device 422 and asecond storage device 424 are operatively coupled to system bus 402 by the I/O adapter 420. Thestorage devices storage devices - A
speaker 432 is operatively coupled to system bus 402 by thesound adapter 430. Atransceiver 442 is operatively coupled to system bus 402 bynetwork adapter 440. Adisplay device 462 is operatively coupled to system bus 402 bydisplay adapter 460. - A first
user input device 452, a seconduser input device 454, and a thirduser input device 456 are operatively coupled to system bus 402 byuser interface adapter 450. Theuser input devices user input devices user input devices system 400. - Of course, the
processing system 400 may also include other elements (not shown), as readily contemplated by one of skill in the art, as well as omit certain elements. For example, various other input devices and/or output devices can be included inprocessing system 400, depending upon the particular implementation of the same, as readily understood by one of ordinary skill in the art. For example, various types of wireless and/or wired input and/or output devices can be used. Moreover, additional processors, controllers, memories, and so forth, in various configurations can also be utilized as readily appreciated by one of ordinary skill in the art. These and other variations of theprocessing system 400 are readily contemplated by one of ordinary skill in the art given the teachings of the present principles provided herein. - Referring now to
FIG. 5 , a power plant with a monitoring system is shown. In particular, apower generator 12 is shown that may represent any kind of power generation mechanism including, e.g., solar power, wind power, hydroelectric power, geothermal power, fossil fuel power, nuclear power, etc. Thepower generator 12 will have a variety of different physical mechanisms and systems, each of which may operate within designed parameters or may malfunction. A plurality ofsensors 13 are therefore used to monitor respective physical systems. Thesesensors 13 may monitor physical properties of the system, including for example temperature, humidity, position, speed, etc., and may also monitor logical properties of the system including, for example, recording logs from software applications and monitoring the status of workloads. - The
sensors 13 provide their sensed information tomonitoring system 14 which performs analyses on the sensor information. It is particularly contemplated that these analyses identify correlations between different sensor time series and further identify anomalous behavior. Themonitoring system 14 reports the results of these analyses to auser 16 and to acontrol system 18. Thecontrol system 18 receives instructions directly from themonitoring system 14 and/or from theuser 16 to perform control actions on thepower generator 12. - The foregoing is to be understood as being in every respect illustrative and exemplary, but not restrictive, and the scope of the invention disclosed herein is not to be determined from the Detailed Description, but rather from the claims as interpreted according to the full breadth permitted by the patent laws. It is to be understood that the embodiments shown and described herein are only illustrative of the principles of the present invention and that those skilled in the art may implement various modifications without departing from the scope and spirit of the invention. Those skilled in the art could implement various other feature combinations without departing from the scope and spirit of the invention. Having thus described aspects of the invention, with the details and particularity required by the patent laws, what is claimed and desired protected by Letters Patent is set forth in the appended claims.
Claims (20)
1. A method for detecting anomalies, comprising:
determining a predictive model for each pair of a set of time series, each time series being associated with a component of a system;
comparing new values of each pair of time series to values predicted by the respective predictive model to determine if the respective predictive model is broken;
determining a number of broken predictive models; and
generating an anomaly alert if the number of broken predictive models exceeds a threshold.
2. The method of claim 1 , wherein determining the predictive model for a pair of time series comprises determining a periodic relationship between the time series and a non-periodic relationship between the time series.
3. The method of claim 2 , wherein determining the predictive model for a pair of time series further comprises performing a Fourier transform on the time series to separate periodic components from non-periodic components.
4. The method of claim 2 , wherein determining the predictive model for a pair of time series further comprises determining a time delay between time series that have a periodic relationship.
5. The method of claim 2 , further wherein determining a non-periodic relationship between the time series comprises determining polynomial bases for modeling a polynomial relationship between the time series.
6. The method of claim 5 , wherein determining a non-periodic relationship further comprises solving a corresponding Sparse Group Lasso problem and correcting coefficients of a solution of the corresponding Sparse Group Lasso problem by linear regression.
7. The method of claim 1 , wherein comparing values of each pair of time series only uses predictive models that have a fitness score higher than a threshold.
8. The method of claim 7 , further comprising calculating a fitness score for each predictive model based on an r-squared error.
9. A method for detecting anomalies, comprising:
determining a predictive model for each pair of a set of time series, each time series being associated with a component of a system, by determining a periodic relationship between the time series and a non-periodic relationship between the time series;
comparing new values of pairs of time series having predictive models that have a fitness score higher than a threshold to values predicted by the respective predictive model to determine if the respective predictive model is broken;
determining a number of broken predictive models; and
generating an anomaly alert if the number of broken predictive models exceeds a threshold.
10. The method of claim 9 , wherein determining the predictive model for a pair of time series further comprises performing a Fourier transform on the time series to separate periodic components from non-periodic components.
11. The method of claim 9 , wherein determining the predictive model for a pair of time series further comprises determining a time delay between time series that have a periodic relationship.
12. The method of claim 9 , wherein determining a non-periodic relationship between the time series comprises determining polynomial bases for modeling a polynomial relationship between the time series.
13. A system for detecting anomalies, comprising:
a fitting module configured to determine a predictive model for each pair of a set of time series, each time series being associated with a component of a system;
an anomaly detection module configured to compare new values of each pair of time series to values predicted by the respective predictive model to determine if the respective predictive model is broken and to determine a number of broken predictive models; and
an alert module configured to generate an anomaly alert if the number of broken predictive models exceeds a threshold.
14. The system of claim 13 , wherein the fitting module is further configured to determine a periodic relationship between the time series and a non-periodic relationship between the time series.
15. The system of claim 13 , wherein the fitting module is further configured to perform a Fourier transform on the time series to separate periodic components from non-periodic components.
16. The system of claim 13 , wherein the fitting module is further configured to determine a time delay between time series that have a periodic relationship.
17. The system of claim 13 , wherein the fitting module is further configured to polynomial bases for modeling a polynomial relationship between the time series.
18. The system of claim 17 , wherein the fitting module is further configured to solve a corresponding Sparse Group Lasso problem and correcting coefficients of a solution of the corresponding Sparse Group Lasso problem by linear regression.
19. The system of claim 13 , wherein the anomaly detection module is further configured to use only predictive models that have a fitness score higher than a threshold.
20. The system of claim 19 , wherein the fitting module is further configured to calculate a fitness score for each predictive model based on an r-squared error.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
US15/680,831 US20180053111A1 (en) | 2016-08-22 | 2017-08-18 | Anomaly Detection for Complex Physical System via Nonlinear Relationship Modeling |
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
US201662377728P | 2016-08-22 | 2016-08-22 | |
US15/680,831 US20180053111A1 (en) | 2016-08-22 | 2017-08-18 | Anomaly Detection for Complex Physical System via Nonlinear Relationship Modeling |
Publications (1)
Publication Number | Publication Date |
---|---|
US20180053111A1 true US20180053111A1 (en) | 2018-02-22 |
Family
ID=61190762
Family Applications (2)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
US15/680,831 Abandoned US20180053111A1 (en) | 2016-08-22 | 2017-08-18 | Anomaly Detection for Complex Physical System via Nonlinear Relationship Modeling |
US15/680,796 Active 2037-11-22 US10340734B2 (en) | 2016-08-22 | 2017-08-18 | Power generation systems with monitoring for anomaly detection via nonlinear relationship modeling |
Family Applications After (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
US15/680,796 Active 2037-11-22 US10340734B2 (en) | 2016-08-22 | 2017-08-18 | Power generation systems with monitoring for anomaly detection via nonlinear relationship modeling |
Country Status (1)
Country | Link |
---|---|
US (2) | US20180053111A1 (en) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110262939A (en) * | 2019-05-14 | 2019-09-20 | 苏宁金融服务(上海)有限公司 | Algorithm model operation and monitoring method, device, computer equipment and storage medium |
CN112928820A (en) * | 2021-01-28 | 2021-06-08 | 临沂大学 | Automatic detection system for power distribution cabinet and detection method thereof |
Families Citing this family (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US11521097B2 (en) * | 2019-02-19 | 2022-12-06 | International Business Machines Corporation | Sparse modeling for optimizing sensor placement |
US11354600B2 (en) * | 2019-03-05 | 2022-06-07 | Booz Allen Hamilton Inc. | System and method for heterogeneous relational kernel learning |
Family Cites Families (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2011142383A1 (en) * | 2010-05-14 | 2011-11-17 | 株式会社東芝 | Output monitoring device for nuclear reactor |
WO2015104691A2 (en) * | 2014-01-13 | 2015-07-16 | Brightsource Industries (Israel) Ltd. | Systems, methods, and devices for detecting anomalies in an industrial control system |
WO2015179778A1 (en) * | 2014-05-23 | 2015-11-26 | Datarobot | Systems and techniques for predictive data analytics |
JP5943357B2 (en) * | 2014-09-17 | 2016-07-05 | インターナショナル・ビジネス・マシーンズ・コーポレーションInternational Business Machines Corporation | Detection apparatus, detection method, and program |
US10539602B2 (en) * | 2016-04-05 | 2020-01-21 | Siemens Energy, Inc. | Active shaft grounding system with diagnostic waveform analysis |
US10534328B2 (en) * | 2016-06-21 | 2020-01-14 | General Electric Company | Methods and systems for enhancing control of power plant generating units |
-
2017
- 2017-08-18 US US15/680,831 patent/US20180053111A1/en not_active Abandoned
- 2017-08-18 US US15/680,796 patent/US10340734B2/en active Active
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110262939A (en) * | 2019-05-14 | 2019-09-20 | 苏宁金融服务(上海)有限公司 | Algorithm model operation and monitoring method, device, computer equipment and storage medium |
CN112928820A (en) * | 2021-01-28 | 2021-06-08 | 临沂大学 | Automatic detection system for power distribution cabinet and detection method thereof |
Also Published As
Publication number | Publication date |
---|---|
US20180054085A1 (en) | 2018-02-22 |
US10340734B2 (en) | 2019-07-02 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
US10340734B2 (en) | Power generation systems with monitoring for anomaly detection via nonlinear relationship modeling | |
Ni et al. | Sensor network data fault types | |
US11797882B2 (en) | Prognostic-surveillance technique that dynamically adapts to evolving characteristics of a monitored asset | |
CN110647414A (en) | Improved analog function security with anomaly detection | |
Kalgonda et al. | Multivariate quality control chart for autocorrelated processes | |
US9122273B2 (en) | Failure cause diagnosis system and method | |
US20180307994A1 (en) | Identifying multiple causal anomalies in power plant systems by modeling local propagations | |
Chen et al. | Cointegration testing method for monitoring nonstationary processes | |
Camacho et al. | The best approaches in the on-line monitoring of batch processes based on PCA: Does the modelling structure matter? | |
US11657121B2 (en) | Abnormality detection device, abnormality detection method and computer readable medium | |
US20160369777A1 (en) | System and method for detecting anomaly conditions of sensor attached devices | |
US10228994B2 (en) | Information processing system, information processing method, and program | |
Zhao et al. | Incipient fault detection for complex industrial processes with stationary and nonstationary hybrid characteristics | |
WO2016195092A1 (en) | Anomaly sensing device | |
Cheng et al. | Using cross-validation for model parameter selection of sequential probability ratio test | |
Ge et al. | Incipient fault detection based on fault extraction and residual evaluation | |
US11144023B1 (en) | Method for PMU data recovery using an improved cubic spline interpolation and singular value decomposition | |
CN111400850B (en) | Equipment fault analysis method, device, equipment and storage medium | |
Jiang et al. | Independent component analysis-based non-Gaussian process monitoring with preselecting optimal components and support vector data description | |
Onel et al. | Integrated data-driven process monitoring and explicit fault-tolerant multiparametric control | |
Rani et al. | Fault detection and isolation using probabilistic wavelet neural operator auto-encoder with application to dynamic processes | |
JP2016045852A (en) | Abnormality diagnostic device and abnormality diagnostic method | |
CN117693747A (en) | Passive inference of signal follow-up in multivariate anomaly detection | |
US11320813B2 (en) | Industrial asset temporal anomaly detection with fault variable ranking | |
Du et al. | Generalized polynomial chaos-based fault detection and classification for nonlinear dynamic processes |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
AS | Assignment |
Owner name: NEC LABORATORIES AMERICA, INC., NEW JERSEY Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNORS:YAN, TAN;SONG, DONGJIN;CHEN, HAIFENG;AND OTHERS;SIGNING DATES FROM 20170812 TO 20170818;REEL/FRAME:043336/0103 |
|
STPP | Information on status: patent application and granting procedure in general |
Free format text: DOCKETED NEW CASE - READY FOR EXAMINATION |
|
STPP | Information on status: patent application and granting procedure in general |
Free format text: NON FINAL ACTION MAILED |
|
STPP | Information on status: patent application and granting procedure in general |
Free format text: FINAL REJECTION MAILED |
|
STCB | Information on status: application discontinuation |
Free format text: ABANDONED -- FAILURE TO RESPOND TO AN OFFICE ACTION |