Multivariate and multi-timescale anomaly detection
Techniques for anomaly detection are disclosed. An anomaly detection system transforms observation signals received from signal generators into multidimensional representations. The system extracts portions of the multidimensional representation indicative of anomalies. Based on the extracted portions of the multidimensional representation, the system determines metrics representing the observation signals and determines whether the metrics satisfy anomaly detection criteria.
1 . An anomaly detection system comprising one or more modules configured to:
receive an observation signal from a first signal generator, the observation signal comprising a first unidimensional series of values;
receive a reference signal from a second signal generator, the reference signal comprising a second unidimensional series of values corresponding to the first unidimensional series of values;
combine the observation signal with the reference signal by transforming the first unidimensional series of values of the observation signal and the corresponding second unidimensional series of values of the reference signal into a multidimensional representation, the multidimensional representation comprising a multidimensional image;
identify a portion of the multidimensional image comprising a region indicative of one or more anomalies and excluding one or more regions that do not contain information indicating the presence of the one or more anomalies;
extract the portion of the multidimensional image;
convert the portion of the multidimensional image into a metric;
determine that the metric satisfies an anomaly detection criteria; and
generate an output indicating an anomaly in the observation signal from the signal generator.
2 . The anomaly detection system of claim 1 , wherein identifying the portion of the multidimensional image comprises: applying a convolutional neural network trained to recognize the region of the multidimensional image indicative of anomalies.
3 . The anomaly detection system of claim 1 , wherein determining the metric comprises:
applying a generative classifier trained to determine the metric by capturing the characteristics of normalcy in the portion of the multidimensional image.
4 . The anomaly detection system of claim 1 , wherein the multidimensional image is a two-dimensional image comprising a first axis corresponding to the observation signal and a second axis corresponding to the observation signal.
5 . The anomaly detection system of claim 1 , wherein the multidimensional image is a three-dimensional image.
6 . The anomaly detection system of claim 1 , wherein:
the first signal generator is a first sensor; and
the second signal generator is a second sensor.
7 . The anomaly detection system of claim 1 , wherein:
the observation signal comprising the first unidimensional series of values has a second time scale; and
the reference signal comprising the second unidimensional series of values corresponding to the first unidimensional series of values and has the first timescale.
8 . The anomaly detection system of claim 7 , wherein the one or more modules of the anomaly detection system are further configured to:
receive a second observation signal from a third signal generator, the second observation signal comprising a third unidimensional series of values;
receive a second reference signal from a fourth signal generator, the second reference signal comprising a fourth unidimensional series of values corresponding to the second unidimensional series of values;
combine the second observation signal with the second reference signal by transforming the third unidimensional series of values of the second observation signal and the corresponding fourth unidimensional series of values of the second reference signal into a second multidimensional representation, the second multidimensional representation comprising a second multidimensional image;
identify a portion of the second multidimensional image comprising a region indicative of one or more anomalies and excluding one or more regions that do not contain information indicating the presence of the one or more anomalies;
extract the portion of the second multidimensional image;
convert the portion of the second multidimensional image into a second metric;
determine that the second metric satisfies the anomaly detection criteria; and
generate a second output indicating a second anomaly in the second observation signal from the third signal generator.
9 . The anomaly detection system of claim 8 , wherein:
the second observation signal comprising the third unidimensional series of values has a second time scale; and
the second reference signal comprising the fourth unidimensional series of values corresponding to the first unidimensional series of values and has the second timescale.
10 . The anomaly detection system of claim 9 , wherein the one or more modules of the anomaly detection system are further configured to:
combine a first output indicating a first anomaly with a second output indicating the second anomaly; and
generate a system anomaly output based on the combination of the first output and the second output.
11 . A method comprising:
receiving an observation signal from a first signal generator, the observation signal comprising a first unidimensional series of values;
receiving a reference signal from a second signal generator, the reference signal comprising a second unidimensional series of values corresponding to the first unidimensional series of values;
combine the observation signal with the reference signal by transforming the first unidimensional series of values of the observation signal and the corresponding second unidimensional series of values of the reference signal into a multidimensional representation, the multidimensional representation comprising a multidimensional image;
identify a portion of the multidimensional image comprising a region indicative of one or more anomalies and excluding one or more regions that do not contain information indicating the presence of the one or more anomalies;
extract the portion of the multidimensional image;
converting the portion of the multidimensional image into a metric;
determining that the metric satisfies an anomaly detection criteria; and
generating an output indicating an anomaly in the observation signal from the signal generator.
12 . The method of claim 11 , wherein identifying the portion of the multidimensional image comprises:
applying a convolutional neural network trained to recognize the region of the multidimensional image indicative of anomalies.
13 . The method of claim 11 , wherein determining the metric comprises:
applying a generative classifier trained to determine the metric by capturing the characteristics of normalcy in the portion of the multidimensional image.
14 . A system comprising:
a plurality of anomaly detection modules configured to detect anomalies in a plurality of observation signals from a plurality of signal generators, wherein:
individual anomaly detection modules of the plurality of anomaly detection modules are configured to:
receive an observation signal of the plurality of observation signals from a respective first signal generator of the plurality of signal generators, wherein the observation signal comprises a first unidimensional series of values;
receive a reference signal from a respective second signal generator of the plurality of signal generators, the reference signal comprising a second unidimensional series of values corresponding to the first unidimensional series of values;
combine the observation signal with the reference signal by transforming the first unidimensional series of values of the observation signal and the corresponding a second unidimensional series of values of the reference signal into a multidimensional representation, the multidimensional representation comprising a multidimensional image;
identify a portion of the multidimensional image comprising a region indicative of one or more anomalies and excluding one or more regions that do not contain information indicating the presence of the one or more anomalies;
extract the portion of the multidimensional image;
convert the portion of the multidimensional image into a metric;
determine that the metric satisfies an anomaly detection criteria; and
determine an output indicating an anomaly in the observation signal from the respective signal generator; and
wherein the system is configured to generate an anomaly report combining the outputs of the plurality of anomaly detection modules.
15 . The system of claim 14 , wherein the individual anomaly detection modules identify the portion of the respective multidimensional image by:
applying a convolutional neural network trained to identify anomalous variations in normal sample values corresponding to the respective signal generator.
16 . The system of claim 14 , wherein the individual anomaly detection modules determine the respective metric by:
applying a generative classifier trained to determine the metric by capturing the characteristics of normalcy in the portion of the respective multidimensional image.
17 . A method comprising:
receiving a plurality of observation signals and a plurality of reference signals by a respective plurality of anomaly detection modules, wherein:
the plurality of observation signals and the plurality of reference signals comprise unidimensional series of values, and
individual observation signals of the plurality of observation signals correspond to individual reference signals of the plurality of reference signals;
combine the corresponding observation signals and reference signals by transforming, by the plurality of anomaly detection modules, the respective unidimensional series of values of the observation signals and the corresponding reference signals into multidimensional representations, the multidimensional representations comprising multidimensional images;
identifying, by the plurality of anomaly detection modules, portions of the respective multidimensional images comprising regions of the multidimensional representations indicative of one or more anomalies and excluding one or more regions of the multidimensional representations that do not contain information indicating the presence of the one or more anomalies;
extract the portions of the respective multidimensional images;
converting, by the plurality of anomaly detection modules, the respective portions of the multidimensional images into metrics;
determining, by the plurality of anomaly detection modules, whether the respective observation signals include anomalies based on to the respective metrics satisfying anomaly detection criteria;
determining, by the plurality of anomaly detection modules, outputs indicating anomalies determined in the observation signals corresponding to the respective metrics; and
generating an anomaly report combining the outputs of the plurality of detection modules.
18 . The method of claim 17 , wherein identifying the portions of the multidimensional images comprise: applying convolutional neural networks trained to identify anomalous variations in normal samples values corresponding to the respective signal generators.
19 . The method of claim 17 , wherein determining the metric comprises:
applying a generative classifier trained to determine the metric by capturing the characteristics of normalcy in the portion of the multidimensional images.