Anomaly detection apparatus, method and computer-readable medium

Patent metadata

Field	Value
Publication number	US-12044714-B2
Application number	US-201817270899-A
Country	US
Kind code	B2
Filing date	Aug 30, 2018
Priority date	Aug 30, 2018
Publication date	Jul 23, 2024
Grant date	Jul 23, 2024

Abstract

Provided a method comprising: obtaining waveform data sets of a periodic electric waveform signal, with a length set to one cycle time; calculating a frequency spectrum for each waveform data set; extracting and separating odd and even frequency harmonics to create odd and even frequency harmonic matrices on which a canonical correlation analysis (CCA) being applied to obtain CCA features; performing linear transformation on the CCA features to obtain linear transformed features; generating a model based on the linear transformed features; performing magnitude quantization of frequency spectrums of waveform data sets to identify normal and anomalous waveform signals.

First claim

What is claimed is:

1. An anomaly detection apparatus comprising:
a processor; and
a memory in circuit communication with the processor, wherein the processor is configured to execute program instructions stored in the memory to:
obtain a plurality of sets of waveform data of a periodic electric waveform signal of an electric appliance, with a length of each of the plurality of sets of waveform data being set to one cycle time of the periodic electric waveform signal;
transform, using FFT (fast Fourier transform) processing, each of the plurality of sets of waveform data from a time domain into a frequency domain representation including a plurality of frequency bins of a fundamental frequency which is a reciprocal of the one cycle time and harmonics of the fundamental frequency;
apply a frequency processing including normalization and filtering to the frequency domain representation of each of the plurality of sets of waveform data to obtain a frequency spectrum for each of the plurality sets of waveform data;
extract and separate odd and even frequency harmonics of the fundamental frequency from the frequency spectrum of each of the plurality of sets of waveform data to create odd and even frequency harmonic matrices;
perform canonical correlation analysis (CCA) on the odd and even frequency harmonic matrices to obtain CCA features;
perform linear transformation on the CCA features to obtain linear transformed features;
generate a cluster model based on the linear transformed features;
perform magnitude quantization of the frequency spectrum of each of the plurality of sets of waveform data to identify normal and anomalous waveform signals; and
output at least the identified result,
wherein the magnitude quantization includes:
applying filtering to the frequency spectrum of each of the plurality of sets of waveform data;
transforming, using inverse FFT processing, the frequency spectrum of each of the plurality of sets of waveform data into a time domain representation;
creating a two bins histogram based on an effective magnitude of each of the plurality of sets of waveform data;
creating a list or a vector, composed of a larger histogram count value out of the two-bins; and
identifying the normal and anomalous waveform signals.

2. The anomaly detection apparatus according to claim 1 , wherein the periodic electric waveform signal is an AC (Alternating Current) current signal, and the plurality of sets of waveform data of the AC current signal are each phase aligned with an AC voltage signal.

3. The anomaly detection apparatus according to claim 1 , wherein the processor is configured to execute the program instructions stored in the memory to model the CCA features by using:
only odd frequency harmonics,
only even frequency harmonics, or
any combination of the odd frequency harmonics and the even frequency harmonics.

4. The anomaly detection apparatus according to claim 1 , wherein the processor is configured to execute the program instructions stored in the memory to calculate an amplitude spectrum for each of the plurality of frequency bins of each of the plurality of sets of waveform data, as the frequency spectrum thereof.

5. The anomaly detection apparatus according to claim 1 , wherein the processor is configured to execute the program instructions stored in the memory to apply high pass filtering to the frequency spectrum of each of the plurality of sets of waveform data as has been normalized to cut off a frequency component not more than the fundamental frequency.

6. The anomaly detection apparatus according to claim 1 , wherein the processor is configured to execute the program instructions stored in the memory to perform the linear transformation that generates a linear combination of the odd and even frequency harmonic matrices.

7. The anomaly detection apparatus according to claim 1 , wherein the processor is configured to execute the program instructions stored in the memory to perform as the magnitude quantization, iteratively, for a number of frequency components in the frequency spectrum of each of the plurality sets of waveform data, RMS (root mean square) quantization to create the two-bins histogram based on an RMS value.

8. The anomaly detection apparatus according to claim 1 , wherein the processor is configured to execute the program instructions stored in the memory to:
obtain a frequency spectrum matrix with a number of rows equal to a number of the waveform data sets and a number of columns equal to a number of frequency bins of the frequency spectrum of each of the plurality of sets of waveform data;
perform, iteratively, for a number of frequency components in the frequency spectrum in the frequency spectrum matrix, the magnitude quantization on the frequency component of each of the plurality of sets of waveform data to create the two-bins histogram;
append a larger count value out of the two-bins histogram to the list;
perform the magnitude quantization by calculating an RMS (Root Mean Square) value of time domain waveform data obtained from the frequency spectrum of the frequency spectrum matrix subjected to filtering to extract frequency components specified;
obtain a minimum frequency component where the list takes a minimum count value;
perform the magnitude quantization on the minimum frequency component;
calculate a quantized label vector by calculating a bin range in which the magnitude falls and assigning 0 to a lower range and 1 to an upper range;
calculate a predicted label vector;
generate first and second matrices using one hot-encoding of the quantized and predicted label vectors;
calculate a dot product of the first and second matrices to generate a matrix with two rows and two columns;
identify a predicted label corresponding to a maximum in the first row of the matrix with the two rows and the two columns, as normal; and
identify a predicted label corresponding a maximum in the second row of the matrix, as anomalous.

9. A computer-implemented anomaly detection method comprising:
obtaining a plurality of sets of waveform data of a periodic electric waveform signal of an electric appliance, with a length of each of the plurality of sets of waveform data being set to one cycle time of the periodic electric waveform signal;
transforming, using FFT (fast Fourier transform) processing, each of the plurality of sets of waveform data from a time domain into a frequency domain representation including a plurality of frequency bins of a fundamental frequency which is a reciprocal of the one cycle time and harmonics of the fundamental frequency;
applying a frequency processing including normalization and filtering to the frequency domain representation of each of the plurality of sets of waveform data to obtain a frequency spectrum for each of the plurality sets of waveform data;
extracting and separating odd and even frequency harmonics of the fundamental frequency from the frequency spectrum of each of the plurality of sets of waveform data to create odd and even frequency harmonic matrices;
performing canonical correlation analysis (CCA) on the odd and even frequency harmonic matrices to obtain CCA features;
performing linear transformation on the CCA features to obtain linear transformed features;
generating a cluster model based on the linear transformed features;
performing magnitude quantization of the frequency spectrum of each of the plurality of sets of waveform data to identify normal and anomalous waveform signals; and
outputting at least the identified result,
wherein the magnitude quantization includes:
applying filtering to the frequency spectrum of each of the plurality of sets of waveform data;
transforming, using inverse FFT p

Assignees

• Nec Corp

Inventors

• Petladwala Murtuza
• Sendoda Mitsuru
• Ohta Yuko
• Suzuki Ryota

Classifications

• G01R23/18 Primary

  with provision for recording frequency spectrum · CPC title

• G01R23/145

  by heterodyning or by beat-frequency comparison with the harmonic of an oscillator · CPC title

• G01R21/06

  by measuring current and voltage (G01R21/08 - G01R21/133 take precedence) · CPC title

• G06N20/00

  Machine learning · CPC title

• G05B23/024

  Quantitative history assessment, e.g. mathematical relationships between available data; Functions therefor; Principal component analysis [PCA]; Partial least square [PLS]; Statistical classifiers, e.g. Bayesian networks, linear regression or correlation analysis; Neural networks · CPC title