# US 12,182,731 B2 · Anomaly detection apparatus, anomaly detection method, and program > Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt - Number: US 12,182,731 B2 - Document ID: US12182731B2 - Status: Granted - Kind: B2 - Application number: 17/260,956 - Filed: 2019-07-01 - Published: 2021-10-21 - Granted: 2024-12-31 - Earliest claimed: 2018-07-20 - Assignees: Nippon Telegraph and Telephone Corporation - Applicants: Nippon Telegraph and Telephone Corporation - Inventors: Yuta Kawachi; Yuma Koizumi; Noboru Harada; Shin Murata - Web page: https://patentgrep.com/patent/US12182731B2 ## CPC - G06N 3/088: Non-supervised learning, e.g. competitive learning - G06N 5/04: Inference or reasoning models - G01M 99/00: Subject matter not provided for in other groups of this subclass - G05B 23/02: Electric testing or monitoring - G06N 3/045: Combinations of networks - G06N 3/0455: Auto-encoder networks; Encoder-decoder networks - G06N 3/047: Probabilistic or stochastic networks - G06N 3/0475: Generative networks - G06N 3/09: Supervised learning - G06N 20/00: Machine learning - G10L 15/10: using distance or distortion measures between unknown speech and reference templates - G10L 15/16: using artificial neural networks - G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS - G01M: TESTING STATIC OR DYNAMIC BALANCE OF MACHINES OR STRUCTURES; TESTING OF STRUCTURES OR APPARATUS, NOT OTHERWISE PROVIDED FOR - G05B: CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS (systems for controlling or regulating non-electric variables G05D; systems for regulating electric or magnetic variables G05F; control devices or systems insofar as characterised by mechanical features only G05G) - G10L: SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING ## Abstract A possible region of encoding results of anomalous samples is limited. An encoder storage unit 14 stores an encoder for projecting an input feature value into a latent space in which the latent space is a closed manifold, a normal distribution obtained by learning normal data and an anomalous distribution obtained by learning anomalous data are held on the manifold, and a decoder for reconstructing the output of the encoder. An encoding unit 15 obtains a reconstruction result output by the decoder when a feature value of target data is input to the encoder. An anomaly score calculation unit 16 calculates an anomaly score of the target data based on distances between the reconstruction result and the normal distribution and distances between the reconstruction result and the anomalous distribution. ## Claims ### Claim 1 An anomaly detection apparatus, comprising: a storage configured to store a neural network comprising an encoder for projecting an input feature value into a latent space, and a decoder for reconstructing the output of the encoder, wherein the latent space is a closed manifold, and holds a normal distribution obtained by learning normal data and an anomalous distribution obtained by learning anomalous data on the manifold; and a processor configured to: obtain a reconstruction result output by the decoder when a feature value of target data is input to the encoder, wherein the target data represents operation data of a machine as waveform data processed by a sensor; and calculate an anomaly score of the target data based on distances between the reconstruction result and the normal distribution and distances between the reconstruction result and the anomalous distribution, wherein the encoder outputs angle parameters and a concentration parameter, and the processor is further configured to: set two different distributions on the manifold as a prior distribution of the normal distribution and a prior distribution of the anomalous distribution; learn the normal distribution and the anomalous distribution with the concentration parameter set to a predetermined constant, among the outputs of the encoder obtained when a feature value of learning data composed of the normal data and the anomalous data is input; and learn the normal distribution and the anomalous distribution by using a KL divergence upper limit value, wherein the KL divergence upper limit value is calculated based on a vector obtained by multiplying a rotation matrix and a fixed vector having any value in common overall, a direction vector of the prior distribution, and a concentration parameter of the prior distribution; and the rotation matrix is constituted by the angle parameters output by the encoder when a feature value of the learning data is input, and is constructed using vectors obtained by adding a value obtained by multiplying a random number generated from a Gaussian distribution and a value obtained by converting the concentration parameter output by the encoder when the feature value of the learning data is input into a standard deviation equivalent of a Gaussian distribution, to the angle parameters output by the encoder; and determine, based on the calculated anomaly score of the target data and a predetermined threshold, an anomalous state of operation of the machine. ### Claim 2 (depends on 1) The anomaly detection apparatus according to claim 1, wherein after the learning of the normal distribution and the anomalous distribution by the third processing circuitry using the learning data is executed a predetermined number of times, the learning of the normal distribution and the anomalous distribution by the fourth processing circuitry using the learning data is executed. ### Claim 3 (depends on 1) The anomaly detection apparatus according to claim 1, wherein in the encoder, a cost function having a high value for a value near 90 degrees is set for the angle parameters. ### Claim 4 (depends on 1) The anomaly detection apparatus according to claim 1, wherein the latent space is a hypersphere, and the normal distribution and the anomalous distribution are von Mises-Fisher distributions. ### Claim 5 (depends on 1) The anomaly detection apparatus according to claim 1, the processor further configured to execute operations comprising: receive first waveform data as the target data from the sensor; receive second waveform data as the normal data; and receive third waveform data as the anomalous data. ### Claim 6 (depends on 5) The anomaly detection apparatus according to claim 5, wherein the first waveform data represent an operation sound of a machine, the second waveform data represent a normal operation sound of the machine, and the third waveform data represent an anomalous operation sound of the machine. ### Claim 7 (depends on 1) The anomaly detection apparatus according to claim 1, the processor further configured to: receive first waveform data as the target data from the sensor, wherein the sensor converts the operation data of the machine into the first waveform data; and determine, based on the calculated anomaly score of the target data, an anomalous state of operation of the machine. ### Claim 8 (depends on 1) The anomaly detection apparatus according to claim 1, wherein the first waveform data represents a type of the operation data of the machine, and the type of the operation data is one of: sound, temperature, pressure, displacement, or traffic data. ### Claim 9 An anomaly detection apparatus comprising: a neural network comprising: an encoder for projecting an input feature value into a latent space, and a decoder for reconstructing the output of the encoder, wherein the latent space is a closed manifold and holds a normal distribution obtained by learning normal data and an anomalous distribution obtained by learning anomalous data on the manifold; and a processor configured to: obtain a reconstruction result output by the decoder when a feature value of target data is input to the encoder, wherein the target data represents operation data of a machine as waveform data processed by a sensor; and calculate an anomaly score of the target data based on distances between the reconstruction result and the normal distribution and distances between the reconstruction result and the anomalous distribution, wherein the encoder outputs angle parameters and a scalar parameter corresponding to a logarithmic variance of a Gaussian distribution, and the processor further configured to: set two different distributions on the manifold as a prior distribution of the normal distribution and a prior distribution of the anomalous distribution; learn the normal distribution and the anomalous distribution with the concentration parameter set to a predetermined constant among the outputs of the encoder obtained when a feature value of learning data composed of the normal data and the anomalous data is input; and learn the normal distribution and the anomalous distribution by obtaining vectors on a unit hyperplane that are the result of multiplying a rotation matrix and a fixed vector having any value in common overall, multiplying the concentration parameter of the prior distribution by the inner product of the vectors on the unit hyperplane and the direction vector of the prior distribution, subtracting the resulting value from an exponent of the scalar parameter output by the encoder when the feature value of the learning data is input, and setting the resulting value as a KL divergence value, wherein the rotation matrix is constituted by the angle parameters output by the encoder when a feature value of the learning data is input, and is constructed using vectors obtained by adding a value obtained by multiplying a random number generated from a Gaussian distribution by the scalar parameter output by the encoder when the feature value of the learning data is input, to the angle parameters output by the encoder; and determine, based on the calculated anomaly score of the target data and a predetermined threshold, an anomalous state of operation of the machine. ### Claim 10 (depends on 9) The anomaly detection apparatus according to claim 9, wherein after the learning of the normal distribution and the anomalous distribution by the third processing circuitry using the learning data is executed a predetermined number of times, the learning of the normal distribution and the anomalous distribution by the fourth processing circuitry using the learning data is executed. ### Claim 11 (depends on 9) The anomaly detection apparatus according to claim 9, wherein in the encoder, a cost function having a high value for a value near 90 degrees is set for the angle parameters. ### Claim 12 (depends on 9) The anomaly detection apparatus according to claim 9, wherein after the learning of the normal distribution and the anomalous distribution by the third processing circuitry using the learning data is executed a predetermined number of times, the learning of the normal distribution and the anomalous distribution by the fourth processing circuitry using the learning data is executed. ### Claim 13 (depends on 9) The anomaly detection apparatus according to claim 9, wherein in the encoder, a cost function having a high value for a value near 90 degrees is set for the angle parameters. ### Claim 14 (depends on 9) The anomaly detection apparatus according to claim 9, wherein the latent space is a hypersphere, and the normal distribution and the anomalous distribution are von Mises-Fisher distributions. ### Claim 15 An anomaly detection method comprising: storing, in a storage of an anomaly detection apparatus, a neural network comprising an encoder for projecting an input feature value into a latent space, and a decoder for reconstructing the output of the encoder, wherein the latent space is a closed manifold, and holds a normal distribution obtained by learning normal data and an anomalous distribution obtained by learning anomalous data on the manifold, the encoder outputs angle parameters and a concentration parameter, and; obtaining, by a reconstruction result output by the decoder when a feature value of target data is input to the encoder, wherein the target data represents operation data of a machine as waveform data processed by a sensor; and calculating, by an anomaly score of the target data based on distances between the reconstruction result and the normal distribution and distances between the reconstruction result and the anomalous distribution, setting, two different distributions on the manifold as a prior distribution of the normal distribution and a prior distribution of the anomalous distribution; learning, the normal distribution and the anomalous distribution with the concentration parameter set to a predetermined constant, among the outputs of the encoder obtained when a feature value of learning data composed of the normal data and the anomalous data is input; and learning, the normal distribution and the anomalous distribution by using a KL divergence upper limit value, wherein the KL divergence upper limit value is calculated based on a vector obtained by multiplying a rotation matrix and a fixed vector having any value in common overall, a direction vector of the prior distribution, and a concentration parameter of the prior distribution; and the rotation matrix is constituted by the angle parameters output by the encoder when a feature value of the learning data is input, and is constructed using vectors obtained by adding a value obtained by multiplying a random number generated from a Gaussian distribution and a value obtained by converting the concentration parameter output by the encoder when the feature value of the learning data is input into a standard deviation equivalent of a Gaussian distribution, to the angle parameters output by the encoder; and determining based on the calculated anomaly score of the target data and a predetermined threshold, an anomalous state of operation of the machine. ### Claim 16 (depends on 1) A non-transitory computer-readable recording medium on which a program recorded thereon for causing a computer to function as the anomaly detection apparatus according to claim 1.