Anomaly detection based on directional data

An anomaly detecting apparatus for detecting an anomaly in monitored data, comprising:

• a directional data generating portion for sequentially generating directional data which are unit vectors indicating a feature of each piece of monitored data corresponding to the monitored data which are input in sequence;
• a dissimilarity calculating portion for calculating a dissimilarity of the directional data to a pre-specified reference vector,
• a moment updating portion for updating a moment of the distribution of the dissimilarity appearing when the directional data is modeled with a multi-dimensional probability distribution, on the basis of a moment already calculated and the dissimilarity of the directional data corresponding to a new piece of the monitored data;
• a parameter calculating portion for calculating a parameter determining a variance of the multi-dimensional probability distribution, based on the moment;
• a threshold calculating portion for calculating a threshold of the dissimilarity based on the multi-dimensional probability distribution the variance of which is determined by the parameter;
• and
• an anomaly detecting portion for detecting an anomaly in the monitored data that corresponds to the dissimilarity if the dissimilarity exceeds the threshold,
• wherein the dissimilarity calculating portion calculates the dissimilarity by using an Expression:
• Z≡l−rTu,
• where z denotes the dissimilarity, r denotes the reference vector, and u denotes the directional data;
• wherein the multi-dimensional probability distribution is a von Mises-Fisher distribution;
• and
• the distribution of the dissimilarity is approximated to a chi-square distribution;
• wherein the parameter calculating portion further calculates a parameter determining an effective dimension of the von Mises-Fisher distribution based on the moment;
• and
• the threshold calculating portion calculates the threshold of the dissimilarity based on the von Mises-Fisher distribution the effective dimension and the variance of which are predetermined;
• wherein:
• the distribution of the dissimilarity is represented by an Expression,
• q
• ⁡
• (
• z
• )
• =
• 1
• Γ
• ⁡
• (
• (
• n
• -
• 1
• )
• /
• 2
• )
• ⁢
• (
• z
• 2
• ⁢
• Σ
• )
• n
• -
• 1
• 2
• ⁢
• exp
• ⁡
• [
• -
• z
• 2
• ⁢
• Σ
• ]
• ⁢
• 1
• z
• ,
• where z denotes the dissimilarity, Γ( ) denotes a gamma function, n denotes the parameter determining the effective dimension of the von Mises-Fisher distribution, and S denotes the parameter determining the variance of the von Mises-Fisher distribution;
• and
• the moment updating portion calculates a new moment based on the moment already calculated and on the dissimilarity of the directional data corresponding to a new piece of the monitored data by using an Expression,
• <zk>(t)=(1−b)<zk>(t−1)+bz(t)k,
• where <zk>(t) denotes the k-th moment to be calculated newly at time t, <zk>(t−1) denotes the k-th moment already calculated at time t−1, z(t) denotes the dissimilarity of the directional data corresponding to a new piece of the monitored data at time t, and b denotes a forgetting factor,
• wherein the percentage of the monitored data that is to be considered normal is input into the threshold calculating portion and the threshold calculating portion calculates the threshold such that the accumulation of the probabilities that the dissimilarity in the distribution of the dissimilarity corresponding to the multi-dimensional probability distribution the variance of which is determined by the parameter becomes less than or equal to the threshold is equal to the percentage;
• wherein the multi-dimensional probability distribution is a multi-dimensional normal distribution and the distribution of the dissimilarity is a chi-square distribution;
• wherein:
• the anomaly detecting apparatus detects an anomaly in a text classifying device which classifies sequentially input text data into a particular category;
• the directional data generating portion sequentially generates the directional data indicating a feature of the text data, correspondingly to each piece of the sequentially input text data;
• said dissimilarity calculating portion calculates the dissimilarity of the directional data to the reference vector specified correspondingly to the category;
• the anomaly detecting portion detects an anomaly in the text classification device if the dissimilarity corresponding to the text data classified by the text classifying device into the category exceeds the threshold;
• the moment updating portion updates the moment of the distribution of the dissimilarity appearing when the directional data is modeled by using a multi-dimensional probability distribution corresponding to the category, if the text data is classified into the category and the dissimilarity does not exceed the threshold;
• the parameter calculating portion calculates a parameter determining the variance of the multi-dimensional probability distribution corresponding to the category, based on the moment;
• and
• the threshold calculating portion calculates a threshold of the dissimilarity corresponding to the category, based on the multi-dimensional probability distribution the variance of which is determined by the parameter, and
• further comprising:
• a category input portion for a user to input an instruction as to whether the text data should be classified into the category, if anomaly in the text classifying device is detected;
• and
• a reconfiguration directing portion for directing the text classifying device to reconfigure the text classifying device so as to properly classify a plurality of pieces of text data classified by a user, if an anomaly in the text classifying device has been detected with a frequency higher than or equal to a predetermined frequency, and
• wherein:
• the anomaly detecting apparatus detects an anomaly in an information system including a plurality of information processing apparatuses;
• transactions sent by each of the information processing apparatuses for requesting for processing of a service provided by another of the information processing apparatuses are input, as the monitored data, into the directional data generating portion and the directional data generating portion sequentially generates the directional data indicating features of the transactions sent in a predetermined period;
• and
• the anomaly detecting portion detects an anomaly in the information system if the dissimilarity exceeds the threshold
• wherein the directional data generating portion comprises:
• a measured variable input portion for inputting time-series data of a plurality of measured parameters measured from a monitored object, as the monitored data;
• a change rate calculating portion for calculating the time-series change rate at which each of the plurality of measured parameters changes over time based on the time-series data of the measure parameters;
• and
• a dependency calculating portion for calculating the degree of association indicating the intensity of association of each of the plurality of measured parameters with each of the other measured parameters, based on the change rates of that parameter and that other parameter to generate the directional data based on the degree of association
• wherein:
• the anomaly detecting apparatus detects an anomaly in an automobile;
• the measured variable input portion inputs, as the monitored data, the time-series data of the plurality of measured parameters measured by a plurality of sensors provided in portions of the monitored automobile;
• the dissimilarity calculating portion calculates the dissimilarity of the directional data to the reference vector determined based on the time-series data of a plurality of measured parameters measured from a properly operating automobile;
• and
• the anomaly detecting portion detects an anomaly in the monitored automobile if the dissimilarity exceeds the threshold.