Sensor unit with on-device learning and anomaly detection
A sensor unit is coupled to a machine and configured to detect anomalous behavior of the machine. The sensor unit includes a low power microcontroller that learns to recognize a plurality of operations of the machine. The sensor unit generates mean vector and inverse of a Cholesky decomposition matrix for each operation. During a detection mode the sensor unit computes a Mahalanobis distance for each feature vector, mean vector and first matrix. The sensor unit detects anomalous behavior or classifies the operation of the machine based on the Mahalanobis distances.
1 . A method, comprising:
generating, with an inertial sensor of the sensor unit, a plurality of samples of first sensor data with a sensor of a sensor unit coupled to a machine;
receiving the plurality of samples of the first sensor data with a feature generator of the sensor unit;
generating, with the feature generator of the sensor unit, a first feature vector for each sample;
receiving the first feature vectors with a learning module of the sensor unit;
receiving, with a detection module of a first microcontroller of the sensor unit, an inverse of a Cholesky decomposition of a covariance matrix based on the first feature vectors;
generating, with the inertial sensor of the sensor unit, second sensor data;
receiving the second sensor data with feature generator;
generating a second feature vector with the feature generator based on the second sensor data;
receiving the second feature vector with the detection module; and
activating an interrupt from the first microcontroller of the sensor unit to a second microcontroller of the sensor unit indicating that the machine is operating in an anomalous state based, in part, on a Mahalanobis distance associated with the second feature vector and the inverse of Cholesky decomposition being greater than a threshold distance.
2 . The method of claim 1 , comprising generating a mean vector corresponding to a mean of the first feature vectors.
3 . The method of claim 2 , comprising generating variance values for each sample of the first feature data, wherein each first feature vector includes the variance values for the corresponding sample.
4 . The method of claim 2 , comprising generating peak-to-peak values for each sample of the first feature data, wherein each first feature vector includes the peak-to-peak values for the corresponding sample.
5 . The method of claim 2 , wherein each first feature vector includes the sensor data from the corresponding sample.
6 . The method of claim 1 , further comprising computing the square of the Mahalanobis distance.
7 . The method of claim 1 , comprising determining that the machine is operating anomalously if the Mahalanobis distance is greater than a threshold value.
8 . A method, comprising:
generating, with an inertial sensor of the inertial sensor unit, a plurality of samples of first sensor data;
receiving the plurality of samples of the first sensor data with a feature generator of the sensor unit;
generating, with the feature generator of the sensor unit, a first feature vector for each sample;
receiving the first feature vectors with a learning module of the sensor unit;
receiving, with a detection module of a first microcontroller of the sensor unit, a plurality of inverse of a Cholesky decompositions of covariance matrices based on the first feature vectors, each inverse of Cholesky decomposition being associated with a respective operation of the machine;
generating, with the inertial sensor of the sensor unit, second sensor data;
receiving the second sensor data with feature generator;
generating a second feature vector with the feature generator based on the second sensor data;
receiving the second feature vector with the detection module; and
activating an interrupt from the first microcontroller to a second microcontroller of the sensor unit indicating that the machine is operating in an anomalous state based, in part, on each of a plurality of Mahalanobis distances associated with the second feature vector and the inverse of Cholesky decompositions being greater than a threshold distance.
9 . The method of claim 8 , comprising:
generating, with the inertial sensor of the sensor unit, third sensor data;
generating a third feature vector with the feature generator based on the third sensor data;
receiving the third feature vector with the detection module;
identifying a smallest Mahalanobis distance from a plurality of Mahalanobis distances associated with third feature vector and the inverse of Cholesky decompositions; and
determining that the machine is performing an operation corresponding to the inverse of Cholesky decomposition associated with the smallest Mahalanobis distance if the smallest Mahalanobis distance is smaller than the threshold distance.
10 . The method of claim 8 , comprising:
generating, with the sensor unit, a respective mean vector for each of a plurality of operations of the machine.
11 . The method of claim 8 , comprising generating variance values for each sample of the first feature data, wherein each first feature vector includes the variance values for the corresponding sample.
12 . The method of claim 8 , comprising generating peak-to-peak values for each sample of the first feature data, wherein each first feature vector includes the peak-to-peak values for the corresponding sample.
13 . The method of claim 8 , wherein each first feature vector includes the sensor data from the corresponding sample.
14 . A sensor unit, comprising:
a plurality of inertial sensors coupled to a machine and configured to generate a plurality of samples of first sensor data in a learning mode and to generate second sensor data in a detection mode; and
a sensor processing unit coupled to the plurality of sensors and including a first microcontroller and a second microcontroller, the first microcontroller including:
a feature generator configured to:
receive, in the learning mode, the plurality of samples of the first sensor data:
generate, in the learning mode, a first feature vector for each sample;
receive, in the detection mode, the second sensor data; and
generate, in the detection mode, a second feature vector based on the second sensor data;
a detection module configured to:
receive, in the detection mode, an inverse of a Cholesky decomposition of a covariance matrix based on the first feature vectors;
receiving the second feature vector; and
activate an interrupt from the first microcontroller to the second microcontroller indicating that the machine is operating in an anomalous state based, in part, on a Mahalanobis distance associated with the feature vector and the inverse of Cholesky decomposition being greater than a threshold distance.
15 . The sensor unit of claim 14 , wherein the plurality of inertial sensors includes at least one accelerometer.
16 . The sensor unit of claim 15 , wherein the plurality of inertial sensors includes at least one gyroscope.
17 . The sensor unit of claim 16 , further comprising at least one temperature sensor.
18 . The sensor unit of claim 14 , wherein the sensor processing unit is configured to determine whether or not the machine is operating anomalously based on a square of the Mahalanobis distance.