Detection of deviation from an operating state of a device

1. A method of detecting deviation from an operational state of a device, the method comprising:

obtaining preprocessed data corresponding to data sensed by one or more sensor devices coupled to the device;

processing the preprocessed data using a trained anomaly detection model to generate an anomaly score, wherein processing the preprocessed data at the trained anomaly detection model includes:

inputting the preprocessed data to an autoencoder; and

generating a residual based on an output of the autoencoder, wherein the anomaly score is generated based on the residual; and

processing the anomaly score using an alert generation model to determine whether to generate an alert.

2. The method of claim 1 , wherein processing the preprocessed data at the trained anomaly detection model further includes:

inputting the residual to a Hotelling test statistics module; and

generating the anomaly score at the Hotelling test statistics module using a multivariate test statistic that is based on residual data and reference residual data.

3. The method of claim 1 , further comprising, prior to obtaining the preprocessed data:

receiving a first time series of data indicative of normal operation of the device from the one or more sensor devices; and

training the autoencoder based on the first time series.

4. The method of claim 1 , wherein the alert generation model includes a sequential probability ratio test that determines whether a set of one or more anomaly scores indicates deviation from normal operation of the device.

5. The method of claim 1 , wherein the data sensed by the one or more sensor devices indicates at least one of a motion or an acceleration associated with vibration of the device.

6. The method of claim 1 , wherein obtaining the preprocessed data includes applying a transform to the data sensed by the one or more sensor devices to generate a set of features in a frequency domain.

7. The method of claim 6 , further comprising determining a contribution of each feature to the anomaly score.

8. The method of claim 1 , further comprising generating a graphical user interface including:

a graph indicative of a performance metric of the device over time;

an alert indication corresponding to a portion of the graph; and

an indication of one or more sets of feature data associated with the alert indication.

9. A system to detect deviation from an operational state of a device, the system comprising:

a memory including a trained anomaly detection model and an alert generation model; and

one or more processors coupled to the memory, the one or more processors configured to:

obtain preprocessed data corresponding to data sensed by one or more sensor devices coupled to the device;

process the preprocessed data using the trained anomaly detection model to generate an anomaly score, the trained anomaly detection model including:

an autoencoder configured to generate a reconstruction of the preprocessed data; and

a residual generator configured to generate a residual based on an output of the autoencoder, wherein the anomaly score is generated based on the residual; and

process the anomaly score using the alert generation model to determine whether to generate an alert.

10. The system of claim 9 , wherein the trained anomaly detection model further includes a Hotelling test statistics module configured to generate the anomaly score based on the residual.

11. The system of claim 9 , wherein the memory further includes a calibration module that is executable by the one or more processors to:

receive a first time series of data indicative of normal operation of the device from the one or more sensor devices; and

train the autoencoder based on the first time series.

12. The system of claim 9 , wherein the alert generation model includes a sequential probability ratio test that determines whether a set of one or more anomaly scores indicates deviation from normal operation of the device.

13. The system of claim 9 , wherein the data sensed by the one or more sensor devices indicates at least one of a motion or an acceleration associated with vibration of the device.

14. The system of claim 9 , wherein the one or more processors are configured to apply a transform to the data sensed by the one or more sensor devices to generate the preprocessed data including a set of features in a frequency domain.

15. The system of claim 14 , wherein the one or more processors are configured to determine a contribution of each feature to the anomaly score.

16. The system of claim 9 , wherein the memory further includes a graphical user interface module that is executable by the one or more processors to generate a graphical user interface to display an alert indication.

17. A self-training sensor apparatus comprising:

a sensor to generate sensor data associated with operation of a device;

an input interface to receive input to select between a training mode of operation and a detection mode of operation;

a memory including model generation instructions;

one or more processors coupled to the memory, the one or more processors configured to:

while operating in the training mode of operation, receive first sensor data from the sensor and execute the model generation instructions based on the first sensor data to generate a trained anomaly detection model; and

while operating in the detection mode of operation:

receive second sensor data from the sensor;

use the trained anomaly detection model to generate an anomaly score based on the second sensor data; and

process the anomaly score using an alert detection model to determine whether to generate an alert; and

an output interface to output an alert indication responsive to the one or more processors generating the alert.

18. The self-training sensor apparatus of claim 17 , wherein before the anomaly detection model is trained, the self-training sensor apparatus is agnostic regarding the device.

19. The self-training sensor apparatus of claim 17 , wherein the memory includes a data structure representing an autoencoder, wherein execution of the model generation instructions causes the autoencoder to be trained, and wherein the anomaly score is generated based on a residual corresponding to an output of the autoencoder.

20. A method of operating a self-training sensor apparatus, the method comprising:

receiving, at an input interface of the self-training sensor apparatus, a first input to select a training mode of operation;

in response to receiving the first input:

receiving, from a sensor of the self-training sensor apparatus, first sensor data associated with operation of a device; and

executing, at one or more processors of the self-training sensor apparatus, model generation instructions based on the first sensor data to generate:

a trained anomaly detection model that includes an autoencoder; and

an alert detection model; and

after generating the trained anomaly detection model and the alert detection model:

receiving, at the input interface, a second input to select a detection mode of operation; and

in response to receiving the second input:

receiving second sensor data from the sensor;

using the trained anomaly detection model to generate an anomaly score based on a residual corresponding to an output of the autoencoder; and

processing the anomaly score using the alert detection model to determine whether to generate an alert.

21. The method of claim 20 , further comprising, in response to a determination to generate the alert, outputting an alert indication at an output interface of the self-training sensor apparatus.