Autoregressive Model Fitting for Time Series Anomaly Detection

A computer-implemented method for anomaly detection in time-series data from a robot, said method comprising:

• providing time-series data for parameters collected during an operation by the robot, where the parameters include a rotational velocity and a torque command for a joint of the robot;
• segmenting the time-series data to retain segmented time-series data from a time range of interest beginning when values of the rotational velocity have dropped below a first threshold amplitude and ending when values of the torque command have dropped below a second threshold amplitude;
• fitting an autoregressive model to the segmented time-series data for the torque command, including defining an order of the autoregressive model as two, formulating a set of equations relating a value of the torque command at each time step to a linear combination of values of the torque command at two previous time steps plus a stochastic term, and using a least-squares regression to determine values of coefficients used in the equations which minimize the stochastic term;
• setting an error term equal to the stochastic term for each time step of the autoregressive model;
• calculating an error metric as a square root of a sum of the squares of the error term for all of the time steps;
• and
• issuing an alert when the error metric has a value exceeding a predefined threshold.

A time-series anomaly detection system, said system comprising:

• a computer having a processor and memory configured to perform steps including;
• receiving time-series data for parameters collected during an operation by a machine;
• segmenting the time-series data to retain segmented time-series data from a time range of interest;
• fitting an autoregressive model to the segmented time-series data for one of the parameters;
• calculating an error term for each time step of the autoregressive model;
• calculating an error metric based on the error term for all of the time steps;
• and
• issuing an alert when the error metric has a value exceeding a predefined threshold.