Unsupervised anomaly detection for autonomous vehicles

A computer-implemented method of detecting and addressing anomalies in autonomous vehicle telemetry data, the method comprising:

• receiving, by an anomaly detection system, a new time series data record from an autonomous vehicle;
• processing, by the anomaly detection system, the new time series data record with a machine learning model to generate an anomaly score;
• and
• in response to determining that the anomaly score is greater than an anomaly threshold value:
• determining, by the anomaly detection system, an action to be taken to address the anomaly, wherein the action to be taken to address the anomaly is at least one of navigating to an emergency repair location, immediately performing a landing procedure, or rescheduling a future trip;
• and
• transmitting, by the anomaly detection system, a command that causes the action to be taken to address the anomaly;
• wherein the machine learning model was trained using a technique that includes:
• initializing, by a computing system, a set of fitting weights for the machine learning model, wherein the fitting weights are used by the machine learning model to determine anomaly scores for new time series data records;
• initializing, by the computing system, a set of time series data record weights associated with a training data set, wherein each time series data record weight is associated with a time series data record of a plurality of time series data records in the training data set, and wherein the time series data record weights are usable to identify one or more anomalous time series data records of the training data set to be ignored during optimization of the set of fitting weights;
• and
• training, by the computing system, the machine learning model by alternating at least two times between using the set of time series data record weights while optimizing the set of fitting weights based on the training data set, and using the set of fitting weights while optimizing the set of time series data record weights, wherein using the set of time series data record weights while optimizing the set of fitting weights includes ignoring at least one of the plurality of time series data records in the training data set based on the time series data record weights.

A system, comprising:

• an autonomous vehicle;
• and
• an anomaly detection system communicatively coupled to the autonomous vehicle, the anomaly detection system having at least one computing device that includes at least one processor and a non-transitory computer-readable medium having logic stored thereon that, in response to execution by the at least one processor, causes the anomaly detection system to perform actions comprising:
• receiving, by the anomaly detection system, a new time series data record from the autonomous vehicle;
• processing, by the anomaly detection system, the new time series data record with a machine learning model to generate an anomaly score;
• and
• in response to determining that the anomaly score is greater than an anomaly threshold value:
• determining, by the anomaly detection system, an action to be taken to address the anomaly;
• and
• transmitting, by the anomaly detection system, a command to the autonomous vehicle that identifies the action to be taken;
• and
• wherein the autonomous vehicle is configured to:
• receive the command from the anomaly detection system;
• and
• autonomously perform the action identified by the command;
• wherein the machine learning model was trained using a technique that includes:
• initializing, by a computing system, a set of fitting weights for the machine learning model, wherein the fitting weights are used by the machine learning model to determine anomaly scores for new time series data records;
• initializing, by the computing system, a set of time series data record weights associated with a training data set, wherein each time series data record weight is associated with a time series data record of a plurality of time series data records in the training data set, and wherein the time series data record weights are usable to identify one or more anomalous time series data records of the training data set to be ignored during optimization of the set of fitting weights;
• and
• training, by the computing system, the machine learning model by alternating at least two times between using the set of time series data record weights while optimizing the set of fitting weights based on the training data set, and using the set of fitting weights while optimizing the set of time series data record weights, wherein using the set of time series data record weights while optimizing the set of fitting weights includes ignoring at least one of the plurality of time series data records in the training data set based on the time series data record weights.