Method for predictive maintenance of satellites

1. A method comprising:

during a first time period:

accessing a first historical timeseries telemetry data and a first timeseries failure data for a first population of satellites at a first frequency;

training a first model to predict a failure of a first satellite of the first population of satellites within a first time window based on the first historical timeseries telemetry data and the first timeseries failure data; and

training a second model to predict the failure within a second time window, shorter than the first time window, based on the first historical timeseries telemetry data and first timeseries failure data; and

during a second time period for a first satellite:

accessing a first set of timeseries telemetry data broadcast by a first satellite at a first frequency during a current time interval;

predicting a first probability of failure of the first satellite within the first time window based on the first model and the first set of timeseries telemetry data;

predicting a second probability of failure of the first satellite within the second time window based on the second model and the first set of timeseries telemetry data; and

in response to the first probability of failure falling below a first threshold probability and the second probability of failure exceeding the first threshold probability:

detecting an external source of failure of the first satellite;

retrieving a script for transferring the first satellite to a safe state; and

uploading the script to the first satellite.

2. The method of claim 1 :

further comprising, during the first time period:

training a third model to predict a first anomaly score characterizing the first historical timeseries telemetry data and the first timeseries failure data within the first time window; and

training a fourth model to predict a second anomaly score characterizing the first historical timeseries telemetry data and the first timeseries failure data within the second time window;

further comprising, during the second time period for the first satellite:

in response to the first anomaly score falling below a first threshold anomaly score, characterizing the first set of timeseries telemetry data for the first time window as nominal; and

in response to the second anomaly score exceeding the first threshold anomaly score, characterizing the first set of timeseries telemetry data for the second time window as anomalous; and

wherein generating the first command to increase downlink frequency and prioritize transmission of telemetry data from the first satellite comprises generating the first command further in response to characterizing the first set of timeseries telemetry data for the second time window as anomalous.

3. The method of claim 1 , wherein in response to the first probability of failure falling below the first threshold probability and the second probability of failure exceeding the first threshold probability comprises accessing a second set of timeseries telemetry data broadcast by the first satellite at a second frequency higher than the first frequency.

4. The method of claim 1 , further comprising:

retraining the first model based on the second set of timeseries telemetry data broadcast by the first satellite at the second frequency to predict a third probability of failure within the first time window; and

retraining the second model based on the second set of timeseries telemetry data broadcast by the first satellite at the second frequency to predict a fourth probability of failure within the second time window.

5. The method of claim 1 , further comprising:

accessing a second set of historical timeseries telemetry data and a second set of timeseries failure data based on ground testing of the first population of satellites;

training the first model to predict a failure of the first satellite of the first population of satellites within a first time window based on the second set of historical timeseries telemetry data and second set of timeseries failure data; and

training the second model to predict the failure within a second time window, shorter than the first time window, based on the second set of timeseries telemetry data and the second set of timeseries failure data.

6. A method comprising:

during a first time period:

accessing a first set of historical timeseries telemetry data and a first set of timeseries failure data for a first population of satellites at a first frequency;

training a first model to predict a failure of a first satellite of the first population of satellites within a first time window based on the first set of timeseries telemetry data and the first set of timeseries failure data; and

training a second model to predict the failure within a second time window, shorter than the first time window, based on the first set of timeseries telemetry data and the first set of timeseries failure data; and

during a second time period for a first satellite:

accessing a first set of timeseries telemetry data broadcast by a first satellite at a first frequency during a current time interval;

predicting a first probability of failure of the first satellite within the first time window based on the first model and the first set of timeseries telemetry data;

predicting a second probability of failure of the first satellite within the second time window based on the second model and the first set of timeseries telemetry data; and

in response to the first probability of failure exceeding a threshold probability and the second probability of failure exceeding the threshold probability:

retrieving script for entering a safe state; and

uploading the script to the first satellite.

7. The method of claim 6 , further comprising, predicting a source of the second probability of failure by:

accessing historical telemetry data, in a set of telemetry domains, for the first satellite;

grouping telemetry data into a set of historical telemetry containers, each telemetry container associated with a discrete historical time window and containing values representing the set of telemetry domains;

populating a current telemetry container with current telemetry data; and

characterizing the current telemetry container as anomalous, in a particular telemetry domain in the set of telemetry domains, based on distances between the current telemetry container and the set of historical telemetry containers in a feature space; and

identifying the source of the second probability of failure, within the second time window, based on the particular telemetry domain; and

wherein retrieving the script comprises selecting the script, from a set of pre-generated scripts, executable by the first satellite to respond to the source of the second probability of failure.

8. The method of claim 6 , wherein predicting the source of the second probability of failure comprises predicting the source of the second probability of failure in response to the second probability of failure exceeding the threshold probability.

9. The method of claim 6 , further comprising, during a second time period:

accessing a second set of timeseries telemetry data broadcast by a first satellite;

predicting a third probability of failure of the first satellite within the second time window based on the first model and the second set of timeseries telemetry data;

predicting a fourth probability of failure of the first satellite within the fourth time window based on the second model and the second set of timeseries telemetry data; and

in response to the third probability of failure exceeding the threshold probability and the fourth probability of failure exceeding the threshold probability:

accessing a set of historical telemetry data, in a set of telemetry domains, for the first satellite;

grouping the set of historical telemetry data into a set of historical telemetry containers, each telemetry container associated with a discrete historical time window and containing values representing the set of telemetry domains;

populating a current telemetry container with the second set of timeseries telemetry data;

for each telemetry domain in the set of telemetry domains, calculating an anomaly score for the telemetry domain in the current telemetry container based on distances between the current telemetry container and the set of historical telemetry containers in a feature space; and

within a user portal, ranking the set of telemetry domains for investigation at the first satellite based on anomaly scores of telemetry domains in the current telemetry container.

10. The method of claim 9 , wherein ranking the set of telemetry domains for investigation at the first satellite comprises ranking the set of telemetry domains based on Shapley values associated with each of the telemetry domains specifying a degree of influence of a feature of the first satellite contributing to the failure of the first satellite.

11. The method of claim 6 :

further comprising, defining an anchor time specifying a start point for the first and second time windows based on a probability of failure exceeding the threshold for a duration of time; and

wherein accessing the first set of timeseries telemetry data broadcast by the first satellite comprises accessing the first set of timeseries telemetry data broadcast by the first satellite based on the anchor time.