# US 12,602,622 B2 · Method and device for training and predicting a conjunction parameter from conjunction data messages

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 12,602,622 B2
- Document ID: US12602622B2
- Status: Granted
- Kind: B2
- Application number: 19/027,530
- Filed: 2025-01-17
- Published: 2025-07-24
- Granted: 2026-04-14
- Earliest claimed: 2024-01-19
- Assignees: Neuraspace, S.A.
- Applicants: Neuraspace, S.A.
- Inventors: Marta Isabel Vaz Guimarães Pinto De Melo
- Web page: https://patentgrep.com/patent/US12602622B2

## CPC

- G06N 3/045: Combinations of networks
- G06N 20/00: Machine learning
- B64G 3/00: Observing or tracking cosmonautic vehicles
- G06N 3/08: Learning methods
- G06N 3/084: Backpropagation, e.g. using gradient descent
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- B64G: COSMONAUTICS; VEHICLES OR EQUIPMENT THEREFOR

## Abstract

The present document discloses a computer-implemented method for predicting a conjunction parameter from conjunction data messages and the respective method for training a transformer-based machine-learning model for predicting a conjunction parameter. It is also disclosed a device comprising a computer-readable medium comprising the trained transformer-based machine-learning model and a computer configured to carry out the training method of the transformer-based machine-learning model.

## Claims

### Claim 1

A computer-implemented method for training a transformer-based machine-learning model for predicting a conjunction parameter from conjunction data messages, CDMs, wherein a conjunction is a pair of orbiting objects, the method comprising:
receiving a dataset comprising a conjunction time series for each conjunction, wherein each said series is irregularly spaced and comprises a plurality of CDM records, each CDM record comprising a time to TCA, Time of Closest Approach, and a conjunction parameter;
generating a CDM array comprising a plurality of cells, one cell for each CDM record, organized in rows, one row for each conjunction time series, and a plurality of columns, where each row is right-aligned such that a rightmost column of the CDM array comprises the most recent CDM record for each conjunction time series;
generating a mask for the CDM array where each CDM record is masked if the time to TCA of said each CDM record is lower than a predetermined time to TCA threshold;
encoding each CDM record of the CDM array by embedding the conjunction parameter into an embedded encoding using an artificial neural-network;
training a transformer-based machine-learning model by using, as training input, the CDM array masked by said mask and using, as training output, the CDM array, wherein each said CDM record is a token for transformer-based machine-learning model,
wherein the conjunction parameter is a spatial conjunction parameter which comprises a position uncertainty at the TCA.

### Claim 2 (depends on 1)

The method according to claim 1, wherein the encoding further comprises:
positionally encoding the time to TCA; and
combining the embedded encoding of the conjunction parameter with the positionally encoding of the time to TCA by summing the embedded encoding of the conjunction parameter with the positionally encoding of the time to TCA,
wherein each said encoded CDM record is a token for transformer-based machine-learning model.

### Claim 3 (depends on 1)

The method according to claim 1, wherein the embedding of the conjunction parameter into an embedded encoding is carried out by a single linear layer artificial neural-network.

### Claim 4 (depends on 2)

The method according to claim 2, wherein the positionally encoding of the time to TCA is carried out by generating a vector comprising values generated as a function of time to TCA.

### Claim 5 (depends on 4)

The method according to claim 4, wherein the positionally encoding of the time to TCA is carried out by generating a vector comprising values generated as a function of time to TCA defined by:
PE
time
,
2
⁢
i
=
sin
(
time
max
⁢
time
2
⁢
i
d
model
)
PE
time
,
2
⁢
i
+
1
=
cos
(
time
max
⁢
time
2
⁢
i
d
model
)
where time is the time to TCA to be positionally encoded, maxtime is a scaling parameter, i is the dimension, and dmodel is the representation dimension, in particular maxtime is the maximum time to TCA received in the dataset.

### Claim 6 (depends on 1)

The method according to claim 1, further comprising a preparatory step of receiving said predetermined time to TCA threshold.

### Claim 7 (depends on 1)

The method according to claim 1, wherein the spatial conjunction parameter comprises the position uncertainty at the time of closest approach and at least one other CDM-related parameter.

### Claim 8 (depends on 1)

The method according to claim 1, wherein the CDM array has at least one of a number of columns equal to a highest number of CDM records for any conjunction time series and a number of rows equal to a number of received conjunction time series.

### Claim 9 (depends on 1)

The method according to claim 1, wherein each masked CDM record of the CDM array is padded with a value or string indicative of masking or padded with a null value or padded with a zero value, and
wherein empty cells of the CDM array are padded with a value or string indicative of padding or padded with a null value or padded with a zero value.

### Claim 10 (depends on 1)

The method according to claim 1, further comprising deleting any row for a conjunction time series of the CDM array if a number of CDM records is lower than a predetermined minimum CDM number or if a number of unmasked CDM records is lower than a predetermined minimum unmasked CDM number.

### Claim 11 (depends on 1)

The method according to claim 1, further comprising applying a self-attention layer within the transformer encoder and decoder to capture contextual dependencies from input data; and applying a multi-head attention layer within the transformer encoder and decoder to capture contextual dependencies within the input data.

### Claim 12 (depends on 1)

A computer-implemented method for predicting a conjunction parameter from CDMs using the transformer-based machine-learning model trained according to claim 1, the method comprising:
feeding as input to the trained transformer-based machine-learning model a conjunction time series for each of the one or more conjunctions to be predicted, wherein each said series comprises a plurality of received CDM records, each received CDM record comprising a time to TCA, Time to Closest Approach, and a conjunction parameter, and a plurality of prediction CDM records to be predicted, each prediction CDM record comprising a periodic time stamp with a predetermined period and a masked conjunction parameter;
outputting predicted CDM records from the output of trained transformer-based machine-learning model,
wherein the conjunction parameter is a spatial conjunction parameter which comprises a position uncertainty at the time of closest approach.

### Claim 13 (depends on 12)

The computer implemented method according to claim 12, wherein the predicted CDM records are sent as CDMs.

### Claim 14

A device comprising a non-transitory computer-readable medium, the non-transitory computer-readable medium having stored thereon instructions that, when executed by the device, cause the device to train a transformer-based machine-learning model for predicting a conjunction parameter from conjunction data messages, CDMs, wherein a conjunction is a pair of orbiting objects, by:
receiving a dataset comprising a conjunction time series for each conjunction, wherein each said series is irregularly spaced and comprises a plurality of CDM records, each CDM record comprising a time to TCA, Time of Closest Approach, and a conjunction parameter;
generating a CDM array comprising a plurality of cells, one cell for each CDM record, organized in rows, one row for each conjunction time series, and a plurality of columns, where each row is right-aligned such that a rightmost column of the CDM array comprises the most recent CDM record for each conjunction time series;
generating a mask for the generated CDM array where each CDM record is masked if the time to TCA of said each CDM record is lower than a predetermined time to TCA threshold;
encoding each CDM record of the CDM array by embedding the conjunction parameter into an embedded encoding using an artificial neural-network;
training a transformer-based machine-learning model by using, as training input, the generated CDM array masked by said mask and using, as training output, the generated CDM array,
wherein each said CDM record is a token for transformer-based machine-learning model,
wherein the conjunction parameter is a spatial conjunction parameter which comprises a position uncertainty at the TCA.

### Claim 15

A non-transitory computer-readable medium comprising computer program instructions that, when executed by a computer, cause the computer to carry out the steps of:
receiving a dataset comprising a conjunction time series for each conjunction, wherein each said series is irregularly spaced and comprises a plurality of CDM records, each CDM record comprising a time to TCA, Time of Closest Approach, and a conjunction parameter;
generating a CDM array comprising a plurality of cells, one cell for each CDM record, organized in rows, one row for each conjunction time series, and a plurality of columns, where each row is right-aligned such that a rightmost column of the CDM array comprises the most recent CDM record for each conjunction time series;
generating a mask for the generated CDM array where each CDM record is masked if the time to TCA of said each CDM record is lower than a predetermined time to TCA threshold;
encoding each CDM record of the CDM array by embedding the conjunction parameter into an embedded encoding using an artificial neural-network;
training a transformer-based machine-learning model by using, as training input, the generated CDM array masked by said mask and using, as training output, the generated CDM array,
wherein each said CDM record is a token for transformer-based machine-learning model,
wherein the conjunction parameter is a spatial conjunction parameter which comprises a position uncertainty at the TCA.

### Claim 16 (depends on 14)

The device according to claim 14, wherein the non-transitory computer-readable medium has stored thereon further instructions that, when executed by the device, cause the device to predict a conjunction parameter from CDMs using the trained transformer-based machine-learning model by:
feeding as input to the trained transformer-based machine-learning model a conjunction time series for each of the one or more conjunctions to be predicted, wherein each said series comprises a plurality of received CDM records, each received CDM record comprising a time to TCA, Time to Closest Approach, and a conjunction parameter, and a plurality of prediction CDM records to be predicted, each prediction CDM record comprising a periodic time stamp with a predetermined period and a masked conjunction parameter;
outputting predicted CDM records from the output of trained transformer-based machine-learning model,
wherein the conjunction parameter is a spatial conjunction parameter which comprises a position uncertainty at the time of closest approach.

### Claim 17 (depends on 15)

The non-transitory computer-readable medium according to claim 15, further comprising computer program instructions that, when executed by a computer, cause the computer to predict a conjunction parameter from CDMs using the trained transformer-based machine-learning model by:
feeding as input to the trained transformer-based machine-learning model a conjunction time series for each of the one or more conjunctions to be predicted, wherein each said series comprises a plurality of received CDM records, each received CDM record comprising a time to TCA, Time to Closest Approach, and a conjunction parameter, and a plurality of prediction CDM records to be predicted, each prediction CDM record comprising a periodic time stamp with a predetermined period and a masked conjunction parameter;
outputting predicted CDM records from the output of trained transformer-based machine-learning model,
wherein the conjunction parameter is a spatial conjunction parameter which comprises a position uncertainty at the time of closest approach.