DETECTING RF EMISSIONS AND CLASSIFYING RF SOURCES USING SPECTRUM TRANSFORMER MODELS

1 . An edge computing unit comprising:

a signal receiver;

at least one server rack;

at least one power unit;

at least one environmental control system; and

at least one isolation system,

wherein the at least one server rack is programmed with one or more sets of instructions that, when executed by the at least one server rack, cause the edge computing unit to perform operations comprising:

capturing a first plurality of RF signals by the signal receiver over a first period of time;

transforming first data from a time domain to a frequency domain, wherein the first data represents the first plurality of RF signals;

generating a first spectrogram representative of the first plurality of RF signals based at least in part on the transformed first data;

extracting a first set of features from the first spectrogram;

encoding the first set of features with a frequency range associated with the first data;

identifying a plurality of sets of features, wherein each one of the plurality of sets of features is associated with one of a plurality of RF events, and wherein each one of the plurality of RF events comprises at least one transmission of at least one RF signal;

generating, by a spectrum encoder of a transformer model, at least a first embedding based at least in part on the first set of features and the plurality of sets of features, wherein generating at least the first embedding comprises performing attention on the first set of features and each one of the plurality of sets of features;

receiving, from a computer device, a query for information regarding at least one of the first plurality of RF signals or the first period of time;

generating, by a text encoder, at least a second embedding based at least in part on the query for information;

identifying at least a third embedding, wherein the third embedding was generated by encoding at least one document associated with one of RF signals or sources of RF signals;

generating, by a spectrum decoder of the transformer model, at least one response to the query for information, wherein generating the at least one response comprises performing attention on the first embedding, the second embedding, and the third embedding, and wherein the at least one response comprises a set of words regarding the at least one of the first plurality of RF signals or a source of the at least one of the first plurality of RF signals;

transmitting second data representing the set of words to the computer device; and

causing the computer device to display or play aloud the set of words based at least in part on the second data.

2 . The edge computing unit of claim 1 , wherein the signal receiver comprises:

at least one superconductor chip, wherein the at least one superconductor chip is cooled to below twenty degrees Kelvin; and

at least one post-processing module comprising a field programmable gate array.

3 . The edge computing unit of claim 1 , wherein the set of words indicates that at least one of the first plurality of RF signals is associated with one of the plurality of RF events.

4 . A method comprising:

capturing first data regarding at least a first transmission of RF signals over a period of time, wherein the first data is captured in the time domain by a first receiver;

transforming the first data regarding the first transmission of the RF signals to a frequency domain;

generating at least a first two-dimensional representation based at least in part on the first data transformed to the frequency domain;

extracting at least a first set of features from the first two-dimensional representation;

encoding at least the first set of features with second data representing at least one range of frequencies associated with the first data;

identifying a first plurality of sets of features, wherein each one of the first plurality of sets of features was extracted from one of a plurality of two-dimensional representations generated based at least in part on data regarding RF signals associated with one of a plurality of RF events transformed to the frequency domain;

generating at least a first embedding based at least in part on the first set of features encoded with the second data and at least some of the first plurality of sets of features; and

identifying at least one attribute of the first transmission of the RF signals based at least in part on the first embedding.

5 . The method of claim 4 , wherein the first receiver is a component of an edge computing unit comprising a containerized system having:

at least one server rack;

at least one power unit;

at least one environmental control system; and

at least one isolation system.

6 . The method of claim 4 , wherein the first receiver comprises:

at least one superconductor chip, wherein the at least one superconductor chip is cooled to below twenty degrees Kelvin; and

at least one post-processing module comprising a field programmable gate array, and

wherein the edge computing unit further comprises one or more rack-mounted redundant arrays of independent disks.

7 . The method of claim 4 , wherein generating at least the first two-dimensional representation based at least in part on the first data comprises:

partitioning the first two-dimensional representation into a plurality of tiles, wherein each one of the plurality of tiles is a portion of the first two-dimensional representation,

wherein extracting at least the first set of features from the first two-dimensional representation comprises:

extracting a second plurality of sets of features from the plurality of tiles, wherein each one of the second plurality of sets of features is extracted from one of the plurality of tiles, and

wherein encoding at least the first set of features with the second data comprises:

encoding each one of the second plurality of sets of features with data representing a range of features corresponding to the one of the plurality of tiles from which the one of the second plurality of features was extracted.

8 . The method of claim 4 , wherein the at least one attribute of the first transmission of the RF signals comprises:

a frequency associated with the first transmission of the RF signals;

an identifier of a source of the first transmission of the RF signals;

an identifier of an entity associated with the source of the first transmission of the RF signals;

an intensity of the first transmission of the RF signals;

a location of the first transmission of the RF signals; or

a purpose of the first transmission of the RF signals.

9 . The method of claim 8 , wherein the at least one attribute of the first transmission of the RF signals is one of a frequency associated with the first transmission of the RF signals, an intensity of the first transmission of the RF signals, an identifier of a source of the first transmission of the RF signals, or a location of the first transmission of the RF signals, and

wherein the method further comprises:

classifying the first transmission of the RF signals as at least a portion of one of the plurality of RF events based at least in part on the at least one of the frequency associated with the first transmission of the RF signals, the intensity of the first transmission of the RF signals, the source of the first transmission of the RF signals or the location of the first transmission of the RF signals; and

storing an indication that the first transmission of the RF signals is at least a portion of the one of the plurality of RF events in at least one data store.

10 . The method of claim 8 , wherein the at least one attribute of the first transmission of the RF signals is one of a frequency associated with the first transmission of the RF signals, an intensity of the first transmission of the RF signals, an identifier of a source of the first transmission of the RF signals, or a location of the first transmission of the RF signals, and wherein the method further comprises:

determining that the first transmission of the RF signals is not one of the plurality of RF events based at least in part on the at least one of the frequency associated with the first transmission of the RF signals, the intensity of the first transmission of the RF signals, the source of the first transmission of the RF signals or the location of the first transmission of the RF signals; and

storing an indication that the first transmission of the RF signals is not one of the plurality of RF events in at least one data store.

11 . The method of claim 8 , further comprising:

receiving at least a second embedding over one or more networks, wherein the second embedding is generated based at least in part on a second set of features extracted from a second two-dimensional representation generated based at least in part on third data regarding the first transmission of RF signals transformed to the frequency domain and at least some of a second plurality of sets of features extracted from one of the plurality of two-dimensional representations;

determining that each of the first embedding and the second embedding was generated based at least in part on the first transmission of RF signals at a common time; and

determining, based at least in part on the first embedding and the second embedding, the at least one attribute of the first transmission of the RF signals, wherein the at least one attribute is a location of the source at the common time.

12 . The method of claim 4 , wherein the first embedding is generated by a spectrum encoder of a transformer model,

wherein the spectrum encoder comprises:

a self-attention module comprising a multi-head attention feature, at least one normalization layer, and at least one feedforward network, wherein the self-attention module is configured to perform attention on at least the first set of features encoded with the second data; and

a multi-modal attention module configured to perform attention on at least the attention-refined first set of features encoded with the second data and each of the first plurality of sets of features,

wherein the first embedding is generated based at least in part on an output of the multi-modal attention module.

13 . The method of claim 4 , wherein identifying the at least one attribute of the first transmission of the RF signals based at least in part on the first embedding comprises:

receiving a query for information;

generating at least a second set of features based at least in part on the query for information;

encoding the second set of features with third data representing at least one range of frequencies associated with the query for information;

identifying a second plurality of sets of features, wherein each one of the second plurality of sets of features was generated based on semantic information regarding RF signals or sources of RF signals; and

decoding the first embedding based at least in part on the second set of features and the second plurality of sets of features to generate a response to the query for information,

wherein the response to the query for information identifies the at least one attribute.

14 . The method of claim 13 , wherein the first embedding is decoded by a spectrum decoder of a transformer model,

wherein the spectrum decoder comprises:

a self-attention module comprising a multi-head attention feature, at least one normalization layer, and at least one feedforward network, wherein the self-attention module is configured to perform attention on at least the second set of features encoded with the third data; and

a multi-modal attention module configured to perform attention on at least the attention-refined second set of features encoded with the third data, the first embedding, and each of the second plurality of sets of features,

wherein the response to the query for information is generated based at least in part on an output of the multi-modal attention module.

15 . The method of claim 13 , wherein the semantic information describes at least one of:

a frequency of an RF signal transmitted prior to the period of time;

an intensity of the RF signal;

a location of a transmission of the RF signal; or

an identifier of a source of the RF signal, and

wherein each one of the second plurality of sets of features was generated by providing the semantic information to a text encoder.

16 . The method of claim 4 , wherein the first two-dimensional representation is a first spectrogram representing at least the first transmission of RF signals, and

wherein each one of the plurality of two-dimensional representations is a spectrogram generated based at least in part on the data regarding the RF signals associated with the one of the plurality of RF events transformed to the frequency domain.

17 . An edge computing unit comprising:

a signal receiver; and

at least one server rack comprising at least one processor unit and at least one data store,

wherein the at least one server rack is programmed with one or more sets of instructions that, when executed, cause the edge computing unit to perform operations comprising:

receiving a query of information from a first computer system over one or more networks;

transforming first data representing a plurality of RF signals captured by the signal receiver to a frequency domain;

generating a first spectrogram based at least in part on the transformed first data;

extracting a first set of features from the first spectrogram;

identifying second data representing a plurality of RF events, wherein each one of the plurality of RF events includes a transmission of at least one RF signal;

extracting a first plurality of sets of features from the second data, wherein each one of the first plurality of sets of features represents one of the plurality of RF events;

executing a spectrum encoder to generate a first plurality of embeddings, wherein each one of the first plurality of embeddings is generated based at least in part on the first set of features and one of the first plurality of sets of features;

generating at least one embedding based at least in part on the query for information;

identifying a second plurality of embeddings, wherein each one of the second plurality of embeddings represents at least a portion of one of a plurality of resources including semantic information regarding RF signals, sources of RF signals or RF events including transmissions of RF signals;

executing a spectrum decoder to generate a response to the query for information based at least in part on the first plurality of embeddings, the at least one embedding generated based at least in part on the query for information, and the second plurality of embeddings; and

providing a set of words generated based at least in part on the output to one of the first computer system or a second computer system.

18 . The edge computing unit of claim 17 , wherein extracting the first set of features from the first spectrogram comprises:

partitioning the first spectrogram into a plurality of tiles, wherein each one of the plurality of tiles is a portion of the first spectrogram; and

extracting the first plurality of sets of features from the plurality of tiles, wherein each one of the first plurality of sets of features is extracted from one of the plurality of tiles.

19 . The edge computing unit of claim 17 , wherein the signal receiver comprises:

at least one superconductor chip, wherein the at least one superconductor chip is cooled to below twenty degrees Kelvin; and

at least one post-processing module comprising a field programmable gate array.

20 . The edge computing unit of claim 17 , wherein the spectrum encoder comprises:

a first self-attention module comprising a first multi-head attention feature, at least a first normalization layer, and at least a first feedforward network, wherein the first self-attention module is configured to perform attention on at least the first set of features encoded with data regarding a frequency range of the first spectrogram; and

a first multi-modal attention module configured to perform attention on at least the attention-refined first set of features encoded with the data regarding the frequency range of the first spectrogram and each of the first plurality of sets of features, and wherein the spectrum decoder comprises:

a second self-attention module comprising a second multi-head attention feature, at least a second normalization layer, and at least a second feedforward network, wherein the second self-attention module is configured to perform attention on the at least one embedding generated based at least in part on the query for information encoded with the data regarding the frequency range of the first spectrogram; and

a second multi-modal attention module configured to perform attention on at least the attention-refined at least one embedding encoded with the data regarding the frequency range of the first spectrogram, the first plurality of embeddings, and the second plurality of embeddings.