Patent Issued for Adaptable on-deployment learning platform for driver analysis output generation (USPTO 11620494): Allstate Insurance Company

2023 APR 26 (NewsRx) -- By a News Reporter-Staff News Editor at Insurance Daily News -- A patent by the inventors Aragon, Juan Carlos (Redwood City, CA, US), Madigan, Regina (Mountain View, CA, US), filed on September 24, 2019, was published online on April 4, 2023, according to news reporting originating from Alexandria, Virginia, by NewsRx correspondents.

Patent number 11620494 is assigned to Allstate Insurance Company (Northbrook, Illinois, United States).

The following quote was obtained by the news editors from the background information supplied by the inventors: “Aspects of the disclosure relate to enhanced processing systems for providing dynamic driving metric outputs using improved machine learning methods. In particular, one or more aspects of the disclosure relate to sensor data that may be used to generate driving metrics.

“Many organizations and individuals rely on driving metrics, such as a driver’s skill level, in performing driver safety and performance. In many instances, sensor data may be analyzed to determine driving metrics such as the output score of a predictive risk model, the output score of a driver evaluation model, and others. However, this situation may present limitations because particular driving factors that reflect particular characteristics of the individual driver may be ignored in the driving analysis due to their complexity or because they are difficult to obtain. There remains an ever present need to develop alternative solutions for determining driving metrics using sensor data that account for the individual driver’s characteristics so that the outputs of the solutions are customized to the individual driver.”

In addition to the background information obtained for this patent, NewsRx journalists also obtained the inventors’ summary information for this patent: “Aspects of the disclosure provide effective, efficient, scalable, and convenient technical solutions that address and overcome the limitations of contemporary driving analysis methods. In accordance with one or more embodiments of the disclosure, a computing platform comprising at least one processor, a communication interface, and memory storing computer readable instructions may receive, from one or more vehicle sensors, sensor data inputs. For one or more of the sensor data inputs, the computing platform may generate a pattern deviation output corresponding to an output of a sensor data analysis model, an actual outcome associated with a lowest time to collision (TTC) value, and a sequence of driving actions that occurred over a prediction horizon corresponding to the pattern deviation output. The computing platform may cluster the pattern deviation outputs to maximize a ratio of inter-cluster variance to intra-cluster variance. The computing platform may train a long short term memory (LSTM) for each cluster. Based on the LSTM for each cluster, the computing platform may verify consistency of the pattern deviation outputs in each cluster. After verifying the consistency of the pattern deviation outputs in each cluster, the computing platform may modify the sensor data analysis model to reflect pattern deviation outputs associated with verified consistency.

“In one or more instances, the prediction horizon may be a duration over which collision prediction is estimated. In one or more instances, the computing platform may determine an error value corresponding to a difference between an estimated TTC value and the lowest TTC value for each of the sensor data inputs. The computing platform may group each of the sensor data inputs based on whether the estimated TTC value exceeds a predetermined TTC threshold. For each group, the computing platform may rank each of the sensor data inputs based on their corresponding error values. The computing platform may determine, for each group and by determining that the error values corresponding to each subset of the sensor data inputs exceeds a predetermined error threshold, a subset of the sensor data inputs.

“In one or more instances, each group may be associated with a different predetermined error threshold. In one or more instances, the one or more of the sensor data inputs may correspond to the subsets of the sensor data inputs.

“In one or more instances, the computing platform may predict, for each cluster (e.g., clusters that may correspond to pattern deviations) and using the trained LSTMs, a subsequent driver action. In one or more instances, the LSTMs may be trained on a subset of the pattern deviation outputs for each cluster based on similarity of their sequences of driving actions, and the similarity of their sequence of driving actions may be determined based on Euclidian distance.

“In one or more instances, after training the LSTMs on the subset of the pattern deviation outputs for each cluster, the computing platform may apply the LSTMs to the remaining pattern deviation outputs in their respective clusters to determine additional error values. The computing platform may incorporate a first subset of the remaining pattern deviation outputs into future LSTM training after determining that additional error values corresponding to the first subset of the remaining pattern deviation outputs are below a predefined LSTM training threshold. The computing platform may discard a second subset of the remaining pattern deviation outputs after determining that additional error values corresponding to the second subset of the remaining pattern deviation outputs exceed the predefined LSTM training threshold.”

The claims supplied by the inventors are:

“1. A computing platform, comprising: at least one processor; a communication interface; and memory storing computer-readable instructions that, when executed by the at least one processor, cause the computing platform to: receive, from one or more vehicle sensors, sensor data inputs; generate a risk model that outputs an assessment of risk associated with a given driver experiences based on the sensor data inputs associated with driving; determine a lowest time-to-collision (TTC) value based on the sensor data inputs; determine an error between an estimated TTC event from the risk model and the lowest TTC value observed over a period of time, wherein the error is associated with a pattern deviation output from the risk model that predicted the estimated TTC event; generate, for one or more of the sensor data inputs, a sequence of driving actions that occurred over a prediction horizon corresponding to the pattern deviation output; cluster pattern deviation outputs including the pattern deviation output to maximize a ratio of inter-cluster variance to intra-cluster variance; train a long short term memory (LSTM) for each cluster on a subset of the pattern deviation outputs for each cluster based on similarity of their sequences of driving actions, and wherein the similarity of their sequence of driving actions is determined based on Euclidian distance; and modify the risk model to reflect pattern deviation outputs.

“2. The computing platform of claim 1, wherein the prediction horizon comprises a duration over which collision prediction is estimated.

“3. The computing platform of claim 1, wherein the memory stores additional computer-readable instructions that, when executed by the at least one processor, cause the computing platform to: group each of the sensor data inputs based on whether the estimated TTC value exceeds a predetermined TTC threshold; rank, for each group, each of the sensor data inputs based on their corresponding error values; and determine, for each group, a subset of the sensor data inputs, wherein determining the subset of the sensor data inputs comprises determining that the error values corresponding to each subset of the sensor data inputs exceeds a predetermined error threshold.

“4. The computing platform of claim 3, wherein each group is associated with a different predetermined error threshold.

“5. The computing platform of claim 3, wherein the one or more of the sensor data inputs correspond to the subsets of the sensor data inputs.

“6. The computing platform of claim 1, wherein the memory stores additional computer-readable instructions that, when executed by the at least one processor, cause the computing platform to predict, for each cluster and using the trained LSTMs, a subsequent driver action.

“7. The computing platform of claim 1, wherein the memory stores additional computer-readable instructions that, when executed by the at least one processor, cause the computing platform to: after training the LSTMs on the subset of the pattern deviation outputs for each cluster, apply the LSTMs to remaining pattern deviation outputs in their respective clusters to determine additional error values; incorporate a first subset of the remaining pattern deviation outputs into future LSTM training after determining that additional error values corresponding to the first subset of the remaining pattern deviation outputs are below a predefined LSTM training threshold; and discard a second subset of the remaining pattern deviation outputs after determining that additional error values corresponding to the second subset of the remaining pattern deviation outputs exceed the predefined LSTM training threshold.

“8. The computing platform of claim 1, wherein training the LSTM for each of the clusters comprises: performing training over a beginning portion of the sequence of driving actions; performing training, after performing the training over the beginning portion of the sequence of driving actions and before performing training over a final portion of the sequence of driving actions, over one or more middle portions of the sequence of driving actions, wherein the middle portions of the sequence of driving actions occur after the beginning portion of the sequence of driving actions and before the final portion of the sequence of driving actions, and wherein the training over the one or more middle portions of the sequence of driving actions occurs one middle portion at a time; and performing training, after performing the training over the middle portions of the sequence of driving actions, over the final portion of the sequence of driving actions, wherein the final portion of the sequence of driving actions occurs after the middle portions of the sequence of driving actions.

“9. The computing platform of claim 8, wherein the memory stores additional computer-readable instructions that, when executed by the at least one processor, cause the computing platform to: determine, after training the LSTM for each of the clusters, clusters associated with different sequence of driving actions but with shared sensor data inputs; and store an indication of the determined clusters.

“10. The computing platform of claim 1, wherein modifying the risk model comprises incorporating the pattern deviation outputs into training of the risk model.

“11. The computing platform of claim 1, wherein the memory stores additional computer-readable instructions that, when executed by the at least one processor, cause the computing platform to: send driving analysis output information corresponding to analysis of the sensor data inputs, wherein sending the driving analysis output information causes generation of a driving analysis output corresponding to feedback on a user’s driving ability.

“12. A method comprising: at a computing platform comprising at least one processor, a communication interface, and memory: receiving, from one or more vehicle sensors, sensor data inputs; generating a risk model that outputs an assessment of risk associated with a given driver experiences based on the sensor data inputs associated with driving; determining a lowest time-to-collision (TTC) value based on the sensor data inputs; determining an error between an estimated TTC event from the risk model and the lowest TTC value observed over a period of time, wherein the error is associated with a pattern deviation output from the risk model that predicted the estimated TTC event; generating, for one or more of the sensor data inputs, a sequence of driving actions that occurred over a prediction horizon corresponding to the pattern deviation output; clustering pattern deviation outputs including the pattern deviation output to maximize a ratio of inter-cluster variance to intra-cluster variance; training a long short term memory (LSTM) for each cluster on a subset of the pattern deviation outputs for each cluster based on similarity of their sequences of driving actions, and wherein the similarity of their sequence of driving actions is determined based on Euclidian distance; and modifying the risk model to reflect pattern deviation outputs.

“13. The method of claim 12, wherein the prediction horizon comprises a duration over which collision prediction is estimated.

“14. The method of claim 12, further comprising: grouping each of the sensor data inputs based on whether the estimated TTC value exceeds a predetermined TTC threshold; ranking, for each group, each of the sensor data inputs based on their corresponding error values; and determining, for each group, a subset of the sensor data inputs, wherein determining the subset of the sensor data inputs comprises determining that the error values corresponding to each subset of the sensor data inputs exceeds a predetermined error threshold.

“15. The method of claim 14, wherein each group is associated with a different predetermined error threshold.

“16. The method of claim 14, wherein the one or more of the sensor data inputs correspond to the subsets of the sensor data inputs.

“17. The method of claim 12, further comprising predicting, for each cluster and using the trained LSTMs, a subsequent driver action.

“18. One or more non-transitory computer-readable media storing instructions that, when executed by a computing platform comprising at least one processor, a communication interface, and memory, cause the computing platform to: receive, from one or more vehicle sensors, sensor data inputs; generate a risk model that outputs an assessment of risk associated with a given driver experiences based on the sensor data inputs associated with driving; determine a lowest time-to-collision (TTC) value based on the sensor data inputs; determine an error between an estimated TTC event from the risk model and the lowest TTC value observed over a period of time, wherein the error is associated with a pattern deviation output from the risk model that predicted the estimated TTC event; generate, for one or more of the sensor data inputs, a sequence of driving actions that occurred over a prediction horizon corresponding to the pattern deviation output; cluster pattern deviation outputs including the pattern deviation output to maximize a ratio of inter-cluster variance to intra-cluster variance; train a long short term memory (LSTM) for each cluster on a subset of the pattern deviation outputs for each cluster based on similarity of their sequences of driving actions, and wherein the similarity of their sequence of driving actions is determined based on Euclidian distance; and modify the risk model to reflect pattern deviation outputs.”

URL and more information on this patent, see: Aragon, Juan Carlos. Adaptable on-deployment learning platform for driver analysis output generation. U.S. Patent Number 11620494, filed September 24, 2019, and published online on April 4, 2023. Patent URL (for desktop use only): https://ppubs.uspto.gov/pubwebapp/external.html?q=(11620494)&db=USPAT&type=ids

(Our reports deliver fact-based news of research and discoveries from around the world.)