Deep Learning Accelerator for Vehicle Sensor Data Compression
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Vehicle systems lack the capability to perform real-time analysis of sensor data due to latency constraints, preventing them from processing information collected during operation.
Innovation Solution
Integration of a deep learning accelerator (DLA) within the zonal computing system to generate run-time analytics and compress sensor data, reducing latency and increasing storage efficiency by performing analytics locally rather than relying on cloud processing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If sensor data is processed and analyzed in the cloud, then comprehensive analytics can be generated, but latency increases and real-time processing capability is lost
Solution Approach 1:
The patent segments the data processing workflow into two distinct parts: (1) real-time processing at the edge device using lightweight machine learning models for immediate analytics, and (2) comprehensive post-processing in the cloud using full-scale models for detailed analysis. This segmentation allows the system to achieve both low latency for time-critical operations and high accuracy for comprehensive analytics by assigning appropriate processing tasks to each location.
Solution Approach 2:
The patent introduces an edge computing device as an intermediary between the vehicle sensors and the cloud infrastructure. This intermediary performs preliminary processing, filtering, and lightweight analytics locally, then selectively transmits only relevant data to the cloud. This intermediary role resolves the contradiction by handling time-sensitive operations locally while maintaining the cloud's capability for comprehensive analysis.
2Loss of information
If all sensor data is transmitted and stored, then complete information is available for analysis, but storage requirements and data transmission bandwidth increase significantly
Solution Approach 1:
The patent extracts and processes only the most relevant features and insights from the raw sensor data at the edge device using machine learning models. Instead of transmitting and storing all raw data, the system extracts key analytics, anomalies, and condensed information that capture the essential content. This extraction approach maintains data completeness in terms of informational value while dramatically reducing the quantity of data that needs to be stored and transmitted.
Solution Approach 2:
The patent transforms raw sensor data into processed analytics and derived parameters through machine learning processing at the edge. By changing the data representation from raw high-volume sensor readings to compressed analytics with lower dimensionality, the system preserves the essential information content while reducing storage requirements. The transformed data maintains its analytical value but occupies significantly less storage space.
3Speed
If lightweight machine learning models are used at the edge, then real-time processing is achieved, but model accuracy may be reduced compared to full-scale models
Solution Approach 1:
The patent segments the machine learning modeling into two tiers: lightweight models deployed at the edge device for real-time processing with acceptable accuracy, and full-scale high-accuracy models running in the cloud for comprehensive analysis. This segmentation allows each model type to be optimized for its specific purpose, resolving the accuracy-speed tradeoff by assigning appropriate models to appropriate tasks.
Solution Approach 2:
The patent applies preliminary processing and feature extraction using lightweight models at the edge before data is sent to the cloud. This preliminary action prepares the data in a way that enhances the effectiveness of subsequent cloud-based full-scale model processing. The preliminary filtering and feature extraction by lightweight models enables the cloud models to focus on more complex analysis, achieving both speed and accuracy across the complete processing pipeline.
Data Source
AI summary
Methods, systems, and devices for management of vehicle system information using a deep learning device are described. The deep learning device of a vehicle (such as a deep learning accelerator (DLA)) may receive information associated with an environment of the vehicle from one or more sensors of the vehicle. The DLA may perform one or more operations using one or more machine learning models. For example, the DLA may compress the information which may reduce a resolution associated with the information, a frame rate associated with the information, or both. The DLA may generate, as part of a run-time operation, a first set of analytics associated with operation of the vehicle using the compressed information. Additionally, or alternatively, the DLA may generate, as part of a post-processing operation, a second set of analytics using the compressed or an uncompressed version of the information.