Deep Learning Crash Prediction Using Wavelet Transform
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for predicting vehicle crash performance are costly and labor-intensive, requiring large amounts of high-quality data, and struggle to accurately represent trends related to passenger injury using conventional crash performance indexes.
Innovation Solution
Applying wavelet transform to vehicle crash acceleration data to pre-process and image it, then using the transformed data in a deep learning model to predict frontal crash performance, incorporating features extracted from pre-trained models and crash performance indexes for injury value or grade prediction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional machine learning methods are used to predict crash performance, then a large amount of high-quality data is required, but it is difficult to secure large amounts of data due to the high costs of crash tests
Solution Approach 1:
The patent applies wavelet transform as a pre-processing step to crash acceleration data before feeding it to the machine learning model. This preliminary transformation extracts important features and patterns from the raw data, enabling the model to achieve high prediction accuracy with limited crash test data by leveraging the enhanced information content in the transformed domain
Solution Approach 2:
The patent transforms the crash acceleration data from the time domain to the frequency domain using wavelet transform. This parameter change in the data representation allows the model to capture different characteristics of crash signals that are not apparent in the original time series, improving prediction performance with fewer data samples
2Measurement precision
If conventional crash performance indexes are used, then the prediction process is simpler, but they have a limit in accurately representing trends related to passenger injury
Solution Approach 1:
The patent segments the crash acceleration signal into different frequency components using wavelet transform. This segmentation allows the model to analyze specific frequency ranges that correspond to different injury mechanisms, providing more accurate representation of passenger injury trends while maintaining a manageable model structure through focused feature analysis
Solution Approach 2:
The patent adds a frequency domain dimension to the conventional time-domain crash analysis by applying wavelet transform. This dimensional expansion enables the model to capture multi-scale characteristics of crash signals, improving injury trend representation accuracy without excessively increasing model complexity through the use of established wavelet bases
3Measurement precision
If real crash tests are conducted to identify crash performance, then accurate crash performance data is obtained, but it results in considerable costs and man-hour
Solution Approach 1:
The patent creates a virtual model that replicates crash performance characteristics by training on transformed crash acceleration data. This virtual copy can predict crash performance for new vehicle designs without conducting physical crash tests, significantly improving development efficiency while maintaining accurate crash performance assessment through the wavelet-enhanced learning model
Data Source
AI summary
A method of predicting crash performance based on deep learning from vehicle crash acceleration data includes performing, by an input unit, wavelet transform on vehicle crash acceleration data. The method also includes applying wavelet transform data to a pre-trained model. The method also includes concatenating a feature extracted from the pre-trained model with a vehicle crash performance index. The method also includes calculating a crash prediction result in a crash performance model through artificial neural network learning from concatenated data of the vehicle crash performance index and the extracted feature.