Emissions Model Feature Selection via Low-Pass Filtering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional emissions models for internal combustion engines require excessive computing time and power due to large feature matrices generated by high sampling frequencies and long histories, often incorporating unnecessary precise measurement values that increase noise and sensitivity, especially under non-stationary conditions.
Innovation Solution
The use of low-pass filters to aggregate past information within measurement series, allowing for automatic feature selection to determine the best combination of features for improved model quality and reduced calculation time, focusing on the order of magnitude rather than exact measurement values, and employing different time constants for filtering to concentrate information and reduce the need for exhaustive searches.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If high sampling frequency and long history are used to improve model prediction quality, then the feature matrix becomes very large, but computing time and computing power increase excessively
Solution Approach 1:
The patent extracts only the essential information from the measurement series by applying low-pass filters that aggregate past values. Instead of using all 1800 columns from high-frequency sampling, the filter extracts the dominant trend information, removing unnecessary detailed fluctuations that do not contribute to prediction quality.
Solution Approach 2:
The patent changes the parameter representation by transforming raw measurement values into filtered values with different time constants. This parameter transformation reduces the effective dimensionality while preserving the essential dynamic behavior needed for accurate predictions.
2Loss of information
If precise measurement values from the past are taken into account, then the feature matrix includes detailed information, but the noise of the model increases and sensitivity rises
Solution Approach 1:
The low-pass filter acts as an intermediary between the raw measurement series and the emissions model. It mediates the information flow by smoothing out high-frequency noise while preserving the essential low-frequency trends, thereby reducing model sensitivity and improving robustness.
Solution Approach 2:
The patent applies parameter transformation through filtering with different time constants (e.g., 1s, 3s, 5s, 10s, 30s). This changes the temporal scale of the input features, allowing the model to capture essential dynamics without being overly sensitive to precise moment-to-moment variations.
3Measurement precision
If all possible combinations of features are tested in model training, then the best model quality is achieved, but the calculation time becomes prohibitively long
Solution Approach 1:
The patent segments the feature selection process into two stages: first, multiple filtered measurement series are pre-computed with different time constants; second, the emissions model is trained on these pre-processed features. This segmentation avoids the need to test all possible combinations of raw features, dramatically reducing computation time.
Solution Approach 2:
The low-pass filtering is performed as a preliminary action before model training. By pre-processing the measurement series with various time constants and selecting the best filtered version, the patent avoids the computationally expensive task of testing all feature combinations during model training.
Data Source
AI summary
A method for creating an emissions model of an internal combustion engine. The method begins with a provision of a plurality of measurement series at an internal combustion engine. There then follows a filtering of the measurement series using various low-pass filters, and ascertaining, from the filtered measurement series, those measurement series, when provided as an input variable for the emissions model during optimization of the emissions model, the smallest deviation from predicted emissions of the emissions model for measured emissions is achieved.