DEF Optimal Condition monitoring
DEF Optimal Condition monitoring
Embedded Files
Evaluation
The effectiveness of the developed algorithm was assessed by measuring the time lag between the initial alert generated by the system and the subsequent appearance of the corresponding fault code. This evaluation method demonstrates the algorithm’s ability to provide early warnings ahead of actual fault occurrences, enabling proactive maintenance and reducing downtime.
Uniqueness
Forecasting as Classification:
Unlike traditional forecasting methods, I approached the problem as a classification task. By deeply understanding the physics of DEF dosing, I realized that prediction could be achieved both through deep learning forecasting models and by using a classifier operating on quantized time series data.The reason for utilizing quantization was to evaluate the system’s effectiveness on quantized time series data. This approach also laid the foundation for the tokenization scheme used in a common time series foundational model framework called PhyFai (Physics-Informed Foundational AI), which we have been developing.
Diesel engines operate with excess air to ensure complete combustion, but this leads to the formation of NOx, which are harmful pollutants. DEF, which is a mixture of about 32.5% urea and 67.5% deionized water, is injected into the hot exhaust gases where it decomposes to produce ammonia (NH3). The ammonia then reacts with NOx, converting these pollutants into harmless nitrogen (N2) and water vapor (H2O), both of which are natural components of the air
Problem
Clogging DEF lines or injectors: Urea crystallization or deposits from degraded DEF can clog DEF lines, filters, or injector nozzles. This blockage restricts DEF flow, resulting in lower than required dosing.
Improper DEF Quality or Dilution: Using DEF that is diluted excessively with water or is of poor quality reduces its effectiveness and can cause dosing faults. DEF exposed to extreme temperatures or stored improperly can degrade, crystallize, or freeze, affecting dosing accuracy.
Solution
Solution
Physics-Informed Feature Engineering: Leveraged Fuel-Rate and Def-Dosing-Rate to come up with a feature engineered value Modified Instantaneous Dosing Ratio (MIDR) - log(Def-Dosing Rate/Fuel- Rate).
Implementation
Data Aggregation: Engine RPM data was segmented into bins, and for each bin, MIDR values were summarized using five statistical descriptors to represent 10-minute intervals of engine operation.
Model Training: A machine learning classification model was trained on these summarized MIDR features, using labeled data that included both healthy and DEF dosing anomaly instances.
Anomaly Detection: The trained classifier was then deployed to analyze incoming 10-minute MIDR summaries, effectively classifying each interval as either normal or anomalous with respect to DEF dosing performance.
Page updated
Google Sites
Report abuse