Embedding Chemical Sensor Signals for Odor Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current computing devices lack the ability to robustly sense smells and interpret chemical sensor data, as they cannot convert raw signals into human-interpretable labels across the entire space of possible odors, and require time-consuming individual training for specific smells, limiting their ability to detect new mixtures and olfactory properties.
Innovation Solution
A computing system equipped with a sensor that generates electrical signals indicative of chemical compounds and a machine-learned model trained to process these signals, generating embedding outputs in an embedding space for tasks such as olfactory property prediction, disease detection, and spoilage identification, using a combination of transformer models and graph neural networks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If individual training is performed for each smell, then the device can accurately identify trained smells, but the training process becomes time-consuming and computationally taxing
Solution Approach 1:
The patent applies universality by training a single machine-learned model to handle multiple smell identification tasks simultaneously. Instead of creating separate trained models for each odor, one universal model learns to identify multiple different smells through exposure to diverse training data, eliminating the need for separate individual training processes for each odor type.
Solution Approach 2:
The patent applies preliminary action by pre-training the machine-learned model on a comprehensive dataset of multiple smells before deployment. This preliminary training phase allows the model to acquire knowledge about various odors in advance, so that when deployed, it can immediately identify multiple smells without requiring additional training time for each specific odor encountered during operation.
2Measurement precision
If individual training is performed for each smell, then the device can identify trained smells, but it fails to determine non-trained properties and new mixtures
Solution Approach 1:
The patent applies universality by designing a machine-learned model that can identify multiple different smells using a single training process. The model is trained on diverse data representing various odors and their properties, enabling it to generalize its knowledge to identify not only the specific trained smells but also new mixtures and non-trained properties, thus achieving both accuracy and versatility.
Solution Approach 2:
The patent applies parameter changes by training the model on varied parameters including different smell compositions, concentrations, and properties. This diverse training exposes the model to a wide range of chemical characteristics, enabling it to adapt to and identify new mixtures and non-trained properties by recognizing patterns across different parameter combinations rather than memorizing specific fixed configurations.
3Productivity
If a machine-learned model is trained to process electrical signals, then computational efficiency is improved, but the model requires extensive training data covering the entire space of possible odors
Solution Approach 1:
The patent applies partial action by training the machine-learned model on a representative subset of odor data rather than attempting to cover the entire theoretical space of all possible odors. The model is trained on carefully selected training data that captures the essential variations and properties of odors, allowing it to achieve good performance on multiple smell identification tasks without requiring exhaustive training data covering every conceivable odor combination.
Data Source
AI summary
Electronic chemical sensors can output raw electrical signal data in response to sensing a chemical compound, but the raw electrical signal data can be difficult to interpret. Processing the electrical signal data with a machine-learned model to generate an embedding output in an embedding space can provide a better understanding of the electrical signal data. Moreover, leveraging preexisting chemical property prediction models to generate other embeddings in the embedding space can allow for more accurate and efficient classification tasks of the electrical signal data.