Sensor Data Classification With Cosine-Based Unknown Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing classification methods using neural networks for sensor data in driverless navigation fail to detect scenarios that were not included in the training data, leading to potential errors in decision-making.
Innovation Solution
A method that utilizes a classifier with a neural network trained on averaged feature vectors, employing cosine similarity and Mahalanobis distance to identify unknown scenarios, and a GAN for generating training samples to enhance detection of new scenarios.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a neural network is trained on predefined classes for sensor data classification, then classification accuracy for known scenarios is improved, but the ability to detect unknown scenarios deteriorates
Solution Approach 1:
The patent pre-calculates averaged feature vectors for all predefined classes before actual classification occurs. These averaged vectors serve as reference points that enable the system to not only classify known objects but also detect when an object does not match any predefined class, thus preparing the system in advance for both classification and anomaly detection tasks.
Solution Approach 2:
The patent introduces cosine similarity as an intermediary metric between the feature vector of a detected object and the averaged feature vectors of predefined classes. This intermediary measurement allows the system to quantify how well an object matches any predefined class, and when the similarity is below a threshold for all classes, it indicates an unknown scenario, thus bridging the gap between classification and unknown detection.
2Productivity
If training data covers only common scenarios, then training efficiency is improved, but reliability in rare or novel scenarios deteriorates
Solution Approach 1:
The patent enables the classification system to self-diagnose its limitations by detecting when an object does not match any predefined class through the cosine similarity mechanism. This self-service capability allows the system to automatically identify scenarios outside its training distribution, providing a foundation for continuous improvement without requiring manual intervention to recognize all possible edge cases during training.
Solution Approach 2:
The patent establishes a feedback mechanism where the cosine similarity results provide information about the confidence of classification. When similarity scores are low across all classes, this feedback indicates the presence of unknown scenarios, allowing the system to trigger appropriate responses such as flagging for manual review or initiating retraining processes, thus creating a closed-loop system that improves reliability over time.
Data Source
AI summary
A method and a device for classifying sensor data are proposed, wherein the method comprises providing a respective averaged feature vector for a multiplicity of classes wherein the method further comprises the following steps by means of a classifier, wherein the classifier comprises at least one neural network trained on the basis of training data:determining a feature vector on the basis of sensor data,respective determining of a cosine similarity between the feature vector and a respective averaged feature vector for the multiplicity of classes,comparing the respective cosine similarity with a threshold value established for each class beforehand, anddetecting a scenario not represented by the multiplicity of classes if the threshold values are not reached for all classes, and whereinthe device comprises a classifier and is designed for carrying out the method.