Over-the-Air Machine Learning for Device Anomaly Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional machine learning systems rely on generic models that are not customized for individual devices, leading to unreliable predictions and high numbers of unresolved performance issues and failures in devices generating large amounts of non-linear, heterogeneous data.
Innovation Solution
An over-the-air machine learning (OTA ML) engine that automatically generates customized machine learning models for each device using real-time data streams, training models to predict device anomalies and recommend resolution actions, thereby improving prediction accuracy and reducing human intervention.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If generic machine learning models are used for all devices, then device complexity is reduced and ease of manufacture is improved, but prediction accuracy deteriorates and reliability worsens
Solution Approach 1:
The system performs preliminary actions by pre-training base models on aggregated device data before deployment. These pre-trained models serve as foundations that are later fine-tuned with device-specific data, eliminating the need to start from scratch for each device and improving prediction accuracy while managing complexity.
Solution Approach 2:
The machine learning model is segmented into modular components: a base model trained on aggregated data from multiple devices, and device-specific adaptations fine-tuned on individual device data streams. This segmentation allows the system to balance generalization benefits with customization needs, improving reliability without overwhelming complexity.
2Reliability
If customized machine learning models are generated for each device, then prediction accuracy and reliability improve, but device complexity and computational resources required increase
Solution Approach 1:
The system implements dynamic model generation where customized models are created on-demand based on incoming data streams. The model complexity and customization level adapt dynamically to the quality and quantity of available device data, allowing the system to achieve high reliability when data is abundant while managing complexity when data is limited.
Solution Approach 2:
Each device automatically generates its own customized model using its own data streams and the pre-trained base model as foundation. This self-service approach eliminates the need for manual model customization and reduces the burden on external systems, improving reliability through device-specific modeling while managing complexity through automation.
3Measurement precision
If real-time data streams are processed continuously, then model accuracy and adaptability improve, but energy consumption and computational load increase
Solution Approach 1:
The system implements periodic processing of data streams at strategically chosen intervals rather than continuous processing. Models are updated periodically when sufficient new data is accumulated, balancing prediction accuracy improvements with energy consumption constraints. This periodic action maintains measurement precision while reducing computational load and energy use.
Solution Approach 2:
The system processes only the necessary portion of data streams required for model improvement rather than analyzing every data point continuously. By selecting representative samples and processing partial data sets, the system achieves adequate prediction accuracy while significantly reducing energy consumption and computational requirements compared to exhaustive continuous processing.
Data Source
AI summary
A system receives a plurality of data streams, including a plurality of data points, associated with properties of a device. The system generates, for each data stream, a data set that includes at least a specified number of data points over a prior time period. The system identifies a first data stream that represents a property to be predicted for a future time period. The system generates a joined data set that includes a subset of the data sets not including a first data set of the first data stream. The joined data set and the first data set are inputted into a trained machine learning model, trained to output a predicted value of the first property for the future time period. The predicted value of the first property indicates an anomaly, and a resolution action for preventing the anomaly from occurring at the future time period is identified.