Edge Anomaly Detection via Model Segmentation and Universality
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current Deep Learning-based anomaly detection models for edge devices are computationally complex, requiring network communication with data centers or cloud nodes and specialized hardware, making them unsuitable for deployment on resource-constrained devices and unable to handle diverse data generated by various edge devices.
Innovation Solution
Customized Deep Learning anomaly detection models are generated by fetching configuration-related information from edge devices, selecting and converting primary models to reduce variables and optimize processing, then packaged in edge device-specific formats like C language binaries for localized execution without network communication or specialized hardware.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If Deep Learning-based anomaly detection models are implemented on edge devices, then anomaly detection capability is improved, but device complexity and resource requirements increase
Solution Approach 1:
The system segments the anomaly detection functionality by separating the training phase (performed on powerful servers) from the inference phase (deployed on edge devices). The complex Deep Learning models are divided into trained weight files that can be loaded and executed on resource-constrained edge devices, allowing anomaly detection capability while reducing on-device computational complexity.
Solution Approach 2:
The system creates simplified copies of the Deep Learning models by exporting trained models as weight files that contain only the essential parameters needed for inference. These copied model representations can be executed on edge devices without requiring the full training infrastructure, thus maintaining detection capability while reducing device complexity requirements.
2Reliability
If comprehensive Deep Learning models are deployed to detect anomalies across all edge devices, then detection coverage is improved, but adaptability to diverse device configurations deteriorates
Solution Approach 1:
The system achieves universality by creating a platform-independent model export format that can be deployed across diverse edge devices with different hardware configurations. The trained models are converted into weight files that can be executed on various devices (mobile, IoT, embedded) without requiring device-specific retraining, thus providing both broad detection coverage and high adaptability to diverse configurations.
Solution Approach 2:
The system changes the parameter representation of the models by converting them from framework-specific formats (TensorFlow, PyTorch) to a universal weight file format. This parameter transformation allows the same model to be deployed across different device configurations and operating systems, improving both detection coverage and configuration compatibility.
3Measurement precision
If Deep Learning models require network communication with data centers, then model accuracy is improved, but processing speed and autonomy deteriorate
Solution Approach 1:
The system performs preliminary actions by training the Deep Learning models offline on powerful servers before deployment. The trained models are then exported as self-contained weight files that can be executed autonomously on edge devices without requiring real-time network communication with data centers, thus achieving both high accuracy and fast local processing speed.
4Power
If specialized hardware is used to execute Deep Learning models, then computational performance is improved, but device accessibility and deployment ease deteriorate
Solution Approach 1:
The system substitutes the requirement for specialized hardware with optimized software implementations that can run on general-purpose processors. By converting models to universal weight files and using efficient inference engines, the system achieves acceptable computational performance on standard edge devices without requiring specialized AI hardware, thus improving deployment accessibility while maintaining performance.
Data Source
AI summary
Customized DL anomaly detection models and generated and deployed on disparate edge devices. Configuration-related information is fetched from the edge devices and, based on the configuration/capabilities of the edge device, at least one primary deep learning-based anomaly detection model is selected, which are customized based on the configuration/capabilities of the edge device. Customization involves limiting the volume of the predictors/variables and optimizing the iterations used to determine anomalies and/or make predictions. The customized models are subsequently packaged in edge device-specific formats, such as a customized set of binaries in C language or the like. The resulting customized DL anomaly detection application is subsequently deployed to the edge device where it is executable without the need for specialized hardware or communication with network entities, such as cloud nodes or servers.