Information Collection Control for Threshold-Based Edge AI Inference

The patent segments the environmental information processing by dividing it into essential parameters (temperature, humidity) that directly affect AI accuracy and non-essential parameters. Only the critical temperature information is acquired and applied to the learned model, while other environmental factors are excluded. This segmentation reduces processing load while maintaining AI processing accuracy by focusing only on the most impactful environmental parameter.

The patent applies parameter changes by adjusting the learned model's parameters (specifically the normalization parameters in the neural network) based on the current environmental temperature. Instead of storing multiple complete learned models for different environments, the system dynamically modifies the model parameters according to the detected temperature, enabling adaptability across various use environments while using a single model structure with adjustable parameters.

The patent implements preliminary action by pre-acquiring temperature information and pre-adjusting the learned model parameters before actual AI processing begins. The system continuously monitors environmental temperature and proactively updates the normalization parameters in advance, so when processing is needed, the model is already optimized for current conditions. This eliminates the need to stop processing for model changes.

The patent applies periodic action by acquiring temperature information at fixed time intervals rather than for every single video frame. The system updates the environmental temperature data periodically and holds this information for multiple processing cycles, reducing the frequency of temperature acquisitions while maintaining sufficient accuracy for AI processing across varying thermal conditions.