Crowd-Sourced Neural Network for Indoor Positioning

The system enables self-service by allowing mobile devices to automatically contribute their own RSS measurements to train the neural network model. Devices autonomously participate in the calibration process without requiring manual intervention, transforming users into active contributors who inadvertently improve the system's positioning accuracy through their normal movement patterns.

The patent transforms the calibration approach by changing from static manual fingerprint collection to dynamic neural network training using RSS parameters. The system collects RSS measurements from multiple devices, processes them through a neural network that learns optimal positioning parameters, and continuously adapts to environmental changes, thereby reducing calibration time while maintaining accuracy.

The patent replaces traditional mathematical signal models with a neural network-based system. Instead of using simplified physics-based models that assume ideal propagation conditions, the system uses a data-driven neural network that learns complex indoor signal patterns from actual measurements, achieving higher reliability without proportionally increasing computational burden through efficient architecture design.

The system performs preliminary training of the neural network using crowd-sourced data during off-peak times or when sufficient data is accumulated. This pre-training phase prepares the model in advance, allowing it to make accurate positioning predictions during actual use with minimal real-time computation, thus balancing model accuracy with computational efficiency.

The system implements feedback mechanisms where the neural network continuously learns from incoming RSS measurements and adjusts its parameters accordingly. The feedback loop includes confidence level calculations that provide information about data quality, allowing the system to automatically weight or discard measurements based on their reliability, thereby managing data processing complexity while maintaining high training data quality.

The patent applies partial action by selectively processing only the most reliable measurements for training. Instead of attempting to process all crowd-sourced data equally, the system uses confidence thresholds and selection criteria to focus computational resources on high-quality data points, reducing processing complexity while maintaining training effectiveness.

The system applies partial supervision by using confidence levels to determine which measurements require full verification versus which can be processed more quickly. High-confidence measurements are accepted with minimal verification, while low-confidence measurements undergo more rigorous validation, thereby improving training efficiency without sacrificing model accuracy.

The patent implements preliminary filtering and validation of crowd-sourced data before it enters the main training pipeline. By pre-processing data to remove obvious errors and validate basic criteria in advance, the system reduces the computational burden during actual neural network training, thereby improving productivity while maintaining model accuracy through subsequent supervised learning.