Neural Network Training for Indoor RSS Fingerprint Navigation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Indoor navigation and positioning for mobile devices is hindered by the dynamic nature of indoor infrastructure, making it challenging to maintain accurate fingerprint maps, as traditional signal models oversimplify the complexities of indoor environments and are resource-intensive.
Innovation Solution
A neural network training system that integrates a postulated mathematical model of wireless received signal strength (RSS) features with machine learning methods, allowing for more accurate calibration of indoor position systems by adjusting initial weights through backpropagation to match measured RSS parameters, thereby reducing computational burden and improving convergence.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If traditional signal models are used for indoor navigation, then the system is simpler to implement, but the positioning accuracy deteriorates due to oversimplification of indoor environment complexities
Solution Approach 1:
The patent transforms the positioning approach by changing from traditional signal strength modeling parameters to neural network weight parameters. The system uses measured RSS values to train and adjust neural network weights, enabling the model to adapt to complex indoor environments while maintaining manageable system complexity through standardized neural network architectures.
2Use of energy by moving object
If manual calibration efforts are used to generate and maintain fingerprint maps, then the system requires less computational resources, but the calibration process becomes time-consuming and error-prone
Solution Approach 1:
The system implements self-service calibration by automatically collecting RSS measurements from mobile devices and using these measurements to train the neural network. The fingerprint map is generated and maintained automatically through the neural network training process, eliminating the need for manual calibration efforts while reducing both time and computational resources required.
Solution Approach 2:
The patent performs preliminary neural network training using collected RSS data before actual positioning operations. This preliminary training phase establishes the fingerprint map and optimizes the model, so that subsequent positioning can be performed efficiently with reduced computational burden during real-time operation.
3Measurement precision
If fingerprint maps are maintained to improve positioning accuracy, then the positioning precision improves, but the system becomes challenging to maintain due to dynamic indoor infrastructure
Solution Approach 1:
The system transitions from static fingerprint maps to dynamic neural network models that can adapt to changing indoor environments. The neural network weights are continuously adjusted based on new RSS measurements, allowing the positioning system to automatically adapt to infrastructure changes such as remodels or added structures without requiring manual fingerprint map updates.
Solution Approach 2:
The patent implements feedback mechanisms where measured RSS values are continuously compared with neural network predictions, and the weight matrices are recursively adjusted through backpropagation. This feedback loop enables the system to automatically correct for environmental changes and maintain positioning accuracy without manual intervention.
4Measurement precision
If full training regimen is used to train neural network from randomly generated weights, then the model accuracy improves, but the computational burden increases significantly
Solution Approach 1:
The system performs preliminary weight initialization based on a postulated RSS model before full neural network training. This preliminary action provides a reasonable starting point that is closer to the optimal solution, reducing the number of training iterations required and thereby decreasing the computational burden while still achieving high model accuracy.
Solution Approach 2:
The patent introduces a postulated RSS model as an intermediary between random weight initialization and full neural network training. This intermediary model provides physically informed initial weights that bridge the gap between arbitrary random values and the complex optimization landscape, enabling more efficient training convergence.
Data Source
AI summary
A method and system of neural network training for mobile device indoor navigation and positioning. The method, executed in a processor of a server computing device, comprises determining, in the processor, at a first location relative to a wireless signal source at a second location, a set of received signal strength (RSS) input parameters in accordance with a postulated RSS model, the processor implementing an input layer of a neural network, receiving, from a mobile device positioned at the first location, a set of RSS measured parameters from the wireless signal source at the second location, computing, at an output layer of the neural network implemented by the processor, an error matrix based on comparing an initial matrix of weights associated with the at least a first neural network layer representing the RSS input feature to an RSS output feature in accordance with the RSS measured parameters of the mobile device at the first location, and recursively adjusting the initial weights matrix by backpropogation to diminish the error matrix until the generated RSS output feature matches the RSS measured parameters.