Acceleration of the learning of artificial neural networks utilizing a thermodynamic and kinetic model with applications in sensor data processing
Signal detection and classification are very important in industrial applications, especially in real-time sensor data processing. Traditional techniques cannot process these sensor data very well in industrial applications because of high noise and a large amount of information requiring parallel processing to achieve sufficient throughput. Artificial neural networks are a new technique to solve this problem. However, the commonly used back-propagation neural network typically has a very slow learning rate, and learning instabilities that make the learning unpredictable. The author studied both problems from a mathematical view point as well as a thermodynamic/kinetics view point, and derived a new learning equation, called the delta-activity rule. The learning time was decreased by two to eight times. The author also applied this new algorithm to two different application problems: high frequency signal detection and weld-seam tracking. In both applications, the results show significant improvement over the traditional sensor data analysis techniques.
- Research Organization:
- Colorado School of Mines, Golden, CO (United States)
- OSTI ID:
- 5587097
- Resource Relation:
- Other Information: Thesis (Ph.D.)
- Country of Publication:
- United States
- Language:
- English
Similar Records
Artificial Neural Network-Based Monitoring of the Fuel Assembly Temperature Sensor and FPGA Implementation
A robust and efficient training algorithm for feedforward neural networks