This paper proposes a layered modular architecture to adaptively perform data mining tasks in large sensor networks. The architecture consists in a lower layer which performs data aggregation in a modular fashion and in an upper layer which employs an adaptive local learning technique to extract a prediction model from the aggregated information. The rationale of the approach is that a modular aggregation of sensor data can serve jointly two purposes: first, the organization of sensors in clusters, then reducing the communication effort, second, the dimensionality reduction of the data mining task, then improving the accuracy of the sensing task . Here we show that some of the algorithms developed within the artificial neuralnetworks tradition can be easily adopted to wireless sensor-network platforms and will meet several aspects of the constraints for data mining in sensor networks like: limited communication bandwidth, limited computing resources, limited power supply, and the need for fault-tolerance. The analysis of the dimensionality reduction obtained from the outputs of the neural-networks clustering algorithms shows that the communication costs of the proposed approach are significantly smaller, which is an important consideration in sensor-networks due to limited power supply. In this paper we will present two possible implementations of the ART and FuzzyART neuralnetworks algorithms, which are unsupervised learning methods for categorization of the sensory inputs. They are tested on a data obtained from a set of several nodes, equipped with several sensors each.
Inter-Stage Output Voltage Amplitude Improvement Circuit Integrated with Class-B Transmit Voltage Amplifier for Mobile Ultrasound Machines
