An intelligent sensor architecture with fuzzy associative memory system

In realizing the intelligent sensor which flexibly transforms input information into the information required by the user employing such knowledge processing as fuzzy inference, the following points are especially important: (1) the representation of the knowledge as well as the acquisition of the knowledge should be easy; and (2) the inference should be executed with a high speed. This paper proposes a sensor construction which satisfies both of the forementioned requirements. The intelligent sensor proposed in this paper employs the if-then type knowledge representation as the fuzzy rule with a high affinity to human thought. The sensor is composed of three independent neural nets: the input (if) part, the output (then part), and the input/output relation (if-then) part. In the input part, the learning vector quantization (LVQ) network evaluates the features of the input waveform, and the result is mapped directly on the concept represented in the condition part of the fuzzy rule. LVQ executes self-learning without a supervisor based on the features of the input waveform and automatically generates the membership function needed in the condition part of the fuzzy rule. This simplifies the knowledge acquisition process. In the input/output relation part, the fuzzy inference is executed with a high speed by the parallel processing of the associative memory network. Thus, the intelligent sensor based on the fuzzy associative memory can improve the forementioned two points (1) and (2). Furthermore, to demonstrate the high versatility of the proposed construction, this paper considers the problem in which two entirely different sample problems are realized by the same construction. The realization example is shown by a simulation.