Neural Network Circuit Design And Core Device Fabrication For Intelligent Modulation Recognition

In non-cooperative wireless communication,the recognition of modulation mode is the premise to correctly realize communication demodulation and ensure the normal reception of communication.It is the correctness of recognition result and demodulation method that can ensure the Radio Frequency(RF)signals from multiple senders be received correctly.Therefore,automatic modulation recognition technology and its circuit research have high significance in military and civil communications.The recognition rate of traditional modulation recognition method is low,due to its dependence on manual intervention,and is inefficient in the complex wireless communication environment nowadays.Compared with traditional recognition methods,modulation recognition based on neural network has been regarded as one of the most important way to realize unmanned and automatic recognition in recent years.However,like other applications of neural network,the development of modulation recognition technology based on neural network has also encountered bottlenecks,as the massive data required by neural network method can barely be processed with limited computing resources in mobile applications which are usually configured in a more traditional von Neumann computing architecture.By contrast,the neural network circuitry based on innovative neural devices has the advantages of high operational efficiency,small footprint and low power consumption.In this dissertation,the first step is to study the underlying neural components.New neural components and circuit units with learning capability are designed for intelligent modulation recognition.Based on these two basic circuit cells,a neural functional circuit and a neural network recognizer circuit are constructed.The sample of the circuit cell capable of learning is fabricated and characterized.The properties of neural devices and neural circuits are analyzed as well,and the application of recognizer in modulation recognition is discussed.The dissertation covers following aspects:1)A one selector one resistor(1 S1R)cell with learning capability is designed.The characteristics is analyzed and the preliminary experimental verification is verified.Theoretically,the 1S1R cell not only has the same learning capability as other memristors,but also has the advantages of anti-sneak-current and self-compliance current.In order to explain its mechanism,a sample of thin-film self-compliance selector was fabricated by magnetron sputtering technology,and special interfaces structure was established based on the annealing process.The test results show that the above method greatly optimizes the performance of the sample.Finally,the theoretical models for the self-compliance selectors were proposed to explain the mechanism of resistive switching properties,self-compliance phenomenon,and endurance failure phenomenon.The self-compliance selector is combined with two memristors in series separately,and thus a circuit cell capable of learning is established.The simulation results show that the performance of 1S1R cell is better than that of the ordinary memristor.2)With the help of Through Silicon Via(TSV)technology in multi-chip 3D integration,a new 3D neuron transistor was designed,whose mechanism is similar to planar MOS floating gate neuron transistor,and can be integrated into the interposer of the 3D integration component.It is beneficial to the implementation of the 3D integration for the high-performance computing circuit and improve the functionality of the component without increasing the chip area.Simulation results show that the device has good neuronal computing performance.3)Furthermore,the 1S1R circuit unit and the novel 3D neuron transistor was combined to construct a 3D neuron recognizer circuit.Compared with the traditional neuron circuit,the resistance value of 1S1R structure can be manipulated from the outside,and it has the capability of keeping the resistance value unchanged during the working process,which means that in practical application,the weights of the neurons can be changed by training,so that it has the capability of learning different knowledge.Based on this characteristic,a recognizer circuit utilizing the principle of neural network is constructed.Simulation results show that the recognizer can intelligently identify multiple input data.The application of the circuit in intelligent recognition of modulation is discussed.