Application Research Of Memristor And Its Crossbar Array In Data Reading-writing And Image Recognition

The proposed memristor and its discovery bring the electronic computing technology and artificial neural networks a greatly potential promotion. Memristor is a kind of basic circuit element which has the excellent nonlinear, nonvolatile, passive property and a nano size. It is hoped to utilize memristor to end the epoch of von Neumann’s computing architecture, then manufacture a brain-like computer which works as real neural network. Meanwhile, it also helps to reduce dissipation of energy. Because of these advantages, it can be widely used in research and application of data storage, artificial neural network, parallel computing, and chaotic circuits etc.Memristor crossbar array architecture is a hot spot of research based on memristor, it can be used in data storing, neural network training and pattern recognizing. Combining the nano scale memristor, crossbar array architecture can save space, and make system chips getting smaller. As memristor can work as a live neuron or other neural parts, namely, memristor can realize calculating and storing at same time, the crossbar array would be of benefit to the realization of neural networks.This thesis mainly discussed the basic mathematical model and several expanded models of memristor and their fundamental characteristics. Then, a brief introduction was made to explain the crossbar’s working process. Based on the binaryzation definition of memristor, we presented a circuit diagram for writing and reading data in the crossbar array architecture, and discussed the main procedure to write into crossbar array and read from it respectively. We noticed that it might have some external noise while the crossbar is working, and the target memristor in crossbar would become unstable while a reading bias drove on it. So, we provided a plan to recover its status. We also noticed that when the scale of crossbar became larger, the data validity should be seriously considered, leading us to utilize a balance pattern to make the circuit more reliable. At last, based on the crossbar array architecture, this thesis presents a kind of memristive neural network, which is used for binary image learning and recognizing.The key point of this thesis is memristor and its crossbar architecture, and their application in data reading-writing and image recognition. The major contributions of the thesis are as follows:① The characteristics research of classical memristor model and expanded memristor modelIn this part, a brief review and discuss were made for classical memristor model presented by Chua. The definitions of charge-controlled memristor and flux-controlled memristor were given based on the relations between several circuit parameters. The main characteristics were discussed at the same time. We also made a mathematical derivation to analyse the HP memristor model and its drift effect, including the comparison of two window functions built for boundary effect. Different from the HP memristor’s assumption, namely, ONR is much less than OFFR, we presented an expanded model that fully considered the influence of ONR. Mathematical analysis proved that even the magnitude are very close between ONR and OFFR, the hysteresis effect is ideal enough in some extenal conditions. Simulation results showed that these memristor have much faster change radio and lower energy dissipation. In addition, researches of dual extended memristor model and spin memeristor model shows the multiformity of memristors from the aspect of status change mechanism.② Researches of memristor crossbar array and its reading-writing policyFocusing on the reading-writing mechanism of the memristor status in crossbar array system, we worked out a system circuit diagram for these operations. To deal with the situation of memristor internal status changed by the reading signal, and the current measured result influenced by sneak pathes, we designed a recovery regime and a balance memory pattern to reduce the reading error, and impove the effective information storage. Then, we brought coding theory into this circuit to reduce the complexity of writing and reading signals. Simulation results verified the reading-writing effects of logic 0 and 1 operation in crossbar under different patterns.③ Researches of crossbar based memristor neural network for binary image learning and recognizingA simple feedforward memristive neural network was built for binary image learning and recognition. The memory layer was used for superposition of image learning, the impact layer memorized the learning procedure, and the decision layer made the decision according to the similarity. This network could eliminate the disturbance of error and noise, and have an ability of fast leaning ratio and highly noise tolerance. The features extracted by impact factors can also used by other kind of neural works for further recognition. Consequently, the memristive structure composed of evaluation function and neural function will simplify the network’s circuit and promote the integration between varieties of networks.