Research On General Technology Of Hardware Accomplishment Of Neural Network In The Electric Drive System

History and trend of Artificial Neural Network (ANN) and Network Control (NNC) were reviewed and the electric drive control systems based on Expert Control (EC), Fuzzy Control (FC) and neural network control were proposed. Seven forms of Neural Network (NN) applied in the electric drive control system were summarized. NN can be used as an internal part of the electric drive system or complete some function. NN can be used to detect signal, estimate parameter or observe state of the electric drive system for the purpose of real-time compensation or close-loop feedback control. NN can be used in model identification to realize Model Reference Adaptive Control (MRAC) of the electric drive system. NN can be used as a controller of the electric drive system. NN can be used to achieve neural network inverse control in AC drives. NN can be used in state detection and fault diagnosis of the electric drive system. NN can be used to achieve compound control of the electric drive system.Based on the control view, a general neuron model and the adjusted learning algorithm of NN connection weights were concluded. The main features and selection principles of the 8 kinds of commonly learning algorithms, such as: Hebb, Perceptron,δ(Delta), Widrow-Hoff, Correlation, Winner-Take-All, Outstar and Boltzmann, were analyzed. Then, the topological structures, learning algorithms and application of Adaline neuron, BP, CMAC and BAM NNs were designed and simulated.The technical method of hardware accomplishment of NN and reconfigurable topological structure of controller based on FPGA was presented. Especially, the general kernel technologies were studied, including: fixed-point multiplication / division operations, floating-point addition / subtraction / multiplication / division operations, the realization of the Sigmoid transfer functions of NN nodes, Hebb learning algorithm, so on. The 4 standard models of NN nodes were defined as following: type 1-1, type n-1, type n-m and type 1-m.The basic steps of NN design and its engineering application were systematically described. After the introduction of NN-identification, six commonly used NNC structures were summarized, which includes: Direct Neural Network Controller (DNNC), Compound Neural Network Controller (CNNC), Adaptive Neural Network Controller (ANNC), Inverse Neural Network Controller (INNC), Supervised Neural Network Controller (SNNC) and Optimized Neural Network Controller (ONNC). The above CNNC, ANNC and INNC were the focus points. Meanwhile, the stability of Neural Network Control System (NNCS) was analyzed and proved. The basic features of the NN learning and self-learning control were analyzed. The learning control rules and the corresponding general topological structures were summarized. The general principle and the rule-based self-learning control system were studied. The concept of coupling variable was introduced for the grey information model, and the novel cooperative control strategy based on quantitative model and qualitative model was put forward, referencing to the three-level hierarchical intelligent control structure used in the complex systems. The basic architecture of foreground-background self-learning neural network control in the electric drive system was presented creatively.By the NNC theory and the general kernel technologies of FPGA-based hardware accomplishment of NN, the detailed process was exhibited through programming BP-NN forward propagation, CMAC-NN control and discrete BAM-NN fault diagnosis of power converter. The comprehensive experimental system based on DSP, FPGA and linear motor was developed. According to the test performance results of linear motor servo system to CNC machine tools, the feasibility and effectiveness of the general kernel technologies were approved.