Sensing-storage-computing architecture based on memristive convolutional neural networks for transformer fault detection

Abstract

The sensing-storage-computing architecture integrated with a convolutional neural network (CNN) leverages memristor crossbar arrays to implement parallel matrix operations and incorporates high-efficiency artificial synaptic and neuronal circuits. These features can markedly improve the computational speed and energy efficiency of neural networks, thereby helping overcome the limitations of the traditional von Neumann architecture in power transformer fault diagnosis. In this paper, an adaptive voltage-threshold memristor model was first developed, and a 1R1M memristive synaptic and neuronal circuit with flexible weight adjustment was designed and optimized to achieve high-precision nonlinear mapping by combining the Tanh and sigmoid activation functions. A double-layer memristive neural network circuit was then constructed based on a backpropagation (BP) neural network, and a memristive CNN architecture was further developed to optimize high-efficiency memristor-based compute-in-memory and activation circuits. By employing a hybrid kernel and variable-dimensional optimization strategy, we drastically reduced the model size while maintaining high recognition accuracy. Furthermore, a vertically integrated sensing-storage-computing system was established by integrating a gas-sensitive sensor array with the memristive CNN and cascading it with the memristive BP neural network, enabling real-time qualitative classification of transformer fault types and quantitative detection of fault gas concentrations. The proposed architecture maintains excellent classification accuracy while significantly reducing model size and circuit power consumption. These advantages provide a promising solution for power system fault monitoring and lay a solid hardware and algorithmic foundation for the practical application of memristors in complex industrial scenarios.