IoT-Based Real-Time Monitoring and Fault Prediction for Oil-Immersed Transformers Using Improved Spatiotemporal Attention Mechanism

Abstract

This project proposes a three-layer monitoring system based on the Internet of Things to solve the problems of data acquisition lag and low efficiency of multi-source information fusion in traditional oil-immersed transformer monitoring schemes. The perception layer uses Pt100 (±0.1℃) temperature-sensitive (accuracy ±0.1℃), electrochemical gas (ppm level) and piezoelectric acceleration sensors to achieve synchronous acquisition of 12 parameters such as oil temperature, seven characteristic gas concentrations, vibration acceleration, etc., up to 100 Hz. The edge layer uses a sliding average filtering and wavelet transform to filter the data, achieving a 35 dB noise reduction effect and compressing the feature extraction time by 50 milliseconds. Then, an improved spatiotemporal attention mechanism algorithm (STA-I) is introduced to dynamically adjust the weight using the time-trend factor, combined with the adaptive fusion strategy of spatial multimodal data. The STA-I algorithm introduces a time-trend factor to dynamically adjust weights, enhancing the capturing of temporal trends in data. Specifically, it assigns 2.3 times the weight to mutation data compared to normal data, improving fault prediction accuracy. Experimental datasets include 1 million data points collected from 10 oil-immersed transformers over three years. Results show that the average absolute error of system data collection is 0.32℃ for oil temperature and 3.2% for hydrogen concentration, surpassing [specific IEC or IEEE standard name] industrial standards. The average packet loss rate in a mixed network environment (which refers to a situation where multiple network types such as 4G, Wi-Fi, and Ethernet are involved simultaneously or in different scenarios during the data transmission process related to oil-immersed transformer monitoring, and the average packet loss rate is calculated based on the packet loss data collected from each of these network types under specific test conditions and then taking an average weighted by the proportion of data transmitted through each network type) is 0.8%, and the system response time is 0.83 seconds. Compared with LSTM, the STA-I algorithm achieves a prediction accuracy of 96.8%, which is 12.3% higher than LSTM. For local overheating faults, the recognition accuracy reaches 98.5%, and the reasoning time is shortened by 40%.