Gas analyzers often use dissolved gas analysis (DGA) in oil to determine transformer faults, and their accuracy is not a fixed value. Due to factors such as analysis methods, instrument performance, and operating standards, the accuracy of traditional methods is mostly around 80% -86%. Optimization schemes with intelligent algorithms can improve the accuracy to over 95%, and some optimization models even approach 100%. The specific differences are as follows:
Traditional analysis methods such as the traditional three ratio method and Rogers ratio method, which rely on gas concentration ratios to determine faults, have limitations in accuracy due to incomplete coding and absolute judgment boundaries. The accuracy of the three ratio method is about 86.32%; However, early diagnostic methods based on simple models performed worse, such as a basic model based on support vector machines (SVM), which had an overall accuracy of only 90% for fault diagnosis and a recognition rate of only 80% for faulty transformers.
Optimization methods with intelligent algorithms: With the integration of algorithms and detection technologies, accuracy has significantly improved. The combination of evolutionary k-means and expert sub models achieved an accuracy rate of 98.29% on the IECTC10 database; The intelligent diagnostic system embedded with random forest algorithm has a typical fault recognition accuracy of ≥ 98.5% for 32 fault modes in 6 major categories; There is also a fuzzy inference system based on the correlation features of characteristic gases, which achieves an accuracy of 100% in specific fault time-series data testing; When gas chromatography is combined with digital automatic comparison and intelligent recognition technology, the detection accuracy can also be improved to 95%.
Other optimization model methods: Some customized models optimized by feature selection and algorithm optimization also have excellent performance. For example, based on the golden jackal optimization algorithm to select feature quantities and the sky eagle algorithm to optimize the AO-RF model of random forests, the diagnostic accuracy can be improved by 1.84% -15.86% compared to ordinary RF, SVM and other models; The model based on Bayesian networks and hypothesis testing has a maximum diagnostic accuracy of 88.9%, which is lower than some top algorithm models, but still significantly improves compared to traditional methods.
However, the above accuracy is the result under ideal conditions. In practical applications, if there are problems such as non-standard sampling, gas leakage caused by oil sample transportation vibration, and insufficient purity of carrier gas, it may result in about 23.7% of misjudgment cases, which will significantly lower the accuracy of final fault diagnosis.