**1. Introduction**

Power transformers that work under harsh environments would experience thermal decomposition of oil and cellulose insulation materials, such as arcing, corona discharge, low energy sparks, severe overloading, overheating of insulation systems and pump motor failures. These conditions alone or in combination can produce combustible and noncombustible gases [1] Detection of anomalies requires an assessment of the amount of gas produced. Gas in oil-immersed transformers can be used to identify fault types, including thermal and electrical interference. Gases obtained from chromatographic analysis of insulating oils may contain dissolved carbon monoxide (*CO*), carbon dioxide (*CO*2), nitrogen (*N*2), hydrogen (*H*2), methane (*CH*4), acetylene (*C*2*H*2), ethylene (*C*2*H*4), and ethane (*C*2*H*6). The composition, formation rate and specific content ratio of dissolved gas can be used to indicate transformer condition.

The composition and content of dissolved gases in oil of transformer insulation can reflect the operation condition of transformer to a great extent thus dissolved gas analysis (DGA) has become an effective method for fault diagnosis of oil-immersed transformers [2].

Organizations such as the Institute of Electrical and Electronics Engineers (IEEE) and the International Electrotechnical Commission (IEC) recommend a variety of diagnostic techniques [3], depending on the type of transformer and operating conditions. Some of the most commonly used techniques include Doernenburg ratio, Rogers ratio, Duval triangle model, etc. These classical diagnostic methods mostly take the ratio of different gases as the characteristic input and then judge the actual operating condition of the transformer by the threshold value formed by experience or statistical methods. Fuzzy network, support vector machine, artificial neural network, and other commonly used artificial intelligence methods

**Citation:** Yu, W.; Yu, R.; Tao, J. An Unsupervised Mutual Information Feature Selection Method Based on SVM for Main Transformer Condition Diagnosis in Nuclear Power Plants. *Sustainability* **2022**, *14*, 2700. https:// doi.org/10.3390/su14052700

Academic Editors: Luis Hernández-Callejo, Sergio Nesmachnow and Sara Gallardo Saavedra

Received: 25 January 2022 Accepted: 16 February 2022 Published: 25 February 2022

**Publisher's Note:** MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

**Copyright:** © 2022 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/).

are also generally introduced into the field of power transformer fault diagnosis [4–7]. However, in the studies of different scholars, the features used as the basis of intelligent diagnosis are often different.

In previous studies, in addition to problems in diagnostic methods, there are the following phenomena in monitoring data: less test data, less available data sets and unbalanced data type distribution, which bring great problems to algorithm verification [8].

Main transformers are important equipment for power generation of nuclear facilities. They are in a high-load long-term condition and are more prone to failure caused by aging [9]. Meanwhile, due to the particularity of nuclear power refueling overhaul and the conservative culture of nuclear power [10], the maintenance strategy of nuclear equipment is more rigorous and conservative, and the failure modes of main transformers may be slightly different. The data of the main transformers are classified separately in the IEC database [2], which shows transformer performance difference in nuclear industry.

Due to the particularity of nuclear power transformers, there are less marked data and more constraints on the monitoring data that can be used for research. The features are important inputs of the diagnostic algorithm. High-dimensional features bring high computational cost and the risk of "over-fitting". Dimensionality reduction or selection is an important research direction.

In this paper, an unsupervised mutual information feature algorithm is proposed for feature selection of different features proposed in the current classical algorithm and intelligent algorithm, as a pattern recognition method, SVM can construct the optimal classification hyperplane under the condition of small sample learning and distinguish transformer conditions according to the input features. The main transformer condition diagnosis model based on support vector machine is adopted for diagnosis in this paper, and the case data of main transformer is verified.
