Table 1 Recent works in infrared thermography for monitoring electrical equipment

Feng et al. [8]

The work presented a deployment of an electric transformer thermal assessment scheme using a matching heat model of the thermal circuit. The model sampled mean temperature values of the oil, coils, core, and ambient as well as thermal capacitors and the status of the ONAN/OFAF cooling schemes, whereas the corresponding output parameters are the oil changes in temperature at the inlet and outlet channels, the temperature of the hotspot, the position of the tap changer, loadability indices, and some evaluation criteria

The model was simulated using an object-based technique in three parts: real-time capturing of temperature inputs, development of the thermal model and monitoring terminal, and internet application

Ali-Younus and Yang [9]

Studied IR thermal image using discrete wavelet at second decomposition level under four machine states: normal, bearing, mass-imbalance, and misalignment faults. The authors presented a mechanism for obtaining machine state indices from field IRT images using DWT. The method involves a setup of a 0.5hp variable dc motor fault simulator with a flexibly coupled 30 mm diameter shaft rotating at 3450 rpm and monitored by a FLIR Thermocam. Features of interest for the final analysis of each of the four machine conditions are the average, standard deviation, kurtosis, entropy, mean absolute deviation (MAD), and skew

Manana et al. [10]

Analyzed various electric motor manufacture-induced defect conditions like turn-turn faults, earth to live winding faults and disconnected winding, among other fault conditions, which can come from manufacturing of field poles and insertion inside the stator

Eftekhari et al. [11]

Presented a measure of inter-turn short circuit fault in induction motor in the stator electric circuit using IRT. Histogram of thermal images provided features of interest gotten and compared between good and abnormal 4pole, 2hp, 380 V IM driving servo motor load.

The stator coils were supplied with 220 V to evade catastrophic harm to the machine, especially in the short circuit condition. In computing the distance of the r color vector to the average vector μ, the Mahalanobis distance was preferred over the Euclidean distance for measuring divergence in pixels of the IR image. The authors used Hull_index, Hot_area, and Histogram_{mean values} to forecast hot spots in the stator

Cui et al. [12]

The work discussed the method of feature extraction, de-noising, and segmentation as used in image processing for improving accuracy of diagnostic models by the use of Backpropagation and Radial Basis Function in neural networks for intelligent solutions of faults

Jadin et al.[13]

The paper focused on a semi-automatic, qualitative IR image analysis for quick thermal fault detection and classification. Normalized Cross-Correlation (NCC) algorithm was adopted to detect related parts of interest in the pictures. Thereafter, important numerical features of the images were collected from each identified region and grouped with multilayer perceptron (MLP) for obtaining the temperature of the electrical unit

Huda and Taib [14]

The authors identified a system for checking the status of power equipment using intelligent networks taking features like component-based intensity features, first-order histogram-based statistics, and gray-level co-occurrence matrix gotten by analyzing the images, that are applied as input for the neural network model. Using four separate training algorithms like Resilient back propagation, Levenberg–Marquardt, Bayesian Regularization, and Scale conjugate gradient, the work trained the multilayered perceptron networks. With the component-based intensity features outperforming the other two features. Later, the Levenberg–Marquardt produced good classification results when training the MLP network algorithm for the classification of the status of electrical equipment

Huda and Taib [15]

The use of intelligent IRT for preventive and predictive maintenance for fault detection in electrical devices was the goal of the research, thermal defects in electrical equipment. The authors applied statistical structures to feed the classification network. And, used both MLP as well as discriminant analysis classification models. With a classification accuracy of 82.4%, the discriminant-based model outperformed that of the neural network classifier

Garcia et al. [16]

IR image segmentation and statistical feature extraction for regions of interest by the application of a motor current signature analysis (MCSA) balancing technique for isolating bearing issues, lop-sided mass, and misaligned rotor

Karvelis et al. [17]

The paper focused on the automatic pattern recognition process in infrared images based on object matching by analyzing four abnormal electro-mechanical conditions of the IM like imbalances in the stator, failure of the cooling fan, defective bearings, and broken rotor structure. The authors divided the IM into the frame, fan cover, and motor coupling as well as using a combination of photometric and geometric invariant descriptors. The feature descriptors were obtained in 6 steps from the various points of interest located in the machine. SIFT key points were gotten from both the training and target image samples. Then, the geometric transformations between the two images were calculated after matching their respective SIFTs. Later, models of the test and train images were matched using earlier geometric transformations. The average intensity of any section of the IM together with its surroundings were obtained. Before classifying the temperature to isolate the nature of the problem. The research captured IR images of a 1.1-kW IM driving an auxiliary DC machine using a FLIR S65 IR camera. Applied the Naïve Bayes and C4.5 Tree-based classification models validated by a tenfold cross-validation model to gauge the classifier’s performance

Janssens et al. [18]

Used two image-processing pipelines for identifying rotor imbalance irrespective of bearing defects by firstly discounting successive image frames that are thereafter abridged to their spread within the image coordinates and secondly identifying bearing defects notwithstanding machine imbalanced by the introduction of the standard deviation of the temperature, the Gini coefficient, and the Moment of Light

Zou et al. [19]

An original IR-based two-stage artificial intelligence classification technique for diagnosis of faulty states of electrical devices.

The initial stage was based on the K-means algorithm for the extraction of region and thermal data. The next stage deals with the selection of IR image-relevant features. A set of seven features of both stages were taken as input and fed to the support vector machine (SVM) classification algorithm. The final SVM model was optimized using a coarse-to-fine parameter optimization method and compared with backpropagation

Munoz-Ornelas et al. [20]

Studied how camera position or location impacts monitoring of induction motors from 6 orientations not limited to distance from the object, the angle from the target and height above the target, etc.

Ramirez-Nunez et al. [21]

A new method of auto fine-tuning of the thermal camera with additional external temperature sensors to standardize IR images for improved accuracy of the actual temperature of the object

Singh et al. [22]

Electric motor transient analysis at different stages of operation using IRT and applicable pseudo-coloring method

Singh et al. [23]

Failure of cooling system recognition using IRT for no-load and loaded operating condition

Dragomir et al. [24]

The authors looked at the effects of dust from electrical assets on IR images and gauged the associated thermal stress. They presented a method of delineation of a copper bar into different six surface areas of various reflexivity with different thicknesses of dust particles on each surface area. Experimented with a 300A applied on the copper bar for 50 min until constant temperature was reached. The setup was accomplished with a FLIR T650^oC alongside a temperature logger and a FLIR Max IR analytical software. And concluded that dust particles limit reflexivity from such metallic surfaces while enhancing its IR emissivity

Khan et al. [25]

Presented the application of a specific IR camera called IRISYS (IRI4010) for condition assessment of electric motors and power transformers used along with the IRISYS ISI 4604–4000 Series Imager Software

Dutta et al. [26]

IR images of power equipment were captured and transformed to the appropriate HSI color model, taking cognizance of the hue region for further processing. Edge detection filters such as Sobel, Prewitt, and Roberts were applied to detect high-temperature areas within the IR image. The Otsu image segmentation algorithm was used for clustering the hue regions and with 27 IR images, analyzed with MSE and Peak SNR; the method provided improved segmentation results in comparison with its monochromatic images

Resendiz-Ochoa [27]

Achieved a segmented thermal image using manual thresholding of the IR images, for detection of hotspots in power equipment under investigation

Liu et al. [28]

Presented a rotor platform defect identification based on infrared image and convolution neural network (CNN) classification algorithm for auto-selection of IR image feature and fault type identification

Lopez-Perez and Antonino-Daviu [29]

Presents an illustrative model based on IR images that hinge on the study of Isotherms by showing the temperature gradient and locating the point of the defect using three case studies: a 4Pole 3ph 58 kW, 380 V motor driving a blower; a 2pole 3ph 10 kW, 380 V motor driving pump in polyol tank; and a 4pole 3ph 75 kW 380 V motor driving a fan cooling tower

Dragomir et al. [30]

The work focused on the application of IRT for heat stress detection in HV busbar under electrical load considering the extraneous conditions and correction indices necessary for getting valid results

Mariprasath and Kirubakaran [31]

The research identified various parameters and methods that can be used to assess the health of power transformers such as internal fault detectors, Recovery Voltage technique, Furan Current Analysis, Expert system based software, Frequency Response Analysis, online PD detection, Dissolved Gas Analysis, and Power quality systems. The authors highlighted the merits and demerits of each method and applied six case studies of IRT to depict different locations of hotspots in power transformers of different ratings and recommended IRT for effective, safe, and efficient CM

Sangeetha et al. [32]

The research obtained single- and dual-dimensional relationships between the distance of image capture, emissivity, and hotspot temperature having derived a relationship between the aforementioned three parameters

Fambrini et al. [33]

The research presented an auto-IRT-based system for fault real-time monitoring of power distribution networks using deep learning image processing-based neural networks. The legacy JSEG IR image segmentation was used and the result proved the method would supersede the manual monitoring method

Sahu et al. [34]

The work presented an IRT methodology for monitoring aging acceleration in transformer insulation, by calculating its per unit life. Thereafter, identified effects of transformer insulation’s Aging Accelerating Factor (AAF_Ti) caused by unusually high abnormal temperature on the equipment windings. Data were sourced from IR images of transformers as well as associated readings of oil temperatures at the top of the tank taken at different intervals. The proposal discussed the effect tank coefficient of reflexivity and oil emissivity on predicting the hotspot in the windings using digital image processing with mathematical expressions found in the IEEE Guide to Loading. The authors developed a model equation for calculating the revised temperature value of the existing hotspot and record any mismatch error found. The model indicated that the actual hotspot in the winding depends on the winding hotspot temperature, the top oil temperature read by the IR camera, and the top oil temperature pointed by the oil temperature indicator

Najafi et al. [35]

The authors proposed an interpretable Machine Learning containing an automated channel for assessing the status of power assets by means IRT dataset. The application of a pre-processing stage divides the images based on the unit’s temperature, that is cold or hot conditions. Finally, a sliding window technique based on AdaBoost and Random Forest (RF)-based classifiers were utilized for segmentation

Vidhya et al. [36]

Used Symlet wave transform to achieve flatter transformer breather image decomposition of typical parameters of IR images. The images were represented in discrete form by the mechanism of discrete transformation. The statistical information derived from the transformation under different states of decomposition reveals the changes in the temperature distribution within the breather piping system at defined functional states of the transformer. The Symlet technique shows very low asymmetric features in most types of Daubechies wavelets. Comparison of the decomposition was done for normal and abnormal operating conditions. Local regions were defined through feature descriptors like histograms. The result shows that wavelets can bring out inherent characteristics of the IR image

Mahami et al. [37]

The work used the Bag-of-Visual Word (BoVW) to capture anomalies in IR features with Speeded-Up Robust Features (SURF) detector and descriptor. Also applied the Ensemble learning-based Extremely Randomized Tree (ERT) to automatically identify anomalies in IM