Application of Fault Overlay Method and CNN in Infrared Image of Detecting Inter-Turn Short-Circuit in Dry-Type Transformer

Abstract

Inter-turn short-circuit (ITSC) faults do not necessarily produce high temperatures but have special heat distribution and characteristics. Therefore, a new recognition solution for diagnosing faults is proposed, based on the fault overlay method, and the convolutional neural network (CNN) is trained to achieve the automatic identification of infrared images. In this method, through the coverage of layers, the proposed image augmentation method is carried out and simulates the fault data of increasing training. We produce 43 fault traces through the fault overlay method on the three-phase winding of a transformer and use 90 infrared images of transformers in normal operation combined with them to enhance the amount of image data. The fault recognition ability is realized based on CNN model training, including analysis of experimental results of grayscale and color images, and Gaussian noise. In the test of the practical case, a short-circuit test of the 11.4 kV dry-type transformer is carried out, and the ITSC fault is identified when the load is about 15%. The fault characteristic block on this thermal image is 36.3 degrees, which verifies the identification available by this method, and has a certain reference value for the development of infrared image diagnosis technology for power equipment.

1. Introduction

Dry-type transformers (DTT) have advantages over oil-immersed transformers, so they are used in many situations. However, the load-bearing capacity of the DTT is limited to operating below the rated capacity, while the oil-immersed type has a better overload capacity. If the operating conditions are not followed, the insulating medium will be damaged. The most critical part of a transformer is insulation, where thermal aging is the leading cause of insulation failure. The insulation failure of the winding accounts for 30% of transformer failures, with a turn-to-turn fault being the most common fault in the transformer [1]. The cause of the ITSC fault is that the conductor insulation layer becomes weakened to an extent. Then, the circuit can no longer withstand the mechanical stress of the fault, so it suffers dielectric damage. In [2], from the statistics of 343 power transformers over 10 years, the most common cause of failure was insulation (36.74%), followed by winding (21.30%), both of which were related to ITSC. Ref. [3] analyzed the thermal analysis of DTT failure through computational fluid dynamics software and verified that the pre-failure symptoms of the transformer were related to heat. Therefore, detecting thermal trends in advance is significant and helpful in reducing accidents. This study focuses on a cast-resin dry-type transformer.

The advantage of infrared thermal imaging technology is that its properties are non-intrusive and non-destructive, along with its fast response time and wide temperature ranges [4]. This is a valuable tool in the industry for monitoring equipment and facilities’ preventive condition-based maintenance (CBM) [5]. In Section 2, featuring our observations, only a few turns are short-circuited in the initial stage of the ITSC. It causes the short-circuited coil to generate a circulating current and to heat up, but it does not reach the set temperature. However, the available thermal image monitoring system only has the function of a fixed temperature alarm [6]. Therefore, it is necessary to identify the fault in the device’s thermal image. Image recognition technology must first collect the object’s data to be tested, but the difficulty of obtaining fault data has become an essential threshold for application.

In the past, electrical equipment’s infrared imaging fault diagnosis mainly tried to segment the region of interest (ROI) [7]. Methods for image segmentation based on temperature can be roughly divided into threshold segmentation, edge detection, and region segmentation [5]. These methods are not available for environmental changes and are not suitable for finding ITSC features. In recent years, deep learning methods have become a new research direction for image fault diagnosis [8,9,10,11,12,13,14,15,16,17], and the convolutional neural network (CNN) has especially excellent feature extraction performance [18]. In electrical engineering, many scientists have applied it to motor fault identification [14,15,16,17,18,19,20,21,22,23,24,25,26,27], oil-immersed transformers [28], gas-insulated switchgear [29], insulators [30], solar panels [31], and electrical equipment [5]. There have been few studies on identifying and locating ITSC using infrared imaging. Most utilize frequency response analysis (FRA) [32], differential admittance [33], and differential current monitoring [34], but are prone to inaccuracies as transformers age. Most utilize differential current, differential admittance monitoring, and frequency response analysis (FRA), but are prone to inaccuracies as transformers age. The simulation of the transformer winding circuit shows that the ITSC will cause the internal loop current to increase [35], and the temperature will inevitably increase.

When using deep learning methods, it is important to obtain data sources. Many fault detection references use a fault simulator to simulate faults and obtain the fault data. Many artificial intelligence methods are only effective under specific training patterns and expert rules for each case; it is difficult to generalize and apply a common framework to any transformer [26]. However, the target discussed in this paper is difficult to obtain data through the simulator. Therefore, we present an innovative method to overcome this problem using deep learning to extract features and identify faults.

The fault overlay method is proposed in this paper. First, the study collects images of the devices’ regular operation. Then, fault images are analyzed, and fault traces are extracted, which are then used to synthesize new fault images. Finally, deep learning is used to classify and establish a fault diagnosis model. In addition to single-type faults, the occurrence of compound faults is also analyzed. The training images are divided into two types, including grayscale and color, and the choice of thermal images is also discussed. The goal is that the user can more easily see temperature changes in color, so we utilize CNN to train the model. In order to promote the robustness of the model, Gaussian noise is added to each image and then trained.

More specifically, this study provides two main contributions. (1) We propose a method to solve the problem of thermal characteristics that do not have high temperature characteristics on the thermal image; called the fault overlay method. Mainly through the coverage of layers, the proposed image enhancement method is carried out and simulates the fault data of increasing training. At the same time, the problem of difficulty in obtaining thermal image fault data is solved. (2) The fault data collection of DTT is not as easy as that of motors. Since transformers are of a high voltage class, making a simulated fault device is a huge undertaking. We provide infrared fault diagnosis system solutions based on condition maintenance. Furthermore, we introduce this application into the actual detection target of fixed-type thermal imaging cameras, as shown in Figure 1.

Section 2 introduces the thermal image monitoring device used in this study and the method for data collection. Section 3 describes the fault type in the DTT and establishes a method for acquiring images for fault cases. Section 4 introduces the training model for fault diagnosis. Section 5 shows the simulation result analysis and verifies the proposed method’s efficiency. Lastly, the conclusion is discussed in Section 6.

2. Data Collection and Monitoring Equipment

This research collects the training data by capturing images of the DTT in the substation. The fixed-type thermal imaging camera is mounted to capture the transformer’s overall appearance, as shown in Figure 2a. In addition, it is placed for a period of time to ensure that the captured image has various transformer load cases, from low to high.

The laboratory develops a thermal image monitoring system, as shown in Figure 2b,c. The thermal image camera, with 160 × 120 pixels, its core is a Raspberry Pi and thermal image sensor for online monitoring equipment. It will send the thermal image in real time using the network and save it to the server. The data sent by the device is a matrix with values in absolute temperature K, which will be converted into a grayscale or color image after normalization, as shown in Figure 3. The collected data can be the raw data from the sensor, not the covered image that has been processed using the color map. This method needs to be executed on grayscale images in advance.

3. Establishment of Fault Traces

Establishing the fault image database requires a rule of thumb because it affects the performance of the diagnostic. This section discusses the cause of the fault through the actual collected fault data. The problem of ITSC in DTT has been discussed [5], but only a few researches have proposed detection schemes for this fault. Next, we use the thermal imaging of a normal transformer as a basis and extract the fault traces to create the fault image database. Finally, the synthesized images are used as the training data.

3.1. Dry-Type Transformers Failure

Transformer loss causes the winding and core temperature to rise [19]. DTT have higher operating temperatures than oil-immersed transformers, especially the temperatures in the coil [4]. Therefore, the transformer needs thermal image monitoring because its temperature significantly influences its durability, and most transformer faults are accompanied by heating.

The ITSC, the most common transformer failure, often occurs in the winding insulation degradation in a DTT. A huge inrush current or long-term high load increases stress on the insulation layer, resulting in insulation degradation. Loads with high harmonic components are likely to increase copper loss and hysteresis loss, and the temperature will rise more obviously, especially power electronic loads or nonlinear loads. In its early stages, the insulation of the high voltage winding will gradually degrade and have inter-layer discharge [20]. Then, when the winding suffers a short-circuit, the current increases, followed by the temperature, causing the transformer to break down, as shown in Figure 4.

The transformer circuit in Figure 5 explains the reason why heat energy is generated. This circuit simulates a transformer with two turns short-circuited to the primary winding and an equivalent model of each winding. Z_n is the impedance of the nth turn winding. E_n is the induced voltage of the nth turn winding. I_wn is the current of the nth turn winding. R_F is the short circuit resistor. I_loop is the internal loop current generated by the ITSC.

In this equivalent circuit, the I_S current is represented by Equation (1). The voltage relationship within the loop of the short fault can be expressed by Equation (2). When a dead short, R_F will be very close to 0, so it simplifies to Equation (3). This is the magnitude of the inner loop current with only two turns of the coil shorted. The direction of I_S does not pass through E₂. So, short-circuit the winding of n turns, and the loop current is Equation (4).

$$I_S=\frac{V_S-{{\displaystyle \sum }}_{i=1}^m{E}_i}{{{\displaystyle \sum }}_{i=1}^m{Z}_i}+\frac{S_{load}}{V_S}$$

(1)

where m is the number of primary winding coils. S_load is the load capacity of the secondary side.

$$I_{Loop}{R}_F=\left(I_S-I_{Loop}\right)Z_2+E_2$$

(2)

$$I_{Loop}=\frac{I_S{Z}_2+E_2}{Z_2+X_2}\cong I_S+\frac{E_2}{Z_2}$$

(3)

$$I_{Loop,n}=I_S+\frac{{{\displaystyle \sum }}_{i=1}^n{E}_i}{{{\displaystyle \sum }}_{i=1}^n{Z}_i}$$

(4)

Therefore, as the number of short-circuit turns increases, the loop current I_Loop will also increase. The I_S will increase as Z_i and E_i decrease in series. Even if maintained below the rated current of the transformer, the loop current can be very high [35]. Ultimately, line losses cause temperature rises, which can damage transformer insulation and cause fires.

Figure 6 uses infrared thermal imaging to detect the ITSC of a DTT. The short-circuited coil transmits thermal energy, and a noticeable red feature surrounds the coil periphery of the transformer, as shown in Figure 7. Because the dielectric strength of epoxy resin is adequate, the winding will not immediately burn after a short circuit.

3.2. Fault Traces

Through the above, we understand the characteristics and causes of the ITSC fault. Therefore, we use the concept of drawing layers and rules of thumb. We will cover a new transparent layer on the thermal image of the object to be monitored and draw the expected fault on the new layer, which is called the fault trace, as shown in Figure 8. We use GNU Image Manipulation Program (GIMP) software to perform this task.

Observe the characteristics of the ITSC fault on the thermal image, and depict it on the captured image, as shown in Figure 9a,b. The fault traces are separated from the captured image layer. The separately stored fault traces will be saved as the fault data, as shown in Figure 9c. The background image of the fault trace is black, and the black pixel in the image matrix is 0, so it can be overlaid with normal images to produce the fault image, as shown in Figure 10. In order to have enough data for training, the same fault trace can be overlaid on the image at different angles and positions. Following the method mentioned above, training data for different fault types is established. This method can only be used on grayscale thermal images because there is only one channel on the grayscale graph, and it is linear. If it is a color thermal image, you must know the color map of the converted color used by the camera. Or use the regression analysis method to convert back to grayscale, which may make the image slightly distorted.

3.3. Fault Overlay Method

First, the Fault Trace needs to be normalized, and the maximum value of the matrix is 255. Then adjust the intensity of the Fault Trace to simulate the fault from the initial stage to the severe stage. The method of simulating intensity variation is to multiply a variable α which is in a set interval. The range here is set from 0.3 to 0.7, and the interval is 0.1. Therefore, α is 0.3, 0.4, 0.5, 0.6, 0.7.

The fault overlay method concept is shown in Figure 11. Take the ITSC in phase R as an example; there are n thermal images of the Normal Image collected from the device. After one fault trace is multiplied by α to produce m images, a total of n × m fault images can be established through the overlay of the normal image n, and the fault traces m. The overlay method is to add the image matrix as Equation (5), and, after overlaying the image, we set the exceeding value, which is over 255 to 255, Equation (6).

$$\left[Normal Image\right]+\alpha \cdot \left[Fault Trace\right]=\left[Fault Image\right]$$

(5)

$$\left[Fault Image\right]={{\displaystyle \{}}_{\mathrm{pixel}, \mathrm{otherwise}. }^{255, \mathrm{if} \mathrm{pixel} > 255,}$$

(6)

Only one Fault Trace is not enough to express ITSC faults in phase R fault types. By increasing the number of fault traces, the model can learn the characteristics of this fault, as shown in Figure 12. Figure 13 shows the fault traces of the three R, S, and T phases in a sequence in which the thicknesses differ. Therefore, if s fault traces are increased, one fault type will have a total of n × m × s images, Equation (7). The number of Fault Traces must be judged based on experience, and try to summarize the characteristics of this fault.

One phase total fault image = n × m × s

(7)

The fault types in the study contain the ITSC in three phases in all three fault types, as shown in Figure 14. Here, the ITSC is divided into three types: R, S, and T. The purpose is to locate the fault in the transformer.

The received data of infrared radiation is normalized in a range of 0 to 255, representing low to high temperatures. The grayscale image is a 120 × 160 × 1 matrix, while the matrix of the color image is 120 × 160 × 3.

3.4. Image Data Augmentation

There are two parts to image data augmentation. One is the Fault Trace before the overlay step, and the other is the Normal Image and Fault Image as training data.

There is a problem that the training results will overfit when the original image is directly input, so data augmentation is performed on the image data. Furthermore, problems are more prone to occur when the number of fault traces is too small. To avoid overfitting the location of the fault trace from the training model results, we will adjust the fault trace before the overlay. The characteristic of the thermal image of the ITSC is that there is a circle of heating on the mold, so it passes through a little up and down displacement.

The normal image is the data monitored over a period, including the change of the device from low load to high load. This paper discusses the monitoring method of a fixed thermal image sensor, so the fault image recommends a small range of rotation, scaling, and displacement. The above setting parameters are shown in

Table 1. The Data Augmentation Parameters.

Fault Trace
vertical displacement	±5%
Normal Image and Fault Image
rotation	±5°
vertical displacement	±10%
horizontal displacement	±10%
scaling	±10%

.

4. Identification Method Based on CNN

4.1. CNN Model

CNN is a deep learning method that uses the characteristics of learning images as a basis for identification. The advantage of this network is that the design of the convolution is close to the behavior of the human eye to see the feature. Furthermore, it has flexibility in terms of translation, zoom, and tilt. Compared with traditional machine learning, CNN can reduce the use of other algorithms. Usually, the CNN general model mainly takes the input image as a square because it is convenient for model design. The input image in this research is a rectangle of 120 × 160, and we are concerned about carrying out too much resizing and causing distortion. Therefore, two methods are provided, one is to select a larger input model and use a preprocessing method of “filling” the input image, and the other is to establish a rectangular structure that fits the input image. This study adopted the latter option. The architecture of the CNN network in the study consists of the input layer, convolution layer, pooling layer, dropout layer, fully connected layer, and output layer. The input layers contain grayscale images or color images. Convolution layers are used to conduct the convolution on the previous layer’s output using kernel maps, and the feature map can be obtained by Rectified Linear Unit (ReLU) [21]. The pooling layers provide an approach for the down sampling of feature maps. The fully connected layers use ReLU as the activation function. The classification layer classifies the output result by the Softmax function, and the item with the largest weight from the previous layer will be the result. The single fault and the compound fault are marked, respectively. In order to increase flexibility and avoid overfitting, the dropout layer was added to the model of this study [22].

The architecture of the CNN in this study is found in

Table 2. The architecture of the CNN network.

Layer	Output Shape
Input	120 × 160 grayscale image
Conv2D	116 × 156 × 25
MaxPooling2D	58 × 78 × 25
Conv2D	54 × 74 × 50
MaxPooling2D	27 × 37 × 50
Conv2D	23 × 33 × 100
MaxPooling2D	12 × 17 × 100
Flatten	20400
FullyConnected	500
FullyConnected	250
FullyConnected	4
Softmax

. The input image is a 120 × 160 × 1 matrix, representing the length and width of the image and the grayscale channel. The first layer is the convolution layer with 25 kernels; the kernel map is 5 × 5. The second layer is the pooling layer, and the pool size is 2. The third layer is the convolution layer with 50 kernels used to extract 5 × 5 of the feature. The fourth layer is the pooling layer and the pool size is 2. The fifth layer is the convolution layer with 100 kernels used to extract 5 × 5 of the feature. The sixth layer is the pooling layer and the pool size is 2. The seventh and eighth layers are fully connected layers and each has 500 kernels and 250 kernels, respectively. The ninth layer is also the fully connected layer which has four kernels and is used for classification.

Categorical cross-entropy is used as a loss function in the model, and it can classify several types after using the Softmax function. The loss function is shown in Equation (8).

$$loss\left(\widehat{y},p\right)=-\frac{1}{n}{\displaystyle \sum _{i=1}^C{\displaystyle \sum _{j=1}^D{\widehat{y}}_{ij}\log \left(p_{ij}\right)}}$$

(8)

C is the number of categories (normal, and R, S, T phase turn-to-turn short circuits), and D is the number of all data. y_i,j is a binary indicator (0 or 1) from a hot encoding, the j-th data belongs to the i-th true category. p_i,j is the predicted probability that the j-th data belongs to the i-th category. In addition, the optimization algorithm used in the model has the advantage of the convergence characteristic in AdaGrad [23] and the concept of the momentum in the Adam optimizer [24,25].

In order to improve the classification performance of CNN, cross-validation is used to find the parameters with the highest precision. The process of cross-validation is repeated k times, with each of the k subsamples used exactly once as the validation data. The k results can be averaged to produce a single estimation and improve the training efficiency. The k parameter is 10 in the model.

4.2. Grayscale Image and Color Image

The different colors of color images represent different temperatures, while grayscale images only change in shades. The sensor’s measurement is in absolute temperature K. The grayscale image can be obtained by normalization, and the grayscale image can be converted into a color image with a specific color map. Compared to shades of color, human vision more easily recognizes the color difference.

Many types of colormaps can be used to convert grayscale thermal images. Many thermal imaging cameras of different brands have different parameters, but there is no way to obtain the color map conversion parameters. So, we choose colormap with richer colors. Therefore, the study uses the JET color map to convert grayscale images to RGB, as shown in Figure 15.

In the same architecture of the model: The matrix dimension of the grayscale image is width × length × channel, and the channel is one, so it can be regarded as an image in two dimensions. The color image is in RGB and has three channels. For the convolutional filters, the dimension of the kernel maps is two and the RGB is three. For the feature map, the gray map image has one feature map obtained by convolution, and the RGB has three feature maps. Above all, although the training architectures for the grayscale and the color images can be the same, the three channels of the color image can be respectively trained, as shown in Figure 16.

5. Results

This section compares the performance of models generated by the grayscale and RGB imaging and discusses the effectiveness of the diagnostic methodology.

5.1. Training Data with Fault Traces

The thermal image collected by the thermal image camera is processed to obtain the fault image, as shown in

Table 3. The database.

Status	Fault Traces	Total Image	Training Data	Testing Data	Label
Normal	X	90	45	45	1
ITSC (phases R)	17	6120	3060	3060	2
ITSC (phases S)	16	5760	2880	2880	3
ITSC (phases T)	10	3600	1800	1800	4

. There are eight types of faults in the experimental thermal image data. The fault overlaying on the image is simulated with four-degree (0.6,0.8,1,1.2) increments in its intensity. The number of each fault overlaying the ITSC is 17, 16, and 10 for phases R, S, and T, respectively.

5.2. The Classification Performance of the CNN Model

In the experiment, data collection is processed by the fault overlay method to produce fault images, and the images contain 37 types of classifications. Image data is divided into a training set and test set, and the CNN model is used to extract the feature and classify the fault type. In the training process, the CNN model learns the image features of each fault to adjust the training parameters and uses cross-validation to improve the training effect. The batch size is set to 256, and the maximum iteration epoch is set to 40. The situation of each iteration in the training process is shown in Figure 17. The loss can effectively converge within 15 epochs. The validation set also converges during the training process, and there is no over-fitting phenomenon. Therefore, in addition to the good performance on the training set, there are better results on the test set. After training, the test samples are used to evaluate the diagnostic performance of the model. The classification results of the CNN model are shown in

Table 4. The classification results of the CNN model.

	Gray		RGB(JET)
	Accuracy	Loss Rate	Accuracy	Loss Rate
training set	100%	0%	100%	0%
test set	100%	0%	100%	0%

.

The results show that the CNN model can accurately classify the fault features that are difficult to find using the human eye. However, when the ambient temperature is higher than the object, the difference in color intensity between the fault trace and the surrounding is not obvious because the infrared thermal imaging is normalized. Using the new failure data which does not exist in the database is accurately identified by the CNN model, as shown in Figure 18.

5.3. The Classification Performance of the CNN Model with Gaussian Noise

The dropout layer is added to the model to prevent over-fitting the training set during training. According to [11], Gaussian noise is one of the most common interferences in infrared thermal imaging. Therefore, Gaussian noise is added to the training images to simulate this situation, as shown in Figure 19. Using the same architecture of the CNN as mentioned above, the classification results with the Gaussian noise are shown in

Table 5. The classification results of the CNN model after adding Gaussian noise.

	Gray + Noise		RGB(JET) + Noise
	Accuracy	Loss Rate	Accuracy	Loss Rate
training set	99.80%	0.21%	99.96%	0.19%
test set	98.90%	1.12%	99.85%	0.93%

.

The results show that the feature of the thermal image with low-intensity fault traces is difficult to identify after adding the Gaussian noise. The noise may be too large and allows the model to regard the feature as high temperature and cause misjudgment. However, it can be seen that the CNN model has stable recognition ability.

5.4. Other Practical Cases

In another case, a transformer test was performed on a cast resin transformer damaged by a fault. This is an accident caused by an ITSC fault, as shown in Figure 20a. The details of the fault are shown in Figure 20b. The ITSC fault occurred on the left side (T phase). The sensor is placed between the S and T cylindrical casings. To easily monitor the temperature variation, the thermal image screen on the system is displayed in color, and the applied color map is also shown in Figure 15.

In this case, the location of the thermal imager is limited by space, so sensors with different angles are placed. The camera is placed between the S and T phases, and the angle is shot from bottom to top. The Fault Trace drawn by this thermal imager is shown in Figure 21.

This DTT is 5000 KVA, three-phase, 50 Hz, and 11.4KV-480V. The rated current of the primary side is 220A. We take the transformer to short circuit for testing by gradually increasing the input relative ratio’s rated current and adding 5%, 10%, 15%, 20%, 25%, and 30% of the rated current, respectively. Figure 22 shows the measured data.

When 5% input of the rated current, the current is 11 A, and there is no obvious point in the thermal image, as shown in Figure 23a. When the rated current is 10%, the current is 22 A, and some features are visible in the thermal image, but are not obvious, as shown in Figure 23b. When the rated current is 15%, the current is 33 A. It can be seen that some coils have uneven temperatures, as shown in Figure 23c. When the rated current is 20%, the current is 44 A, and the fault characteristics of the ITSC can be clearly seen as shown in Figure 23d. When the rated current is 25% and 30%, the current is 55 A and 66 A. It can be clearly seen that the fault characteristic is an arc, as shown in Figure 23e,f.

The designed CNN model can detect faults at 15% of the rated current. We observe the time curve at one temperature point on the fault feature, as shown in Figure 24. From Figure 24, it can be known that the temperature of the turn-to-turn short circuit is actually not high in the low load stage, so it is hard to detect using the general detection method.

In fact, factors such as emissivity, reflected temperature, humidity, etc., need to be considered in order to be precise when using a thermal imager. However, this study has low dependence on the accuracy of the image temperature, and we mainly focus on the characteristic shape of the image. Therefore, reducing the overall execution complexity is necessary. In addition, this method benefits not only DTT, but also the high-voltage equipment because the thermal characteristics of high-voltage equipment are often a precursor to failure, which is of great significance for improving the intelligence level of substations.

6. Conclusions

The ITSC fault is a common problem for high voltage DTT. This fault type has sufficient thermal characteristics. Therefore, this research proposes fault diagnostics based on the fault overlay method and CNN. The primary purpose of the fault overlay method is to have enough training samples and to cover the possibility of various fault features. We use the normal transformer’s thermal imaging as the basis and add fault traces to establish the fault database. The fault classification is based on the deep learning architecture of CNN.

The high voltage DTT is suitable for the fault diagnosis method of fixed online monitoring to achieve pattern recognition. The results show that the fault overlay method effectively simulates the fault on the original thermal image and is based on the rule of thumb of equipment fault characteristics. The classification performance of the CNN model is satisfactory, even if the noise is added. Internet of Things and labor cost reductions will be the trend in the future. This research will help to improve the efficiency of online infrared fault diagnosis.