KR-20260062556-A - A method for diagnosing transformer faults using thermal imaging and time-series data, and a multimodal system utilizing the same

KR20260062556AKR 20260062556 AKR20260062556 AKR 20260062556AKR-20260062556-A

Abstract

A transformer abnormality detection device is disclosed. The transformer comprises an OLTC (On Load Tap Changer), a housing equipped with an optical window capable of photographing the interior of the transformer, and a set of sensors installed in the transformer to output a time-series signal regarding the state of the transformer. The transformer abnormality detection device comprises a non-volatile storage device; a processing unit including a processor; and a thermal imaging camera that provides an output image of the interior of the transformer, which is optically exposed through the optical window. The processing unit is configured to read and execute a first set of instructions from the non-volatile storage device to determine whether there is an abnormality in the transformer by analyzing the output image, and is configured to read and execute a second set of instructions from the non-volatile storage device to determine whether there is an abnormality in the transformer by analyzing a set of time-series signals obtained from the set of sensors.

Inventors

김호철
김동현
황호성

Assignees

을지대학교 산학협력단

Dates

Publication Date: 20260507
Application Date: 20241029

Claims (10)

A transformer abnormality detection device configured to detect abnormalities in the transformer, The above transformer includes an OLTC (On Load Tap Changer), a housing equipped with an optical window capable of photographing the interior of the transformer, and a set of sensors installed in the transformer and configured to output a time-series signal regarding the state of the transformer. The above transformer abnormality detection device includes a non-volatile storage device; and a processing unit including a processor. The processing unit is configured to read from the non-volatile storage device and execute a first set of commands configured to determine whether there is a state abnormality of the transformer by analyzing an output image output by a thermal imaging camera that provides an output image of the interior of the transformer optically exposed through the optical window, and to read from the non-volatile storage device and execute a second set of commands configured to determine whether there is a state abnormality of the transformer by analyzing a set of time-series signals acquired from the set of sensors. Transformer abnormal condition detection device.
In paragraph 1, When the above second set of instructions is executed, the step of inputting the above one set of time series signals into a deep learning network of the input-based technology; and the step of the deep learning network outputting whether there is a state abnormality of the transformer; are executed, and The deep learning network mentioned above is, N stacked residual blocks, wherein each of the above residual blocks is configured to output an output head and a residual head based on input information input to itself, said N residual blocks (N is a natural number greater than or equal to 2); N attention blocks, each receiving the output head output by the N residual blocks; Sum operation layer; and A convolution layer that receives the output feature map of the above sum operation layer and outputs the output data of the above deep learning network; Includes, The output head output by the k-th residual block is input to the k-th attention block (k=1, ..., N), the feature map output by the p-th attention block is input to the (p+1)-th residual block (p=1, ..., N-1), and the N residual heads output by the N residual blocks and the feature map output by the N-th attention block are input to the sum operation layer. Transformer abnormal condition detection device.
In paragraph 2, Each of the above-mentioned attention blocks includes: a compression operation unit that performs a compression operation on data input to the attention block; and an excitation operation unit that performs an excitation operation on data output by the compression operation unit. The above compression operation unit is configured to compress a feature map of size HxWxC input to the attention block into a vector of size 1x1xC through a global average pooling operation, and The above readjustment operation unit is configured to normalize the above 1x1xC size vector, assign weights to it, and output an HxWxC size feature map as the output value of the above attention block. Transformer abnormal condition detection device.
In paragraph 1, The above processing unit is, As a result of reading and executing the first set of instructions from the non-volatile storage device, a value is generated indicating that there is a state abnormality of the transformer, or If, as a result of reading and executing the above second set of instructions from the above non-volatile storage device, a value indicating that there is a state abnormality of the transformer is generated, Determined to determine that there is a status abnormality in the above transformer, Transformer abnormal condition detection device.
In paragraph 1, The above set of time series signals is, Time series data measured and output by the voltage sensor installed in the above transformer, Time series data measured and output by the current sensor installed in the above transformer, Image data output by the thermal imaging camera capturing the above OLTC, and Time series data measured and output by the vibration sensor installed in the above transformer, Composed of one or more of the following, Transformer abnormal condition detection device.
In paragraph 1, When the above first set of instructions is executed, The transformer abnormality detection device generates a 2D token map (510) using data converted from an output image output by the thermal imaging camera; The transformer abnormality detection device comprises the step of converting the 2D token map to generate a plurality of images (511, 512, 513) having different dimensions; The transformer abnormality detection device comprises the step of enhancing the contrast of each of the plurality of images to generate a plurality of contrast-enhanced images (511', 512', 513'); The transformer abnormality detection device comprises the step of generating a contrast-enhanced 2D token map (510') by combining the plurality of contrast-enhanced images; The transformer state abnormality detection device comprises the step of generating an adjusted token map (510``) from the comparison-enhanced 2D token map using the convolution token embedding layer (30) of a predetermined convolutional vision transformer; The transformer state abnormality detection device uses the convolutional transformer block (40) of the convolutional vision transformer to generate a 2D token map (520) for another stage from the adjusted token map; and The transformer abnormality detection device analyzes the output image based on a 2D token map (520) for the other stage; Established to determine whether there is a status abnormality of the above transformer by executing Transformer abnormal condition detection device.
In paragraph 1, When the first set of instructions above is executed, a method for converting a 2D token map into a convolutional vision transformer including a convolutional token embedding layer (30) and a convolutional transformer block (40) is executed, and The method for converting the above 2D token map is, A step of converting a 2D token map (510) to generate multiple images (511, 512, 513) having different dimensions; A step of generating a plurality of contrast-enhanced images (511`, 512`, 513`) by enhancing the contrast of each of the plurality of images; A step of generating a contrast-enhanced 2D token map (510') by combining the above plurality of contrast-enhanced images; The step of the convolution token embedding layer (30) generating an adjusted token map (510``) from the contrast-enhanced 2D token map; and A step in which a convolutional transformer block (40) generates a 2D token map (520) for another stage from the adjusted token map; including, Transformer abnormal condition detection device.
In paragraph 6, The step of generating the above-mentioned multiple contrast-enhanced images is, A step of generating a first weighted image (I2) by multiplying a predetermined weighting matrix (W) on a pixel-by-pixel basis by one selected image (I1) among the plurality of images; A step of generating a second weighted image (J2) by multiplying the inverse matrix (1-W) of the weighting matrix pixel by pixel by the corresponding image (J1) obtained by enhancing the contrast of the underexposed area of the selected image (I1); and A step of generating a contrast-enhanced image for the selected single image by adding the first weighted image (I2) and the second weighted image (J2) in pixel units; including, Transformer abnormal condition detection device.
In Paragraph 7, The above plurality of images are composed of a first image (511), a second image (512), and a third image (513), and The width dimension and height dimension of the second image are each half the width dimension and height dimension of the first image, and The width dimension and height dimension of the third image are each half the width dimension and height dimension of the second image, and The above plurality of contrast-enhanced images are composed of a contrast-enhanced first image (511'), a contrast-enhanced second image (512'), and a contrast-enhanced third image (513'). The width dimension and height dimension of the first image, which is enhanced compared to the above, are each the same as the width dimension and height dimension of the first image, and The width dimension and height dimension of the second image, which is enhanced in comparison to the above, are each the same as the width dimension and height dimension of the second image, and The width dimension and height dimension of the third image, which is enhanced in comparison to the above, are each identical to the width dimension and height dimension of the third image, Transformer abnormal condition detection device.
In Paragraph 7, The step of generating a 2D token map improved compared to the above is, A step of generating a first set of tensors (11, 12, 13) by performing a 2D convolution operation on each of the plurality of contrast-enhanced images; A step of generating a second set of tensors (21, 22, 23) by performing a ReLU operation on each of the first set of tensors; A step of generating a third set of tensors (31, 32, 33) corresponding to the second set of tensors; and A step of concatenating the above third set of tensors to generate the above contrast-enhanced 2D token map (510'); Includes, The number of tensors constituting the third set of tensors and the number of tensors constituting the second set of tensors are the same, and the dimensions of any one tensor among the third set of tensors and the corresponding tensor among the second set of tensors are identical to each other, and any one tensor among the third set of tensors is obtained by performing an element-wise sum operation on the tensors obtained by 2D convolution of each of the second set of tensors, and For a tensor having a dimension smaller than any one of the tensors in the second set above, an upsampling operation and a 2D convolution operation with a stride of 1 are executed sequentially, and For tensors among the second set of tensors that have a dimension larger than any one of the tensors, a 2D convolution operation with a stride of 2 or more is executed, and A 2D convolution operation with a stride of 1 is performed on a tensor having the same dimension as any one of the tensors in the second set above. Transformer abnormal condition detection device.

Description

A method for diagnosing transformer faults using thermal imaging and time-series data, and a multimodal system utilizing the same The present invention relates to a technology for diagnosing transformer failures using internal images of the transformer captured by a thermal imaging camera and time-series data obtained from separate sensors installed on the transformer. Thermal Imaging-Based Equipment Condition Monitoring Technology Thermal imaging cameras can capture temperature data without physical contact with equipment, thereby reducing the risk of mechanical damage or interference during inspections. A high level of inspection efficiency is ensured as the visual representation of temperature distribution allows for the easy identification of potential equipment issues, such as hot spots, heat leaks, or abnormal thermal patterns. Furthermore, the ability to detect abnormal temperatures at an early stage facilitates preventive maintenance, potentially reducing costly downtime. However, there is a drawback in that interpreting thermal imaging data requires specialized knowledge and experience. Additionally, the spatial resolution of thermal imaging cameras is often limited, making it difficult to accurately identify minute defects or detailed anomalies. Distinguishing the precise shape of parts can be particularly challenging when diagnosing small or complex equipment components. Nevertheless, thermal imaging-based equipment condition monitoring systems offer various advantages compared to monitoring systems utilizing other sensory information, such as pressure, vibration, flow rate, and acoustics. Equipment fault detection technology utilizing thermal imaging can be used in various fields, such as fault detection in photovoltaic (PV) power plants, fault detection in BLDC motors, fault detection in building exterior walls, fault detection in mine air compressor pipelines, fault detection in electric heating devices during railway transport, fault detection in power transformers overheating, and fault detection in belt conveyor idlers. The spatial resolution of thermal imaging cameras is inherently limited, which can make it difficult to detect small defects or diagnose anomalies in their early stages. Against this backdrop, it is necessary to enhance the contrast of the input image to address the loss of edge information in thermal images. Deep learning network technology based on brute-force techniques suitable for multivariate sensor data-based anomaly detection In actual industrial systems, the current state of the system is monitored using numerous sensors that generate significant amounts of time-series data during factory operations. By performing outlier detection on this multivariate time-series data, failures can be identified immediately and malicious attacks prevented, thereby ensuring the stable operation of industrial systems. Outlier detection is the process of identifying data patterns that deviate significantly from the overall data structure, and the detection of unexpected behavior in multivariate time-series data, where values change over time, is referred to as outlier detection in multivariate time series. Anomaly detection based on data acquired from current, voltage, and vibration sensors installed on transformers within multivariate time series aims to identify anomalous patterns that deviate significantly from the main segments of the data. This is primarily used to detect anomalies in industrial systems or monitor data errors in large-scale datasets. Recently, deep learning-based technologies have achieved superior performance in anomaly detection and are receiving significant attention from both academia and industry. These technologies can be summarized into two main categories: prediction-based methods and reconstruction-based methods. The former constructs a prediction model that infers future data using historical data and determines anomalies based on the prediction error between the estimated and actual values. When using prediction-based methods, historical information is utilized to predict future values of the time series, and the prediction error is used as an indicator for anomaly detection. However, since future time series values can exhibit high levels of uncertainty and variability, accurately predicting them in complex real-world systems can be inherently difficult. Relying solely on prediction-based methods can have a detrimental effect on anomaly detection performance. In contrast, the reconstruction-based approach reconstructs test data based on training data and detects anomalies based on reconstruction errors. When using reconstruction-based methods, the entire time series is encoded into an embedding space, and anomaly labels are inferred based on the reconstruction error between the original data and the reconstructed version. Since these approaches must operate on and reconstruct the entire time series, performance largely depends on the capabilities of t