AI application in the detection of power lines
(segmentation-based approach), power poles (object-detection-based approach),
and abrupt electrical discharge aka electric-field anomalies or sparks,
respectively.
Introduction
Visual
inspection of power transmission & distribution networks is often carried
out by various Utility companies regularly, to maintain the reliability,
availability, and sustainability of electricity supply. Since the inception,
the most widely used techniques for inspection are the manual inspection which
includes either foot patrol or/and conventional helicopter-based inspection.
However, due to the recent widespread application of Drones/Unmanned Aerial
Vehicles (UAV), an increase in demand to automate the visual inspection of the
infrastructure is observed.
In
this article, we have described the two primary approaches required in the
detection of two primary types of power distribution infrastructure – Power
Lines and Power Poles. Besides, we have also described our approach to detect
anomalies in power poles.
Needless
to say, before any detailed inspection, the detection of the power
infrastructure is mandatorily required. In total, two models were developed,
the segmentation model for power lines (cables) and the object detection model
for power poles. In addition to these, for anomaly detection on power poles, we
have trained our algorithm to detect electrical discharge or sparks. The
detailed approaches used for segmentation, object detection, and anomaly
detection are described in the following section.
Powerlines
Segmentation
The
Powerlines Segmentation is used to separate the powerlines (the power
transmission cables) from the entire image so that an inspection algorithm can
just focus and analyze only the power lines while staying unaffected by the
surrounding or the background of the powerlines. In the following section, a
technical overview of the Deep Learning method used to create the solution for
powerlines segmentation is presented.
·
An Overview of the Network Architecture
The
Nested U-Net Network used for powerline segmentation is quite similar to the
original Convolutional Neural Network (CNN) architecture based conventional
U-Net Network. It uses the Dense block of DenseNet Network to improve the
overall performance of the U-Net Network. It consists of convolutional layers
on skip pathways which bridges the semantic gap between encoder and decoder
feature maps, thus aiding in improving the gradient flow.
·
Datasets
·
Dataset 1: We have obtained 200 Visible Light (VL)
spectrum images from the https://data.mendeley.com/datasets/twxp8xccsw/1. This dataset is provided
for segmentation purposes. It consists of 200 images of size 512×512, along
with the binary wired image masks for all the input images.
·
Dataset 2: The other set of images was obtained
from
https://data.mendeley.com/datasets/n6wrv4ry6v/8. This dataset has a total
of 2000 Visible Light (VL) spectrum images of resolution size 128×128.
·
Dataset Preparation
·
Dataset 1: This dataset consists of binary masks,
so there was no need to generate masks. Unfortunately, this dataset lacks a
substantial count of images or samples necessary for training the model. To
counter such a problem we used data augmentation techniques such as image flip,
image rotation, etc. to generate substantial data synthetically.
·
Dataset 2: Images in this dataset are low-quality.
Hence, we used image super-resolution to improve image quality
further. All the images are super-resolved to 512×512 image size so that we can
extract more features (i.e. additional details about power lines) necessary for
effective learning. For this dataset, a VGG annotator has been used to annotate
images. It gives coordinates of wires in images. Based on these coordinates,
the binary ground truths are generated. Further, data augmentations were
carried out on Dataset 2 same as Dataset 1.
·
Training and Segmentation Methodology
Nested
U-Net architecture, as the name implies, makes use of nested and dense skip
connections between encoder and decoder apart from the typical skip connection
used in U-Net Network. A dense skip connection is used to improve the flow of
the gradient. Nested U-Net Network consists of a dense convolution block
helpful in collecting the semantic level of the feature map from the encoder
part. The Nested U-Net is trained for 200 epochs with a Step size = 20, and the
Learning rate = 0.00001. Additionally, Adam optimizer is used. The powerline
dataset along with binary masks was prepared. CNN is used for training
purposes. Images and its ground truths were inputted to the CNN and predicted
masks were obtained from the trained model.
·
Sample Input Images: Some Sample input images from our
dataset are shown here. These contain the power lines which are captured using
the UAV mounted cameras.
·
Sample Output Images: The predicted power
lines using our model weights are shown here.
·
Evaluation Metric
The
evaluation metric used in the performance evaluation of the segmentation
solution is the Dice Coefficient. The Dice coefficient is very similar to the
IoU. Both are positively correlated, meaning if one says model A is better than
model B during image segmentation, then the other will say the same. Like the
IoU, they both range from 0 to 1, with 1 signifying the greatest similarity
between predicted and truth. Dice Coefficient is formulated as,
Electric Poles
Detection
A
system that can recognize an electric pole in a given image and further
runs an analysis over the detected pole would form the basis of the power
distribution pole inspection solution. Our currently developed solution uses a
“Convolutional Neural Network” architecture to detect power distribution poles
in the image. To develop this solution, it requires a set of data consisting of
images of poles from various angles. The images are then labelled using a tool
that draws a rectangle around the image where the pole is located. These
labelled images are resized to a standard size before training the neural
network. The neural network trained on these images can successfully detect
poles on images kept as testing data, which was a subset of the total labelled
dataset.
·
An Overview of Network Architecture
We
have used YOLOv3, a Darknet detector for power pole detection as it makes
predictions with single network evaluations. YOLOv3 is the faster object
detection algorithm, an optimal choice for real-time detections without much
loss of accuracy. Darknet used here is a variant of Darknet53, which originally
has 53 convolutional layers, but we have used 106 convolutional layers in our
detector which helps to improve the performance. Another advantage of our
system is that it detects better on smaller electrical poles as well.
·
Dataset
To
develops a deep learning model for detecting poles from the street view, we
collected the dataset for the same. We gathered the dataset from an open-source
Github repository https://github.com/kabrabharat/Electric-Pole-detection-using-darknet/tree/master/dataset. It has 100 images with different
image sizes.
·
Data Preparation Methods
The
Data preparation here includes Image resizing. All the images were resized to a
size of 432×768. For the training purpose, we need to have an annotated
dataset. Image Annotations (ROI) i.e Poles from Images were done using an
open-source VGG annotator to get coordinates of bounding boxes xmin, ymin,
xmax, and ymax. These are Normalized and converted into the coordinates x, y,
w, h which is a format to train YOLOv3. Each image should have a corresponding text
file with the same file name as that of the image in the same directory.
The generated text should be in the format:
·
Data Augmentation
Unfortunately,
this dataset lacks substantial count of images or samples necessary for
training the model. To counter such problem we used data augmentation technique
such as image flip to generate substantial data synthetically, along with its
annotated CSV files.
·
Training and Detection Methodology
YOLOv3
is a state-of-the-art, real-time object detection system. It looks at the whole
image at the time of test, so its predictions are informed in the global
context of the image. It also makes predictions with a single network
evaluation which makes it extremely fast vis-à-vis R-CNN and Fast R-CNN. The
Darknet, an open-source neural network framework is used to train the detector.
For this, we have used electrical poles annotated images and corresponding text
files consisting of normalized coordinates of bounding boxes to train the
Darknet model. Predicted pole images were obtained using the trained models’
weight file.
·
Results of Electric Pole Detection
Sample
input images, containing electric poles, from our dataset, are presented in
images below. Also, the predicted power poles using our model weights are
presented in the same.
·
Performance Evaluation
We
have evaluated the performance of the object detection solution against the
entire prepped test dataset. Further, for this whitepaper, we have randomly
chosen some images and focused on the confidence of the bounding boxes found by
our model. Following are the results, vis-à-vis the confidence of the bounding
boxes.
Anomaly
Detection: Abrupt Electrical Discharge or Sparks on Power Poles
The
utility poles of electric power distribution lines are very vulnerable to many
natural hazards, while power outages due to pole failures can lead to adverse
economic and social consequences. Utility companies, therefore, need to monitor
the conditions of poles regularly and predict their future conditions
accurately and promptly to operate the distribution system continuously and
safely. Our system presents a novel pole monitoring method that uses computer
vision methods to meet the need. It processes the input image which is captured
by any angle, detects the power pole at any distance and at lastly classifies
those poles into two classes: Sparking near Pole, No sparking.
·
Training and Spark Detection Methodology
The
Power Poles were detected using pole detection algorithms. Those detected poles
are cropped and passed as input to the spark detection system. For spark
detection, our system architecture uses Computer Vision approaches for noise
removal, binarization and detection of regions of interest. Further, we go for
the image binarization to get the ROI based on intensity values. As we get ROI
we use noise removal techniques again to reduce the noise. As we find out the
spark in between the wires and pole, our system marks it as an anomaly.
·
Results of Spark Detection
In
the following figures, it can be seen that the Red Bounding Box around the Pole
indicates there is sparking near to the pole and Green box indicates no
sparking.
Conclusion and
Future Work
The
electrical power lines, poles, and electrical discharge detection presents a
starting point for the development of the completely automated inspection
solution. The ability to do more targeted inspections in the future will allow
Utility companies to go out and conduct maintenance efficiently as well as in a
more targeted fashion. Currently, we are engaged in further improving in-house
developed neural network models. Also, in parallel, we are working on expanding
Utility solutions by adding more electrical discharge anomaly solutions. Our
goal is to build an AI governed Utility suite which can be used by the Utility
companies in order to inspect their assets. In short, with the main focus on
safeguarding the world’s energy infrastructure, the following are the prime
objectives we are working and tend to include in upcoming days:
·
Inclusion
of different types of anomaly detection such as broken power lines, bent
electrical poles, multiple objects hanging on the overhead power lines etc.
·
Improving
the detection performance further by experimenting with various algorithms for
object detection and segmentation.
·
Improving
performance by collecting more diversified training data.
·
Improving/Optimizing
the algorithm and making it more robust by adding more diversified data.
·
Experimenting
in miniaturizing these algorithmic models so that they can be ported easily
onto the laptops, and there will be no dependency on high-performance servers.
