Designing an AI-based Diagnostic Platform for Heat Transfer Pipelines

Image and Signal Sensor Data Processing Skills
Enhanced ability to process various data modalities by developing a model that diagnoses the risk of underground pipelines using thermal/RGB image data and vibration sensor data.

Application of AI to Industrial Problems
Demonstrated the potential for applying AI technology to diagnose the state of heat transfer pipelines accurately and detect anomalies, showing its applicability in diagnosing various types of infrastructure.

Introduction

Starting with district heating energy projects in the 1980s to improve resource utilization efficiency, more than 2,000 km of heat transfer pipeline networks were established nationwide as of 2021.
As of 2020, 26% of the national heat transfer pipelines are over 20 years old, posing safety concerns due to the unpredictability of pipeline ruptures.
However, due to a shortage of skilled personnel and limitations in inspection technologies, there is a growing need to develop artificial intelligence techniques to complement these shortcomings.

Develop an AI algorithm that diagnoses heat transfer pipeline anomalies based on more than two types of data (image and sensor data).
Develop a model that enhances the efficiency of detecting high-risk areas and diagnosing anomalies in heat transfer pipelines in real-time.

Thermal/RGB Image Data Collection
- Collect buried pipeline data from six cities in South Korea using RGB and thermal cameras.
- Data was gathered around the buried pipelines at the edges of roads, and labels were created by integrating with pipeline IDs.
- Labeled data with risk levels such as normal, moderate risk, and leakage and created overlap images linked with underground maps.

Data Composition
- Built a dataset of approximately 10,000 pairs of thermal and RGB video data.
- Created ground truth labels indicating the presence of heat and leakage.

Sample dataset

A U-Net-based model consisting of a thermal image encoder and an RGB image encoder.
Utilized Depthwise Separable Convolution to reduce computational complexity while maintaining performance.
Detected suspected risk areas and classified actual risk levels using thermal and RGB data, with risk detection performed through an EfficientNet-based encoder and a Residual Block-based decoder.

Structure of image-based detection model

Both IoU (0.8373) and Dice (0.9024) scores surpass the performance of existing models.
Produces more precise image segmentation results and demonstrates robustness against false recognition.
Unlike baselines that generate inaccurate segmentation results for images without risk areas, this model accurately identifies normal areas with no risk.

Risky area segmentation examples

Risky area segmentation result

DenseNet 201 and EfficientNet B7 recorded high performance in the heat and leakage detection stages.

- Sensor data can detect finer cracks compared to image data.
- ACC and AE sensors were attached to heat transfer pipelines, and datasets were collected by simulating the actual buried environment.

ACC Data: 2,880 normal cases, 4,883 leakage cases.

AE Data: 540 normal cases, 1,078 leakage cases.

Segmented sensor data at 2-second intervals.
Used MFCC (Mel-Frequency Cepstral Coefficients) for feature extraction in signal processing.

Examples of sensor dataset

Combined CNN and Bi-LSTM to learn the time-frequency characteristics of vibration data.
Applied Residual Block to improve learning efficiency.
Utilized 12 layers with residual blocks.
Optimized using Binary Cross Entropy loss function and Adam optimizer.

Overview of sensor-based detection model

Achieved 99.93% accuracy in abnormal detection.
Recorded high performance in Precision, Recall, F1-Score, and Accuracy, making it suitable for real-time detection.

Developed an AI model for diagnosing anomalies in heat transfer pipelines using thermal and sensor data.
Demonstrated high detection performance by applying the model to actual heat transfer pipeline environments.
Due to data collection limitations, experiments were conducted on specific conditions of heat transfer pipelines.
Further research is needed to generalize the model across diverse environments.
Enhanced field applicability through the development of real-time diagnostic and alert systems

Page updated

Google Sites

Report abuse