Helping the future of green energy, National Taiwan University of Science and Technology uses GPU to improve the performance of AI fault detection of wind turbines

Wind power is a widely used renewable energy source with many advantages, including low carbon emissions and no pollution. However, wind power also has some challenges, including the difficulty of maintenance and upkeep. Wind turbines are made up of complex components that are susceptible to environmental factors, such as high temperatures and humidity in offshore areas, which can lead to malfunctions.

The Artificial Intelligence Power System Laboratory at National Taiwan University of Science and Technology is using the GPU performance of the NVIDIA RTX A4000 professional graphics card to accelerate the development of AI fault detection solution for three-phase motors. The aim is to shorten the time required for detecting faults in large motors, quickly diagnose the cause of the faults, and assist wind power operators in maintaining and operating wind turbines more effectively, thereby promoting the sustainable development of wind power generation.

AI fault detection based on vibration signal analysis

Three-phase induction motors are commonly used in industrial and commercial applications and are widely used in wind power generation. Fault detection is an important aspect of maintaining three-phase induction motors because faults can not only lead to reduced equipment efficiency but also pose safety concerns. The Artificial Intelligence Power System Laboratory at National Taiwan University of Science and Technology is seeking new AI detection methods to accelerate fault detection accuracy and speed up troubleshooting.

After trying voltage, current, and other methods, the vibration signal is currently set as the research direction of the new generation of detection. Since specific vibration patterns are generated when motors malfunction, researchers use image recognition technology to convert motor vibration signals into waveform diagrams and diagnose problems using deep learning models, which can efficiently classify issues including rotor faults, stator faults, bearing faults, and misalignment faults.

Multiple challenges: Difficult to access model training information and insufficient computing power

Photo: Graduate student of Artificial Intelligence Power System Laboratory, National Taiwan University of Science and Technology

The research team faced numerous challenges in the early stages of their research. First of all, due to the limited number of motor fault signals accumulated from large wind turbines, the model's classification performance on most existing waveform label data was poor, making it difficult to train the AI model. After considering multiple factors, the Artificial Intelligence Power System Laboratory at National Taiwan University of Science and Technology decided to use transfer learning to solve this problem.

Transfer learning can use a pre-trained source model to transfer the learned knowledge to another similar task, to increase the training speed of the new model, reduce time costs, and achieve research goals faster. In this case, researchers can train the model on three-phase induction motors with different horsepower and perform cross-domain training and fault diagnosis between different loads and capacities to increase the accuracy of fault detection.

However, as AI models become increasingly complex, the computational demands also increase. A graduate student from the Artificial Intelligence Power System Laboratory at the National Taiwan University of Science and Technology shared that ResNet, a commonly used deep learning model for image recognition, extracts useful features through multiple layers of processing, typically requiring significant computational power. Additionally, during research, there are multiple steps such as repeatedly measuring data, labeling features, and training models, resulting in large amounts of data and computation, and taking a considerable amount of time.

Speed up the training of new models with GPU acceleration

A graduate student from the Artificial Intelligence Power System Laboratory of National Taiwan University of Science and Technology added: To address these challenges, we use NVIDIA GPUs. Compared with integrated graphics on a computer, using the NVIDIA RTX A4000 professional graphics card, a model that previously took 1 hour to train now only takes 1 minute. Its Tensor core and 16GB GDDR6 memory support most calculations. After evaluation, we found that the performance has improved by more than 15 times.

GPUs are highly efficient in compute-intensive tasks. NVIDIA RTX A4000 with NVIDIA Ampere GPU architecture has 6144 NVIDIA CUDA cores and 192 NVIDIA Tensor cores, providing efficient AI acceleration capabilities. It also comes with 16GB of GDDR6 memory with up to 448GB/s bandwidth, providing powerful computing capabilities. It's worth noting that NVIDIA RTX A4000 is a single-slot GPU suitable for various workstations and server systems. In addition, the maximum power consumption of RTX A4000 is 140W, and its low power design is more suitable for long-term research work than general gaming graphics cards.

As for the after-sales service that users care about most, Leadtek provides an exclusive 3-year warranty, and has established a global GPU technical consulting center, so that users from all walks of life can enjoy the most professional and reliable services and AI computing performance.

Photo: Leadtek NVIDIA RTX A4000 with 16GB GDDR6 memory and third-generation Tensor core