A New Hydrogen Sensor Fault Diagnosis Method Based on Transfer Learning With LeNet-5.

Area of Science:

• Sensor technology
• Machine learning
• Artificial intelligence

Background:

• Effective fault safety monitoring of hydrogen sensors is crucial for reliable operation.
• Traditional machine learning methods require extensive fault data, which is difficult to obtain in complex, real-world environments.
• Environmental factors like temperature, humidity, shock, and vibration complicate sensor fault data acquisition.

Purpose of the Study:

• To propose a novel transfer learning (TL) method using LeNet-5 for improved hydrogen sensor fault diagnosis.
• To address the challenge of limited fault data in complex operating conditions.
• To enhance the accuracy and reliability of hydrogen sensor fault detection.

Main Methods:

• LeNet-5 was initially trained on a data-rich dataset of gas sensor faults in a normal environment.
• Transfer learning (TL) was employed to migrate the learned parameters from the normal environment task to a complex environment task.
• The migrated LeNet-5 model was then used for fault diagnosis with limited data in a complex environment.
• Experimental verification was conducted using a prototype hydrogen sensor array.

Main Results:

• Traditional LeNet-5 achieved a fault diagnosis accuracy of 88.48 ± 1.04%.
• The proposed TL with LeNet-5 method achieved a significantly higher accuracy of 92.49 ± 1.28%.
• The results demonstrate the effectiveness of TL in handling limited fault data under complex conditions.

Conclusions:

• Transfer learning with LeNet-5 offers an excellent solution for hydrogen sensor fault diagnosis with minimal data in complex environments.
• The proposed method enhances the safety and practicality of hydrogen sensor applications.
• This approach overcomes limitations of traditional methods in dynamic and challenging operational settings.