Research On Fault Diagnosis Methods Of Semiconductor Hydrogen Sensor

Hydrogen is considered to be an important clean source of energy,which is widely used in the fields of metal smelting,building heating,urban gas,or other trades.Recently,applications of hydrogen in power industries,such as automobile power,ship power,and aviation power,have been emerging and shown a great development trend.Hydrogen is characterized by colorless,tasteless,flammable and explosive.When the concentration of hydrogen leaked in the air reaches the range of 4 % to 75 %,it may explode easily.The semiconductor hydrogen sensor could efficiently detect the hydrogen concentration in the atmosphere,also could display concentration value and alarm through the secondary instrument,which is necessary for the safe use of hydrogen.However,in engineering practice,due to the influence of temperature,humidity,vibration,impact,wind speed,harmful gases,and other factors in the environment,semiconductor hydrogen sensor is susceptible to degradation in sensitivity or further breakdown.It is necessary to study the methods of sensor fault diagnosis in depth,thereby ensuring that the sensors could work reliably.Therefore,in this thesis,we focus on the semiconductor hydrogen sensor,and systematically discuss the sensor fault diagnosis methods under the conditions of steadystate,non-steady state and unbalanced fault signal samples.The specific works are as follows:(1)To verify the validity and advantage of the sensor fault diagnosis methods proposed in this study,we firstly analyze the performance,mechanism and failure modes of the sensors,summarize the fault types of the sensors,and design and build an experimental device for obtaining fault signal data of hydrogen sensors.Then,the faulty sensors are put into the gas chamber,and are injected hydrogen.The fault signal data is collected from the experimental device under different conditions.Finally,the obtained fault signal data is used for the experimental verification of the proposed algorithm.(2)Under steady-state conditions,there are problems in the traditional sensor fault diagnosis methods,such as the separation of feature extraction and pattern recognition.The fault diagnosis accuracy is affected by the quality of feature extraction,so it is quite essential for improvement of diagnosis accuracy.Therefore,this study proposes a fault diagnosis method based on the improved convolutional neural network in the data-driven aspect.This method uses gray image conversion to preprocess the data,and retains the original features of the fault signal data,thereby decreasing the influence of interference information,such as noise.Furthermore,the improved convolutional neural network shows better robustness and generalization ability.The kernels and feature maps of each convolutional layer of convolutional neural network are visually analyzed.The experimental results show that the fault diagnosis accuracy of this method is increased by more than 5.31% compared with the fault accuracy of traditional fault diagnosis methods,such as support vector machine and K-nearest neighbor.(3)Owing to the limited data of sensor fault signal under non-steady state conditions,the ability of convolutional neural network to learn signal features is restricted,which in turn affects the accuracy of fault diagnosis.Based on the view of knowledge transfer,our study proposes a sensor fault diagnosis method featured by cross-distribution of convolutional neural network.This method adopts the knowledge transfer mechanism to transfer the trained convolutional neural network from steady-state conditions to nonsteady state conditions for fault diagnosis.The algorithm proposed in this study not only improves the accuracy of fault diagnosis under non-steady state conditions,but also accelerates the convergence speed of the training accuracy or loss rate of the convolutional neural network algorithm.Therefore,it can reduce the number of training iterations,and overcome the need for a large number of training samples of fault signal data with labels.The experimental results show that the fault diagnosis accuracy of the sensor under non-steady state conditions is increased by 5.14% compared to that of the convolutional neural network without cross-domain distribution.(4)Aiming at the problem of limited data samples of heating wire virtual welding fault signal,this paper proposes a sensor fault diagnosis method based on deep convolutional generative adversarial network from the perspective of data enhancement.This method uses a deep convolutional generative adversarial network to learn the characteristics of the fault signal data of small samples,and generates artificially synthesized data samples.The similarity of sample generation is improved by analyzing the feature growth of the generated pictures and setting the MMD threshold.This method could efficiently expand small sample data sets.The experimental results show that the fault diagnosis accuracy of this method is increased by more than 9.9% compared to the traditional artificial intelligence fault diagnosis methods,such as random forest and support vector machine.