Research On Fast Measurement With Chemical Sensors Based On Recurrent Neural Network

Chemical sensors are demanded for broad applications,including environment monitoring,clinical diagnostics,health care,safety alarms,and food quality inspection.In many application scenarios of chemical sensors,it is vital to obtain the sensing results rapidly,such as the alarm of flammable and explosive gas or the detection of toxic substances.Reducing detection time can also improve user experience and reduce power consumption.However,many chemical sensors have the problem of long response time.An economic and effective solution for that is to predict the final result(the concentration or type of the analyte)based on early transient response data through data processing algorithms.But current methods of related work have the problem of not universal or requiring a large amount of measurement data.For these problems in the existing work,this thesis proposes a fast measurement method that can be used for a variety of sensors.Using less measurement data,the relationship between the analyte concentration and the early transient response sequences can be established,which greatly improves detection speed,and the method also has the advantage of flexible detection periods.To further shorten the collecting time of training data for the method,a response curve prediction method is designed.Finally,a handheld ammonia gas detection system based on the fast measurement method is built,and a real-time ammonia concentration prediction experiment is carried out.The details of the research work are as follows:1.In the proposed fast measurement method with chemical sensors,the first stage is training data collection.A sliding window sampling approach with data augmentation is developed to generate more sequence sets for training from limited amount of response curves.With this method,any sequence on the early transient response curve can be used for detection,without the need to accurately control the sampling time and thus simplifying the implementation of the sensing system.Then,the prediction model combining long short-term memory(LSTM)neural network and polynomial fitting is designed for mixed feature extraction,which are then input to multilayer perceptron,which can effectively reduce overfitting under limited data.The method has been experimentally verified on a public dataset.The results show that,with a small amount of measurement data,a mean relative error of 4.69% is achieved,and the time required for detection is reduced from up to 70 seconds to 4.96 seconds.Compared with other common regression algorithms,this method significantly improves the prediction accuracy.2.By predicting the response curve,the collection time of the training data of the above concentration prediction model is shortened.This method is mainly based on the LSTM network and learns the curve law from the response data of the same type of sensor.After the curve prediction model is obtained,only one-third of the original response data needs to be collected,and the subsequent curve can be obtained through multi-step prediction of the model iteratively.Still using the previous public dataset to verify the method,the predicted response curve is close to the actual curve.Using the predicted curve as the training set of the concentration prediction model mentioned above,a high prediction accuracy can still be obtained(MRE is6.11%).3.Coating/printing processes are developed based on a composite of poly(3,4-ethylene-dioxythiophene): poly(styrenesulfonate)(PEDOT:PSS)and silver nanowires(Ag NWs)for manufacturing ammonia sensors of good uniformity.Using the early transient response data of a small number of sensors in the same batch,the concentration prediction model is trained and deployed on the mobile phone application software,combined with the selfmade sensor data acquisition circuit to build a handheld detection system.Based on the uniformity of sensors,fast and accurate detection of ammonia concentration is demonstrated by applying the system to other fresh sensors in the same batch without calibration,thus the whole system is verified and an example of real-time chemical detection on site is shown.