Research On Machine Olfactory Technique For Anomaly Detection In Enclosed Space

With the development of modern science and technology,the closed and semi-closed space constructed by reinforced concrete has become the main place for people to carry out social activities.This paper proposes a technology that can accurately detect anomalies within the space scope,which is helpful to protect people’s life and property safety and maintain the continuous progress of social production.Because there are many obstacles in the closed and semi-closed space and the gas is not easy to dissipate,the traditional computer vision technology has visual blind spots due to the influence of the occlusions,and the electronic nose can well capture the gas information in the current state for algorithm recognition.Therefore,for the closed space,the use of machine smell technology can be an important anomaly detection and recognition method.In this paper,the abnormal classification of the normal state,cigarette burning,cardboard burning,cloth burning,sponge burning,electronic circuit burning and natural gas leakage in the enclosed space is studied.Firstly,an appropriate gas sensor is selected according to the abnormal gas composition and the sensor array is formed.Then,a prototype machine olfactory system with real-time data acquisition function is designed,and the data is collected by using this system under different experimental variables.Finally,an anomaly detection algorithm is proposed to conduct experiment and analysis on the data set.In the algorithm part,the ConvLSTM structure,which performs convolution operation on the input part of the LSTM unit,is firstly discussed.Then,a method of frame slicing is proposed to process the collected gas data in two-dimensional matrix format,so as to adapt to the ConvLSTM input format.A ConvLSTM anomaly detection model was constructed by batch standardization,random discard algorithm and network parameter adjustment.Experimental results show that ConvLSTM model not only outperforms traditional machine learning algorithms,but also outperforms CNN and LSTM algorithms in accuracy,precision,recall and F1 scores.ConvLSTM embedded convolution operations into the LSTM structure to simulate the process of convolution kernel sliding in space,requiring more data to train the model.For the situation that the time sequence of gas sample data in two-dimensional matrix format is short and the data set is small,other combination methods of CNN and LSTM can be considered as a method to further improve the accuracy of anomaly detection.Using CNN to extract spatial features from two-dimensional data,and then using BiLSTM units to learn time sequence information,this combination can make better use of data,so as to effectively deal with the situation of insufficient data.Then the CNNBi LSTM-Attention anomaly detection model was designed by adding self-attention mechanism and using Bayesian optimization to adjust the parameters of neural network.After experimental analysis,the performance of the CNN-Bi LSTM model is comparable to that of the ConvLSTM model,and the CNN-Bi LSTM-Attention model has further improved in performance,reaching 96.19% classification accuracy.