Research And Implementation Of Flammable Liquid Classification Algorithm Based On Deep Learning

With the increasing contacts between different places,the transportation of personnel and materials is becoming more frequent.Security inspection technology is very important for the safety of transportation.Existing security inspection techniques have low security screening efficiency,high cost and unhealthy defects.It is difficult to apply in a wider range of application scenarios.By analyzing the behavior of liquid molecules in different frequency bands,ultra-wideband centimeter waves are selected as a new flammable liquid security inspection method.Ultra-wideband centimeter-wave transmission signals for multiple flammable liquids are classified using deep learning techniques.Through the self-built data set,the algorithm is trained and verified.Improve the efficiency of the algorithm in embedded devices.The main contents and innovations of the full text can be summarized as data collection,algorithm design and algorithm transplantation.The details are as follows:1)We used the beam focusing system to collect three ultra-wideband centimeter wave transmission signals of common liquids in the range of 8 GHz to 18 GHz.For the training and testing of the algorithm,the data set is established.After data cleaning,squelch and padding,it is found that the data in the data set has two kinds of defects: uneven distribution of information and insufficient data.2)A two-layer feature classification algorithm is proposed for the two defects of the data set.The main content is to decompose the characteristics of the signal into two levels,through multi-scale window.Shallow feature extractors are used to extract shallow features from a single window.A self-encoder that adds a classification constraint to implement this feature extractor.The deep feature extractor is used for feature extraction and classification of feature sequences composed of shallow features of all windows.LSTM is used to do this part of the work.The two-layer feature classification algorithm has a simple structure and adopts some measures to alleviate the over-fitting.Its classification performance is good.3)The Roof-line model evaluates the operational efficiency of the two-layer feature classification algorithm and predicts the speed of the algorithm in different hardware platforms.Based on the calculation results and equipment cost,the Raspberry Pi was selected as the migration target device.Algorithm uses Tensorflow to read the Pytorch model weights.Four optimization methods were tested for the low-level problem of the two-layer feature classification algorithm in the Raspberry Pi.Four optimization methods were tested.The optimization method of replacing the underlying computing library and the parallel computing works well,and the optimization effect of changing the convolution kernel size and replacing the convolution layer with the separable convolution layer is not obvious.In this paper,a two-layer feature classification algorithm for ultra-wideband centimeter wave signals is implemented and transplanted.The accuracy of the three classifications is up to 84.37%,and the reasoning time in the Raspberry Pi is less than 0.3 seconds.A new identification idea has been proposed for the field of liquid dangerous goods security inspection.