Study Of Flame Recognition And High-temperature Molten Aluminum Leakage Monitoring Based On Infrared Image Analysis

Fire and high-temperature molten aluminum leakage are safety incidents of society and industrial production,usually resulting in casualties and huge economic losses.Radiation imaging is the basic principle of thermal imaging technology which has natural advantage to detect high-temperature objects such as early flame and leaking molten aluminum.Therefore,FLIR-A310 thermal imaging camera is used to carry out early flame recognition and high-temperature molten aluminum leakage monitoring researchs in this paper.The video fire detection algorithm based on traditional pattern recognition method is difficult to break through the accuracy bottleneck and has problems of high false and missed alarm.We introduce deep convolutional neural network to replace the traditional artificial feature extraction method,extracting the deep abstract salient features of infrared flame automatically and efficiently,which could achieve the early accurate flame detection in infrared videos.In addition,the high-temperature molten aluminum leakage monitoring is still in the original artificial stage.We design a simulation leakage monitoring algorithm for high-temperature molten aluminum based on infrared image feature fusion to improve the intelligence and automation level of molten aluminum leakage monitoring.The main research contents are listed as follows:1.An early flame detection algorithm in infrared video based on deep convolutional neural network and SVM is proposed.We established the standard data set of early flame in infrared videos and adopted data augmentation strategies to increase the scale of training data set.The Adam optimization algorithm instead of Mini-batch gradient descent was used to train the designed 9-layer IRCNN convolutional neural network.Then,the extracted 2048-dimensional abstract salient features from infrared image were sent into SVM classifier for early flame recognition.The performance of proposed algorithm is optimal and the precision,recall and F1 reach 98.82%,98.58%and 98.70%,respectively.At the same time,this algorithm could meet the early flame detection requirements of conventional thermal imaging camera,predicting 20 infrared images per second.2.Build a compact convolutional neural network named LightflameNet to deal with the problem of deep neural network model deployment on the video fire detection systems.A multi-branch lightweight convolutional network unit,Flame module,was designed to replace the traditional standard convolution layer,which could not only reduce the weight parameters and FLOPs,but also expand the single-layer network width and enhance the ability of infrared flame feature extraction.LightflameNet requires only 2.5 MB of storage space and 89 million FLOPs while almost no loss of model accuracy.Furthermore,it consumes just 6ms to predict a single infrared image with CPU and 2ms with GPU.3.A simulation leakage monitoring algorithm for high-temperature molten aluminum based on infrared image feature fusion is proposed.We constructed a simulation experiment platform of the high-temperature molten aluminum leakage,used hot water to simulate the molten aluminum leakage scenes and established a simulated leakage standard data set consisting of 1200 training samples and 600 testing samples.The LSS descriptor was introduced to compensate for the lack of single HOG feature in this paper.Geometric features such as image gradient edge and similar shape were extracted and fused,and then fed into RBF kernel function support vector machine for classification and recognition.When LSS step size is set as 15 and the camera distance is 3.5m(corresponding to the actual distance of about 10.5m),the proposed algorithm achieves the best performance and the precision,recall and F1 are 96.73%,96.50%and 96.61%,respectively.