Research On Flame Detection Method Based On Random Vector Functional Link Network

As one of the most frequent and most destructive disasters,fires seriously endanger human life,property safety and natural ecological environment.Image-based flame detection technology effectively compensates for the disadvantages of traditional sensor flame detection devices,such as slow propagation speed,high false alarm rate,and inability to apply to large outdoor spaces,and has received extensive attention and research.Extracting flame features and using pattern recognition algorithms to detect flames is a mainstream method of image-based flame detection.However,most of the existent image-based algorithms are focused on the analysis and identification of the flames that have already formed a fire,but the research ability of the early flame high-risk points that have not formed a fire is very limited.Moreover,it is difficult for ordinary cameras to detect through obstacles blocking the lens,which reduces the detection capability of the detector.The use of infrared cameras to collect flame data can effectively solve the above problems.Infrared thermal imaging technology is widely used in flame detection technology.Its advantage lies in that it is not restricted by the vision of shielding objects,has strong anti-interference ability and can add one-dimensional temperature information according to the detected object,which has valuable value for early fire detection and is helpful for early fire warning.In this paper,an infrared flame detector based on Random Vector Functional Link（RVFL）network is proposed by using infrared camera as the acquisition equipment of flame data and combining with the method of multi-feature fusion of flame.The algorithm is improved to improve the accuracy of flame recognition.The research content of this article is arranged as follows:（1）To solve the model instability problem caused by the completely random allocation of input weight and hidden bias,an incremental random vector functional link network with random parameter optimization is proposed.Firstly,the exponential weighted average algorithm is used to stabilize the sequence and reduce the influence of outliers to optimize the input weights and bias,reduce the weight of outliers that may be generated during random allocation,and strengthen the stability of the network.Then,the descending gradient ratio of the convex function at equal intervals is applied to the error sequence of the network to speed up the descending speed of the model error.The optimization algorithm proposed in this paper is verified by UCI datasets.The experimental results show that the optimized network can make the root mean square error of the fitting smaller,the classification accuracy is higher,the stability is better,and the convergence speed of the model is faster.（2）Aiming at the problem of outliers in the network training data and reducing the generalization performance of the model,a regularized random vector functional link network is proposed.Firstly,according to theL₂₁ norm can reduce the influence of noise on the model and reduce the inherent complexity of the learning model,theL₂₁ norm is introduced as the regularization constraint of the model loss function.Then,in order to reduce the influence of empirical risk on the model,the weighted algorithm is used to adjust the weight of the output weight parameters to reduce the influence of outliers on the model and improve the robustness and generalization ability of the model.The performance of the algorithm is verified by function fitting and UCI binary classification datasets.The experimental results show that the generalization performance of the proposed regularized network is better,the robustness is stronger,and the effect of dealing with outliers is remarkable.（3）Using an infrared camera as a flame collection device,an infrared flame detector based on the random vector functional link network is proposed,and the hardware circuit and software program are designed at the same time.By extracting the dynamic and static characteristics of the infrared flame image,combined with the multi-feature fusion method,two optimized networks are used to train and classify the collected data samples.The experimental results show that the optimized network can effectively improve the recognition accuracy and anti-interference ability of the flame detector.