Objective In field of fire protection, the discoloration marks on steel plates during heating provide important information as evidence for fire investigation work, which can effectively reflect the heating temperature. However, the intelligent recording and identification of surface mark changes on steel plates rely on empirical databases for judgment or laboratory testing, lacking efficient means and methods. Spectral imaging technology provides a viable method to record and analyse through the surface spectrum of steel. The snapshot spectral imaging camera chip filted by Fabry-Perot interferometer provides a reliable approach for commercial use in the visible near infrared range. Spectral imaging camera have the advantages of portability, secondary integration, and non-contact multidimensional information acquisition of target. However, due to limitations in multi-layer thin film structures and materials, the free spectral range of this spectral imaging system is limited. To enhance the spectral detection range of the snapshot based spectral imaging system, a handheld spectral imaging system consisting of cascaded dual spectral cameras is used for fast detection under complex light conditions. Spectrum, as the inversion basis for color information, has the potential to more accurately record the surface characteristics of steel plates and identify the corresponding heating temperature changes of feature changes. Therefore, the research of intelligent recognition of steel plate heating traces using handheld spectral imaging equipment is carried on.

Methods  Firstly, handheld spectral imaging system based on snapshot spectral cameras with two different spectral response ranges is integrated (Fig.5), with a response spectral range of 475-975 nm and 40 spectral channels. Secondly, Q235 steel plate samples were used as experimental subjects to prepare experimental samples. The samples were heated at 100 ℃, 200 ℃, 300 ℃, 320 ℃, 380 ℃, 400 ℃, 480 ℃, 500 ℃, 590 ℃, 600 ℃, 700 ℃, 800 ℃, 870 ℃, 900 ℃, 980 ℃, 1000 ℃, 1100 ℃, 1200 ℃ respectively for 30 minutes through muffle furnace constant temperature heating method and 20 ℃ room temperature (Fig.7). Spectral image data collected by the two snapshot spectral camera of the handheld spectral imaging system in 5 different sunlight source conditions were fused, cleaned, and preprocessed to obtain 23500 sets of target sample data. Next, divide 19 types of heated samples into 5 temperature gradients according to application requirements, build a classification and recognition algorithm model, and randomly select 80% of the data for training. Finally, proposing a target weighted discrimination method, with classification and recognition algorithms as the core, using image feature information to divide the same target area into 9 subsets for discrimination (Fig.9). Based on the probability of the recognition results of the 9 regions, the final recognition category of the target can be defined.

Results and Discussions Through cross validation of the remaining 20% of the sample set, linear discriminant analysis and back propagation neural network were used for 5-class classification, achieving accuracies of 91.5% and 94.2% respectively (Fig.13). The linear discriminant and back propagation neural network fusion method guided by decision tree can achieve a cross validation accuracy of 97.6%. Selecting 19 types of heated temperature samples outside the sample set for system application verification, through target weighted discrimination, each sample can be accurately identified as its own temperature gradient with 100% accuracy (Tab.2).

Conclusions The preliminary research has achieved intelligent discrimination of heat trace on Q235 steel plates by handheld spectral imaging system with 475-975 nm spectral range response, laying a technical foundation for the application and promotion of handheld spectral imaging systems in special industries. In the future, the sample heating conditions and categories will be further expanded to assist in the intelligent development of fire investigation work.