With the continuous development of light sources based on light-emitting diodes (LEDs), the demand for illumination has shifted from initial environmental protection and energy conservation to healthiness and comfort. Sunlight is commonly considered to be a perfect lighting source due to its full spectrum characteristic, which not only is most suitable for human visual and non-visual needs but also is widely used in plant growth experiments, phototherapy equipment research, photovoltaic cell testing, camera fill light, etc. In recent years, realizing a light source close to the solar spectrum has become a new development goal of semiconductor lighting. However, due to low algorithm accuracy and operation speed, the existing spectral reproduction methods cannot flexibly reproduce the spectrum to meet industrial needs in practical applications. Therefore, we propose a method based on a fully connected neural network (FCNN) to reproduce natural spectra with high accuracy and speed.

In this work, in order to reproduce the natural spectrum accurately and quickly, a neural network with strong nonlinear fitting ability is proposed to complete spectrum matching. First of all, according to the characteristics of the continuous wide band for natural spectra and spectral distribution for monochromatic LEDs, 23 monochromatic LEDs with different peak wavelengths and full widths at half maximum are selected to make up for the natural spectrum. Then, according to the modified Gaussian distribution spectrum fitting model and spectral superposition principle, different spectral data as the training set and test set for the FCNN model are generated by using monochromatic LEDs' spectra. On this basis, the trained FCNN model that fully reflects the proportional relationship between the synthetic spectrum and the light intensity coefficient of each monochromatic LED is constructed, which can reversely obtain LED ratio parameters from the synthetic spectrum. In other words, the method based on FCNN can obtain the corresponding proportional coefficient of monochromatic LED light intensity for the input target spectrum and then realize spectral reproduction.

Firstly, the wavelength of 380-680 nm of the standard solar spectrum as the target spectrum is reproduced by using the proposed method based on FCNN. The results demonstrate that the fitting correlation indexes of the spectrum reproduction results for the standard spectra AM1.5, CIE-D65, and CIE-A are 0.9670, 0.9812, and 0.9815, respectively (Fig. 3). In order to verify the applicability of the proposed method for different spectra, the proposed network model is used to reproduce more natural spectra measured in different time periods, which reveals that the fitting correlation indexes of the reproduction results for the spectra at 6:00, 11:00, 18:00, and 19:30 are 0.9520, 0.9627, 0.9855, and 0.9726, respectively (Fig. 4). In other words, whether it is for the measured spectrum or the standard solar spectrum, the synthetic spectrum obtained by the proposed method based on FCNN can be highly similar to the target spectrum, and the correlation index can reach above 0.95. In addition, the fitting accuracy and time cost of natural spectrum reproduction using the proposed method are further compared with that using an intelligent optimization algorithm. As a result, the method based on FCNN not only has higher accuracy stability but also requires less fitting time than the genetic algorithm (GA) for spectrum reproduction (Fig. 5). This is because the FCNN can save the network parameters and fully reflect the relationship between target spectrum and the monochromatic LED scale coefficient, and the model after training can be used directly to reproduce natural spectrum with small time costs. The results show that the average running time of the proposed method is 0.04 s, which is several times faster than the method based on GA in reproducing different natural spectra.

In this work, a method of natural spectrum reproduction based on FCNN is proposed to overcome the weakness of long fitting time and low accuracy stability of the current matching algorithms. After successfully training and testing the superimposed spectral data of 23 monochromatic LEDs with different peak wavelengths and full widths at half maximum, the FCNN model for natural spectrum reproduction can be constructed. On this basis, the standard solar spectrum and measured natural spectrum at different time are reproduced using the FCNN model and compared with that using the spectral matching method based on the GA in terms of the fitting time and fitting accuracy. The results show that the correlation index of the reproduction results using the proposed method based on FCNN can all reach above 0.95, and the running time required for reproduction is all less than 50 ms for different natural spectra. Furthermore, the proposed natural spectrum reproduction method based on FCNN has higher accuracy stability and requires less fitting time than the GA. Therefore, the method proposed in this work can reproduce different natural spectra stably, efficiently, and accurately, which can provide a new solution for the development of light sources in full-spectrum illumination.