Goaf water is acidic, corrosive, and usually rich in high concentrations of harmful gases such as hydrogen sulfide. Goaf water is the most important and harmful source of water inrush in coal mines. This experiment mainly takes goaf water mixed with sandstone water as the research object. The experimental material was goaf water and sandstone water from a coal mine in Huainan City, Anhui Province, in July 2022. The goaf water and sandstone water were mixed at different volume ratios. 75 water samples were selected from each group to form the following sample set: 1) Group A: Mixed water with a volume ratio of goaf water and sandstone water of 4:1. 2) Group B: mixed water with a volume ratio of 3:1 between goaf water and sandstone water. 3) Group C: mixed water with a 2:1 volume ratio of goaf water and sandstone water. 4) Group D: mixed water with a 1:1 volume ratio of goaf water and sandstone water. 5) Group E: mixed water with a volume ratio 1:2 of goaf water and sandstone water. 6) Group F: mixed water with a volume ratio of goaf water and sandstone water of 1:3. 7) Group G: mixed water with a volume ratio of 1:4 of goaf water and sandstone water.

In order to ensure that the experimental data is more accurate and reliable, the water samples collected at the site are placed in a dark room, sealed, and stored away from light. According to the different mixing ratios, a total of 525 sets of spectral data were obtained as experimental samples.

The structure of the multispectral acquisition system adopted in this paper is shown in Figure 1. The main components of the system include laser, spectrometer, fluorescence probe, and a computer equipped with spectral acquisition software. A 405 m semiconductor laser (Beijing Huayuan Toda Laser Technology Co., LTD.) was used to excite the fluorescence of the mine water source. The spectrometer is a USB200+ mini-spectrometer (Ocean Optics, United States), which is equipped with a 2048-dimensional linear CCD for fluorescence spectra measurement. The immersion micro-fluorescent probe model FPB-405-V3 (Guangdong Koskai Company) can be inserted into the sample to obtain fluorescence signals. Spectra Suite software is installed on the computer for the acquisition, display, and saving of fluorescence spectra. The algorithm simulation is run in Matlab R2021a environment.

The fluorescence spectrum of mine water source is collected in the same environment to reduce the interference of external factors on the fluorescence spectrum. During the experiment, the power of the laser was set to 100 mW, the spectrum acquisition range of the spectrometer was set to 340–1,021 nm, and the integration time was set to 1 ms. After the equipment was ready, the fluorescence spectra of mine water source samples were collected by the LIF system, and 75 fluorescence spectra of seven kinds of mine water sources were obtained, totaling 525 mine water fluorescence spectra. In the process of establishing the mine water source identification model, 60 samples of each mine water sample are randomly selected as the training set and the rest as the test set. That is, the training set contains a total of 420 mine water fluorescence spectrum samples, and the test set contains a total of 105 mine water fluorescence spectrum samples. In addition, in the process of establishing the fluorescence spectrum identification model of mine water source, ten-fold cross validation is introduced to make the classification model more reliable.

Due to the interference of system noise and external noise in the collection process of laser-induced fluorescence spectrum, the original fluorescence spectrum of mine water source collected contains useless noise interference information, which has a great impact on the experimental results. Therefore, it is necessary to preprocess the original fluorescence spectral data. Common spectral preprocessing methods include Moving average smoothing (MA), Savitzky-Golay smoothing (SG) Schettino et al. (2016), First derivative (FD)Jin et al. (2012) and Second derivative (SD) Czarnecki (2015). These methods are used to denoise the fluorescence spectral data. According to the evaluation index of the selected classification model, the prediction ability of the original spectrum and the denoised spectrum is compared, and the appropriate denoising method is selected.

Random forest consists of many decision trees. It is a supervised algorithm for classification and regression, also known as Classified And Regression Tree (CART), which was proposed by Breiman Parcha et al. (2007). The random forest algorithm process is shown in Figure 2. In the calculation process, the binary tree segmentation rule is used to divide the trained sample set into two subsample sets each time, and each non-leaf node has two branches. The subsample set repeats the action until it can no longer split. The randomness of the random forest is reflected in the fact that when training each tree from the full sample size of N, a dataset that may have repetitions of the same size of N is selected. We call it Bootstrap sampling. At each node, a subset of all features is randomly selected to calculate the optimal segmentation method Goehry et al. (2021). Bagging is a put-back sampling technique based on Bootstrap. According to the sampling probability, about 36.79 % of the original data will not be selected by Bootstrap sampling and will not participate in the establishment of a decision tree. These data constitute the out-of-bag (OOB) data set Kotsiantis (2011). This part of data can be used to evaluate the performance of the decision tree and calculate the prediction error rate of the model, which is called out-of-bag error. The algorithm of each single decision tree has low precision and is prone to overfitting. If the accuracy is improved, multiple decision trees need to be gathered together to form a random forest model. The structure of its prediction model is R = h ( x , ) , k = 1,2,3 , … K , where k represents the number of decision trees, θ 1 , θ 2 , θ 3 , are each random vector independent of each other and in the same direction. The independent variable x is determined, and the final prediction is decided by voting Quadrianto and Ghahramani (2014).

In the random forest model, the feature importance can be measured by the OOB error. For each decision tree, the corresponding out-of-bag data error is recorded as errOOB1. Then, randomly add noise interference to the feature x of all samples of the OOB data, calculate the error of the data outside the bag again, and record it as errOOB2. If there are N trees in the random forest, The importance of feature x W is expressed as W = ∑ ( e r r O O B 2 − e r r O O B 1 ) / N Gupta et al. (2022). Suppose the accuracy rate outside the bag is greatly reduced after noise interference is added. In that case, the feature has a significant influence on the classification result, and the feature is of high importance.

In order to ensure that no other factors affect the results of the experiment, the experiment was conducted in a laboratory without a light source and at a constant temperature. The fluorescence spectrum data of mine water source was collected by micro-optical fiber and fluorescence spectrometer. After laser irradiation, the fluorescent substances in mine water source absorb light energy, stimulate and release energy, produce fluorescence, and form fluorescence spectrum. The spectrum peaks between 420 and 650 nm and differences in this range are concentrated. The original fluorescence spectrum is shown in Figure 3. Figure 3A shows the original fluorescence spectra of all mine water samples, and Figure 3B shows the fluorescence spectra of one sample randomly selected from each type of A-G goaf water and sandstone water mixed samples. The morphology and wave peaks of the original fluorescence spectra are very similar, and there are crosses between different spectra. The spectral differences are small, so it is difficult to observe and distinguish the differences between experimental samples. Therefore, the original fluorescence spectral data should be preprocessed.

In order to eliminate noise interference in the original fluorescence spectrum, reduce errors, and retain useful information in the fluorescence spectrum, MA, SG, FD, and SD were used to preprocess the original fluorescence spectrum data of mine water samples. Respectively, the preprocessed fluorescence spectra are shown in Figure 4.

The four groups of diagrams in Figure 4 show that the fluorescence spectra of MA and SG preprocessed spectra have many advantages, although there are still overlapping parts. Compared with the original fluorescence spectra, they are more dispersed, with less noise interference, and the water samples are easier to distinguish. While the FD and SD processed data are redundant, with large noise interference, which affects the accuracy of the spectral data.

The classification accuracy and training time of the four preprocessing methods were obtained by RF classification. The results were shown in Table 1. The classification accuracy of MA reached 99.24%, and the training time was 0.3434s. The classification accuracy of SG is 99.05%, and the training time is 0.3431s. Both methods have improved the classification accuracy of 98.10% and the training time of 0.3660 of the original spectrum. Overall, the pretreatment effect of MA is the best. The data also showed that the classification accuracy of FD and SD were 80% and 61.52%, and the training time was 0.3975s and 0.3884s, which showed poor processing effect.

Random forest is an ensemble learning algorithm that improves classification and regression accuracy by combining multiple decision trees. In the random forest, we need to adjust the parameters, and the first is to set the minimum number of leaves. Each decision tree divides the data set into smaller subsets through continuous segmentation, each subset corresponds to a node in the tree and the leaf node is the final subset Karabadji et al. (2023). According to the setting of the minimum number of leaves, each node must have a certain number of samples in the subset after segmentation to continue downward segmentation. Otherwise, it will become a leaf node. How to set the minimum number of leaves greatly affects the classification and regression results of random forest. If the minimum number of leaves is set too large, the depth of the decision tree will be shallow, resulting in an underfitting phenomenon. If the minimum number of leaves is set too small, the depth of the decision tree will be large, and overfitting will occur Santra et al. (2020). Out-of-bag (OOB) data can be used as generalization error to evaluate the model. After training, RF's out-of-bag error rate is shown in Figure 5. The minimum number of leaves is set to 6, the number of trees is set to 24, and the out-of-bag error rate is reduced to 0, achieving 100 % identification accuracy and ensuring the stability and reliability of results.

There are 2048 attributes in the original fluorescence spectral data, each of which contains different spectral information, and the importance of different attributes to spectral analysis is obviously different. Non-critical attributes with low importance will affect the establishment of the classification model for mine water inrush samples and thus fail to achieve a good prediction effect. Based on random forest, the feature importance is analyzed. As shown in Figure 6, the importance of most attributes is 0. After MA preprocessing, some attributes eliminate interference information and make their importance prominent. The fluorescence spectrum with wavelengths between 420 and 620 nm has obvious characteristics, and the feature importance is between 0.1–0.3.

The RF classification model can select features from their importance through threshold selection for model optimization Hwang et al. (2023). Try to remove unimportant features at lower levels to optimize the accuracy and efficiency of the model. When the threshold is 0.25, the number of selected feature wavelengths is 2, the minimum classification error is 1.71 % , the number of trees is 60, and the training time is 0.202s. Although the training time is short, the minimum classification error value is too large. When the threshold is set to 0.10 and 0.05, the number of selected feature wavelengths is 281, and the number of decision trees is 18, the minimum classification error can reach 0. Then, the training time is 0.3790s and 0.3951s, respectively. Overall, when the threshold is 0.1, the classification effect is good, and the training time is short.

A total of seven groups of A-G samples mixed with different volume ratios of goaf water and sandstone water were trained. Since the test set and training set were generated randomly, the results would be different each time. The RF prediction model of a certain run was fitted, and the results are shown in Figure 7. After MA preprocessing, the RF model threshold was set to 0.10, and the number of characteristic wavelengths was 281. The predicted results were good, and the predicted values were basically consistent with the actual values. This is consistent with the results under different feature importance thresholds in Table 2.

In order to verify the evaluation effect and reliability of the RF algorithm model for mine water inrush prediction, Support Vector Machine (SVM) Ding et al. (2017), Principal Component Analysis (PCA) Zhou et al. (2020), and Moving average smoothing (MA) algorithm are used to identify mine water inrush independently or combined with algorithm modeling. In the experiment, the training data was randomly selected, and the experiment was repeated three times independently. The experimental results are the average of three times, as shown in Figure 8.

In Figure 8, when only SVM algorithm is used to identify mine water inrush, the spectral data has the interference of redundant information, and the recognition accuracy is the lowest 98.1 % . With the combination of PCA-SVM algorithm, the recognition accuracy has been improved, reaching 99.05 % . After MA preprocessing, the MA-SVM and MA-PCA-SVM algorithms reduce the noise interference in the spectrum and greatly improve the recognition accuracy, reaching 99.81 % . Finally, the prediction and evaluation performance of MA-RF algorithm is the best, and the accuracy reaches 100 % . It has the highest accuracy and stable results.