2.1. Overall Framework Design of Disaster Nursing System

The system mainly includes an information acquisition module, transmission module, decision support module, and visualization module. The overall framework design of the disaster nursing system is designed as shown in Figure 1.

As is shown in Figure 1, (a) collect the prevention information of geological disasters. The main function of the data information management program is the collection of geological disaster prevention information. It can also input, store, manage, query, display, and update all kinds of geological disaster prevention information. In terms of function, it can be managed through an information set, so as to realize data entry and editing, system library editing, etc. The information set is composed of six submodules, which browse and search together to achieve data output. The information set realizes the data information structure management represented by the directory tree window, which is the core of the information data management of the whole subprogram. Through the construction of the data source directory tree to reflect the data information and its correlation, the multilayer structure is used to realize the collection of geological disaster prevention information. (b) Build the development environment of the decision support platform. The overall goal of this system design should meet the auxiliary management of “normal emergency management” and “combat structure analysis” of geological disasters for analysis and decision-making. In addition, in the linkage analysis function of national transition from normal emergency and the emergency management and decision-making analysis of geological disaster rescue, the emergency plan, emergency inspection, planning and training, emergency coordination, emergency management document preparation, etc., must be completed in advance and meet the information processing and analysis functions required for this, as the decision-making measures of the platform. Make full use of the risk source monitoring data and emergency patrol information of geological disasters, analyze, evaluate and predict according to the established model, and give early warning of abnormal conditions and risk levels. Once problems occur, it is necessary to formulate reasonable solutions according to the emergency plan and take preventive measures immediately in combination with the dynamic changes of geological disasters. (c) Establish the database of the decision support platform. To build the database of the whole decision support system, we should start from the specific data of geological disaster prevention information, select the area where the data is located, establish a general platform for data management, input the information of the platform into the database program, and build it in groups, so as to divide different types of information into different categories for subsequent processing. (d) Visualize geological disaster prevention decision-making information. GML is organized and developed by OGC, which can realize the output of decision information of the whole decision support platform. When it is used, it needs to exchange data in advance to transform spatial information. In addition, it does not pay attention to the visualization function and can directly realize the output and visualization of decision information.

In addition, the main goal of the development of this system is to realize a national music classification system, which can provide batch opera classification and annotation services for any user who has the need for national music sorting and classification. This system selects the Mel-cepstrum coefficient that reflects the timbre and the pitch frequency, formant, and band energy distribution that reflect the melody. Combined with the deep confidence network model, it trains a large number of labeled samples to obtain the ability to classify national music and saves the classification model that meets the classification requirements through the test set to the system, so as to achieve the classification goal of unknown style national music.

2.2. Analysis and Design of System Hardware Architecture

This paper describes a measurement system for data fusion processing, so the system requirement analysis considered is relatively comprehensive and complete. It not only involves the hardware platform requirements of data fusion but also needs to match the specific requirements of the industrial site and expand the necessary communication interfaces. To sum up, the demand analysis of system design is as follows.

(1) It is necessary to implement the part based on a multisensor data fusion algorithm and complete the model establishment process of the Dempster Shafer fusion algorithm, that is, the whole process of data fusion processing. (2) It needs an input channel for external sensors to collect data, so as to obtain the industrial field measurement data in real time and accurately, and then hand it over to the algorithm processing module for calculation and processing. (3) A certain operation interface is required for input of external operation, manual parameter setting of algorithm, start and stop of system, switch and adjustment of each function module, query of calculation result information, etc. (4) A certain display interface is required to display the current operation settings and parameter settings and provide a complete human-machine interface in Chinese, which is convenient for users to operate and set. (5) Various communication interfaces are required to ensure the universality and platform of the measurement platform and facilitate the interconnection with other control and detection equipment, including the network interface, serial communication interface, USB interface, and various bus interfaces.

Considering the design requirements of the measurement platform studied in this paper, modular design is the most suitable implementation method. The hardware structure diagram is shown in Figure 2. That is, the whole hardware platform is divided into four submodules according to different design requirements. Each module completes certain system functions, and at the same time, good communication and connection between modules are maintained. According to this idea, the hardware platform of the system is mainly composed of four modules: core board, sensor board, I/O interface board, and man-machine interface board. As is shown, the core of the soft measurement platform described in this paper is a dual-core processor chip OMAP-l137 of TI Company, as shown in Figure 2. It is a processor launched by TI Company that can be used for industrial control, network communication, high-speed coding, and professional audio processing, so it is very suitable for the application environment of soft measurement. The 300 MHz ARM processor is mainly responsible for the process scheduling of the whole algorithm and the control functions of various signals on the industrial site. If the operating system is transplanted, it can also realize the functions of file management, process scheduling, memory management, etc., while the 300 VII 4Z DSP floating-point processor is mainly responsible for the processing of digital signals and high-precision floating-point operations, which effectively makes up for the shortage of the ARM926EJ-S processor in the calculation accuracy.

3.1. Multisensor Information Fusion Technology

For information fusion, the information representation and processing methods used come from these fields. From the functional model of information fusion, we can see that the basic functions of fusion are correlation, estimation, and recognition, with emphasis on estimation and recognition.

Therefore, the design of the fusion algorithm based on joint Kalman filter is the core content of information fusion technology. Remember that the state estimation vector, system covariance matrix, and state vector covariance matrix of local filter I are $$, Q₁, P_i. The global filter is $$, Q_m, P_m. Then, the calculation structure of the joint filter is shown in Figure 4.

3.2. Research on Disaster Nursing Prediction Algorithm Based on DSBP

Although the BP network has been widely used, it also has some defects and shortcomings, mainly including several aspects: (1) Because the learning speed is fixed, the convergence speed of the network is slow and it takes a long time to train. For some complex problems, the training time of the BP algorithm may be very long. This is mainly because the learning rate is too small and can be improved by changing the learning rate or adaptive learning rate [19]. (2) The BP algorithm can make the weight converge to a certain value, but it cannot guarantee that it is the global minimum of the error plane. This is because the gradient descent method may produce a local minimum. The additional momentum method can be used to solve this problem [20].

Firstly, the characteristic parameters of fire are detected by the fire detection sensor group, and then, the measured data are normalized and processed as the input parameters of BP. These parameters are then processed by the BP neural network and output. The range of output results is between [0, 1]. These output results are the input of DS evidence theory. At this time, the results have certain reliability and accuracy after the initial fusion of neural networks. Then, the data are fused again by using DS evidence theory to obtain the unique probability distribution function. Finally, the judgment result of whether the fire occurs is obtained according to the judgment rules. The specific algorithm structure is shown in Figure 5.

Decision level fusion is based on the independent decision-making or classification of the information collected by each local sensor after preprocessing, feature extraction, and recognition, and then, the fusion system transmits the preliminary decision-making information to the decision center for comprehensive evaluation with the help of certain criteria and algorithms, so as to make a globally optimal decision [21]. The data processing of decision level fusion is concentrated in the fusion center, which has strong system stability and fault tolerance, and will not greatly affect the performance of the fusion system due to the effectiveness of several sensors. At the same time, due to the small amount of data, the calculation amount will be much smaller than that of data level fusion and feature level fusion. However, the fusion of the decision-making layer only deals with the associated features, which will ignore some important information in the original data and feature data, resulting in a large amount of information loss and thus reducing the accuracy of loss identification.

The BP neural network adopts the common three-layer neural network, which will not be described in detail here.

4.1. Effectiveness of Sensor Information Fusion Algorithm

Taking the combination of camera information (CS), GPS, and GIS as an example, this paper simulates the above joint Kalman filter algorithm. Taking GIS as the public reference subsystem and CS and GPS as the auxiliary subsystem, the two systems include two subsystems: CS/GPS and GPS/GIS. If the information fusion algorithm is adopted, to verify the effectiveness of this information fusion algorithm, this paper compares the information fusion results with wavelet filter and Kalman filter algorithm. The fusion result is shown in Figure 6.

Figure 6 shows the computer simulation results of the above system. The dotted line in the figure is the positioning error under the condition of conventional Kalman filter. The solid line is the positioning error after using the joint Kalman filter technology. It can be seen that the accuracy of the fused system has been significantly improved. It shows that the proposed scheme is very effective in improving the accuracy and reliability of the whole system.

4.2. Convergence Verification of the Model

We apply the BP network in the neural network toolkit in MATLAB 2018b to realize the training algorithm. This toolbox contains many tool functions for BP network analysis and design. In this system, the forward network initialization function is mainly used: newff gives the initial threshold and weight of network training. The BP network is trained by the function: train, and the threshold and weight matrix of the trained BP neural network are obtained. Finally, the simulation function: Sim is used to simulate the trained network. At the same time, D-S evidence theory is applied to the BP neural network to prove the effectiveness of this method. We use the membership degree as the input of the BP neural network. The safety status of the disaster system is divided into four levels: safety S1, general state (S2), dangerous (S3), and relatively dangerous (S4). The safety status of the disaster system is coded as the following: S1 level: 0.1, 0.1, 0.1, 0.9; S2 grade: 0.1, 0.1, 0.9, 0.1; S3: 0.1, 0.9, 0.1, 0.1; and Grade S4: 0.9, 0.1, 0.1, 0.1. Since the data is designed when the disaster system is in the S1 state, the expected output of the neural network corresponding to the training samples is the following: design a BP neural network with 6 inputs and 4 outputs, using a three-layer neural network, with 6 inputs, 10 nodes in the negative layer, 4 nodes in the output layer, a learning rate of 0.01, a log sigmoid transfer function in the negative layer, and a purelin linear transfer function in the output layer. The expected error is 0.001, and the transfer function is trained by MATLAB software.

Collect 50 groups of sample data and 20 groups of test data; establish BP neural networks, CEF-BP, PSO-BP, and DSBP, namely, 3 evidence bodies E1, E2, and E3; train 50 groups of sample data and BP neural network; and then optimize them through the BP algorithm. The BP length is set to 5, the neural network error threshold is 0.05, and the probability distribution threshold is 0.8. When the BP neural network is optimized, input 20 groups of test data to the BP neural network for disaster diagnosis. The disaster prediction results of three BP neural networks corresponding to some test samples are obtained. Only from multiple BP neural networks, there will be uncertainty on the dial, and accurate diagnosis results cannot be obtained. When the results obtained from BP neural networks and DS evidence theory are fused, the accuracy of disaster prediction can be greatly improved (Figure 7).

Bring the test data into the optimized neural network diagnosis model and the nonoptimized neural network diagnosis model, respectively. After 20 iterations of training, it can be found that the recognition ability of the optimized neural network to the data is much stronger than that of the nonoptimized neural network. Figure 8 shows the error curves of BP, CEF-BO, PSO-BP, and optimized DSBP neural network. After 15 operations, the nonoptimized BP neural network will not change the error value for 6 consecutive operations. At this time, the neural network training will automatically end. The validation sample error is 0.15 > 0.01, which cannot meet the set neural network prediction error requirements and cannot accurately diagnose disasters. The optimized DSBP neural network validation sample error can reach around 0.0021, which can meet the accuracy requirements of the neural network for disasters.

For such a complex problem, it is unrealistic to use a back propagation network (BP network) to store knowledge and judge the state with it in terms of learning speed and reasoning ability. When a single neural network (ANN) is used to judge the target state, it will inevitably lead to a large network structure and an increase in training samples, which makes it difficult for the network training to proceed smoothly. Because the network has poor robustness and is sensitive to sample error, it is easy to cause a series of problems such as low classification accuracy, long network training time, difficult network convergence, low diagnosis accuracy, and unreliable conclusions. Therefore, other models need to be explored. Combining the advantages of neural networks and D-S evidence theory, this paper proposes a multisensor data fusion algorithm based on the multineural network and D-S evidence theory. Firstly, the characteristic values of multiple sensor data are extracted by the wavelet analysis method, and then, the membership degree of each sensor signal and the coal mine safety state are obtained by the membership function method in fuzzy mathematics. Then, effective grouping is carried out, and a neural network is designed for each group of data. Secondly, the output of the neural network is normalized as the basic probability distribution function of D-S evidence theory; then, we use the fusion principle of D-S evidence theory to fuse multiple evidence spaces and finally draw a conclusion.

To verify the prediction accuracy of the algorithm, this paper compares the Mae index, max error index, and RMSE index. The performance index is shown in Figure 9, the distribution probability corresponding to the fused disaster category is greater than the threshold value 0.8, and the diagnosis results can be given immediately, and the diagnosis result is completely consistent with the actual disaster category. Therefore, the method in this paper obviously improves the deficiency of using only a single disaster prediction method and has a good fusion effect for the disaster prediction of uncertain results. Therefore, it has strong feasibility. According to the advantages and disadvantages of D-S evidence theory and BP neural network, the improved D-S evidence theory and BP neural network algorithm are introduced, and a fusion algorithm combining the two methods is proposed; that is, the advantages of the two methods are used, and the method of separating the evidence is used to solve the limitation of D-S evidence theory that requires the evidence to be relatively independent, and the problem of insufficient accuracy of the neural network due to large data volume is solved. This method effectively improves the accuracy of underground safety state judgment.