Museum intelligent warning system based on wireless data module

Weixian Wang

doi:10.1515/nleng-2025-0158

Article Open Access

Museum intelligent warning system based on wireless data module

Weixian Wang

Published/Copyright: September 24, 2025

Published by

Become an author with De Gruyter Brill

Submit Manuscript Author Information

From the journal Nonlinear Engineering Volume 14 Issue 1

Abstract

Fire can cause significant damage to the personal safety of museum personnel and the integrity of exhibits. Many museums are choosing to install intelligent fire warning systems. In response to the low accuracy of fire prediction and recognition in most museums’ intelligent warning systems, this study proposes a museum intelligent warning system based on wireless data modules. This study first establishes an environmental monitoring system within the museum, using temperature sensors distributed throughout the museum for real-time monitoring, and transmitting the monitoring data to the upper computer for processing through a wireless data module. On this basis, a warning system based on an environmental prediction model and a fire recognition model is constructed. The results indicated that the temperature threshold range for fire was greater than 28°C, with humidity less than 48%, smoke concentration greater than 1,800 ppm, CO concentration greater than 1,830 ppm, and toxic gas concentration greater than 2,000 ppm. All the indicators tested in actual situations met the threshold range. This indicates that the proposed warning system can effectively respond to fire situations within museums, providing effective measures for the safety of museum personnel and the protection of cultural relics.

Keywords: wireless data; fire; early warning system; LSTM

1 Introduction

The preservation and protection of cultural relics have always been a topic of great concern in museums. The collection of cultural relics is of great significance for the exploration of historical culture. Museums are not only a microcosm of the historical development of a region or a country, but also a primary place for the general public to acquire knowledge and explore history [1,2]. Due to the different materials and ages of cultural relic collections, the requirements for preservation environment also vary. Therefore, the protection of cultural relics is an important and arduous task. Environmental factors such as temperature and humidity greatly affect the preservation quality of cultural relics [3]. At the same time, museums have abundant collection resources and high foot traffic. Once a fire accident occurs, it will cause unpredictable losses to the entire society and even human development [4,5]. In recent years, many Museum intelligent warning systems (MIWS) have been constructed to monitor museum environments in real-time and prevent museum fire accidents. However, the current warning systems in museums often suffer from inaccurate monitoring data and low real-time performance. Specifically, the accuracy of monitoring data in predicting and identifying fires in many existing systems is often limited by the performance of sensors and environmental factors, such as temperature and humidity changes, which makes the assessment of fire risk unreliable. Many monitoring systems are unable to achieve real-time data updates and respond promptly to sudden fire risks, resulting in ineffective protection of personnel and exhibits. Traditional fire detection techniques mostly rely on a single monitoring indicator and fail to fully utilize modern Deep Learning (DL) and Internet of Things technologies, making it impossible to achieve intelligence and efficiency. Based on this, this study proposes a Wireless data module (WDM)-based MIWS. First, an environmental monitoring system (EMS) is designed, followed by the proposal of an environmental prediction model (EPM) and a fire identification model (FIM) to construct MIWS.

With the continuous development of science and technology, wireless data technology has been widely utilized in museum warning systems. Nguyen et al. [6] proposed a novel wireless artificial intelligence architecture for the evolution of future intelligent networks and demonstrated its effectiveness. Tang et al. [7] proposed a highly integrated passive wireless sensing system to address the problem of simultaneous streaming of multi-sensor data. The system was capable of simultaneously collecting and identifying concurrent data streams from multiple sensor tags. Mabrouki et al. [8] proposed an automatic weather monitoring system to address practical application issues of monitoring systems. This system could effectively collect regional climate data. To replace the current agricultural monitoring system technology, Suji Prasad et al. [9] established an intelligent agricultural management and monitoring system based on LoRa. This intelligent agricultural platform has improved the efficiency of agricultural technology. Yao et al. [10] classified simultaneous wireless power and data transfer systems into four categories to address real-time communication issues in wireless power transmission systems. Finally, its application effectiveness was demonstrated based on the classification results. Adams et al. [11] adopted a sepsis early warning system to address the issue of early identification of sepsis. This early warning system had the potential to identify patients with early sepsis. Tao et al. [12] proposed a smart identification and real-time warning approach for various complex events in horizontal well fracturing, which provides support for decision-making. To improve the safety of smart cities, Wang et al. [13] proposed a new rear-end collision prediction mechanism (RCPM) based on DL methods. RCPM effectively improved the performance of predicting rear-end collisions. Cho et al. [14] put forward and implemented a DL-based early warning scheme to address the issue of the low performance of traditional tracking and triggering systems in rapid response systems. This system helped to make precise decisions in daily practice. Ko et al. [15] used a multi-agent driving simulation method to address the safety benefits of cooperative intelligent transportation systems when providing in-vehicle warning information for quantitative evaluation. The method performed well.

In summary, numerous industry researchers have designed various warning systems for the preservation of museum artifacts, which also incorporate many new technologies. However, few scholars have applied wireless data technology to early warning systems. Given this issue, this study constructs a WDM-based MIWS for environmental monitoring and fire prevention in museums, aiming to enhance the museum’s collection protection capabilities. The innovation of the research lies in combining wireless data technology with DL models to design an efficient and accurate museum environment monitoring and warning system. The contribution of the research lies in achieving comprehensive real-time monitoring of the museum environment through a comprehensive environmental monitoring system. The combination of DL models and environmental monitoring systems has solved the prediction problem of long sequence time data, improving the ability to identify fires under complex environmental changes. The system has good accuracy and real-time performance in fire prediction and environmental monitoring, providing an effective fire safety solution for museums.

2 Methods and materials

Cultural relics in museums need to be stored properly. Abnormal temperature and humidity changes on site can cause damage to cultural relics, affecting both their economic and cultural value [16]. Due to the insensitivity of most museums to temperature changes and the lack of consideration of fire warning, this study proposes a WDM-based MIWS.

2.1 Museum EMS design

The museum has a large number of visitors and high mobility, making it difficult to prevent fires. The harm of fire to the collections in museums is enormous. The collections in museums are often highly sensitive to temperature changes [17]. Based on this, EMS is first established within the museum. The hardware design framework of EMS is shown in Figure 1.

Figure 1

Hardware design in environmental monitoring system.

In Figure 1, the hardware of the museum EMS includes a wireless transmission module, control module, energy module, console, alarm device, etc. The various modules work together to form the museum EMS. Among them, sensors collect information from within the museum and transmit it to the control console. The console processes and displays it through the monitor. By setting a threshold, it is possible to determine whether the indoor environment temperature meets the control standards, which will be reported and processed by technical personnel. In addition, the system also has an energy module and a wireless transmission module, which are mainly used to supply power to each module through voltage reduction and upload the data processed by the console to the upper computer through ZigBee transmission technology. ZigBee wireless transmission technology is a short-range wireless communication protocol, similar to WiFi and Bluetooth [18]. The ZigBee protocol framework is exhibited in Figure 2.

Figure 2

ZigBee protocol architecture.

In Figure 2, the functions of each structural layer are as follows. The application layer provides interfaces and services for computer users. The function of the presentation layer is for data processing, such as encoding and decoding, encryption and decryption, etc. The session layer is responsible for establishing, maintaining, and reconnecting communication sessions. The transport layer can manage end-to-end communication connections. The network layer decides the data path in the network. The data link layer manages data communication between adjacent nodes. The physical layer represents the optoelectronic properties of data communication. The EMS hardware driver program of the museum includes the abovementioned hardware modules. Among them, the wireless transmission module is one of the core modules, which aims to form a wireless network and complete wireless communication. The period of networking the wireless communication module is shown in Figure 3.

Figure 3

Wireless communication module networking process diagram.

In Figure 3, the wireless communication module first initializes the hardware, and then the sub-nodes scan the channel and join the network to determine whether the networking is successful. If unsuccessful, the child node scanning process needs to be returned to re-scan. If successful, it will check if there are any data packets. If not, it will return to the network steps. If it is determined that there is a data packet, the child node sends the data packet to the coordinator node. The data are transmitted to the upper computer for processing through a wireless transmission module. The upper computer will determine whether the data have reached the set threshold. The display of data information can provide technical personnel with real-time environmental status, help people judge the current environmental condition of the museum, and predict the trend of changes in the future to assist staff in making corresponding decisions. Data storage can store actual sample data from environmental monitoring and preserve samples for model training. The setting of thresholds can assist personnel in better monitoring the environment and making corresponding decisions based on actual situations. When the environmental monitoring data reach the threshold, an alarm device will be triggered to remind staff to take action to prevent situations where temperature, humidity, and other indicators in the museum exceed expected values, which may have adverse effects on the exhibits.

2.2 Establishment of museum EPM

Due to the high mobility of personnel within the museum and the high risk of fire, this study designs a fire identification and warning system built on the EMS design in the museum to reduce the degree of damage to personnel and exhibits caused by fires. This study first constructs an EPM based on the obtained museum environmental monitoring data. Subsequently, based on this, an FIM is constructed to jointly identify museum fires using two models. Given the data collection characteristics within the museum, a long short-term memory (LSTM) suitable for long-term equidistant time series is selected. LSTM is a type of recurrent neural network (RNN). The emergence of LSTM is mainly used to address the long-term dependency issue of traditional RNN [19]. The structure of the LSTM cell is shown in Figure 4.

Figure 4

LSTM cell structure diagram.

In Figure 4, compared to RNN, there is an additional C_t between cells in addition to the previous hidden state ht, and the internal structure is relatively complex. The LSTM cell is divided into four parts: Keep gate (KG), Update gate (UG), Write gate (WG), and Output gate (OG). Among them, KG is used to control the forgetting or retention of information in memory units. The operating formula of KG is shown in Eq. (1).

(1) f t = σ ( W f ⋅ [ h t − 1 , x t ] + b f ) ,

where f t is the output of KG. σ is the Sigmoid function. W f and b f are the weight and deviation. h t − 1 denotes the previous hidden state, and x t is the current input. UG decides to update the information of memory units, including Sigmoid and activation function Tanh, both of which contain the input at the current time and the hidden state from the previous time. The parameters of the Sigmoid part include W i i , W c i , W h i , and b h i . After obtaining the two outputs of UG, the updated state can be obtained by calculating the output obtained by KG with the previous state, as shown in Eq. (2).

(2) i t = σ ( W i i x t + W h i h t − 1 + W c i c t − 1 + b i ) .

In Eq. (2), i t is the output of UG. W i i means the weight from the input layer to UG. W c i is the weight from the previous memory unit to UG. W h i is the weight from the previous hidden layer to UG. The other output g t of UG is calculated as shown in Eq. (3).

(3) g t = tanh ( W i g ⋅ x t + b i g + W h g ⋅ h t − 1 + b h g ) .

The update status calculation is done using Eq. (4).

(4) c t = f t ∗ c t − 1 + i t ∗ g t ,

where c t refers to the updated state. c t − 1 is the state from the previous moment. The function of OG is to read the newly updated neural network state and output it to memory units, with specific information output controlled by OG. Output o t is obtained through OG, and finally the current hidden state h t is obtained by o t and c t , which can be passed on to the next cell. The output o t of OG is calculated as Eq. (5).

(5) o t = σ ( W i o ⋅ x t + b i o + W h o ⋅ h t − 1 + b h o ) .

The calculation of the current hidden state h t is done using Eq. (6).

(6) h t = o t × tanh ( c t ) .

In response to the problem that traditional LSTM can only forward process data, this study introduces Bidirectional LSTM (Bi-LSTM) to better construct prediction models. Bi-LSTM is an extension of traditional LSTM aimed at improving the model’s understanding of context in sequence prediction tasks [20]. The forward processing of data by Bi-LSTM is shown in Eq. (7).

(7) h → t = f ( U 1 h → t − 1 + W 1 x t + b 1 ) ,

where h → t is the hidden layer state of forward processing data. U 1 is the hidden layer weight coefficient. W 1 is the input layer weight, and b 1 is the bias vector. Bi-LSTM reverse processes the data as shown in Eq. (8).

(8) h t ← = f ( U 2 h t + 1 ← + W 2 x t + b 2 ) ,

where h t ← is the hidden layer state of the reverse processed data. U 2 is the hidden layer weight coefficient. W 2 is the input layer weight. b 2 is the bias vector. The final output result is shown in Eq. (9).

(9) h t = h → t ⊕ h t ← ,

where the forward and reverse outputs are concatenated to output the hidden state. The method for constructing a museum fire prediction model based on Bi-LSTM is exhibited in Figure 5.

Figure 5

Museum environment prediction model framework.

In Figure 5, the environmental data collected through EMS are normalized. Subsequently, the processed data are segmented into a training set for training and a testing set for evaluation in an 8:2 allocation. The processing method adopts the range transformation method. The initial data are linearly transformed and mapped to the same interval, aiming to reduce data differences and improve prediction accuracy. The processing method is shown in Eq. (10).

(10) Y = M N ,

where Y is the processed data. M is the difference between the minimum feature and the initial data. N is the difference between the maximum feature and the minimum feature. The calculation of M and N is done using Eq. (11).

(11) M = X − X min N = X max − X min ,

where X is the initial data. X min and X max are the minimum feature and maximum feature.

2.3 Establishment of museum FIMs

EPM can prevent fires to a certain extent, but it cannot be relied on solely. Convolutional neural network (CNN) can automatically learn and extract key features from data in feature extraction, while maintaining a certain degree of invariance to the position and deformation of the data. Therefore, based on CNN feature extraction, this study adds an improved LSTM model, Gated recurrent unit (GRU), to construct a Museum FIM (CNN-GRU). The model first uses CNN for feature extraction, and then maps the output features into sequence vectors that are input to GRU for training, obtaining prediction results. GRU is an improved model of LSTM that integrates KG and UG into a single update gate, while mixing neuron states and hidden states, effectively alleviating the problem of “gradient vanishing” in RNN. It can also reduce training parameters while maintaining training effectiveness. The calculation of each state in GRU is shown in Eq. (12).

(12) r t = σ ( W r ⋅ [ h t − 1 , x t ] + b r ) ,

where r t , W r , and b r are the gating, input weights, and bias vectors of the reset gate. The calculation for updating the door is shown in Eq. (13).

(13) z t = σ ( W z ⋅ [ h t − 1 , x t ] + b z ) ,

where z t is the update gate control. W z and b z are the input weights and bias vectors for updating the gate. The current memory information h ˜ t is shown in Eq. (14).

(14) h ˜ t = tanh ( W h ˜ ⋅ [ r t × h t − 1 , x t ] + b h ) ,

where W h ˜ is the input weight from the hidden layer to the reset gate at the current time. b h is the bias vector of h ˜ t . The updated expression is shown in Eq. (15).

(15) h t = ( 1 − z t ) × h t − 1 + z t × h ˜ t .

The schematic diagram of CNN-GRU is shown in Figure 6.

Figure 6

CNN-GRU diagram.

CNN consists of two convolutional layers and two pooling layers. The method is chosen as the same convolution, and the activation function is selected as Sigmoid. After convolution, pooling is performed and input to the GRU. GRU learns the extracted feature vectors. Building a two-layer GRU structure can get the most excellent prediction performance, and eventually, the output of the fully connected layer is denormalized to gain the final forecasting value. When training the GRU, the Adam algorithm is used to iteratively update weights. Through momentum and adaptive learning rate, the weights and biases of each neuron are continuously updated to gain the optimal output of the function. Research data collection involves real-time monitoring of environmental data through temperature, humidity, smoke, carbon monoxide, and toxic gas concentration sensors distributed throughout the museum. The sensor transmits the collected data to the console through a ZigBee wireless module. Data cleaning is carried out according to the following steps: noise removal, standardization processing, and temporal sorting. The cleaned data are transmitted to the upper computer for centralized processing and storage. On this basis, thresholds are set to monitor data in real-time. In the model input stage, the cleaned and normalized data are divided into a training set and a testing set. The data are divided using an 8:2 ratio. During the environmental monitoring phase, the system will regularly collect new real-time data from sensors and use this latest data as input to gradually update the Bi-LSTM model for environmental prediction.

3 Results

To verify the effectiveness of the proposed WDM-based MIWS, the functionality of the designed museum detection system was first tested. Subsequently, an analysis of the EPM performance of the museum was conducted, and finally, the performance of the museum’s FIM was tested and analyzed.

3.1 Museum EMS design analysis

The design and development environment for museum EMS software is microcontroller development kit for advanced RISC machine (MDK-ARM). The evaluation indicators for EPM are RMSE and pearson correlation coefficient (PCC). Among them, RMSE is the square root of the sum of the squares of the difference between the observed and true values, divided by the number of observations. PCC is a statistical indicator used to measure the degree of linear correlation between two variables, and can also be used to evaluate the predictive performance of a model. When the PCC value is 1 or −1, it indicates that the two variables are completely positively correlated or completely negatively correlated. PCC = 0 indicates that there is no linear correlation between the two variables. The basic environment of the model relies on PyTorch 1.6. 1,688 sets of environmental parameters are collected in the FIM, divided into non-fire and fire categories, and preprocessed. The data are randomly selected into a training set and a test set in an 8:2 ratio, and then trained. The model evaluation index adds the Accuracy (ACC) index on the basis of EPM. ACC is the degree to which the mean of multiple measurements under certain conditions matches the true value, utilized to indicate the magnitude of the error. Table 1 lists the settings for the prediction model and recognition model.

Table 1

Model hyperparameter settings

–	Number of hidden layers	Number of single-layer neurons	Iterations	Batch size	Learning rate
Bi-LSTM	1	30	600	33	0.001
CNN-GRU	—	—	100	128	0.001

To verify the performance of EMS, the experiment analyzes 1,000 pieces of data collected by a museum on May 22, 2024. Table 2 contains partial data.

Table 2

Partial monitoring data of environmental monitoring system

Time		2024.5.22 13:36:01	2024.5.22 13:36.12	…	2024.5.22 17:54:28
Temperature	Monitor value (°C)	27	27	…	27
Temperature	Actual value (°C)	27	27	…	27
Humidity	Monitor value (%)	28	28	…	28
Humidity	Actual value (%)	28	28	…	28
Smoke density	Monitor value (ppm)	1,550	1,760	…	1,890
Smoke density	Actual value (ppm)	1,552	1,756	…	1,894
CO concentration	Monitor value (ppm)	930	966	…	975
CO concentration	Actual value (ppm)	934	960	…	980
Toxic gas	Monitor value (ppm)	1,510	1,525	…	1,540
Toxic gas	Actual value (ppm)	1,514	1,522	…	1,540
Illumination	Monitor value (LUX)	350	352	…	366
Illumination	Actual value (LUX)	354	350	…	368

In Table 2, the monitoring system monitored the museum from 13:36:01 to 17:54:28 on May, 22, 2024. The temperature is roughly stable at 27°C, and the humidity is around 28%. The smoke concentration ranges from 1,550–1,890 ppm, the CO concentration ranges from 930–975 ppm, and the toxic gas fluctuates between 1,510 and 1,540 ppm. The illuminance is between 350 and 366 LUX. Compared with the actual values, the monitored values are basically consistent. The data collected by the monitoring system show that the proposed museum EMS can effectively carry out environmental monitoring, with accurate and clear monitoring indicators and high real-time efficiency.

3.2 Performance analysis of museum EPM

To verify the performance of the museum EPM, a comparative experiment is conducted between Bi-LSTM and LSTM models, with RMSE and PCC indicators as evaluation criteria. Subsequently, experiments are conducted on the prediction accuracy of the Bi-LSTM model. Figure 7 shows a comparison between Bi-LSTM and LSTM models.

Figure 7

Comparison of performance between two models. (a) Evaluation of RMSE indicators for two models. (b) Evaluation of PCC indicators for two models.

In Figure 7, the RMSE and PCC indices of LSTM are around 0.0439 and 0.9846, while the RMSE and PCC indices of Bi-LSTM are around 0.0423 and 0.9862. This indicates that Bi-LSTM has better prediction performance and superior model performance compared to LSTM. To verify the accuracy of Bi-LSTM in environmental prediction, a prediction experiment is conducted on the environmental data inside the museum, and the data are processed. The results are shown in Figure 8.

Figure 8

Comparison between Bi-LSTM model prediction results and actual values. (a) Prediction results of Bi LSTM model for temperature and humidity. (b) Bi LSTM prediction results for smoke and CO gas.

In Figure 8(a), the predicted values of temperature by Bi-LSTM are almost identical to the true values. The predicted values for humidity at sample sizes of 25 and 30 are 0.9 and 1, respectively, while the actual values are 0.8 and 0.9. There are only minor differences. In Figure 8(b), the predicted values of Bi-LSTM for smoke are not significantly different from the true values, with the error < 0.1 between the sample sizes of 15 and 35. In the detection of CO gas, the difference between the predicted and actual values is also small, and the two curves are basically consistent. This indicates that Bi-LSTM has high accuracy and good predictive performance for predicting the environment inside museums, and can be applied to environmental prediction inside museums.

3.3 Performance analysis of museum FIMs

To validate the performance of the museum FIM, the results of model training and testing are first analyzed. Subsequently, RMSE, PCC, and ACC are used to evaluate the CNN-GRU, CNN-Bi-LSTM, and CNN-LSTM models, respectively. Finally, based on the recognition situation, the threshold of the fire warning system indicators is obtained, and the threshold is compared with the experimental indicators to prove the correctness of the threshold indicators of the warning system. Figure 9 shows the training and testing results of the CNN-GRU.

Figure 9

ACC of CNN-GRU model on training and testing sets.

In Figure 9, the CNN-GRU identifies 118 non-fire samples and 0 fire samples in the test set. Among the samples with fire in the test set, 12 are identified as non-fire samples and 126 are identified as fire samples. The data show that the ACC of the CNN-GRU on the test set is 95.28%, indicating that the model has a good recognition effect on whether there is fire. To further validate the performance of the CNN-GRU, comparative models and evaluation metrics are selected, and comparative experiments are conducted. The results are shown in Figure 10.

Figure 10

Evaluation of three models by different indicators. (a) Evaluation results of RMSE index on three models. (b) Evaluation results of PCC index on three models. (c) Evaluation results of ACC index on three models.

In Figure 10(a), the experimental values in the RMSE evaluation index show that CNN-LSTM is around 0.3113, CNN-Bi-LSTM is around 0.2891, and CNN-GRU is around 0.2183. In Figure 10(b), the experimental values of the three models in the PCC evaluation index are around 0.8385, 0.8571, and 0.9198, respectively. In Figure 10(c), the ACCs of the three models are around 91.99, 93.38, and 96.38%, respectively. This indicates that CNN-GRU has smaller RMSE values and larger PCC and ACC values compared to other models, demonstrating better reliability, higher recognition accuracy, and stronger performance of the model. By correlating the results of CNN-GRUFIM with the indicators of the actual environment, the threshold indicators of the museum fire warning system can be obtained. The comparison between the obtained threshold indicators and the actual detection indicators can prove the correctness of the proposed indicators, as displayed in Figure 11.

Figure 11

The threshold range of the warning system and the actual indicator results. (a) Fire temperature threshold and measured value. (b) Fire humidity threshold and measured value. (c) Fire smoke threshold and measured value. (d) Fire CO threshold and measured value. (e) Threshold and measured values of toxic gases with fire.

In Figure 11(a), the temperature threshold range with fire is greater than 28°C, and the actual detected temperatures with fire are 38, 36, 35, and 35°C. In Figure 11(b), the humidity threshold range is less than 48%, and the detected humidity with fire is 44, 42, 42, and 41%. In Figure 11(c), the threshold range for smoke concentration with fire is greater than 1,800 ppm, and the smoke concentrations are 1,920, 1,910, 1,860, and 1,940 ppm. In Figure 11(d), the threshold range for CO concentration is greater than 1,830 ppm, and the actual detected CO concentrations are 1,920, 1,940, 1,960, and 1,940 ppm. In Figure 11(e), the threshold range for toxic gas concentration is greater than 2,000 ppm, and the actual detected toxic gas concentrations are 2,240, 2,120, 2,140, and 2,110 ppm. The data show that the threshold of fire occurrence indicators obtained from the warning system can effectively reflect whether a fire has occurred inside the building.

4 Discussion and conclusion

One of the main purposes of establishing a museum is to properly preserve museum artifacts. However, due to environmental factors such as temperature and humidity, the problem of damage to cultural relics is still very serious. This study proposed a WDM-based MIWS for environmental monitoring and fire warning in museums. The effectiveness of MIWS was verified through the design of EMS, EPM, and FIM. In the experiment, the temperature monitored by EMS remained roughly stable at 27°C, and the humidity was around 28%. The smoke concentration ranged from 1,550–1,890 ppm, the CO concentration ranged from 930–975 ppm, and toxic gases fluctuated between 1,510–1,540 ppm. The illuminance was between 350 and 366 LUX. Compared with the actual values, the monitored values were basically consistent. The RMSE of LSTM and Bi-LSTM models were around 0.0439 and 0.0423, and the PCC index was around 0.9846 and 0.9862. In the evaluation indicators, the RMSE of CNN-LSTM, CNN-Bi-LSTM, and CNN-GRU were around 0.3113, 0.2891, and 0.2183, PCC was around 0.8385, 0.8571, and 0.0.9198, and ACC was around 91.99, 93.38, and 96.38%, respectively. The experiment showed that the proposed museum EMS could effectively carry out environmental monitoring, and the monitoring indicators were relatively accurate and clear, with high real-time monitoring efficiency. Bi-LSTM had better prediction performance and superior model performance compared to LSTM. Compared with other models, the RMSE value of CNN-GRU was smaller, while the PCC and ACC values were larger, indicating that the CNN-GRU has better reliability, higher recognition accuracy, and stronger performance. Reddy et al. [21] proposed an artificial intelligence-based recursive neural network and whale optimization framework for early estimation of fire risks in smart cities. This framework uses Internet of Things sensors to collect environmental parameters, saves the data to a cloud storage system, and uses whale optimization algorithm to adjust the parameters of the prediction model. The experimental results show that the highest accuracy of the model reaches 99.5%. Regarding the issue of fire detection and early warning, researchers such as Zheng et al. [22] have developed a real-time analysis system for fire detection and early warning. The system is based on the most advanced object detection algorithm and has constructed a power system fire image dataset containing 7,634 fire images for training object detection models. In addition, genetic algorithms are used to optimize the hyperparameters in the model. The results showed that the accuracy and recall of the model were 0.985, and the recognition speed was about 207 fps. Compared with the proposed methods, the methods proposed in the related studies have higher accuracy. For example, using a whale optimization algorithm can improve prediction accuracy by intelligently adjusting hyperparameters, thereby effectively improving the performance of the model. However, this method requires a large amount of computing resources, especially when training cloud storage and DL models, which have high infrastructure requirements. In addition, the performance of the proposed systems is limited by the diversity and quality of image data. If there is not enough diversity, it will affect the generalization ability of the model. Overall, the proposed warning system uses multiple environmental parameters such as temperature, humidity, and smoke concentration for comprehensive evaluation, which can comprehensively reflect the potential risks of museum fires. Fast data transmission and real-time monitoring have been achieved through the ZigBee wireless protocol, improving the system’s response speed. By combining Bi-LSTM and CNN-GRU models, and utilizing historical and sensor data for DL, the ability to predict and identify fires has been improved. Therefore, the proposed method is more practical in addressing museum warning issues. However, this study only selected commonly used environmental indicators for fires, and the next step could be to consider adding more targeted indicators in museum scenes.

Funding information: The research is supported by: Social Science Planning Project of Henan Province, Research on Henan Red Gene Inheritance (No. 2023 ZT 047).
Author contributions: Author has accepted responsibility for the entire content of this manuscript and approved its submission.
Conflict of interest: Author states no conflict of interest.
Data availability statement: All data generated or analyzed during this study are included in this published article.

References

[1] Song HJ. Terahertz wireless communications: Recent developments including a prototype system for short-range data downloading. IEEE Microw Mag. 2021;22(5):88–99.10.1109/MMM.2021.3056935Search in Google Scholar

[2] Tadeus DY, Yuwono T. Prototype of wireless sensor network based on open industrial protocol: Modbus. In IOP Conference Series: Materials Science and Engineering. Vol. 801, IOP Publishing; 2020. No. 1, p. 012140.10.1088/1757-899X/801/1/012140Search in Google Scholar

[3] Ye H, Li GY, Juang BH. Deep learning based end-to-end wireless communication systems without pilots. IEEE Trans Cogn Commun Netw. 2021;7(3):702–14.10.1109/TCCN.2021.3061464Search in Google Scholar

[4] Yang L, Huang J, Feng B, Zhang F, Zhang Y, Li X, et al. Undersea wireless power and data transfer system with shared channel powered by marine renewable energy system. IEEE J Emerg Sel Top Circuits Syst. 2022;12(1):242–50.10.1109/JETCAS.2022.3140954Search in Google Scholar

[5] Yao Y, Cheng H, Wang Y, Mai J, Lu K, Xu D. An FDM-based simultaneous wireless power and data transfer system functioning with high-rate full-duplex communication. IEEE Trans Ind Inform. 2020;16(10):6370–81.10.1109/TII.2020.2967023Search in Google Scholar

[6] Nguyen DC, Cheng P, Ding M, Lopez-Perez D, Pathirana PN, Li J, et al. Enabling AI in future wireless networks: A data life cycle perspective. IEEE Commun Surv Tutor. 2020;23(1):553–95.10.1109/COMST.2020.3024783Search in Google Scholar

[7] Tang K, Wang M, Liu M, Zhang C, Xu S, Ji S, et al. A highly integrated passive wireless sensing system with synchronized data streaming of multiple tags. IEEE Internet Things J. 2022;9(17):15525–37.10.1109/JIOT.2022.3179410Search in Google Scholar

[8] Mabrouki J, Azrour M, Dhiba D, Farhaoui Y, El Hajjaji S. IoT-based data logger for weather monitoring using Arduino-based wireless sensor networks with remote graphical application and alerts. Big Data Min Anal. 2021;4(1):25–32.10.26599/BDMA.2020.9020018Search in Google Scholar

[9] Suji Prasad SJ, Thangatamilan M, Suresh M, Panchal H, Rajan CA, Sagana C, et al. An efficient LoRa-based smart agriculture management and monitoring system using wireless sensor networks. Int J Ambient Energy. 2022;43(1):5447–50.10.1080/01430750.2021.1953591Search in Google Scholar

[10] Yao Y, Sun P, Liu X, Wang Y, Xu D. Simultaneous wireless power and data transfer: A comprehensive review. IEEE Trans Power Electron. 2021;37(3):3650–67.10.1109/TPEL.2021.3117854Search in Google Scholar

[11] Adams R, Henry KE, Sridharan A, Soleimani H, Zhan A, Rawat N, et al. Prospective, multi-site study of patient outcomes after implementation of the TREWS machine learning-based early warning system for sepsis. Nat Med. 2022;28(7):1455–60.10.1038/s41591-022-01894-0Search in Google Scholar PubMed

[12] Tao J, Wu H, Deng S, Qi Z. Overview of intelligent home security and early warning system based on internet of things technology. Int Core J Eng. 2022;8(5):727–32.Search in Google Scholar

[13] Wang X, Liu J, Qiu T, Mu C, Chen C, Zhou P. A real-time collision prediction mechanism with deep learning for intelligent transportation system. IEEE Trans Veh Technol. 2020;69(9):9497–508.10.1109/TVT.2020.3003933Search in Google Scholar

[14] Cho KJ, Kwon O, Kwon JM, Lee Y, Park H, Jeon KH, et al. Detecting patient deterioration using artificial intelligence in a rapid response system. Crit Care Med. 2020;48(4):e285–9.10.1097/CCM.0000000000004236Search in Google Scholar PubMed

[15] Ko J, Jang J, Oh C. Assessing the safety benefits of in-vehicle warning information by vehicle interaction analysis in C-ITS environments. J Korean Soc Transp. 2021;39(1):1–13.10.7470/jkst.2021.39.1.001Search in Google Scholar

[16] Becker JS, Potter SH, Vinnell LJ, Nakayachi K, McBride SK, Johnston DM. Earthquake early warning in Aotearoa New Zealand: A survey of public perspectives to guide warning system development. Humanities Soc Sci Commun. 2020;7(1):1–12.10.1057/s41599-020-00613-9Search in Google Scholar

[17] Hermans TDG, Šakić Trogrlić R, van den Homberg MJC, Bailon H, Sarku R, Mosurska A. Exploring the integration of local and scientific knowledge in early warning systems for disaster risk reduction: A review. Nat Hazards. 2022;114(2):1125–52.10.1007/s11069-022-05468-8Search in Google Scholar

[18] Luo N, Yu H, You Z, Li Y, Zhou T, Jiao Y, et al. Fuzzy logic and neural network-based risk assessment model for import and export enterprises: A review. J Data Sci Intell Syst. 2023;1(1):2–11.10.47852/bonviewJDSIS32021078Search in Google Scholar

[19] Thirugnanam H, Ramesh MV, Rangan VP. Enhancing the reliability of landslide early warning systems by machine learning. Landslides. 2020;17(9):2231–46.10.1007/s10346-020-01453-zSearch in Google Scholar

[20] Chung AI, Meier MA, Andrews J, Böse M, Crowell BW, McGuire JJ, et al. ShakeAlert earthquake early warning system performance during the 2019 Ridgecrest earthquake sequence. Bull Seismol Soc Am. 2020;110(4):1904–23.10.1785/0120200032Search in Google Scholar

[21] Reddy PDK, Margala M, Shankar SS, Chakrabarti P. Early fire danger monitoring system in smart cities using optimization-based deep learning techniques with artificial intelligence. J Reliable Intell Environ. 2024;10(2):197–210.10.1007/s40860-024-00218-ySearch in Google Scholar

[22] Zheng L, Zhang X, Wang H. Big data approach for fire prevention and warning for power systems. J Signal Process Syst. 2023;95(12):1391–403.10.1007/s11265-023-01857-9Search in Google Scholar

Received: 2024-09-12

Revised: 2025-05-23

Accepted: 2025-05-30

Published Online: 2025-09-24

This work is licensed under the Creative Commons Attribution 4.0 International License.

Articles in the same Issue

https://doi.org/10.1515/nleng-2025-0158

Keywords for this article

wireless data; fire; early warning system; LSTM

Creative Commons

BY 4.0