Design and Implementation of an Edge Computing-Based Underground IoT Monitoring System

Abstract

Underground mining operations face increasing challenges due to their complex and hazardous environments. One key difficulty is ensuring real-time safety monitoring and disaster prevention. Traditional monitoring systems often suffer from delayed data acquisition and rely heavily on cloud-based processing. These factors limit their responsiveness during emergencies. To address these limitations, this study presents an underground Internet of Things (IoT) monitoring system based on edge computing. The system architecture is composed of three layers: a perception layer for real-time sensing, an edge gateway layer for local data processing and decision-making, and a cloud service layer for storage and analytics. By shifting computation closer to the data source, the system significantly reduces latency and enhances response efficiency. The system is tailored to actual mine-site conditions. It integrates pressure monitoring for artificial expandable pillars and roof subsidence detection in stopes. It has been successfully deployed in a field environment, and the data collected during commissioning demonstrate the system’s feasibility and reliability. Results indicate that the proposed system meets real-world demands for underground safety monitoring. It enables timely warnings and improves the overall automation level. This approach offers a practical and scalable solution for enhancing mine safety and provides a valuable reference for future smart mining systems.

1. Introduction

Over the past decade, population growth, technological advancement, and economic expansion have sharply increased the global demand for mineral resources. This has placed mounting pressure on the mining industry. As shallow deposits are depleted, the industry is shifting toward deeper, increasingly inaccessible ore bodies to meet future demand [1,2]. As the mining depth increases, the ground pressure of the underground stope rises significantly. The high-stress concentration makes natural pillars prone to brittle damage or even rockbursts, thus failing to provide long-lasting support for the stope roof. To improve safety conditions for workers, underground stopes usually retain widely spaced continuous pillars, along with intermittently arranged circular or rectangular point pillars. To address the high loss rate and brittle load-bearing capacity of natural pillars in hard rock stopes, many mining companies are developing low-cost, high-performance artificial mine pillars. These pillars support the roof load originally borne by the natural pillars, thereby ensuring stope stability [3]. The mine pillars carry the overburden load in the upper part of the mining area and limit the deformation and damage of the roof. The domino effect from mine pillar failure can trigger widespread roof collapse and stope wall failure. This poses serious risks to personnel and equipment safety [4]. Therefore, it is necessary to comprehensively monitor the entire stope. This allows real-time tracking of mining progress and environmental conditions.

However, obtaining timely and comprehensive monitoring data in deep, complex underground environments remains a significant challenge. Extreme geological conditions, limited space, unreliable connectivity, and the distributed nature of operations make traditional manual inspections and wired monitoring methods inefficient, error-prone, and costly. Time delays between on-site events and data reporting can critically hinder timely response and decision-making, especially in emergencies. Moreover, traditional underground mining process data management still faces numerous problems. Many monitoring and management processes rely on manual writing or manual input for data recording, resulting in low efficiency and increased labor costs [5]. Such manual operations inevitably reduce data accuracy and significantly increase the rate of data loss. To address these challenges, mining researchers and practitioners have turned to advanced sensing and communication frameworks. Among these, the Internet of Things (IoT) has emerged as a transformative solution. Distributed sensors and connected devices enable real-time acquisition and transmission of environmental, personnel, and equipment data, laying the essential groundwork for smarter, safer deep mining operations [6,7,8]. With the wide application of IoT in underground mines [9,10,11,12], the ability of mines to sense, transmit, and process data has greatly improved, enhancing mine safety [13]. In addition, advances in smart sensors, embedded TCP/IP protocol stacks, GPRS technology, and wireless sensor networks (WSN) have enabled many mines to establish monitoring systems for collecting essential data in underground construction environments [14,15,16]. General monitoring systems typically upload data to the server layer for event discrimination and processing, a process susceptible to network delays during transmission to the cloud. Moreover, processing vast amounts of IoT-generated data predominantly in the cloud imposes heavy computational and storage burdens on servers, exacerbating latency and limiting real-time responsiveness [17]. With the emergence of edge computing, sensor data can be pre-processed at the gateway layer in real time. This enables event discrimination and exception handling, reducing decision-making response times and improving real-time event responsiveness [18]. In mining, edge computing has been applied to underground safety monitoring, equipment fault diagnosis, and environmental parameter prediction, enabling timely warnings and efficient resource allocation in complex underground environments [19,20,21].

Currently, several relatively mature mining IoT systems have been developed, such as the dynamic information platform proposed by Wu et al. [22] and the integrated monitoring frameworks reviewed by Zhang et al. [23]. In these systems, core data fusion and analysis are predominantly performed in the cloud. While this centralized architecture supports complex processing and cross-site data integration, it requires transmitting large volumes of sensor data to and from the cloud. This results in considerable latency and reduces the real-time performance of hazard detection in mining applications. More recent designs, such as the robust end-to-end IoT system by Vlachos et al. [24], have introduced edge computing to reduce transmission loads and partially improve real-time capabilities. However, in these systems, most multi-source data fusion and decision-making still occur at the cloud level. As a result, their responsiveness remains constrained by transmission delays and centralized processing bottlenecks. These limitations highlight the need for a more balanced architecture, where critical data fusion, anomaly filtering, and event discrimination are executed at the edge. This approach ensures rapid hazard detection and operational resilience while still leveraging cloud resources for long-term storage, large-scale analytics, and model training.

To address these challenges, this paper applies Internet of Things (IoT) technology to underground mining process data acquisition and designs an edge computing-based underground pillar monitoring platform. The platform adopts a three-layer architecture consisting of the perception layer, edge gateway layer, and cloud service layer. It supports real-time monitoring of stress and strain in the underground mining process, as well as sensing device management, authority control, and historical data queries, thereby improving the efficiency of underground mining management [25]. A key feature is the data preprocessing function of the edge computing gateway, which serves as a critical implementation of edge computing in this system. At the gateway level, raw multi-source sensor data are aggregated, synchronized, and processed using anomaly detection algorithms before being transmitted to the cloud. This local processing approach allows the system to filter erroneous or irrelevant data, perform preliminary event classification, and generate early warnings directly at the edge. Consequently, emergency events can be identified in near real time, and only high-quality, relevant data are uploaded to the server. This reduces bandwidth consumption, prevents cloud dataset distortion, and enhances decisionmaking accuracy [26]. The three-layer architecture, consisting of the perception layer, edge gateway layer, and cloud service layer, enables seamless coordination between real-time data acquisition, on-site edge processing, and centralized long-term analytics. This integrated approach can be effectively applied to monitoring various stages of the mining process and helps overcome the fragmented nature of Industrial IoT applications in the mining sector.

2. Materials and Methods

2.1. Design of the System

Due to the special conditions of underground mining, the engineering involved is intricate and the space is confined. Stopes often use pillars to support the roof during mining. The stope is a high-risk location for accidents, which can easily cause catastrophic accidents such as collapse, water penetration, and the risk is relatively high. Therefore, a safety monitoring system with high accuracy and real-time warning is required [27]. This paper proposes a pillar IoT safety monitoring system based on edge computing, with its overall architecture shown in Figure 1. The model consists of three layers: the terminal node layer, edge computing layer, and cloud computing service layer. These are interconnected through interfaces to form an edge computing network. The terminal node layer comprises data collection devices and sensors, selected according to actual application scenarios, such as measuring stress and strain. The edge computing layer is the core of the system. It leverages computing, networking, and storage resources at the network edge to deploy local tasks on edge servers, enabling rapid computation, real-time responses, security warnings, and anomalous data filtering. This ensures timely alerts and addresses potential hazards at an early stage [28].

The system can integrate multi-dimensional data to identify events more accurately. Because event decisions are made at the edge gateway, network transmission delays are reduced, and data loss at the monitoring center due to network failures is avoided, ensuring timely warnings. For events with high real-time requirements, the system ensures rapid security responses, while events with lower real-time demands are uploaded to the cloud for further processing.

2.2. Sensor Design

2.2.1. Pressure Sensor

The pressure sensor is a lightweight pressure measuring instrument suitable for mines, mainly used for ground pressure observation. It is a flat box welded from two thin steel plates of the same shape, as shown in Figure 2. The shape and size can be adjusted according to the conditions of use. The sensor features a simple structure, small size, easy processing, common materials, and light weight. Its stability and sensitivity meet the requirements of mine measurement.

The working principle is that the force plate directly bears the pressure of the roof rock layer and transmits it to the pillow pressure box, which is filled with oil. This causes the internal hydraulic pressure to rise gradually, which is then detected by the sensor. The hydraulic pressure is measured by detecting changes in the deformation of an elastic diaphragm using a resistance strain gauge, thus providing the measured pressure value.

The hydraulic pillow (pressure sensor) used in this system has a measurement range of 0–60 MPa, an accuracy of 1.6% FS, a resolution of 2 MPa, and a load capacity of 30 tons with a minimum resolution of 0.25 tons. In the field application, the sensors were installed between the expansion unit and the support unit of the artificial expandable pillars and operated continuously for 22 days. The sampling frequency was set to once every 30 min to capture pressure variations while optimizing power consumption. Data acquisition was carried out under actual production conditions in the stope, ensuring the monitoring captured both stable operational loads and dynamic changes caused by mining activities.

2.2.2. Displacement Sensor

The system uses photosensitive displacement sensors, as shown in Figure 3. When the system is operational, the laser diode emits laser pulses aimed at the target (stope roof), scattering light in all directions upon reflection, with part of the scattered light returning to the sensor receiver, being processed by the optical system, and then im-aged onto the avalanche photodiode. The photosensitive sensor utilizes photosensitive elements internally, which are capable of converting the received light signals into electrical signals of the corresponding size. The avalanche photodiode is an optical sensor with an internal amplification function; thus it can detect extremely weak light signals. The processor records and calculates the time elapsed from the emission of light pulses to the return of the received to determine the deformation of the stope roof [29].

The displacement sensor used in this system has a measurement range of 0.03–80 m, a resolution of 0.1 mm, and a repeatability of ±1 mm. Its absolute accuracy is ±(1 mm + D × 0.0005), where D represents the measurement distance. In the field application, the sensors were installed in the stope to monitor the roof subsidence and operated continuously for 22 days, with a sampling frequency of 30 min. This setup ensured that the displacement variations of the stope roof were effectively captured under actual production conditions, providing accurate and reliable data for deformation monitoring.

2.3. ZigBee Terminal Node

This layer is mainly composed of perceptual node modules such as the CC2530 core chip, ZigBee antenna module, RS-232 serial module, LED and sensor module, A/D conversion module, and POWER/RESET module, as shown in Figure 4. Those equipment were all sourced from Ebyte Electronic Technology Co., Ltd. (Chengdu, China). The expansion performance of the gateway is relatively good and can support a variety of types of sensing device access, including wired connections such as serial, Modbus, and wireless methods like ZigBee. Because of the number of sensors deployed in this system and easy to failure, the topology of WSN will be changed frequently. Considering the requirements of low power and low cost, the wireless front-end sensing nodes mainly use the ZigBee self-organizing network to form a wireless sensing network [30]. In this system, the ZigBee network adopts a star topology, with the edge gateway serving as the central coordinator and all sensing nodes communicating directly with it. A total of five sensing nodes are deployed in the monitoring area, each transmitting a small data payload to the gateway once every 30 min, which keeps network traffic low and power consumption minimal. This configuration reduces routing complexity, simplifies network maintenance, and ensures stable communication in the confined and obstacle-rich underground stope environment. ZigBee technology is often used in remote control applications, does not require huge amounts of communication, and can be embedded in a wide range of devices [31]. ZigBee is a wireless sensor technology based on IEEE802.15.4 [32] with low price, low data rate, and high efficiency [33]. Compared with other wireless communication technologies such as Bluetooth, WiFi, and LoRa, ZigBee offers significantly lower power consumption, greater scalability for large-scale node deployment, and more robust performance in interference-prone environments [34]. Previous comparative analyses [35] have shown that ZigBee’s ability to form self-organizing networks, combined with its high node capacity and resistance to electromagnetic interference, makes it a more reliable choice for long-term monitoring in challenging industrial environments such as underground mines. Because of these features, this technology is more suited for handling data transmission tasks in underground mining environments [36]. The gateway is installed near the stope, close to the terminal sensing devices, to maintain a stable connection and minimize transmission delays. Processed data from the gateway is forwarded to the cloud platform through the mine’s existing wired Ethernet network, ensuring reliable, high-bandwidth communication in the underground environment.

Figure 5 shows the core control board of the edge gateway used in the system. The board integrates multiple functional modules, including the CC2530 core chip for protocol processing, a ZigBee communication module for wireless data transmission, and an RS-232 serial interface for wired connections. It also contains an A/D conversion module for sensor signal digitization, power management circuits for stable operation, and multi-pin connectors for peripheral sensor access. This hardware design enables flexible connectivity, supporting both wired and wireless communication with front-end sensing nodes, and ensures reliable data acquisition and transmission in underground mining environments.

The main workflow of the ZigBee terminal node is shown in Figure 6. The system initializes after the ZigBee terminal node starts up, then proceeds to search for and connect to the ZigBee network, and periodically reads sensor data to send it to the edge server.

The sensing nodes are powered by a lithium battery with a capacity of 21 Ah, allowing the system to operate continuously for approximately six months under typical monitoring conditions. This low-power design ensures stable operation in underground environments where frequent battery replacement is impractical.

2.4. Edge Server Hardware Selection

The edge server is the core component of data collection in this system. It is mainly composed of the core processor and peripheral function modules. The edge server of this system not only serves as the protocol conversion function for sensing information aggregation but also has fast computing and processing performance as well as rich peripheral resources. The CPU of the Edge Server is an Intel Celeron G5905, from Intel Corporation, Santa Clara, CA, USA. In addition to the core board, it also has interfaces of HDMI, LVDS × 2, LCD, single 1080p LVDS or dual LVDS below 1366 × 768, Gigabit Ethernet, USB master + slave, PCIE, CAN, SD × 2, UART × 4, which can be configured as 232 or 485, as well as supporting audio input and output, as shown in Figure 7. This high-performance edge server effectively addresses the computing power constraints of edge devices, enabling the execution of complex data processing and anomaly detection algorithms locally.

Since early warning mechanisms require a high degree of real-time performance, it is necessary to move these mechanisms, traditionally deployed on the cloud, forward to the front-end gateway. In addition to the simple routing and forwarding functions of traditional gateways, this gateway also provides data storage, processing, caching, reception, and transmission. Its processing capability is used to aggregate, pre-process, and filter data, thereby enhancing overall data handling efficiency. By combining edge computing with the data monitoring platform, the gateway can perform complex analytics and support incident handling and decision-making. The functionality of the gateway is illustrated in Figure 8.

If data is transmitted to the cloud and an incident occurs, decision packets must be returned to each IoT device, which inevitably causes network latency. Edge computing addresses this issue by adapting to the processing requirements of each application. To meet low-latency requirements, decision-making algorithms are executed at the edge gateway, thereby enabling rapid data processing at the edge and effectively reducing network latency [37].

As an efficient and stable edge gateway server, its software architecture is particularly important. An effective architecture should feature clear code logic, minimal interface encapsulation redundancy, strong scalability, and high system stability. The gateway functions as an aggregate node, receiving all data from the lower sensing nodes. An important consideration is how to efficiently receive large volumes of data in a one-to-many communication mode. In addition, the gateway must support data caching, as the system needs to retain data for a period of time rather than forwarding it immediately. Under normal operating conditions, data are transmitted to the cloud as soon as they are received by the gateway. However, in underground mining environments, network stability is often limited and bandwidth is constrained. To ensure data integrity, the gateway temporarily caches the collected data and forwards them to the cloud once the connection is stable, with a default transmission interval of 30 min. As the core prediction and decision-making logic operates locally at the gateway, this scheduled transmission strategy does not compromise the system’s responsiveness. Due to the gateway’s limited forwarding capacity, caching is also essential to prevent data loss. Most importantly, as the front-end node of edge computing in this system, the gateway must possess strong data processing capabilities [38].

To address the above issues, the gateway architecture in this system consists of four main modules: the receiver module (data reception), the entry module (adaptation), the handle module (data processing), and the transmit module (data forwarding), as shown in Figure 9.

Receiver module—This module receives data from sensing nodes using a multiplexing approach. Whenever a sensing node is connected, a communication channel is allocated. At its core is a selector poller implemented as a dedicated polling thread within the gateway’s application-level thread pool. The gateway runs on an embedded Linux operating system, and each functional module operates as an independent thread, with scheduling handled by the OS thread management mechanism. The selector poller periodically (every 100 ms) checks all registered communication channels for I/O readiness events and dispatches active channels to processing threads, enabling efficient handling of up to 64 concurrent connections. This design combines multi-threaded concurrency with OS-level scheduling, ensuring low latency and high throughput in underground mining conditions.

Entry module—Designed for system scalability, this module manages other modules through dynamic library loading. Each functional module is compiled as a separate dynamic library. When the program runs, it loads the appropriate library according to the configuration file and calls the entry module to manage execution. Different versions of a module can be used; for example, several versions of the handle module may exist for different data processing methods, and switching between them requires only a configuration file change.

Handle module—This module processes incoming data through open interfaces and applies algorithms suited to specific application environments. In this system, anomaly detection is performed using residual analysis with a moving average filter. The raw sensor data are smoothed using a fixed-size moving average window to generate a baseline signal, and the residual is computed as the difference between the measured value and the baseline. If the residual exceeds a threshold determined from historical monitoring data, the event is flagged as an anomaly. This approach effectively filters random noise while enabling timely detection of significant deviations.

Transmit module—Based on the TCP/IP protocol, this module uploads data from the gateway to the server. It packages the data, adds the source address, performs routing and forwarding, and sends it to the server.

2.5. Monitoring Platform Software Design

The data collected at the lower layer must be transmitted to the data monitoring platform after being processed by the edge gateway. This platform supports real-time querying of characteristic data changes during the mining process and provides essential functions such as equipment management and authority management. For example, during mining operations, if the roof strain exceeds the threshold value, the edge gateway makes the decision locally and uploads the relevant data to the platform. The platform then monitors these changes, enabling managers to promptly assess the situation and intervene to prevent adverse developments. The authority management function ensures that sensitive data are not accessible to unauthorized personnel. Users with different permission levels are presented with different interfaces upon login. Equipment management is another key feature of the platform. It records the status and specifications of equipment currently deployed in the mining section, providing managers with a comprehensive view of ongoing operations. In addition to real-time monitoring, the platform includes a history module that allows users to view and analyze historical data over a defined period, supporting long-term operational assessment and decision-making.

The data monitoring platform is designed to enable administrators to monitor, in real time, the data collected by the front-end sensing layer. Its main functions include equipment management, environmental monitoring, personnel positioning, history recording, and system management, as shown in Figure 10.

Equipment management module—This module manages front-end sensing hardware, such as sensors and gateways. It supports the registration of device information, summarizes anomaly and maintenance records, and provides device query functions to quickly locate specific equipment when large numbers of devices are deployed.

Environmental monitoring module—This module is divided into two sub-modules. The first provides an overview of basic data by displaying multiple parameters on a single page, reducing the need for page switching. The second presents specialized data views to enable focused monitoring of parameters of interest.

Personnel positioning module—This module is currently under development and will enable real-time location tracking of personnel within the mining site.

History module—Complementing the environmental monitoring module, this function allows users to view historical monitoring data, which are otherwise unavailable in the dynamic real-time view. It supports analysis of past data trends for operational assessment.

System management module—This module manages user access rights. Through page, module, role, and user management, it assigns different permission levels to different users, thereby preventing data leakage and ensuring data security.

The data monitoring platform represents the application layer where administrators directly interact with the system. The data displayed on the upper-layer platform originate from the underlying hardware, including sensor nodes and aggregation gateways. The gateway receives data from neighboring nodes and preprocesses multi-source inputs. This system applies a sliding-window method, referencing data from adjacent nodes for fault discrimination, event detection, and anomaly identification. When an incident occurs, it is processed according to its specific characteristics, with event discrimination performed at the gateway layer to reduce network transmission time and improve response time. If abnormal data are detected, filtering operations are carried out to alleviate bandwidth pressure and prevent corrupted data from contaminating the server’s dataset.

2.6. Intelligent Prediction of Stope Roof Subsidence and Deformation

Stope subsidence is a key indicator of stope dilation and deformation. Therefore, accurate and intelligent prediction of subsidence is crucial for maintaining the stability of surrounding rock in deep mining. To address this challenge, many researchers have employed deep learning techniques, with commonly used algorithms including the Backpropagation (BP) neural network and the Recurrent Neural Network (RNN). In this study, considering the convergence deformation characteristics of the stope, the variation in subsidence over time is treated as an equally spaced time series. Based on this representation, a Long Short-Term Memory (LSTM) neural network is employed to perform intelligent prediction of stope subsidence.

The LSTM model is trained on the cloud, leveraging the abundant computational resources available in cloud environments to significantly reduce training time. After training, the model is deployed to edge intelligent devices, enabling localized prediction services and reducing data transmission latency. This approach fully utilizes the limited computational capacity of edge devices for inference, while alleviating the computational burden on the cloud. By combining cloud-based training with edge-based inference, the system achieves both high training efficiency and low-latency prediction.

Long Short-Term Memory (LSTM) is a special type of recurrent neural network (RNN). Based on the RNN architecture, LSTM introduces a gating mechanism that enables the model to effectively transmit and represent information in sequences, thereby avoiding the loss or omission of useful prior information. This mechanism effectively addresses the issues of vanishing and exploding gradients. Compared with standard RNNs, LSTM introduces three types of gates: the input gate, the forget gate, and the output gate. Among them, f_t denotes the forget gate, which determines whether the information from the previous time step should be retained or discarded; it denotes the input gate, which controls the extent to which new information from the current input is stored in the cell state; and Ot denotes the output gate, which decides how much of the updated information should be transferred to the next time step. Figure 11 illustrates the internal structure and data flow of a Long Short-Term Memory (LSTM) network. The inputs consist of the current time step input vector X_t and the previous hidden state h_t−1, which are processed through three distinct gating mechanisms:

Forget gate (f_t): Generates a coefficient between 0 and 1 via a sigmoid activation function, determining the proportion of information from the previous cell state C_t−1 to retain or discard.

Input gate (i_t): Also uses a sigmoid activation to produce a gating coefficient, which is combined with a candidate memory vector generated by a tanh activation. This combination controls how much new information is written into the cell state.

Cell state update (C_t): The new cell state is computed by combining the retained portion of the previous state with the scaled candidate memory, using element-wise addition. This additive update path helps mitigate the vanishing gradient problem.

Output gate (O_t): Determines the portion of the updated cell state to output as the new hidden state. Specifically, the updated cell state C_t is passed through a tanh activation and multiplied by the output gate coefficient, producing the current hidden state h_t, which serves as both the time step’s output and an input to the next step.

Through the coordinated operation of these three gates, LSTM networks can selectively forget irrelevant information, retain essential context, and output the necessary information, enabling them to effectively capture long-term dependencies while alleviating the gradient vanishing issue present in standard RNNs.

The performance of an LSTM neural network is highly sensitive to its parameter settings, and arbitrary selection can lead to significant variability in training outcomes. To address this issue, this study employs Particle Swarm Optimization (PSO), an evolutionary computation technique that initializes a population of candidate solutions (particles) and iteratively refines them by simulating the social behavior of swarms. In each iteration, the fitness of each particle is evaluated according to a predefined objective function, and particles adjust their positions in the search space based on both their own best historical performance and the best performance within the swarm. Through this cooperative mechanism, PSO efficiently converges toward the global optimum and, owing to its strong adaptability, can be seamlessly integrated with other algorithms for complex optimization tasks. PSO offers strong capability in optimizing models trained on small, noisy, and irregularly sampled datasets, which are typical in underground mine monitoring scenarios. By automating the search for optimal LSTM hyperparameters, it reduces manual trial-and-error, accelerates convergence, and improves prediction stability, making it highly effective for real-time safety monitoring in complex mining environments. In the proposed PSO–LSTM framework, PSO is used to automatically identify the optimal set of LSTM training parameters, thereby enhancing the model’s predictive accuracy and generalization performance. The overall workflow of the PSO–LSTM optimization process, including data preprocessing, parameter optimization, model training, and validation, is illustrated in Figure 12.

In this study, three key parameters are selected for optimization: the number of epochs, the number of layers, and the time step size.

Epoch refers to one complete pass through the entire training dataset. In practice, a single epoch is insufficient for the model to achieve convergence, and multiple epochs are generally required. However, a greater number of epochs does not necessarily result in better performance, as excessive iterations not only increase the training time but may also lead to overfitting.

Layer denotes the number of layers in the neural network architecture. Moderately increasing the number of layers can enhance model accuracy, but it may also raise the risk of overfitting if not properly controlled.

The time step, determined based on the characteristics of LSTM networks, defines the number of previous time points used to predict the value at the next step. In this study, the input data is treated as an equally spaced time series, where data from the preceding n steps is used to forecast the value at step n + 1. Since the choice of n significantly influences the predictive performance of the model, PSO is employed to determine its optimal value.

The PSO optimization process yielded the following optimal hyperparameter configuration for the LSTM model: the number of hidden layers was set to 2, the time step size was set to 8, and the number of training epochs was set to 150. This configuration ensured both rapid convergence and stable prediction performance while avoiding overfitting, which is particularly important in small-sample and irregular underground monitoring datasets.

In practical engineering applications, convergence deformation of surrounding rock is often not recorded as uniformly spaced time series data. In addition, the available measurement samples are frequently insufficient for training neural network models, and missing data is common. To overcome these issues, we employ interpolation methods to expand the original dataset and transform it into an equally spaced time series. This transformation enables more effective subsequent analysis. Considering the strengths and weaknesses of various interpolation techniques, and based on empirical tests, this study adopts cubic spline interpolation for data processing. This method is particularly suited to underground mine monitoring scenarios, where sensor readings are often sparse, irregularly timed, and subject to occasional data loss, as it can reconstruct a smooth and continuous dataset while preserving the essential characteristics of the original measurements. Cubic spline interpolation divides the entire interval into multiple subintervals, with each fitted by a cubic polynomial. The coefficients of these polynomials are determined by two constraints. The first is the interpolation condition, ensuring that the polynomial passes through the given data points. The second is the smoothness condition, which guarantees the continuity of the function values as well as their first and second derivatives at the boundaries between adjacent segments. By solving these constraints, the polynomial coefficients are obtained and the interpolation process is completed. Cubic spline interpolation preserves the original data characteristics while ensuring smoothness, making it highly suitable for preprocessing in time series prediction tasks.

The raw monitoring data contains significant noise and outliers. To improve data quality before training the neural network, a filtering algorithm is applied. Based on practical experiments and evaluation of multiple criteria, including effectiveness and computational simplicity, the moving average filter is selected for data smoothing. This filter uses a fixed-length queue following the first-in, first-out (FIFO) principle. Each new data point entering the queue replaces the oldest one, and the output is calculated as the arithmetic mean of the N data points in the queue. This method effectively suppresses periodic noise and produces a smooth processed signal.

In this study, the sampling interval for cubic spline interpolation was set to 30 min, consistent with the monitoring frequency of the sensors. For the moving average filter, the window size N was set to 24, corresponding to a 12 h smoothing span. This ensured effective noise reduction while retaining the essential deformation trends.

To further verify the optimization effects of different optimizers, 80% of the tunnel deformation displacement data was used as the training set and 20% as the test set. The baseline was the PSO–LSTM model with the original SGD optimizer. For comparison, the same model was trained with Adagrad, RMSprop, and Adam optimizers, each for 200 epochs, and the mean absolute error (MAE) was selected as the evaluation metric. The training and validation loss curves of the four optimizers are shown in Figure 13.

As shown in Figure 13a, the SGD optimizer exhibited a relatively stable convergence pattern, with the MAE gradually decreasing and stabilizing after approximately 50 epochs. However, its convergence speed was relatively slow. In Figure 13b, the Adagrad optimizer displayed a similar convergence trend to SGD, but with higher fluctuations after 25 epochs and an overall larger MAE value, suggesting a risk of overfitting. In Figure 13c, RMSprop achieved a faster convergence speed, with the MAE loss decreasing sharply within the first 25 epochs. Nevertheless, its loss curve showed significant oscillations, indicating instability that may negatively affect prediction accuracy. In contrast, Figure 13d demonstrates that Adam achieved the best overall performance: the model reached full convergence in about 25 epochs, with the lowest MAE among all tested optimizers. The Adam loss curve was also the smoothest, showing minimal fluctuations, which indicates that the Adam optimizer effectively avoided both overfitting and underfitting. These results confirm that Adam provided the most stable and accurate optimization performance for the PSO–LSTM framework.

3. Field Application Evaluation

This system was applied on-site at the Chifeng Chaihulanzi Gold Mine, located in Chutoulang Town, Songshan District, Chifeng City, Inner Mongolia Autonomous Region. The ore body is a typical horizontal, gently inclined, broken thin ore body. The development method combines adit development with shaft–adit integration, and the main mining methods include the room-and-pillar method, shrinkage method, and dry filling using externally retrieved stone. Currently, the resources in the #1 and #2 mine wells are nearly depleted, while other mining areas remain in the infrastructure construction stage, making it difficult to maintain continuous production. To address this issue, the production departments conducted on-site investigations and found a large number of recoverable high-grade mine pillars in the stopes of the #1 and #2 wells. To ensure continued production in these areas, recycling the residual natural pillars is necessary. At present, the stope support scheme mainly includes two temporary support methods: wood pillars and spiral pillars. These provide localized support but have limited strength, making them unsuitable for replacing natural pillars on a large scale. To maintain production continuity and extend the mine’s service life, we designed a scheme to use artificial expandable pillars to replace residual natural pillars in the stopes. This approach aims to maximize ore recovery while ensuring safety, thereby enabling the safe and economic mining of natural pillars. In the experimental stope, pressure monitoring was performed on the artificial expandable pillars, with a pressure sensor positioned between the expansion unit and the support unit, as shown in Figure 13. The recorded pressure data and corresponding curves are presented in Figure 14, where users can view real-time variations and adjust the observation frequency as needed. In this system, the measured pressure and displacement data are visualized in real time using line charts, enabling operators to clearly observe parameter variations and trends. Threshold values for each monitored parameter are set based on historical operational experience and engineering safety requirements. When the measured value approaches or exceeds the predefined threshold, the system provides a clear visual indication on the chart and issues a corresponding alert, ensuring that operators can take timely action to prevent potential hazards.

The software system allows users to select appropriate models and parameters based on their own data and perform customized training to obtain corresponding model weights. Users can either create a new project or open an existing one, with relevant data displayed in the interface to facilitate parameter adjustment and model training. Upon completion of training, the system visualizes the training and testing losses, along with the predicted and actual values of the test set in the form of curves, while simultaneously presenting key information such as predicted values, actual values, absolute errors, and error percentages on the right side of the interface in textual form. The trained model weights are automatically saved in the specified directory in .h5 format. Before prediction, the monitoring data must be preprocessed to match the data type and format used during model training. By organizing early-stage monitoring data from the stope and selecting suitable models and parameters, intelligent prediction of subsequent convergence deformation can be achieved.

During the monitoring process, displacement sensors were used to measure the deformation of the surrounding rock at monitoring points A, B, and C, and the recorded measurement data are presented in

Table 1. Deformation measured data.

Date	Point A	Point B	Point C
8 August	3.050 m	2.883 m	3.130 m
10 August	3.047 m	2.880 m	3.126 m
12 August	3.042 m	2.877 m	3.121 m
14 August	3.038 m	2.873 m	3.116 m
16 August	3.035 m	2.870 m	3.111 m
18 August	3.029 m	2.866 m	3.107 m
20 August	3.026 m	2.861 m	3.103 m
22 August	3.024 m	2.857 m	3.100 m
24 August	3.021 m	2.852 m	3.098 m
26 August	3.017 m	2.849 m	3.094 m
28 August	3.014 m	2.846 m	3.091 m
30 August	3.011 m	2.843 m	3.088 m

.

As shown in Table 1, the deformation of the stope roof becomes increasingly pronounced over time, with the displacement gradually increasing in a linear manner, reaching an average subsidence of approximately 3.7 cm. Using the neural network prediction model developed in this study, the deformation data at monitoring points A, B, and C, which are located on the stope roof, were predicted based on the measurements obtained from the monitoring system. The prediction results were then compared with the measured data in Table 1, and the comparison is illustrated in Figure 15.

As shown in Figure 16, the predicted and measured curves at monitoring points A, B, and C exhibit good overall agreement, with prediction deviations at all points remaining within 5 mm. This threshold is generally recognized as acceptable in actual field work for roof deformation monitoring. Experimental results demonstrate that the PSO–LSTM neural network model developed in this study can effectively meet the requirements for predicting stope roof deformation, as the predicted values closely match the actual measurements and can serve as a reliable reference for ensuring safe mining operations.

In addition, it should be noted that the evaluation in this study was primarily guided by field engineering practice. Based on long-term on-site monitoring experience, a deviation within 5 mm is generally recognized as acceptable for roof deformation monitoring in underground mines. While quantitative statistical metrics such as MAE and RMSE were not included due to current dataset limitations, future work will expand the monitoring data and incorporate these standard error measures to provide a more comprehensive and comparable evaluation of model performance.

4. Summary and Conclusions

In the underground mining process, mining environment monitoring is a key part of safe mining. To address this problem, this paper enhances the traditional cloud-based decision-making platform by incorporating edge computing, enabling sensor data to be preprocessed at the gateway layer in real time and decisions to be made locally. Field applications showed that, after this enhancement, the system effectively reduced network latency and thus shortened the response time for decision-making in actual mining operations.

The overall design of this system is based on edge computer technology, including hardware circuit design, gateway design, front-end sensing node design, communication protocols, data monitoring platform development and design, and database design. According to the technical characteristics of edge computing, data is preprocessed at the gateway nodes, and event discrimination, error discrimination, and fault discrimination are performed by analyzing the multimodal data.

Comprehensively consider the front-end sensing nodes and gateways needed for data collection in this system, and the working environment, communication distance, power consumption, scalability, cost, and other problems of the sensing nodes in the underground stope, to select a suitable sensing module. Under the premise of limited cost and computing resources, choose the appropriate gateway chip.

Field applications have shown that through the data monitoring platform, it is convenient to display the data collected by the pressure sensors, and the data can be displayed in the form of line charts to show the changes, making it easy for the monitoring personnel to quickly locate the danger and make decisions.

To address the problem of roof subsidence in deep stopes, this study developed a PSO-LSTM prediction model by integrating the Particle Swarm Optimization (PSO) algorithm with the Long Short-Term Memory (LSTM) neural network. A supporting software system was also developed, which includes functions such as project creation and saving, comprehensive parameter configuration, custom training, and intelligent prediction of roof subsidence. Based on the stope roof subsidence data collected by displacement sensors, the practicality and effectiveness of the proposed ultra-short-term convergence prediction model for edge intelligence were successfully verified. The system provides sufficient response time and technical support for addressing various engineering problems caused by stope roof subsidence.

Despite the demonstrated effectiveness of the proposed system, certain limitations remain. In particular, challenges in heterogeneous data fusion persist, arising from the diversity of sensing modules, communication protocols, and data formats, as well as variations in data quality and synchronization. These issues can affect the accuracy and timeliness of decision-making in complex mining environments. In future work, we plan to develop and implement more robust and efficient fusion strategies to address these challenges, thereby enhancing the adaptability, reliability, and real-time performance of the system. Furthermore, we aim to extend the system’s analytical framework to incorporate human–machine–environment collaborative analysis, enabling a more holistic understanding of operational conditions and improving safety and productivity in underground mining.