A Dynamic Security Assessment Method for Ironmaking Plants Based on Cloud-Edge Collaboration in Reconstructed Networks

Abstract

Security assessment of ironmaking plants is one of the crucial means to promote their sustainable development. However, the disparate nature of subsystems within these plants, along with network inconsistencies and isolated data, obstruct a thorough and timely security assessment. At the same time, it is impossible to achieve the sustainable development goals of reducing the adverse impact of safety on the environment, ensuring economic benefits and the health of employees. This study addresses the complexities of heterogeneous networks, disparate systems, and segregated data that are prevalent in traditional ironmaking plants; and a method to reconstruct the plant’s network and execute security assessments is proposed. This method involves coupling existing systems with new ones to create comprehensive data and resource pools by aggregating information from diverse sources. Subsequently, employing multiple regression models and optimized neural network models at the edge and central cloud facilitates dynamic assessment of security concerns. This method enables concurrent consideration of both regional and overall security analysis and decision-making within the plant. Through simulation testing of 27 functionalized module indicator datasets over the preceding 12 months at a specific ironmaking plant, the efficacy of the proposed theoretical methods and technological approaches in constructing security systems for ironmaking plants is substantiated.

1. Introduction

As an important industrial facility, ironmaking plants consume a lot of resources in the production process and have an impact on the environment. Security assessment not only involves personnel safety, but also focuses on equipment operation, chemical use, and waste treatment. Through reasonable security assessment and management, the risk of interruption to production can be reduced, the loss and cost can be reduced, and the continuity and stability of production can be ensured. At the same time, the negative impact on the environment can be minimized, the waste of resources can be reduced, and the sustainable development of the environment can be protected. In addition, a safe working environment can improve employee productivity, reduce the incidence of work-related injuries and occupational diseases, and ensure the health and safety of employees; thus, there is a positive impact on the sustainable development of enterprises. However, heterogeneous networks, disparate systems, and segregated data within ironmaking plant networks pose significant challenges to security assessments. Consequently, there is an urgent need to surmount the impediments posed by the independent isolation of network systems and data in these plants, and to build an integrated IIoT fusion scheme for a comprehensive evaluation of regional and overall plant security issues. Nonetheless, the prevailing research indicates constraints in the current methods of security assessment in these plants.

Despite some progress in recent years in developing dynamic safety assessment methods for ironmaking plants, existing research still faces significant limitations that, from a scientific perspective, urgently need to be addressed through more in-depth and comprehensive studies. These limitations primarily include insufficient assessment indicators, oversimplified method designs, and relatively subjective assessment results. For instance, the research conducted by Jianhou Gan et al. [1] introduced an intelligent monitoring network for coking processes based on Internet of Things (IoT) technology in Zigbee mesh cluster networks, which focused mainly on forming a perception layer network. Although this study provided a new perspective for monitoring the coking process, it was mainly limited to the construction of the perception layer and failed to consider a multidimensional set of safety assessment indicators comprehensively, resulting in a less comprehensive assessment system. In their study, Liu Wei et al. [2] focused on safety factors in iron transportation, and devised a comprehensive iron transport behavior safety monitoring system. While this system made progress in enhancing safety monitoring for specific production processes, its design was still confined to specific production processes and failed to establish an integrated safety architecture, limiting its application in the full process of safety assessment of iron manufacturing. Zhu Junjie et al. [3] proposed an emergency prevention and control system based on neural networks, utilizing Keras deep learning models for facial recognition, which significantly improved the level of enterprise in emergency prevention and control. However, although the system performed well in emergency situations, it was overly specialized for regular safety assessment and management, especially in evaluating safety performance in dynamic and complex environments; this overlooked the dynamic and systemic nature of the safety assessment process. In summary, while previous works have provided valuable perspectives and technical support for the safety assessment of ironmaking plants, they still lack a comprehensive, in-depth, and integrated safety assessment system, especially in the face of increasingly complex production environments and variable safety threats. Therefore, this paper aims to overcome these limitations in existing research by developing a dynamic safety assessment method for ironmaking plants based on cloud-edge collaboration. This offers a more comprehensive, profound, and dynamic safety assessment framework.

These studies demonstrate the innovative use of IoT and AI technologies in enhancing the security of ironmaking plants. However, their designs often focus narrowly on individual production areas without integrating a broader, enterprise-wide industrial IoT framework. This lack of a unified approach highlights a gap in current research, particularly the absence of a robust IoT architecture foundational for comprehensive security assessments. Furthermore, existing studies on security assessment in ironmaking plants show limitations. For instance, Shirali et al. [4] explored the interplay between security spending and operational performance but did not offer a method for real-time security evaluation. Nazaripour et al. [5] developed a Customized Predictive Risk Index (CPRI) to forecast accidents from 350 unsafe observations, providing valuable insights into accident prediction but not encompassing overall security assessment. Wu Qibing et al. [6] proposed a security assessment model focusing on the impact of various factors on security, utilizing the Analytic Hierarchy Process (AHP) for determining the weight of indicators. However, this approach’s credibility is undermined by its limited indicators and simplistic evaluation methodology. Miao Xiaoyu et al. [7] introduced an early warning model for security production, employing a detailed hazard assessment for major risks in the industry and fuzzy logic for result interpretation; yet the subjective nature of its outcomes persists. Lastly, Wang Biao et al. [8] offered a dynamic security risk assessment model that identifies and evaluates risks dynamically within industrial enterprises. Although this marks a step forward, it still requires advancements in data processing for comprehensive security analysis. These findings indicate a pressing need for a more integrated, dynamic, and technologically sophisticated approach to security assessment in the ironmaking industry, addressing the current research’s limitations and the subjective nature of existing evaluation methods. Although this model captures dynamic security situations in enterprises, there is room for improvement in computer data processing principles.

Overall, the construction of ironmaking plant security systems has undergone three developmental stages: architecture, single-system design, and whole-system design. However, there is currently no mature construction scheme for overall design. The deficiency in security assessment research in ironmaking plants lies in the scarcity, lack of specificity, overly simplistic design, and the relatively subjective nature of assessment results; this necessitates more in-depth research.

To address these deficiencies, this study proposes a cloud-edge collaboration reconstructed network method for dynamic security assessment evaluation in ironmaking plants. This method involves edge intelligence upgrades, coupling of decomposed functionalized modules, and other means to reconstruct the ironmaking plant’s network infrastructure, forming a new cloud-edge system. At the edge, leveraging heterogeneous network data enables comprehensive analysis and processing, designing multivariate regression analysis models suitable for the actual security conditions in ironmaking plants, effectively addressing security issues encountered by various subsystems. At the central cloud, employing particle swarm optimized BP neural networks facilitates assessing and monitoring security situations. Compared to existing research, this method significantly enhances comprehensive and timely security monitoring and evaluation in ironmaking plants, holding substantial importance for accident prevention and security emergency responses.

2. Materials and Methods

2.1. Ironmaking Plant Dynamic Security Assessment System Design

Traditional Ironmaking Plant System and Network Reconstruction

Traditional ironmaking plants revolve around blast furnaces and encompass multiple associated systems such as the blast furnace system, feeding and loading system, pulverized coal injection, air supply system, iron slag treatment system, gas dust removal system, slag treatment, and gas dust removing, among others [9]. However, control systems within these traditional plants operate in a relatively dispersed and independent manner, and lack interactive information and behavioral influence. This disjointedness presents a challenge to allowing these systems to effectively collaborate. Concurrently, data pertaining to the ”people” component, which includes both management and operational staff; the “things” component, covering equipment, products, and infrastructure; and the “environment” component, addressing both the production and natural environments, have not been consolidated with the pertinent data associated with the ironmaking production system. As a result, comprehensively assessing overall security based on the various system operations within existing ironmaking production becomes intricate. Figure 1 depicts a simplified diagram of the traditional ironmaking production process.

Construction methods, standards, and technology implementation in traditional ironmaking plants vary, leading to a network characterized by heterogeneous integration and ubiquity. To integrate heterogeneous networks and multiple-source heterogeneous data, network reconstruction involves the following steps:

Step 1: Construction of an intelligent monitoring and security warning wireless sensor network. In the ironmaking plant dynamic security assessment system, the sensor network covers critical areas of the ironmaking plant to monitor temperature, pressure, humidity, gas concentrations, and other essential parameters. These sensors collect and transmit real-time data, aiding in understanding the factory’s operational status promptly. In case of potential security risks, they rapidly issue alerts, enabling immediate countermeasures to be taken. This sensor network not only monitors equipment operation and identifies normal functioning but also provides historical data records [10], facilitating post-event analysis and tracing, and effectively preventing potential faults and accidents. The wireless sensor network is composed of a 5G Radcap network, Zigbee network, Bluetooth network, and Wifi network. Each network terminal is connected to the data collection (DC) gateway, and further connected to the multisource heterogeneous fusion monitoring platform of the central cloud through the IoT gateway, as shown in Figure 2.

Step 2: Pooling of systems and device data collection through IoT gateways. In traditional ironmaking production processes, the fixed hardware leads to the formation of fixed production patterns within systems [11]. Constructing new systems by mobilizing hardware resources becomes challenging and fails to meet diverse requirements. Leveraging wireless sensor networks, hardware resources of various functionalized modules are decoupled from the original fixed system through IoT gateways or protocol bridging. These resources are integrated into the plant’s equipment resource pool via gateways or established protocols, and controlled and combined uniformly by the central cloud server. This enables the fusion of heterogeneous devices in the plant and provides flexibility in hardware deployment. The integrated hardware resources form a “hardware resource pool”, enabling application construction on the resource pool and facilitating dynamic hardware resource allocation.

The hardware systems in the device pool can be controlled and implemented by intelligent Industrial Control Systems (ICS) under edge cloud nodes of various subsystems based on the plant’s security status. Automatic notifications are triggered when the system detects signs of abnormalities or declining equipment performance. In case of a security incident in the ironmaking plant, without affecting production operations, the system flexibly coordinates the original functionalized modules, reconfiguring a targeted ironmaking plant security accident response system in a coordinated manner.

Data collection is critical for the construction of the dynamic security assessment system [12]. Terminal sensors of the system collect three types of data: equipment status (ES), measurement data (MD), and operation data (OD). These foundational data are generated by sensors of wireless networks integrated by gateway devices. The gateway simultaneously collects data from various physically connected functionalized modules and requires monitoring during the equipment production process. These data include equipment operating time, start–stop frequency, and equipment status; as well as temperature, humidity, gas concentrations, and other fundamental measurement data. Periodically, the gateway reads, processes, stores, and displays data at set time intervals. This is depicted in Figure 3.

The security status data of all functionalized modules and equipment in the ironmaking plant are processed by the data acquisition gateway, and the data are uniformly uploaded to the cloud platform through the Message Queuing Telemetry Transport (MQTT) interface protocol of the Internet of Things gateway. Data visualization occurs through the monitoring display page of the cloud platform and is stored in the central cloud backend server, forming a vast “data pool”. This “data pool” provides local access at the edge, enabling mathematical model analysis to explore correlations and patterns hidden in the data, enhancing localized security within the ironmaking plant. Simultaneously, at the central cloud, it serves as the data source for artificial neural network models, forming the basis for an overall security assessment of the ironmaking plant. For instance, in the case of an electric motor, this study collects measurement data including current, voltage, temperature; and operational data such as operating time and start–stop frequency, and real-time data on the motor’s status. Based on fuzzy algorithms or expert learning methods, a score is assessed to reflect the equipment’s security condition, yielding a single output indicative of the overall motor security. Finally, this derived data is encapsulated and uploaded to the cloud platform. Thus, on the cloud platform, instead of raw data such as voltage, current, start–stop counts, or operational statuses, what we obtain is a sole data point directly reflecting the equipment’s security status. By monitoring equipment, personnel, and environmental statuses via gateways and wireless sensor networks, the IoT system can identify signs of possible equipment failures, and predict potential faults in advance, avoiding sudden security incidents. This method is data-driven, relying on a large amount of historical performance and real-time monitoring data. It not only reduces the likelihood of security incidents but also enhances equipment availability, enabling the ironmaking plant to operate more safely and steadily.

Step 3: Construction of a multisource heterogeneous network dynamic security assessment system. Traditional security management systems focus solely on parameters closely related to the ironmaking plant’s production process. The safety dynamic perception assessment system, as described in this paper for the ironmaking plant area, is specifically designed to surveil systems integral to the ironmaking production process. It does so by leveraging frameworks in “human resource management”, “materials and facilities management”, and “environmental management”; succinctly termed “people”, “things”, and “environment”, respectively. The system’s core aim is to accurately capture and convey state information pertinent to potential hazards. This method constructs an ironmaking plant dynamic security assessment system using technologies such as the Internet of Things, edge computing, and artificial neural networks. The system architecture is depicted in Figure 4.

Firstly, the production-related systems of the ironmaking plant, as well as the non-production-related systems, are disassembled into functionalized modules. Subsequently, through gateways, the heterogeneity among various system devices is integrated, relationships between functionalized modules are analyzed, and they are coupled with each other. Real-time data are collected from various terminals and uploaded to the central cloud, forming a pool of usable data, functionalized modules, and equipment pools. By establishing the coupling relationship between the “Iron Production Related Systems” and the “Non-Iron Production Related Systems”. Device, material, and process data from major systems are aggregated via wireless sensor networks and IoT gateways, forming a multisource heterogeneous network system.

Step 4: Construction of an integrated cloud-edge-end platform. Ironmaking production consists of the blast furnace body system and a series of auxiliary systems, characterized by a “one center, multiple subsystems” feature [13], aligning with the edge computing architecture. To alleviate substantial pressure on the central cloud server caused by the wireless sensor network uploading massive data, a distributed security perception platform with high computing speed, accurate data processing, and decisive decision-making capabilities was designed during the ironmaking plant dynamic security assessment system. This platform consists of an edge service management platform deployed in the central cloud and a Multi-Access Edge Computing (MEC) node deployed in the edge cloud. It integrates the core control parts of the original systems in the ironmaking plant production process with the new “Person”, “Things”, and “Environment” IoT systems into the newly added MEC nodes. By utilizing MEC for managing individual system nodes while integrating them into the ironmaking plant’s edge service management center in the central cloud, information among the original systems is shared and controlled.

By integrating the core control functionalities of the cloud platform into newly added edge cloud nodes to construct an edge cloud platform, a system environment composed of computation, storage, and applications is formed in the edge cloud. This setup breaks down the barriers to information sharing among various functionalized modules through interaction between edge clouds, facilitating collaborative linkage and resource sharing, enabling prompt local processing within the system [14]. The system’s design assumes a “plus” pattern, vertically achieving control of the original subsystems and interacting with the central cloud for information exchange while horizontally communicating and collaborating with peer core control subsystems. Through both vertical and horizontal interconnections, it reduces data transmission relay and processing time, decreases end-to-end latency, and eases network bandwidth pressure. On the one hand, it achieves connectivity from the end to the edge computing node and the cloud. On the other hand, MEC nodes connect with each other, breaking information exchange barriers, forming an edge information exchange network, ensuring data flow and closed-loop control [15]. MEC nodes, as intermediate links between terminal gateways and clouds, not only realize intelligent control within functionalized modules vertically but also upload data to the cloud platform for overall decision-making, providing robust support for constructing the overall security assessment system of the ironmaking plant. The cloud-edge-end three- layer architecture deployment is shown in Figure 5.

2.2. Establishment of Security Assessment Index System

Ensuring security within the factory area is imperative to preempt security incidents within the ironmaking plant [16]. Beyond averting conventional security issues associated with production equipment, it is vital to proactively mitigate potential hazards arising from personnel and materials to prevent their escalation into accidents [17]. In this study, the “Dynamic Safety Risk Assessment Model for Industrial Enterprises” [8] was adopted and appropriately modified within the framework of a novel system developed to tackle the dynamic safety assessment issues prevalent in ironmaking facilities. A comprehensive safety assessment indicator system was constructed, enabling the identification and specification of entities requiring evaluation, along with their pertinent evaluation metrics. Through the extraction of safety-related state data pertaining to these entities from the data repository, a thorough analysis was performed to assess potential hazards and the nature of their impact on the overall system safety. The culmination of this process was the formulation of a detailed safety analysis table, as delineated in

Table 1. Security analysis.

Number	Subsystems	Functionalized Modules	Hazardous Mode	Equipment Defects	Environmental Defects	Duration of Hazardous Exposure
1	Blast furnace system	Detection	24	5	5	144
		Electrical	19	5	2	160.8
		Cooling	17	4	3	54
2	Feeding and loading system	Screening	21	2	5	58.9
		Weighing	4	0	2	130.4
		Transmission	23	3	2	20
3	Pulverized coal injection	Spray gun	6	1	0	132
		Bin pump	18	1	1	223.5
		Coal grinding	22	5	0	146.5
4	Air supply system	The fan system	8	2	1	80
		Wind furnace	13	3	4	80
		Combustion	15	0	3	120
5	Iron slag treatment system	Granulating	20	1	2	115
		Dehydration	17	3	2	54
		Cinder flushing	8	1	3	176.25
6	Gas dust removal system	Gravity	13	3	2	112
		Venturi	21	2	3	135
		Bag	8	3	1	85
7	People	Recognition	4	0	3	90
		Video surveillance	13	1	0	40
		Positioning	2	1	1	106
8	Things	Vehicle	18	3	5	86
		Material	6	0	2	103
		Equipment	4	5	1	46.2
9	Environment	Natural	16	3	3	114
		Generate	28	5	3	170
		Energy	5	3	2	67

.

In order to eliminate the influence of the scale between the indicators, the raw data need to be normalized so that the indicators are in the same order of magnitude to facilitate comprehensive comparative evaluation. The formula is as follows:

$$R_j^{\prime }=\frac{R_j-min{R}_j}{max{R}_j-min{R}_j}$$

(1)

where $R_j$ is the $j$th indicator value, $min{R}_j$ and $max{R}_j$ are the minimum and maximum values of the $j$th indicator respectively and a raw value $R_j$ is mapped to the value $R_j^{\prime }$ in the interval [0, 1] by max-min normalization.

The subsequent step involves calculating the following risk assessment indicators based on the data from the security analysis table:

• Calculate the equipment hazard factor $k_1$

Equipment hazard factor calculation formula:

$$k_1=\mathrm{N}\mathrm{u}\mathrm{m}\mathrm{b}\mathrm{e}\mathrm{r} \mathrm{o}\mathrm{f} \mathrm{e}\mathrm{q}\mathrm{u}\mathrm{i}\mathrm{p}\mathrm{m}\mathrm{e}\mathrm{n}\mathrm{t} \mathrm{d}\mathrm{e}\mathrm{f}\mathrm{e}\mathrm{c}\mathrm{t}\mathrm{s}\times 5\mathrm{\%}$$

(2)

2.

Calculate the environmental hazard factor $k_2$

Environmental hazard factor calculation formula.

$$k_2=\mathrm{N}\mathrm{u}\mathrm{m}\mathrm{b}\mathrm{e}\mathrm{r} \mathrm{o}\mathrm{f} \mathrm{e}\mathrm{n}\mathrm{v}\mathrm{i}\mathrm{r}\mathrm{o}\mathrm{n}\mathrm{m}\mathrm{e}\mathrm{n}\mathrm{t}\mathrm{a}\mathrm{l} \mathrm{d}\mathrm{e}\mathrm{f}\mathrm{e}\mathrm{c}\mathrm{t}\mathrm{s}\times 5\mathrm{\%}$$

(3)

3.

Judgment of the substance hazard factor $k_3$,

Referring to the “Industrial Enterprise Dynamic Security Risk Assessment Model” [8] we list the substance risk factors were divided into five categories: A, B, C, D, and E, and the evaluation is shown in

Table 2. Substance risk factor assessment table.

Production or Storage of Substance Fire Hazard Level	Physical Hazard Factor
A	0.20
B	0.15
C	0.10
D	0.05
E	0.00

.

4.

Unsafe people behavior $k_4$

5.

Number of personnel $k_5$

6.

Duration of hazardous exposure $E$

7.

Security management level $C_1$

8.

System’s automatic control risk capability $C_2$

9.

Computation of the intrinsic security level $h_s$, which encompasses the efficacy in averting accidents at their source through diverse means. The assessment evaluates intrinsic security levels from three perspectives: primarily, people as pivotal in achieving intrinsic security within an enterprise; secondly, things as the guarantor of intrinsic security; and lastly, the environment as a driver for intrinsic security within an enterprise. This multidimensional evaluation of security intrinsic levels is depicted in Figure 6 and Figure 7.

According to the risk points of the intrinsic safety level, the combinations of safety levels are then listed out, as in

Table 3. Safety level combinations.

Intrinsic Security Level Assessment Chart	Safety Level Combination Situation
1	A1
2	A2, A1 + B1
3	A3, A1 + B2, A2 + B1, A1 + B1 + C1
4	A2 + B2, A3 + B1, A1 + B1 + C2, A1 + B2 + C1, A2 + B1 + C1
5	A3 + B2, A1 + B1 + C3, A1 + B2 + C2, A2 + B1 + C2, A2 + B2 + C1, A3 + B1 + C1
6	A1 + B2 + C3, A2 + B1 + C3, A2 + B2 + C2, A3 + B1 + C2, A3 + B2 + C1
7	A2 + B2 + C3, A3 + B1 + C3, A3 + B2 + C2
8	A3 + B2 + C3

.

2.3. Overall Process of Dynamic Security Assessment System

The production processes within the ironmaking plants encompass various hazardous factors spanning high temperatures, high pressures, toxicity, and harmfulness; thus, multiple extensive hazard sources are engendered [18].

2.3.1. Security Perception Process

During the routine operational phase within the ironmaking plant, the security assessment system, leveraging data alterations across diverse subsystems, generates instantaneous risk assessments through mathematical models deployed at the edge cloud and artificial neural network models operating at the central cloud. This iterative process continually disseminates warning notifications or emergency directives. The specific operational sequence involves:

Step 1: Real-time surveillance by IoT gateways encompassing functionalized modules across various subsystems to verify the equipment status, measurement data and operation data. Anomalies trigger alarms, leading to localized interventions. Under normal circumstances, local security assessments are conducted to ascertain the security status of functionalized modules.

Step 2: If deemed safe, data are transmitted to the multisource heterogeneous convergence monitoring platform and the security assessment platform. Conversely, if considered unsafe, data are relayed to the respective subsystem’s MEC to evaluate its impact on subsystem security.

Step 3: Subsystem MEC analysis determines the security impact. If no adverse impact is discerned, localized interventions are triggered; otherwise, data are transferred to the edge service management platform for assessment of the factory’s overall security.

Step 4: Evaluation by the edge service management platform concludes non-impact on security or potential impact; corresponding localized interventions or security assessments ensue.

Step 5: The security assessment platform determines emergency control necessity. If warranted, it disseminates security emergency operation strategies to the edge service management platform, which then allocates specific measures to subsystem MECs. Subsequently, the ICS module is scheduled for execution based on received security emergency strategies.

Step 6: Post-execution, the system resumes monitoring the status, measurement data, and operational parameters of each terminal. The comprehensive system perception and evaluation process are depicted in Figure 8.

2.3.2. Security Assessment Process

Dynamic security assessment within the factory area necessitates comprehensive data fusion processing of pertinent events, entailing the identification of correlations across temporal, spatial, and procedural dimensions. The assessment process can be delineated into several pivotal steps (as delineated in Figure 9).

Step 1: Decomposition of objects and checklist formulation. The overarching system is segmented into subsystems, with a comprehensive checklist developed for each, to ascertain the requisite depth and scope of analysis.

Step 2: Data preprocessing and readiness. Upon aggregating an extensive dataset, pertinent information is extracted concerning the primary systems of the ironmaking process, alongside data from the human–material–environment nexus within the ironworks. These crucial data, which are integral to the safety protocols of the ironworks, are subjected to rigorous preprocessing to ensure their readiness for in-depth analysis.

Step 3: Deployment of analytical engines. Employing a suite of analytical tools, including anomaly detection, event analysis, correlation, and metadata analysis engines, the analysis synthesizes empirical data with industry-specific accident trends to evaluate system impacts and potential risks, facilitating scenario prediction through analogical reasoning.

Step 4: Identification of hazard sources and safety analysis compilation. Leveraging signal analysis methodologies, subsystem models, diagnostic rules for equipment, and fault data annotation, the framework identifies potential hazards that might precipitate system failures, material damages, or personal injuries. Through algorithmic processing, it delineates accident archetypes, extracts anomalies/metadata pertinent to safety incidents within the ironworks, and compiles a detailed safety analysis report.

Step 5: Safety evaluation and diagnosis via a dual-level cloud-edge assessment framework. This phase entails the execution of safety assessments and diagnostic procedures, culminating in the formulation of evaluative outcomes, through a nuanced cloud-edge safety assessment model.

Step 6: Categorization of fault types and prioritization for remedial action. This entails the enumeration and classification of fault types based on their potential impact on the system, followed by a stratification according to severity, to facilitate a prioritized approach towards mitigation and prompt resolution.

Step 7: Industrial control system (ICS) scheduling reflective of safety awareness and alert levels, accompanied by the implementation of emergency safety interventions. This final step involves the strategic scheduling of ICS operations, grounded in an acute awareness of safety and predefined alert thresholds, with a proactive stance towards the enactment of emergency safety measures.

2.4. Ironmaking Plant Edge Security Assessment Model Design

2.4.1. Construction of the Ironmaking Plant Edge Security Assessment Model

A multivariate regression analysis method is employed at the periphery of the ironmaking plant to select pertinent independent variables and establish a regression model. The principal steps are as follows.

1. Determination of independent and dependent variables: Independent variables encompass the count of equipment hazard factor, environmental hazard factor, substance hazard factor, unsafe people behavior, number of personnel, duration of hazardous exposure, security management level, system’s automatic control risk capability. The dependent variable is the “Comprehensive Security Index at the Ironmaking Plant Edge”.

2. Data collection and preprocessing: Relevant data are sourced from the system’s “data pool” and undergo preprocessing operations such as data cleansing, filling missing values, and handling exceptional data points.

3. Model fitting: A regression equation is formulated, employing regression analysis methods, assuming a sample of ‘$n$’.

The independent variable of the ‘$i$’-th sample is $X_i=\left(x_1,x_2...,x_p\right)$and the dependent variable is $Y_i$. The regression equation can be expressed as:

$$Y_i={\beta }_0+{\beta }_1{x}_1+{\beta }_2{x}_2+...{\beta }_p{x}_p+{\epsilon }_i$$

(4)

Here, ${\beta }_0,{\beta }_1,{\beta }_2,\dots ,{\beta }_p$ represents the regression coefficients. ‘$P$’ in the regression equation refers to the number of independent variables. ${\epsilon }_i$ is the error term, assumed to conform to a normal distribution and be independently and identically distributed.

4. Model evaluation: The least squares method is used for estimating model parameters. Regularization techniques are applied to prevent overfitting. Model performance is gauged using cross-validation and Mean Squared Error (MSE) to ascertain predictive capabilities.

5. Model application: The established multivariate regression model is deployed for prediction and decision-making in the edge cloud. This includes assessing the security status of managed areas at the edge and taking measures to mitigate hazardous accidents within specific time frames.

2.4.2. Edge Cloud Model Parameterization and Performance Evaluation Based on Multiple Linear Regression

We utilized a multiple regression model for data collection related to security assessment in ironmaking plants and estimated model parameters using the least squares method. This method minimizes the sum of squared errors and provides reliable estimates for model parameters. During this process, the following assumptions are made:

1. Independence and identically distributed (IID) assumption: Sample data concerning ironmaking plant security are independently and randomly drawn from a population and should be identically distributed.

2. Linearity assumption: The independent variable is linearly related to the dependent variable.

3. Normality Assumption: The error term follows a normal distribution.

4. Homoscedasticity Assumption: The error term maintains constant variance.

The formula for minimizing the sum of squared residuals is represented by Equation:

$${\mathrm{m}\mathrm{i}\mathrm{n}}_{{\beta }_0,{\beta }_1,\dots ,{\beta }_p}{\sum }_{i=1}^n{(y_i-{\widehat{y}}_i)}^2$$

(5)

where $y_i$ denotes the actual observed value, ${\widehat{y}}_i$ denotes the value predicted by the model, ‘$n$’ denotes the sample size, and ‘$p$’ denotes the number of independent variables.

The multiple regression model for the ironmaking plant security perception data collection stage is given by Equation:

$$\widehat{y}={\beta }_0+{\beta }_1{x}_1+{\beta }_2{x}_2+\cdots +{\beta }_p{x}_p+ϵ$$

(6)

where $\widehat{y}$ denotes the predicted value of the dependent variable, ${\beta }_0$ denotes the intercept, ${\beta }_1{\beta }_2\dots {\beta }_P$ denotes the coefficient of the independent variable, $x_1{x}_2\dots x_p$ denotes the dependent variable, and $ϵ$ denotes the error term.

Moreover, ridge regression is employed to forestall overfitting issues. Cross-validation is utilized for model evaluation and optimization. The ridge regression objective function is represented by Equation:

$$J(w)=\frac{1}{2n}{\sum }_{i=1}^n{(y_i-{\widehat{y}}_i)}^2+\lambda ||w|{|}_2^2$$

(7)

where ‘$w$’ is the vector of regression coefficients, $\lambda$ is the regularization intensity parameter, and $\left|w\right|{|}_2^2$ is the L2 paradigm square of the vector of regression coefficients.

By minimizing the objective function, the ridge regression coefficient estimation formula is given by Equation (8):

$$\widehat{w}={(X^{T}X+\lambda I)}^{-1}{X}^{T}y$$

(8)

where ‘$I$’ is the unit matrix. ‘$X$’ is a design matrix which contains the data of the independent variables for all the samples. Each row represents a sample and each column represents an independent variable.

Finally, K-fold cross-validation was employed to assess the model’s performance and generalizability. The parameter vector $\beta$ was derived by employing the least squares method, which was applied to the multiple linear regression model.

To achieve optimal parameter estimation, the objective was to minimize the sum of squared residuals, represented as:

$${\mathrm{m}\mathrm{i}\mathrm{n}}_{\theta }\left|\right|y-X\theta |{|}^2$$

(9)

where ‘$X$’ represents the design matrix containing all sample features, and $\theta$ denotes the parameter vector.

The solution using the least squares method yields the closed-form solution of $\theta$ as:

$$\theta ={\left(X^{T}X\right)}^{-1}{X}^{T}y$$

(10)

Here, ${\left(X^{T}X\right)}^{-1}{X}^T$ signifies the pseudoinverse matrix, attainable through matrix decomposition.

Ultimately, the obtained parameter vector $\theta$ is reintegrated into the model, formulating the multiple linear regression equation:

$$y={\theta }_0+{\theta }_1{x}_1+{\theta }_2{x}_2+··· +{\theta }_n{x}_n$$

(11)

Here, ${\theta }_0$ is the intercept term, while ${\theta }_1,{\theta }_2,\dots ,{\theta }_n$ represents the coefficients associated with independent variables. This equation predicts dependent variable values based on inputting feature values of independent variables.

This process elucidates the construction of the multiple linear regression model at the ironmaking plant’s edge. Based on the described equations, parameter estimates are derived as:

$$\begin{gathered} {\widehat{\beta }}_0=\overline{y}-{\widehat{\beta }}_1{\overline{x}}_1-{\widehat{\beta }}_2{\overline{x}}_2-{\widehat{\beta }}_3{\overline{x}}_3-{\widehat{\beta }}_4{\overline{x}}_4-{\widehat{\beta }}_5{\overline{x}}_5-{\widehat{\beta }}_6{\overline{x}}_6-{\widehat{\beta }}_7{\overline{x}}_7-{\widehat{\beta }}_8{\overline{x}}_8 \\ \widehat{\beta }1=\frac{\sum i=1^n(x_{i1}-{\overline{x}}_1)(y_i-\overline{y})}{\sum i=1^n{(x_{i1}-{\overline{x}}_1)}^2} \\ \widehat{\beta }2=\frac{\sum i=1^n(x_{i2}-{\overline{x}}_2)(y_i-\overline{y})}{\sum i=1^n{(x_{i2}-{\overline{x}}_2)}^2} \\ \widehat{\beta }3=\frac{\sum i=1^n(x_{i3}-{\overline{x}}_3)(y_i-\overline{y})}{\sum i=1^n{(x_{i3}-{\overline{x}}_3)}^2} \\ \widehat{\beta }4=\frac{\sum i=1^n(x_{i4}-{\overline{x}}_4)(y_i-\overline{y})}{\sum i=1^n{(x_{i4}-{\overline{x}}_4)}^2} \\ \widehat{\beta }5=\frac{\sum i=1^n(x_{i5}-{\overline{x}}_5)(y_i-\overline{y})}{\sum i=1^n{(x_{i5}-{\overline{x}}_5)}^2} \\ \widehat{\beta }6=\frac{\sum i=1^n(x_{i6}-{\overline{x}}_6)(y_i-\overline{y})}{\sum i=1^n{(x_{i6}-{\overline{x}}_6)}^2} \\ \widehat{\beta }7=\frac{\sum i=1^n(x_{i7}-{\overline{x}}_7)(y_i-\overline{y})}{\sum i=1^n{(x_{i7}-{\overline{x}}_7)}^2} \\ \widehat{\beta }8=\frac{\sum i=1^n(x_{i8}-{\overline{x}}_8)(y_i-\overline{y})}{\sum i=1^n{(x_{i8}-{\overline{x}}_8)}^2} \end{gathered}$$

(12)

Each ${\widehat{\beta }}_j$ is an estimated measure of the influence of the independent variable $x_j$ on the dependent variable ‘$y$’. These coefficients are calculated by minimizing the sum of squared differences between the actual observations and the model’s predictions, known as the least squares method. Specifically:

${\widehat{\beta }}_0$ is the intercept term, representing the expected value of the dependent variable ‘$y$’ when all the independent variables $x_1,x_2\dots x_p$ are zero.

${\widehat{\beta }}_1,{\widehat{\beta }}_2\dots {\widehat{\beta }}_p$ are slope coefficients, representing the expected change in ‘$y$’ for each unit increase in the corresponding independent variable.

Where $\overline{y}$ is the sample mean of the dependent variable ‘$y$’, and ${\overline{x}}_1$, ${\overline{x}}_2$, ${\overline{x}}_3$, ${\overline{x}}_4$, ${\overline{x}}_5$, ${\overline{x}}_6$, ${\overline{x}}_7$, ${\overline{x}}_8$ are the sample means of $x_1$, $x_2$, $x_3$, $x_4$, $x_5$, $x_6$, $x_7$, $x_8$, respectively. ${\overline{x}}_i$ is the sample mean of the ‘$i$’-th independent variable.

This method of calculation can address potential collinearity issues in the model and provides stable coefficient estimates when there is a high degree of correlation among the independent variables. Additionally, these formulas are also used to prevent overfitting when there are few data points, ensuring the model has good predictive performance and interpretability.

Finally, leveraging the estimated parameters, the multiple regression equation forecasts the evaluation index of security incidents at the ironmaking plant.

$$y={\widehat{\beta }}_0+{\widehat{\beta }}_1{x}_1+{\widehat{\beta }}_2{x}_2+{\widehat{\beta }}_3{x}_3+{\widehat{\beta }}_4{x}_4+{\widehat{\beta }}_5{x}_5+{\widehat{\beta }}_6{x}_6+{\widehat{\beta }}_7{x}_7+{\widehat{\beta }}_8{x}_8$$

(13)

When provided with values for equipment and facility defects, operational environment defects, hazardous exposure duration, and unsafe human behavior, this equation computes the corresponding number of security incidents at the ironmaking plant.

2.4.3. Establishment and Optimization of a Multivariate Regression Model for Comprehensive Security at the Edge Cloud of an Ironmaking Plant

By introducing multiple regression models and interaction terms, the edge cloud constructs a more accurate model for the safety data collection stage of the ironmaking plant for the safety assessment of the ironmaking plant. This model is principally aimed at forecasting crucial variables such as the security assessment index associated with the ironmaking facility.

The target variable, designated as ‘$y$’, signifies the security assessment index attributed to the ironmaking plant. It encompasses independent variables $x1$ to $x8$, which respectively correspond to equipment hazard factor, environmental hazard factor, substance hazard factor, unsafe people behavior, number of personnel, duration of hazardous exposure, security management level, system’s automatic control risk capability. $\epsilon$ symbolizes the error term. The formulated edge cloud regression model is outlined as follows:

$$y=\beta 0+\beta 1x1+\beta 2x2+\beta 3x3+\beta 4x4+\beta 5x5+\beta 6x6+\beta 7x7+\beta 8x8+\epsilon$$

(14)

For this model, the resolution of the following system of equations is imperative:

$$\begin{gathered} \sum \left(y-\beta 0-\beta 1x1-\beta 2x2-\beta 3x3-\beta 4x4-\beta 5x5-\beta 6x6-\beta 7x7-\beta 8x8\right)=0 \\ \sum \left(y-\beta 0-\beta 1x1-\beta 2x2-\beta 3x3-\beta 4x4-\beta 5x5-\beta 6x6-\beta 7x7-\beta 8x8\right)x1=0 \\ \sum \left(y-\beta 0-\beta 1x1-\beta 2x2-\beta 3x3-\beta 4x4-\beta 5x5-\beta 6x6-\beta 7x7-\beta 8x8\right)x2=0 \\ \sum \left(y-\beta 0-\beta 1x1-\beta 2x2-\beta 3x3-\beta 4x4-\beta 5x5-\beta 6x6-\beta 7x7-\beta 8x8\right)x3=0 \\ \sum \left(y-\beta 0-\beta 1x1-\beta 2x2-\beta 3x3-\beta 4x4-\beta 5x5-\beta 6x6-\beta 7x7-\beta 8x8\right)x4=0 \\ \sum \left(y-\beta 0-\beta 1x1-\beta 2x2-\beta 3x3-\beta 4x4-\beta 5x5-\beta 6x6-\beta 7x7-\beta 8x8\right)x5=0 \\ \sum \left(y-\beta 0-\beta 1x1-\beta 2x2-\beta 3x3-\beta 4x4-\beta 5x5-\beta 6x6-\beta 7x7-\beta 8x8\right)x6=0 \\ \sum \left(y-\beta 0-\beta 1x1-\beta 2x2-\beta 3x3-\beta 4x4-\beta 5x5-\beta 6x6-\beta 7x7-\beta 8x8\right)x7=0 \\ \sum \left(y-\beta 0-\beta 1x1-\beta 2x2-\beta 3x3-\beta 4x4-\beta 5x5-\beta 6x6-\beta 7x7-\beta 8x8\right)x8=0 \end{gathered}$$

(15)

Matrices can represent this system of equations:

$$X\beta =y$$

(16)

where ‘$X$’ denotes the $n\times 9$ matrix, with the initial column comprising a column vector filled with ones, representing the intercept, followed by eight columns embodying independent variables $x1-x8$. ‘$n$’ denotes the count of observed samples, $\beta$ represents the coefficient vector (9 × 1), and ‘$y$’ denotes the vector of dependent variables ($n$ × 1).

Employing the least squares method to solve this system of equations yields the coefficient vector as follows:

$$\beta =[12.725,-0.186,-0.134,-0.156,-0.017,-0.203,-0.223,0.145,0.116]$$

Once the coefficient vector $\beta$ is determined, this model can forecast the comprehensive security index attributed to the ironmaking plant. To enhance the accuracy of predictions, augmenting the volume of data collection while continuously adjusting model parameters is essential.

Within this framework, the quantity of equipment defects, operational environment deficiencies, material imperfections, duration of exposure to hazards, incidence of unsafe practices among staff, overall workforce size, efficacy of safety management protocols, and the system’s automated control proficiency for mitigating risks profoundly influence the ironmaking plant’s overarching safety index. The predictive efficacy of the multivariate regression model hinges on the linear relationships between independent variables and the selection thereof. Consequently, comprehensive data analysis and meticulous variable selection during the model establishment phase are pivotal to ensure robust predictive and explanatory capabilities within the model.

R-squared values (R²) and Mean Squared Error (MSE) are utilized for model evaluation. R² signifies the extent to which independent variables explicate the dependent variable, varying between 0 and 1, where values closer to 1 denote stronger explanatory power within the model. On the other hand, MSE indicates the scale of the model’s prediction error, with smaller values denoting enhanced predictive capability.

In terms of model optimization, methodologies such as cross-validation and regularization are implemented. Cross-validation assists in evaluating the model’s generalization and circumvents issues of overfitting or underfitting. Regularization techniques such as ridge regression and Lasso regression avert excessive complexity in the model, mitigating overfitting concerns.

2.5. Establishment and Algorithm Design of the Ironmaking Plant Security Assessment Central Cloud Model

2.5.1. Construction Process of Particle Swarm Algorithm-Optimized Artificial Neural Network Model

The security model implemented in the ironmaking plant integrates an Artificial Neural Network (ANN) deployed within the central cloud. This ANN model, drawing from previous expertise, engages in the identification, analysis, and comprehension of various subsystems within intricate systems. Subsequently, it updates the status information of the system and undertakes an iterative process to comprehend, analyze, and formulate an overall assessment of the system’s security [19]. The system employs the gathered indicator data from the aforementioned “data pool” as inputs to the neural network and uses the resulting assessment values as anticipated outputs for training purposes. By harnessing its memory functionality, the neural network generates evaluation values when re-input with identical data, thereby reflecting the security condition of the plant, catering to predictive needs [20].

Neural network models exhibit outstanding capabilities in managing vast datasets, integrating data for comprehensive analysis, adapting to new information, and detecting anomalies, which render them exceptionally suitable for evaluating safety in the intricate settings of steel plants that involve data from multiple, diverse sources.

Initially, our consideration spanned decision trees, random forests, support vector machines, and autoregressive integrated moving average models. However, neural network models present several advantages over these alternatives:

• Comprehensive management of large and diverse data: Neural networks are adept at processing and integrating information from a broad array of sensors and data sources, which is an essential capability for the extensive analysis of surveillance footage and sensor data in steel plant safety evaluations.
• Automatic feature extraction: Neural networks can autonomously learn and extract high-level features from large datasets, adeptly handling complex data and patterns suitable for the analysis of images, videos, and high-dimensional sensor data.
• Dynamic adaptation and continuous learning: Properly trained neural networks exhibit excellent generalization to unseen data, continually learning and adapting post-deployment. This ensures that, once trained, the model is well equipped to manage new and unforeseen safety risks; this is crucial for dynamically evolving industrial settings.
• Flexibility and scalability: Neural network models can be customized and scaled to meet the intricate and variable requirements of steel plant environments.
• Enhanced predictive power: Deep learning models are particularly effective at identifying non-linear relationships and patterns in data; this is crucial for anticipating future safety incidents and identifying potential hazards.

In conclusion, neural network models are adept at navigating the complexities and dynamics of steel plant environments, proficiently identifying and predicting safety risks. This positions them as a potent and adaptable instrument for assessing safety situations in steel plants.

Regarding the optimization of neural network weights, a fusion of the particle swarm optimization (PSO) algorithm, which is notable for its global search capabilities; and the back-propagation (BP) algorithm, which is recognized for its robust local search capabilities, is employed. This amalgamation aims to leverage their individual strengths to enhance the neural network’s generalization and learning capacities, ultimately boosting predictive accuracy [21]. The PSO algorithm, an evolutionary computing optimization algorithm, operates with ‘$n$’ particles randomly distributed in an ‘$d$’-dimensional space. The ‘$i$’-th particle’s position in the ‘$d$’-dimensional space, denoted by $X_i=(x_{i1},x_{i2}\dots \dots x_{id})$, serves as an independent variable input into the fitness function. The current fitness value for $\mathrm{f}\left(X_i\right)$ is derived.

The velocity of particle ‘$i$’ is symbolized as $V_t=(v_{t1},v_{t2}\dots \dots v_{td})$, encompassing the particle’s position and velocity components in ‘$m$’ dimensions. $Pbes{t}_i=(p_{i1},p_{i2}\dots p_{id})$ signifies the best position experienced by particle ‘$i$’, while $Gbes{t}_i=(g_{i1},g_{i2}\dots g_{id})$ represents the best position encountered by the entire population. Generally, in ‘$d$’-dimensional space, the search range for particles is confined within a certain [$X_{MIN}$,$X_{MAX}$], and the velocity’s change range is constrained within predefined [−$V_{MIN}$,$V_{MAX}$].

Throughout iterations, the update of particle flight speed and position ensues via individual and global best values. The formula for updating the ‘$i$’-th dimension of particle ‘$d$’ velocity is as follows:

$$V_{id}^k=w{V}_{id}^{k-1}+c_1{r}_1(Pbes{t}_i-X_{id}^{k-1})+c_2{r}_2(Gbes{t}_i-X_{id}^{k-1})$$

(17)

The formula for updating the ‘$i$’-th dimension of particle ‘$d$’ position is expressed as:

$$X_{id}^k=X_{id}^{k-1}+V_{id}^{k-1}$$

(18)

Here, ‘$w$’ represents the inertia weight, used to adjust the search range within the solution space. ‘$c_1$’ and ‘$c_2$’ denote acceleration constants regulating the maximum learning step size. ‘$r_1$’ and ‘$r_2$’ are random functions within the range [0, 1], enhancing the search’s randomness.

The neural network will undergo training for the ironmaking plant’s security-associated ‘$h$’ index using particle swarm optimization, wherein the BP error function ‘$e$’ serves as the PSO optimization function [22]. Each parameter from the optimization results will be allocated to the initial weights and thresholds of the BP neural network. Employing the PSO algorithm for optimizing the initial weights and thresholds of the neural network can accelerate convergence. Iterations within the group will focus on computing the minimum error.

Specifically, the error calculation formula will act as the optimization function for the PSO algorithm. Through group iterations, the minimum error will be determined. This minimum error will be utilized within the BP neural network as the minimum error for back-propagation, enabling the training of weights and thresholds between network layers for optimization. If the corresponding energy function (related to the activation function in the PSO algorithm) remains unchanged within a specific number of evolutionary generations, the BP algorithm will be utilized for local optimization. If the resulting energy function is lower after the BP algorithm intervention compared to the previous one, the outcomes from the BP algorithm will be considered the best position in the PSO algorithm. Otherwise, the results obtained by the BP algorithm will replace the worst position in the PSO algorithm. This iterative process continues until the specified evolutionary generation is reached. The training process of the neural network is delineated as follows:

Step 1: Network initialization occurs, assigning random numbers within the range (−1, 1) as initial values for various connection weights. An error function ‘e’ is defined, alongside the number of learning cycles ‘n’.

Step 2: ’k’ samples are randomly selected, and their corresponding outputs are utilized as the training set for the neural network, $x\left(k\right)=\left(x_1\left(k\right),x_2\left(k\right),\dots ,x_n\left(k\right)\right)$, $d_o\left(k\right)=\left(d_1\right(k),d_2(k),\dots ,d_q(k\left)\right)$, where $x\left(k\right)$ represents the input vector, and $d_o\left(k\right)$ signifies the expected output vector.

Step 3: Involves computing the inputs and outputs of each neuron in the hidden layer using the following formulas:

$$h{t}_n(k)={\sum }_{t=1}^n{w}_{th}{x}_t(k)-b_h,h=\mathrm{1,2},\dots ,p$$

(19)

$$h{o}_h\left(k)=f\right(h{i}_h(k\left)\right),h=1,2, \dots ,p$$

(20)

$$y{i}_o(k)={\sum }_{h=1}^p{w}_{ho}h{o}_h(k)-b_o,o=\mathrm{1,2},\dots ,q$$

(21)

$$y{o}_o\left(k\right)=f\left(y{i}_o\right(k\left)\right),o=\mathrm{1,2},\dots ,q$$

(22)

Here, ${hi}_h\left(k\right)$ represents the input to the hidden layer, ${ho}_h\left(k\right)$ is the output of the hidden layer, ${yi}_o\left(k\right)$ denotes the input to the output layer, ${yo}_o\left(k\right)$ represents the output of the output layer, $w_{ih}$ indicates the connection weights between the input layer and the hidden layer, $b_h$ signifies the threshold of each neuron in the hidden layer, $w_{ho}$ denotes the connection weights between the hidden layer and the output layer, and $b_o$ is the threshold of each neuron in the output layer. $f\left(x\right)$ represents the activation function.

Step 4: Involves computing errors based on expected and actual outputs, followed by error back-propagation. The back-propagation utilizes the gradient descent method, initially computing the partial derivatives of the error function for each neuron in the output layer:

$$\frac{{\partial }_{\phi }}{{\partial }_{who}}=\frac{{\partial }_{\phi }}{{\partial }_{yfo}}\frac{{\partial }_{\gamma f_{\phi }}}{{\partial }_{who}}$$

(23)

$$\frac{{\partial }_{y{i}_o}\left(k\right)}{{\partial }_{w_{ho}}}=\frac{\partial \left({\Sigma }_h^p{w}_{ho}h{o}_h\right(k)-b_o)}{{\partial }_{w_{ho}}}$$

(24)

$$\frac{{\partial }_e}{{\partial }_{y{i}_o}}=\frac{\partial \left(\frac{1}{2}{\Sigma }_{o=1}^q{\left(d_o\right(k)-y{o}_o(k\left)\right)}^2\right)}{{\partial }_{y{i}_o}}$$

(25)

$$\frac{{\partial }_e}{{\partial }_{h{i}_h}\left(k\right)}=\frac{\partial \left(\frac{1}{2}{\Sigma }_{o=1}^q{\left(d_o\right(k)-y{o}_o(k\left)\right)}^2\right)}{{\partial }_{h{o}_h}\left(k\right)}\frac{{\partial }_{h{o}_h}\left(k\right)}{{\partial }_{h{i}_h}\left(k\right)}$$

(26)

Step 5: Involves adjusting the weights and thresholds of each connecting layer’s neurons using the error partial derivatives:

$$\Delta w_{ho}(k)=-\mu \frac{{\partial }_e}{w_{ho}}=\mu {\delta }_o(k)h{o}_h(k)$$

(27)

$$w_{ho}^{N+1}=w_{ho}^N+\vartheta {\delta }_o\left(k\right)h{o}_h\left(k\right)$$

(28)

$$\Delta w_{ih}(k)=-\mu \frac{{\partial }_e}{w_{ih}}=-\mu \frac{{\partial }_e}{h{i}_h(k)}\frac{\partial h{i}_h(k)}{\partial w_{ih}}={\delta }_h(k)x_i(k)$$

(29)

$$w_{ih}^{N+1}=w_{ih}^N+\vartheta {\delta }_h\left(k\right)x_i\left(k\right)$$

(30)

After updating the weights and thresholds, the global error is recalculated:

$$E=\frac{1}{2m}{\sum }_{k=1}^m{\sum }_{o=1}^q(d_o(k)-y_o(k){)}^2$$

(31)

Finally, the algorithm checks if the error meets the required standard. If the error attains the predefined accuracy or the learning cycles exceed the maximum set count, the algorithm concludes; otherwise, it returns to the third step to initiate the subsequent learning cycle.

2.5.2. Simulation of Neural Network Evaluation Model

We designed a three-layer BP neural network to accomplish the dynamic security assessment of the ironmaking plant. The specific configuration includes: an input layer with 6 nodes, a hidden layer with 27 nodes, and an output layer with 2 nodes employing the Sigmoid activation function. The BP network was trained using the gradient descent with momentum and adaptive learning rate algorithm (‘traingdx’).

The PSO learning factors were set as $c_1$ = 1.49445 and $c_2$ = 1.49445. The maximum weight for the PSO algorithm was set to $W_S$ = 0.9, and the minimum weight to $W_e$ = 0.4. The PSO algorithm employed a population size of 40, and conducted 100 iterations for optimization, with a maximum optimization speed $V_{MAX}$ = 1 and a minimum optimization speed $V_{MIN}$ = −1. The particle search space was within [−5, 5]. The neural network underwent 1000 training ‘$epochs$’ with a learning rate $I_r$ = 0.01 and a training goal of minimizing the error to ‘goal’ = 0.000001. The fitness curve of the particle swarm optimization BP neural network obtained through Matlab2020a software is shown in Figure 10. From the graph, we observed effective convergence in 57 iterations for the hybrid neural network model optimized by the particle swarm optimization algorithm.

Based on the aforementioned neural network model and the combined particle swarm optimization algorithm, incorporating data collected from the ironmaking plant, the training of the neural network was conducted. Figure 11 demonstrates the prediction error curve of the PSO−BP neural network.

To further validate the effectiveness of the PSO neural network in the security assessment of the ironmaking plant, this study compared the output results of the network predictions between the PSO neural network and the traditional BP neural network. The comparison of network prediction outputs is shown in Figure 12.

From Figure 10, Figure 11 and Figure 12, the experimental results confirm that, relative to the traditional BP neural network, the output results of the hybrid neural network optimized by the PSO algorithm exhibit better performance, requiring fewer iterations, demonstrating lower errors, achieving higher precision in the security perception assessment of the ironmaking plant, and yielding more favorable outcomes.

Finally, we compared the prediction errors of the two types of neural networks. As illustrated in

Table 4. Comparison of three types of prediction errors between PSO–BP neural network and BP neural network category.

Category	PSO-BP Neural Networks	BP Neural Networks
Mean Absolute Error (MAE)	0.052103	0.060356
Mean Squared Error (MSE)	0.0028932	0.0051007
Root Mean Squared Error (RMSE)	0.053788	0.071419

, the particle swarm optimization (PSO) neural network demonstrates reduced mean squared errors.

3. Results

3.1. Evaluation of Comprehensive Safety Model for Edge Cloud Part of Ironmaking Plant

This model was finally utilized to predict new data for assessing the comprehensive security index of the ironmaking plant. The least squares method was applied to solve the equation represented as follows:

$$\beta ={\left(X^{T}X\right)}^{-1}{X}^T\mathrm{y}$$

(32)

Here, $\beta$ represents a column vector encompassing the intercept and coefficients of independent variables, ‘$X$’ denotes the design matrix, and ‘$\mathrm{y}$’ signifies the column vector of dependent variables.

Based on the data we have collected, we can construct ‘$X$’ and ‘$\mathrm{y}$’ and use the NumPy library in Python3.6 to perform matrix calculations to solve for the coefficients of the model. See Appendix A for detailed procedures.

In this representation, the first item represents the intercept, while the subsequent eight items signify the coefficients of independent variables. These coefficients can be substituted into the model to derive the final regression model for the edge-side comprehensive security index:

Edge-side comprehensive security index = 9.998 + 0.241 × equipment hazard factor + 0.277 × environmental hazard factor + 0.221 × substance hazard factor + 0.384 × duration of hazardous exposure − 1.081 × unsafe human behavior − 0.334 × number of personnel + 0.336 × security management level − 0.270 × system’s automatic control risk capability.

This regression model incorporates eight independent variables and is suitable for evaluating the comprehensive security index of the ironmaking plant’s edge cloud management area. Through this regression model, we gain deeper insights into variable relationships and leverage it for predicting and analyzing future data, thereby offering substantial support for decision-making.

This model can be employed to evaluate the comprehensive security production index of the ironmaking plant’s edge cloud based on diverse indicators. Focused on accident indicators (preventive, occurrence, or accident equivalent), the security production index incorporates data from the past 12 months of the ironmaking plant’s edge-side coverage area into the regression model. Analytical calculations are conducted as necessary to scientifically evaluate and analyze security production trends. The mathematical representation for the security production index is:

$$Y=(R_1/R_0)\times 100$$

(33)

where, ‘$Y$’ denotes the vertical comparative index (year-on-year index), reflecting the continuous improvement level of security production (accidents) in the company or local region. ‘R’ signifies the industry security production characteristic index or comprehensive index; ‘$R_1$’ represents the index for the current year, and ‘$R_0$’ symbolizes the reference (comparative) index (previous year index, base year index, or average index of the last ‘n’ years).

Moreover, this model is instrumental in analyzing the extent of impact that various indicators have on the comprehensive security index of the ironmaking plant’s edge cloud, thereby identifying the indicators with the most significant impact on security index; thus a scientific basis for ironmaking plant security management is provided.

3.2. Application of Optimized Neural Network Algorithm in Security Assessment of Central Cloud in Ironmaking Plant

3.2.1. Application of Optimized Neural Network

The dynamic security assessment by the central cloud of the ironmaking plant mandates comprehensive consideration of data collected from various terminal gateways and the interrelation between the ironmaking process system and the “People”, “Things”, and “Environment” systems. This study extracted security data from the designed data pool for various segments and functionalized modules within a year and employed statistical methods to analyze these security data to statistically determine the number of hazard modes occurring in various systems throughout the year. It identified equipment and material defect counts using the equipment defect analysis method. The operational environment defect count was determined by statistical analysis of data collected from the environment system within the people, things, and environment systems, extracting equipment log information to calculate the hazard exposure duration.

Calculate the hazard index ‘h’ for each functionalized module according to formula:

$$h=h_s(1+k_1)(1+k_2)(1+k_3)(1+k_4)(1+k_5)$$

(34)

After calculation, the risk assessment indicators for each system are shown in the

Table 5. Table of risk assessment indicators for each system.

Number	Subsystems	Functionalized Modules	${\mathit{h}}_{\mathit{s}}$	${\mathit{k}}_1$	${\mathit{k}}_2$	${\mathit{k}}_3$
1	Blast furnace system	Detection	5	0.25	0.25	0.2
		Electrical	1	0.25	0.1	0.2
		Cooling	7	0.2	0.15	0
2	Feeding and loading system	Screening	6	0.1	0.25	0.15
		Weighing	3	0	0.1	0.1
		Transmission	8	0.15	0.1	0.1
3	Pulverized coal injection	Spray gun	4	0.05	0	0
		Bin pump	4	0.05	0.05	0
		Coal grinding	6	0.25	0	0.1
4	Air supply system	The fan system	8	0.1	0.05	0.15
		Wind furnace	6	0.15	0.2	0.2
		Combustion	1	0	0.15	0.15
5	Iron slag treatment system	Granulating	1	0.05	0.1	0.1
		Dehydration	3	0.15	0.1	0
		Cinder flushing	8	0.05	0.15	0.15
6	Gas dust removal system	Gravity	3	0.05	0.1	0.2
		Venturi	6	0.15	0.15	0.15
		Bag	5	0.1	0.05	0.15
7	People	Recognition	4	0	0.15	0.2
		Video surveillance	3	0.05	0	0
		Positioning	2	0.05	0.05	0.1
8	Things	Vehicle	6	0.15	0.25	0.15
		Material	6	0	0.1	0.1
		Equipment	4	0.25	0.05	0.15
9	Environment	Natural	1	0.15	0.15	0.1
		Generate	2	0.25	0.15	0
		Energy	5	0.15	0.1	0.15

.

1. Input layer: The input layer encompasses eight indicators: equipment hazard factor, environmental hazard factor, substance hazard factor, unsafe people behavior, number of personnel, duration of hazardous exposure, security management level, system’s automatic control risk capability, with a node count of 8. An input matrix of 8 rows and 297 columns was constructed by truncating the eight classes of indicators for 27 functionalized modules per month, forming the neural network training set, as depicted in

Table 6. Neural network training set input (January data for example).

Indicators	January
	Blast Furnace System			Environment
	Monitoring	Electrical	Cooling	Natural	Generate	Energy
Hazardous Mode	11	21	3	12	7	17
$k_1$	0	1	5	3	0	1
$k_2$	5	2	3	5	2	2
E	144	160.8	54	132	90	57
$h_s$	5	1	7	6	3	8
$k_3$	0.2	0.2	0	0	0	0.1

.

The calculated hazard index ‘h’ was transformed into a numerical matrix of 1 row and 297 columns, serving as the expected output for the neural network training set, as depicted in

Table 7. Expected output of neural network training set (January data for example).

Indicators	January
	Blast Furnace System			Environment
	Monitoring	Electrical	Cooling	Natural	Generate	Energy
Hazard index $h$	850.8	158.7	146.11	204.93	853	278.9

.

The data for the 27 functionalized modules in the 12th month were utilized as the test set for testing, as illustrated in

Table 8. Neural network test set input.

Month	Subsystems	Functionalized Modules	Hazardous Mode	Equipment Defects	Environmental Effects	Duration of Hazardous Exposure	${\mathit{h}}_{\mathit{s}}$	${\mathit{k}}_3$
December	Blast furnace system	Detection	24	5	5	144	5	0.2
		Electrical	19	5	2	160	1	0.2
		Cooling	17	4	3	54	7	0
	Feeding and loading system	Screening	21	2	5	58	6	0.15
		Weighing	4	0	2	130	3	0.1
		Transmission	23	3	2	20	8	0.1
	Pulverized coal injection	Spray gun	6	1	0	132	4	0
		Bin pump	16	1	1	223	4	0
		Coal grinding	22	5	0	146	6	0.1
	Air supply System	Fan system	8	2	1	18	8	0.15
		Wind furnace	13	3	4	120	6	0.2
		Combustion	15	0	3	115	1	0.15
	Iron slag treatment system	Granulating	20	1	2	54	1	0.1
		Dehydration	17	3	2	176	3	0
		Cinder flushing	8	1	3	26	8	0.15
	Gas dust removal system	Gravity	11	0	5	144	5	0.2
		Venturi	21	1	2	162	1	0.15
		Bag	3	5	3	54	7	0.15
	People	Recognition	4	0	3	40	4	0.2
		Video surveillance	13	1	0	106	9	0
		Positioning	2	1	1	86	2	0.1
	Things	Vehicle	18	3	5	40	6	0.15
		Material	6	0	2	103	6	0.1
		Equipment	4	5	1	46	4	0.15
	Environment	Natural	16	3	3	115	1	0.1
		Generate	28	5	3	170	2	0
		Energy	5	2	2	67	5	0.15

.

Due to the different dimensions of the constructed training and test sets, normalization was performed to ensure data consistency, normalizing each indicator data within the range of (0, 1).

2. Hidden layer: To ensure accuracy in training and shorten training time, the number of neurons in the hidden layer was chosen as 16, with the activation function between neurons selected as the ‘Sigmoid’ function.

$$f\left(x\right)=\frac{1}{1+e^{-x}}$$

(35)

The training was set for 1000 epochs, with a learning rate of 0.01 and a target minimum error of 0.000001.

3. Output layer: The neural network output resulted in a hazard index ‘h’ for 27 functionalized modules, with a dimension of 1. The output numerical matrix scale ranged between (0, 1), necessitating reverse normalization to obtain data for the system’s hazard index ‘h’.

Upon training and testing the neural network, real-time hazard index ‘h’ for each functionalized module can be obtained. However, the hazard index of each functionalized module alone cannot serve as a basis for evaluating the system’s security level. Further calculations are necessary to reflect the overall system’s security condition. Nonetheless, the application of neural networks streamlines the processing of a large volume of collected indicator data into hazard index data reflecting functionalized modules, facilitating subsequent overall system security perception.

Algorithmically computed, the hazard indices for 27 functionalized-modules in the test set for the 12th month are displayed in

Table 9. The test set of 27 functionalized-modules hazard index table.

Month	Subsystems	Functionalized Modules	$\mathbf{Hazard} \mathbf{Index} \mathit{h}$
December	Blast furnace system	Detection	1350
		Electrical	265.32
		Cooling	521.64
	Feeding and loading system	Screening	558.81
		Weighing	473.35
		Transmission	222.64
	Pulverized coal injection	Spray gun	554.4
		Bin pump	985.64
		Coal grinding	1208.63
	Air supply System	The fan system	830.07
		Wind furnace	794.88
		Combustion	158.7
	Iron slag treatment system	Granulating	146.11
		Dehydration	204.93
		Cinder flushing	1957.96
	Gas dust removal system	Gravity	855.02
		Venturi	156.73
		Bag	142.58
	People	Recognition	496.8
		Video surveillance	126
		Positioning	257.15
	Things	Vehicle	853.11
		Material	747.78
		Equipment	278.93
	Environment	Natural	165.84
		Generate	488.75
		Energy	487.34

.

3.2.2. Implementation of Security Assessment in the Ironmaking Plant

Step 1: Calculating the overall system hazard index ‘H’.

Predicted hazard indices ‘h’ from the neural network cannot independently serve as a basis for evaluating the system’s security level. To further assess the system’s security level, the hazard index ‘H’ needs computation according to the following formula:

$$H={\sum }_{i=1}^N{h}_i{n}_i/{\sum }_{i=1}^N{E}_i{n}_i$$

(36)

where $h_i$ is the hazard index for functionalized modules, $n_i$ is the number of functionalized modules, and $E_i$ is the exposure duration of risks. By calculation, the calculated hazard index for the example demonstrated in this study is $H=\frac{\mathrm{14,154.65}}{2517.55}=5.6224$.

Step 2: Evaluating the ironmaking plant’s risk control capability ‘C’.

This involves a comprehensive assessment of the ironmaking plant personnel’s integrated security management level ‘$C_1$’ and the ironmaking plant system’s automated risk control capability ‘$C_2$’ on a percentage basis. By conducting surveys based on the study from the enterprise risk control assessment table for various systems, a score is given for the enterprise risk control capability $C=C_1+C_2$. In this experiment, the ironmaking plant’s risk control capability is assumed to be $C=66$.

Step 3: Calculating the ironmaking plant’s control capability $B\left(k\right)$.

The specific formula for calculating the control capability index is:

$$B\left(k\right)=\alpha C-\beta H$$

(37)

Within this context, $\alpha$ and $\beta$ are constants, with $\alpha =0.068$ and $\beta =0.55$.

This control index is used to represent the dynamic changes in the enterprise’s risk level. When $B\left(k\right)$ is greater than 0, it indicates an increase in the system’s risk level, whereas if $B\left(k\right)$ is less than 0, it signifies a decrease in the system’s risk level. After calculation, based on the experimental data mentioned earlier, the index for the ironmaking plant’s control capability is $B\left(k\right)=1.40$.

Step 4: Calculating the Initial Security Risk Level of the Enterprise $S(K-1)$.

The initial security risk level of the enterprise can be calculated using the following formula:

$$S(K-1)=G\mathrm{l}\mathrm{n}\frac{1}{P}$$

(38)

where: ‘$G$’ is a constant (10), and ‘$P$’ is the rate of injuries per thousand employees. Assuming the ironmaking plant’s injury rate is 3/1000, the initial security risk level of the ironmaking plant is 58.09. In the context of safety within ironmaking plants, a higher injury rate per thousand personnel is indicative of a lower initial safety risk level.

Step 5: Calculating the comprehensive security risk level of the enterprise $S\left(K\right)$.

The specific formula for calculating the comprehensive security risk level is:

$$S\left(K\right)=S(K-1)+B\left(k\right)$$

(39)

After calculation, the comprehensive security risk level for this experiment is $S\left(K\right)=59.49$.

Upon performing these five steps of calculation, the results of the system’s security risk level are tabulated in

Table 10. The results of calculating the risk level of the system.

Initial Risk Level $\mathit{s}(\mathit{k}-1)$	Hazard Index  $\mathit{H}$	Risk Management Capability  $\mathit{C}$	Management Capability Index  $\mathit{B}\left(\mathit{k}\right)$	Comprehensive Risk Level  $\mathit{S}\left(\mathit{k}\right)$
58.09	5.6224	66	1.40	59.49

.

Step 6: Division of security risk level.

The security risk level is categorized as shown in

Table 11. Security risk level division.

Failure	Criticality	Passing	Excellent
S < S2	S2 ≤ S < 65	65 ≤ S < S1	S1 ≤ S

below.

$S1$ and $S2$ are interval critical values, calculated by the following formulas:

$$S1=60-G\mathit{ln}\left[1+\frac{1}{2MP}(1+\sqrt{1+4MP})\right]$$

(40)

$$S2=60-G\ln \left[1+\frac{1}{2MP}(1-\sqrt{1+4MP})\right]$$

(41)

where: ‘$M$’ is the actual number of employees, ‘$P$’ is the expected rate of injuries per 1000 employees. Assuming the enterprise has 3500 employees, with an expected injury rate of 0.002, after calculation, $S1=67.912$ and $S2=63.725$. However, the comprehensive risk level of our system is $S=59.49$, indicating an inadequate level.

Through neural network predictions and security risk value calculations, the risk level of the ironmaking plant’s environment can be determined. If the relative security risk level is low, it is necessary to analyze the security analysis table to identify the abnormal indicator data leading to a lower security risk level and conduct an analysis of potential faults or possible faults. Upon analyzing the results, commands are issued through the constructed edge IoT service management system to control or repair the fault areas. Through the aforementioned series of operational steps, the realization of ironmaking plant security assessment and response to security incidents can be achieved.

Neural network models enhanced with PSO exhibit notable performance advancements in safety perception assessment tasks within steel manufacturing environments, outshining conventional approaches such as standard neural networks, decision trees, or support vector machines. This section delves into the distinctive improvements and underscores the benefits of utilizing such models:

Comparative Analysis of Prediction Outcomes

• Accuracy: Models refined by PSO consistently achieve superior prediction accuracy. This enhancement stems from PSO’s ability to unearth optimal solutions within the parameter space—solutions that may be bypassed during traditional training methodologies.
• Generalization Capability: PSO-enhanced models demonstrate an augmented ability to generalize. This is attributed to the PSO algorithm’s promotion of a comprehensive parameter search during the optimization phase, which effectively minimizes the propensity for overfitting.
• Convergence Efficiency: In certain instances, PSO-refined neural network models reach optimal solutions more swiftly. This efficiency is a result of the PSO algorithm directing the search with a collective intelligence approach, thereby circumventing the pitfalls of local optima that plague conventional methods.

Advantages and Rationale behind the Model

• Parameter Optimization: The PSO algorithm excels in fine-tuning neural networks’ weights and biases by mimicking the communal behaviors observed in bird flocks. This strategy surpasses traditional gradient descent methods in avoiding local optima traps.
• Adaptability: Boasting remarkable versatility, the PSO algorithm can tackle a wide array of optimization challenges. In the context of safety perception assessments in steel mills, where data often exhibit complex nonlinear patterns, PSO’s global search capability proves invaluable.
• Implementation and Adjustability: Unlike more intricate optimization algorithms, PSO is straightforward to implement, with intuitive parameter adjustments. This accessibility enables model developers to tailor models effectively to specific predictive scenarios.
• Robustness: PSO-optimized neural networks are characterized by their robustness, maintaining consistent performance despite minor variations in input data. This stability is due to PSO’s consideration of multiple solution avenues during optimization, bolstering the model’s resilience to uncertainties.

In summary, neural network models optimized via particle swarm optimization stand out in safety perception assessment within steel mills, offering enhanced prediction accuracy, generalization, and convergence rates over traditional models. These benefits are primarily linked to PSO’s global search capability, its adaptability in processing complex data, and the straightforward, robust nature of its implementation. Consequently, PSO-optimized neural network models emerge as a formidable tool in managing intricate, nonlinear prediction tasks.

4. Discussion

In this study, we innovatively proposed a dynamic safety assessment method for iron-making plants, integrating cloud-edge collaboration technology. This method effectively achieves efficient data integration and processing through network reconstruction technology. Compared to the centralized platform concept of the Industrial Internet of Things proposed by Cong Liqun et al. [23], the highlight of our study lies in emphasizing safety perception assessment at the edge layer and data analysis processing through cloud technology, significantly enhancing response efficiency and the accuracy of safety risk assessment. Utilizing a particle swarm optimized neural network model, our study surpasses traditional models in aspects such as prediction accuracy, generalization ability, and convergence speed. Compared to the platform design concept based on the Internet, cloud computing, and big data analysis proposed by Zhou Xiaoge et al. [24], our model not only emphasizes the importance of information collection and data processing but also enhances the global search capability and adaptability to complex data processing through the particle swarm optimization algorithm. Additionally, this study validated the effectiveness and feasibility of the proposed method through a 12-month simulation test on 27 functional module indicators within a specific ironmaking plant. Compared to the safety production accident risk early warning model for steel enterprises proposed by Dang Guangyuan [25], our model achieves a more comprehensive safety perception and assessment, reducing reliance on subjective experience and significantly improving the accuracy of safety assessment. Although Junjie Wang et al. [26] proposed a neural network-based method for assessing the safety of building structures that could effectively predict the safety levels of structures with the advantage of rapid and mass processing, by introducing the particle swarm optimization algorithm, our research further improves the model’s prediction accuracy and convergence speed in dealing with complex nonlinear problems. Miao Xiaoyu [7] discussed the application of cloud computing in safety management of iron-making plants, and our study further expands the application scope of cloud computing by constructing a cloud-edge collaborative architecture, significantly improving the timeliness of data processing and the efficiency of safety assessment. Valentina Casola et al. [27] proposed a new model for cloud-edge collaborative resource allocation for the Industrial Internet of Things, introducing new ideas in resource allocation, cost, performance, and security strategies. However, our study, by comprehensively utilizing multivariate regression analysis and artificial neural networks, not only achieves real-time data collection and monitoring but also conducts in-depth analysis of safety risks and predictions, providing a more comprehensive safety assessment and risk management solution for ironmaking plants. Bouafia Abderraouf et al. [28] researched a dynamic safety assessment method that offers improvement options for the safety of EMS. In contrast, our study, by adopting a cloud-edge collaborative framework and software-defined networking technology, achieves more flexible and efficient safety assessments, especially performing better in dynamic environments.

In summary, by adopting novel technical means and methods, this study opens new vistas for dynamic safety assessment and risk management in ironmaking plants, significantly contributing to enhancing the safety production level and management efficiency of ironmaking plants.

5. Conclusions

This investigation adeptly addressed the prevalent issues of network heterogeneity and data isolation in ironmaking facilities through a comprehensive network reconstruction. On this foundation, it introduced a cutting-edge framework for dynamic security perception assessment, leveraging cloud-edge collaboration.

• Incorporating intelligent monitoring and safety-alerting wireless sensor networks alongside IoT gateways, the research innovatively reengineered the ironmaking plant’s multisource heterogeneous network system. This reengineering facilitated device pooling and data acquisition. Moreover, it established a multisource heterogeneous security perception system coupled with a cohesive cloud-edge platform, markedly bolstering security assessment capabilities. Additionally, this study devised a bespoke set of safety assessment indicators and a tailored security assessment procedural framework specific to ironmaking operations.
• Utilizing a synergistic cloud-edge strategy, the study employed multisource linear regression models at the edge cloud for localized safety concerns, and at the central cloud, it harnessed particle swarm optimized artificial neural network models to address complex safety issues spanning the entire facility. This approach not only reinforced safety management and mitigated potential risks but also furnished a holistic perspective for tackling safety analysis challenges at both regional and plant-wide levels.
• The practicality and efficacy of the proposed methodology were corroborated through a 12-month simulation trial of functionalized module indicator data within a select ironmaking plant. The outcomes affirmatively indicated that this methodology outperforms conventional safety systems in ironmaking facilities, especially in enhancing response velocity, the stability of intelligent manufacturing systems, and the overall breadth and promptness of safety monitoring and evaluation.

The study underscores the imperative for dynamic security perception assessments in ironmaking plants to be a continual, rigorously monitored, evaluated, and refined process to foster the long-term progression of these facilities. Future research endeavors will concentrate on augmenting data integrity to minimize false alarms and omissions, and on refining algorithms and technologies to elevate the precision and dependability of safety perception assessments within ironmaking environments, thereby advancing safety standards. The methodologies and findings of this paper offer novel insights and methodologies for the conception and enhancement of safety frameworks in ironmaking plants, presenting viable strategies for realizing more intelligent and secure iron production processes.

Moreover, the research outputs detailed in this document have been acknowledged through the awarding of two invention patents, specifically, “An Ironmaking Plant Management System and Its Architecture Method Based on Edge Internet of Things” (Patent No.: ZL202111194007.8, Authorization No.: CN 114,118,678 B, Authorization Date: 2023.10.27, Certificate No.: 6432615) and “Ironmaking Plant Security Situation Awareness Method Based on Edge Internet of Things Technology and Neural Network” (Patent No.: ZL202111194017.1, Authorization No.: CN114048952 B, Authorization Date: 17 February 2023, Certificate No.: 5741725). These accolades not only underscore the paper’s significant contribution to addressing security challenges in ironmaking plants through a dynamic cloud-edge collaborative approach but are also a testament to its empirical validation and patent recognition.