Identifying Safety Technology Opportunities to Mitigate Safety-Related Issues on Construction Sites

Abstract

Although safety technology has recently been shown to prevent occupational incidents, a systematic approach to identifying technological opportunities is still lacking. Incident report documents, containing large volumes of narrative text, are considered valuable resources for predetermining incident factors. Additionally, patent data, as a form of big data from technological sources, is widely utilized to explore potential technology solutions. In this context, this study aims to identify technology opportunities by integrating two types of textual big data: incident documents and patent documents. Text mining and self-organizingmaps are employed to discover applicable technologies for incident prevention, grouping them into five categories, as follows: machine tool work, high-place work, vehicle-related facilities, hydraulic machines, and miscellaneous tools. A gap analysis between incidents and patents is also conducted to assess feasibility and develop a technology strategy. The findings, derived from both types of big data, provide technology solutions that are essential for improving workplace safety and that can be used by business owners and safety managers.

1. Introduction

Industrial and occupational sites are increasingly exposed to various hazards and risks, necessitating the adoption of solutions aimed at mitigating these risks and fostering sustainable workplaces. Since the advent of industrial innovation, resilience to workplace dangers has been strengthened through advancements in technology solutions, work design, and system safety [1,2]. Both academic research and practical applications have focused on addressing occupational safety through technological operations, workers’ behavior, and sustainable systems [3,4]. In particular, technological innovation plays a crucial role in ensuring workplace safety and environmental protection [5,6]. For instance, information and communications technology (ICT) has been highlighted as a tool to reduce occupational safety and health risks in industries such as machinery, chemicals, and construction. The integration of facility sensors and wearable technologies has contributed to a reduction in accident rates by enabling both workers and managers to implement preventive measures [7]. Additionally, the use of big data and the digitalization of operations has automated production processes, thereby indirectly enhancing safety management practices [4].

Innovative solutions have been adopted to improve safety performance and ensure a sustainable workplace [8]. Technologies such as wearable devices, biosensors, augmented and virtual reality (AR/VR), and unmanned aerial vehicles (UAVs) are employed to mitigate workplace hazards [9]. These technologies have been particularly examined in the construction industry, where their current usage, benefits, and limitations have been assessed [10]. Furthermore, studies have identified critical barriers to adopting these technologies for safety and health management, along with strategies to overcome them.

To prevent incidents arising from system failures in industrial workplaces, it is necessary to analyze the characteristics of past incidents. These incidents are documented in reports that include valuable narrative information such as the location, cause, and type of incident, which result from thorough incident investigations [4,11]. Thus, the use of unstructured text analytics on the large volume of incident reports offers an opportunity to develop alternative means of reducing risks through technology solutions designed for inherent safety.

Numerous studies on unstructured data, particularly patent data, have employed text analytics, such as titles, abstracts, and claims [12,13,14]. Text mining and natural language processing techniques are widely used to extract meaningful keyword information from patent data [15,16]. This keyword information is then applied to technology forecasting and planning [17]. Many studies have focused on identifying innovative technology opportunities with the aim of enhancing corporate productivity and sales, driven by market demands [18,19].

However, academic studies linking technology planning based on text analytics to the resolution of current occupational risks in industrial workplaces remain limited. Most prior research on safety management has focused either on incident investigation or technology development in isolation [2,3,9]. In summary, there is a significant research gap in technology management for safety, as systematic approaches that span identifying technology opportunities to developing safety technologies are lacking. In this regard, text analytics has proven valuable for identifying and linking keywords contained in both incident reports and patent data [3].

Safety systems are increasingly being recognized as critical societal and political issues within business administration. Research and development (R&D) on smart technologies, such as sensors, big data, and machine learning, adopted in factory and construction site systems has flourished in the interest of workplace safety [7,20]. This shift reflects broader societal changes, emphasizing safety over productivity, and fostering the introduction of technological innovations for safer workplaces. These innovations aim to prevent incidents intelligently and automatically, ultimately safeguarding workers’ lives and health.

This paper aims to identify safety technology opportunities and develop strategies for technology innovation to reduce hazards and prevent incidents in the workplace. The central thesis of this study is to apply computational algorithms for matching hazards identified in narrative texts from incident reports with technological solutions documented in patent data. Incident reports provide real-world data regarding the need for technology to mitigate risks, and offer valuable insights into incident causes and processes [4,21]. This approach is novel, focusing on the practical risks highlighted in incident databases as opportunities for technological advancement. By matching incident risks with technological solutions, this study provides an effective means of directly applying technological innovations to actual hazardous situations.

The research process is divided into three phases. First, a text mining algorithm is employed to extract key terms from both incident and patent data. The incident data are gathered to identify the types and causes of incidents, while the patent data are used to pinpoint applicable technologies for reducing or preventing these incidents. Second, the self-organizing map (SOM) algorithm, a clustering algorithm, is applied to develop a map between risks and technologies by integrating the incident and patent data. Compared to other clustering algorithms, SOM provides an intuitive and clear distinction for interpreting the features of grouped data. Finally, the integrated approach is used to formulate a technology strategy by comparing incident and technology data.

A case study of a construction site is presented, these sites having the highest mortality rate among various industrial sectors. Risks within construction sites arise from both facility-related hazards and worker behaviors, such as operating heavy machinery and working at elevated heights. Given the prevalence of these incidents worldwide, investigating effective technologies for risk reduction within the construction sector is of paramount importance.

2. Background

2.1. Technology Innovation in Safety

Previously, emergency sensors and interlock systems were among the most prominent technological applications used in production safety systems. Recently, intelligence algorithms have been developed for automated surveillance and alarm systems, such as machine learning, sensors, and image processing [22,23]. In addition, numerous companies have invested in research and development (R&D) to explore advanced technologies, including sensors, augmented and virtual reality (AR/VR), drones, and machine learning, to enhance the structure of safety systems, as shown in

Table 1. Cases of technology innovation in safety.

Technology	Case of Innovation
Sensor	Worker’s location tracking [24], communication technology [25], bio-sensor and health monitoring [26,27]
AR/VR	Emergency plan [28], automotive drivers’ safety [29]
Drone	Workplace surveillance [30], crack detection [14]
Machine learning	Fire detection [23], worker’s surveillance [22,31], protective equipment [25,26]

.

For example, sensor-based safety systems in construction sites have been developed to detect anomalies arising from unsafe conditions and behaviors. CCTV-based surveillance systems have also been implemented for real-time monitoring to identify hazardous situations, such as workers not wearing safety helmets or fires occurring. These technological innovations have recently been advanced to create more efficient and intelligent data systems, primarily due to the growing societal focus on safety management. Among these innovations, machine learning algorithms based on big data have been predominantly used to identify data features. As a result, it is anticipated that these technology innovations will significantly contribute to improving worker safety across various industries.

2.2. Big Data Analytics for Safety Technology Identification

2.2.1. Text Analytics for Technology Innovation Studies

Technology planning is expected to consider both the technology or supplier perspective and the market or demand perspective. Previous research has made several attempts to discover and analyze data suitable for these different perspectives, interests, and purposes. The literature on technology forecasting commonly identifies three main types of big data sources: intellectual properties (such as patents and trademarks), publications, and opinions [32]. Recent studies have developed data-driven approaches to investigate abundant data from various sources, extracting knowledge to enhance the understanding of technology development patterns [13,17] and market needs structures [33,34].

Among the various methods used to analyze technology innovation-related data, text mining has become a fundamental tool for processing unstructured textual information and extracting useful insights from it [35]. While each data source has its own distinct features, it can be examined from the viewpoints of technology supply and market demand, as shown in

Table 2. Previous research on text analytics based on technology documents.

Market Demand Sources	Technology Supply Sources	Research Design
Patent	Publication	Index analysis [36], regression analysis [37]
Trademark	Patent	Survey [38], topic modeling [19]
Product/service specification	Patent	Collaborative filtering [39], association mining, and textmining [40]
Web review data	Product/service specification	Textmining, topic modeling [41]
Web review data	Patent	Textmining, opinion mining, and structural equation model [33], textmining, opinion mining, and SAO2Vec [34]

.

Patents and academic papers are commonly recognized as primary data sources for the technology supply side [42]. However, some studies have utilized patents not only to assess technological capabilities but also to uncover business opportunities. For example, patents have been leveraged to evaluate the technological strengths of companies and to identify opportunities at both the industry and product levels [43]. Assignee-product portfolios have been constructed using patent data, and new product suggestions for a target company have been made through collaborative filtering techniques [39]. These approaches are possible because patents provide not only the technical details of inventions, but also descriptions of the product attributes to which the technology can be applied.

Trademark data serve as another important data source for obtaining market-related insights and exploring business opportunities. While patent data focuses on the industrial applicability of technological inventions, trademarks offer information about products that have already been commercialized or are likely to be commercialized soon [38]. This makes trademark data particularly valuable for generating more practical knowledge in a real-world business context. Business opportunities can be identified by analyzing patent data to examine an organization’s technological capabilities and the competitive landscape, while trademark data can be used to investigate targeted goods and services [40]. To uncover such opportunities, methods like collaborative filtering, association mining, and text mining are commonly integrated.

2.2.2. Text Analytics for Incident Analysis

Narrative texts offer the advantage of providing detailed insights that capture incidents as a combination of multiple factors rather than isolated elements [21]. Numerous studies have highlighted the utility of narrative text analysis for various purposes. For instance, several studies have validated the accuracy and value of the information contained in narrative texts [44,45]. Other research has focused on categorizing accident factors based on the narratives within incident reports [4,46].

The unstructured data within narrative texts, derived from natural language, are particularly valuable because they contain additional contextual information about the accident process, such as circumstances, causes, contributing factors, culpability, and injuries [47]. As such, narrative texts offer crucial insights into the type and nature of accidents. In the field of industrial safety, safety managers have used the narrative texts in accident reports to investigate the accident process and understand underlying patterns and contributing factors [48]. Natural language-based text written by safety managers serves as raw data that are instrumental in identifying patterns within accident processes and classifying accident types [4,46]. The analysis of textual information has gained increasing attention for identifying accident processes in sectors such as manufacturing, construction, automotive, and aviation, where narratives play a critical role in providing new insights into occupational safety. Notably, these narratives are valuable for understanding the context surrounding the accident process [4,21].

Moreover, narrative texts allow for in-depth analysis by revealing the underlying causes, contributing factors, and injuries associated with the accident process [4]. Based on the latent risks extracted from incident data via narrative text analysis, these risks can be identified as technology opportunities for mitigating and responding to incidents. According to the principle of “as low as reasonably practicable” (ALARP) based on “Hierarchy of Controls”, as depicted in Figure 1 [49,50], this model has been widely accepted in many occupational safety organizations, including HSE (Health and Safety Executive in Great Britain), OSHA (Occupational Safety and Health Administration in US), and NIOSH (National Institute for Occupational Safety and Health in US). Technology can be applied across all stages, from elimination to PPE; however, technological measures are recognized as the most effective way to eliminate underlying risks, ensuring inherent and permanent safety in industrial environments.

Therefore, the focus of this study is placed on identifying and addressing the issues described in the narrative texts, with the goal of developing technologies that eliminate these problems. Compared to other approaches such as protective equipment and administrative controls, technological solutions offer the advantage of delivering inherent and permanent safety. While some studies have identified and classified technologies for occupational safety [2,4], few have linked these efforts to actual occupational and industrial accident cases. This underscores the need for an approach that identifies technology needs derived directly from incident data.

2.3. Research Gap and Focus

As reviewed, few studies have explored the potential of incident contexts as a technology opportunity. By combining incident reports, which contain accident-related information, with patent data, which offer technology-related information for preventive safety, the proposed approach addresses four key aspects of the research gap, distinguishing it from previous research efforts.

First, with regard to the textual data analyzed, previous studies have typically focused on using either incident reports or patent documents independently, without integrating both types of data [2,4]. This study, however, proposes an integrated approach that combines both data sources, enabling the simultaneous analysis of incident-related and technology-related keywords. Second, in terms of the significance of the results, the proposed method not only identifies existing accident causes or technology characteristics, but also directly presents integrated results that can facilitate the discovery of response technologies related to accident causes or the prediction of preventive technologies needed in the future. Third, while existing methods are primarily concerned with structuring or categorizing textual data, this study introduces an algorithm that specifically matches incidents with patents. This approach enables the dynamic portfolio analysis of accidents and technologies, considering their timing and the missing links in their associations. Finally, in terms of application, the primary advantage of the proposed approach is its ability to foster collaboration between safety managers and technology developers based on the relationship between accidents and technologies. In contrast, previous studies have typically focused on either accidents or technologies in isolation, limiting the potential for such collaboration.

3. Proposed Approach

3.1. Integrating Incident Data with Patent Data

This study aims to develop a systematic tool for identifying practical safety needs and technology opportunities on construction sites, as depicted in Figure 2. The hazard- and risk-related keywords found in incident data represent practical safety demands that need to be addressed on construction sites [51], while the technical keywords from patent data serve as a basis for applying ICT solutions to enhance construction safety. As previously discussed, there have been numerous efforts to integrate ICT, such as sensors (IoT), big data, and machine learning, into the process of detecting and monitoring risk conditions and behaviors on construction sites. These technologies have thus emerged as a promising market for safety solutions in the construction industry [52]. By interacting with data on incidents and ICT technologies, this study aims to identify effective technological solutions for improving safety, and to explore the optimal ways to implement these solutions on construction sites.

The analytic procedure for matching incident data and patent data is illustrated in Figure 3. The first step involves gathering databases of incident data and patent data, with OSHA and USPTO (United States Patent and Trademark Office) serving as valuable sources for this research. After collecting the data, text mining is applied to extract keyword information from OSHA incident reports related to construction sites and USPTO patent data pertaining to the ICT sector. Subsequently, the Self-Organizing Map (SOM) algorithm is utilized to create a matching map that integrates both incident and patent information. This matching map enables the identification of applicable technology opportunities based on the distribution of data between incident and patent documents. Finally, technology planning is developed through technology roadmapping, which visually represents technology opportunity trends from both technological and market perspectives, taking into account the lead and lag of various technologies.

3.2. Applied Methodology

3.2.1. Textmining

Text analysis is a technical term that has re-emerged with artificial intelligence in recent years, as well as being a pattern or knowledge analysis technique that extracts previously unknown, computer-comprehensible information of practical value from a large number of unstructured texts. This field is becoming increasingly important because of the large amounts of data available in documents, news articles, research papers, and accident reports. The text analysis process consists of data collecting, text preprocessing, text feature extraction, and knowledge discovery. Text preprocessing is realized by the basic techniques of text segmentation, such as word segmentation, incomplete data removal, and syntactic analysis. One of the main goals of textmining is to characterize the contents of the documents through pattern discovery for keywords. These patterns can be used to improve information retrieval or to input into predictive models. In addition, most textmining begins with vector space models, where data are represented by document-term matrices. These matrices use terms as headers for the rows and documents as headers for the columns. The values in the cells give the counts or frequencies of a term (column) in a document (row).

3.2.2. Self-Organizing Map (SOM)

The SOM is a type of artificial neural network that uses unsupervised learning and trains its network through a competitive learning algorithm to build a two-dimensional map of a problem space. It is also known as a self-organizing feature map (SOFM) and is used to produce a low-dimensional representation of a higher-dimensional dataset while preserving the topological structure of the data [53]. The neurons, also called nodes, in the network are organized in a two-dimensional grid, and each neuron is connected to its neighboring neurons. The neurons are updated during the training process based on their distance from the input data.

A distinctive feature of SOM is that it uses topological information in learning. In other words, the spatial relationships between the input data points are preserved in the output space of the self-organizing map. This means that similar input data points are mapped to nearby neurons in the output space, while dissimilar input data points are mapped to distant neurons. This topological property of SOMs makes them useful for the clustering and visualization of high-dimensional data. The fundamental idea behind the SOM is to map high-dimensional data into a lower-dimensional grid (typically 2D) while maintaining the topological structure of the original data. This grid is usually composed of a set of neurons (also called units or nodes), each of which represents a weight vector. Here is a step-by-step breakdown of how it works:

• Step 1. Initialization

The SOM consists of a grid of neurons, and each neuron is initialized with random weights of the same dimension as the input data;

• Step 2. Competition

Each input vector (data point) from the dataset is compared to the weight vectors of the neurons. The neuron whose weight vector is most similar to the input vector (often using the Euclidean distance) is called the Best Matching Unit (BMU) in the Vote cell. This neuron “wins” the competition;

• Step 3. Cooperative Step

The BMU and its neighboring neurons (in the grid) are adjusted to move closer to the input vector. This adjustment is determined by a neighborhood function, which ensures that neurons closer to the BMU are adjusted more than those farther away. The degree of adjustment decreases with time;

• Step 4. Adjustment

The weight of each of neurons is updated according to Formula (1),

$$w_i\left(t+1\right)=w_i\left(t\right)+\eta \left(t\right)h_{i,BMU}\left(t\right)(x-w_i\left(t\right))$$

(1)

The $w_i\left(t\right)$ is denoted as the weight of the neuron i at time t, and $\eta \left(t\right)$ is denoted as the learning rate at time t. Here, $h_{i,BMU}\left(t\right)$ stands for the neighborhood function that decreases with distance from the BMU to represent the similarity;

• Step 5. Convergence

Through many iterations, the map starts to organize itself such that similar data points in the high-dimensional space are mapped to nearby neurons in the 2D grid, called the U-matrix. The learning rate and neighborhood size gradually decrease over time to allow the SOM to stabilize. The size of the SOM can be determined by referring to two metrics. The quantization error measures the average distance between each data point and its closest BMU as the neuron in the SOM grid that best represents the data point. On the other hand, the topographic error measures how well the SOM preserves the topology (neighborhood relationships) of the input space.

The U-matrix is a widely used method for visualizing the results of SOM [53]. It measures the distance between neighboring neurons in the SOM map by calculating the difference between their weight vectors. The resulting distances are then represented using a color scale image, where lighter colors indicate greater distances between neurons and darker colors indicate smaller distances. This allows for the easy identification of clusters of neurons that are closely spaced and boundary regions that are more distant.

The SOM has been used in patent analysis for various purposes such as research trend identification [54], patent quality classification [55], and technology opportunity analysis [39,56]. More recent years have witnessed some studies using SOM to categorize incidents and identify major characteristics based on keywords from the texts of incident reports [57,58,59]. However, there have been few attempts to link incident data with patent data, despite the potential utility of this approach in finding applicable technology for addressing practical problems that require solutions.

4. Results

4.1. Data Collection: Incident and Patent Data

For this study, unstructured databases were utilized, including incident data from OSHA and patent data extracted from the USPTO. Specifically, 2598 incident reports from construction sites classified under NAICS (North American Industry Classification System) code 23, covering the period from 2015 to 2017, were collected from the OSHA website. Specifically, we use the narrative text column contained in OSHA dataset. Additionally, 2374 patent records related to safety technologies from 2014 to 2016 were gathered. These patents were identified by searching the IPC (International Patent Classification) codes in the H (electricity) and G (physics) categories, using the keywords “safety” and “technology” with the “AND” logic. This search strategy was employed because most ICT-related patents are classified under these categories, as technological advancements have shifted the ICT-IPC matching taxonomy over time. To maximize the scope of patent data, only these two keywords were used.

For text mining, the Natural Language Toolkit (NLTK), a Python package, was employed. During preprocessing, meaningless words, such as articles and stopwords, were removed to structure the keyword data as input for the Self-Organizing Map (SOM) algorithm. Each document was also coded to abstract the information. For instance, consider a document with the code “P160075”, which follows a specific structure—the first letter “P” indicates a patent document, while “A” would denote an incident report; the next two digits (16) refer to the year (2016), and the final four digits (0075) represent the document number. Therefore, “P160075” corresponds to a patent document from 2016 with the number 0075 in the database.

4.2. Constructing Matching Map

The Self-Organizing Map (SOM) was executed using the SOMtoolbox within MATLAB. To ensure both visibility and accuracy, the SOM was structured as a 10-by-8 matrix. The performance of the SOM was evaluated based on two key metrics, quantization error and topographic error, which were found to be 2.229 and 0.031, respectively. These values indicate a good level of accuracy in the SOM’s ability to map the incident and patent data.

The integration of both incident and patent data led to the creation of a matching map, as shown in Figure 4. Each node in the map is associated with a label, which corresponds to the representative document code for the node. These labels were automatically generated using the “som_autolabel” function available in the SOMtoolbox. From the available options within this function, the “vote” method was selected. This approach ensures that only the label with the most frequently input data is assigned to each node, thus representing the most dominant document for each respective area of the map.

Based on the distance between nodes, the data in the SOM were divided into five clusters, each representing distinct patterns in the incident and patent data. In

Table 3. Results of incident–technology clusters.

Cluster	Incident	Patent	Keywords Contained	Main Work	Related ICT	Type of Incident
Ⅰ	737	333	pipe, concrete, metal, machine, piece, cut, cover, hand, finger, electrical	Machine tool work	Automatic machine, Interlock system	pinch, stuck
Ⅱ	1537	64	fell, roof, head, concrete, fractured, scaffold, pipe, ribs, breaking, slipped, weight	High place work	Weight monitoring, Head defense	fall, hit, slip
Ⅲ	36	583	truck, vehicle, control, signal, sensor, module, controller, struck, electrical, motor, alarm	Vehicle	Sensor in vehicle, Signal processing	fall, hit, pinch
Ⅳ	145	945	battery, lift, belt, vehicle, pressure, layer, electrical, syringe, valve, control	Hydraulic machine	Battery monitoring, Valve control	hit, fire. explosion
Ⅴ	143	448	position, needle, valve, seat, locking, support, power, connected, door, electrical	Miscellaneous; Others	Interlock system, Power monitoring, Safety seat	hit, electric shock
Total	2598	2374

, these clusters are summarized with the number of documents and types of matching incidents and patents.

• Cluster I: This cluster contains 737 incident cases and 333 patent cases. The main keywords in both incident and patent data are related to “machine tool work”, with keywords such as pipe, concrete, metal, machine, piece, cut, cover, hand, and finger. The primary incidents in this cluster are amputation and being stuck. These incidents are typically associated with machinery, indicating a need for safety technologies related to machine tool operations.
• Cluster II: This cluster is heavily biased toward incident data, with 1537 cases of incidents. The incidents are primarily related to “high place work”, with keywords such as fell, roof, head, concrete, fractured, scaffold, pipe, ribs, breaking, and slipped. The high-risk nature of working at heights is evident, and the lack of technology support for these types of incidents is highlighted. This cluster emphasizes the need for safety technologies to prevent falls and injuries during high-place work.
• Cluster III: This cluster contains 583 patent cases and focuses on “vehicle” safety technology. The keywords in this cluster include truck, vehicle, control, signal, sensor, module, controller, struck, electrical, and motor. The emphasis here is on vehicle-related safety technology, with patents aimed at improving vehicle safety and preventing accidents related to construction vehicles.
• Cluster IV: This cluster deals with safety technologies or incidents related to “hydraulic machines”, which are commonly used on construction sites for tasks like lifting and moving objects. The keywords in this cluster include battery, lift, belt, vehicle, pressure, and layer. The data indicate the potential for improving safety technologies around hydraulic machinery.
• Cluster V: This cluster presents ambiguous incident and patent data related to various mechanical components, such as position, needle, valve, seat, locking, support, and housing. The specific safety concerns in this cluster are less clearly defined, but they suggest a need for safety technologies around mechanical systems and components.

These five clusters provide a comprehensive view of the safety risks on construction sites, highlighting areas where technology could be applied to reduce incidents and improve worker safety.

Each of the five clusters identified in the SOM analysis corresponds to specific safety concerns in construction sites, and for each, relevant ICT technologies can be applied to mitigate risks and enhance worker safety. These ICT solutions offer practical applications that can reduce risks, improve operational safety, and enhance the overall safety management system on construction sites, tailored to the specific hazards identified in each cluster. The breakdown of the ICT solutions is presented based on the keywords contained in patents associated with each cluster.

• Cluster I (Machine Tool Work): The automatic control and interlock systems are essential technologies for improving safety in machine tool work. The interlock system helps prevent dangerous situations by ensuring that certain operations cannot happen simultaneously or in unsafe sequences. For example, an interlock might prevent a machine from starting unless safety guards are in place, reducing the risk of pinch points and injuries.
• Cluster II (High-Place Work): For high-place work, facility strength monitoring technologies are crucial. These technologies can monitor the integrity of scaffolds, ladders, and other temporary structures to ensure they are safe for workers. Sensors can detect signs of potential structural failure or instability, issuing alerts to prevent accidents caused by collapses. These systems can be integrated into construction management platforms to offer the real-time monitoring of workplace safety.
• Cluster III (Vehicle Safety): This cluster focuses on vehicle safety technologies, particularly proximity sensors and alarm systems that detect vehicles near workers on construction sites. In environments where construction vehicles like forklifts, trucks, and backhoes are frequently in operation, vehicle-to-worker detection systems can prevent accidents by issuing real-time warnings to workers when vehicles are in close proximity. These sensors can be installed on vehicles and worn by workers to ensure both vehicle and worker safety.
• Cluster IV (Hydraulic Machines): For hydraulic machinery, battery monitoring and valve control systems are critical. Hydraulic machines often operate with high force and energy, which can be dangerous without proper control systems. ICT solutions in this cluster can include the real-time monitoring of battery health and performance, ensuring that hydraulic systems are operating efficiently. Additionally, valve control systems can ensure that fluid pressure and flow rates are appropriately maintained to prevent malfunction or accidents.

4.3. Identifying Matching Information

The analysis of matching nodes within each cluster highlights important connections between incidents and patents, revealing technology solutions or opportunities that address specific safety concerns. The matching nodes, which contain both incident and patent data, are identified as areas where incidents and corresponding technologies overlap, suggesting potential opportunities for technology application to mitigate risks. Linking keywords common to both incident and patent documents helps pinpoint the areas of overlap and guide the identification of relevant technological solutions. The circle in Figure 5 illustrates the ambiguous BMUs that share a boundary between clusters. It is highly likely that these BMUs contain both types of documents, thus providing values for comparing incident and patent information. By selecting representative matching nodes from each cluster, concrete examples of how technology can address specific risks are provided. The related documents in these nodes are summarized in

Table 4. Technology keywords for applying incident-responsive technology.

Label	Cluster	Documents Classified in Nodes		Linking Keywords
		Incident	Patent
A151449	Ⅰ, Ⅳ	A140659 A151449 A150735 A161580 A170176	P150034 P150745	machine, member, operate, move, open, position, part, connect, switch, control, remote, lock, guard
A151241	Ⅰ, Ⅱ, Ⅳ	A150322 A150361 A150158 A151241 A170081	P140566 P150259 P150842 P150850	work, move, remove, position, step, lift, load, attach, assembly, truck, member, ladder, support, assembly, mount
P150298	Ⅰ, Ⅲ, Ⅳ	A150058 A150930 A151449 A161647	P140276 P140835 P150007 P150032 P150298	truck, vehicle, control, signal, sensor, module, controller, struck, electrical, motor, pressure, fence, strap, implant, line, location
P150578	Ⅰ, Ⅳ	A150950 A160013 A160883 A161450	P140058 P140041 P150578	lift, belt, pressure, layer, electrical, valve, control, crane, shock, line, level
P140746	Ⅳ, Ⅴ	A150019 A150454 A160182 A160348	P140708 P140746 P150147 P150199 P160084	power, electrical, vehicle, burns, line, shock, switch, signal, circuit, voltage, load

. This approach systematically identifies technology solutions tailored to specific safety challenges on construction sites by examining matching nodes and their linking keywords.

First, the node labeled A151449 groups together five incidents and two patents, extracting information related to machine switches, control, and guards for “machine tool work”. By monitoring the keywords, we can identify the need to reduce risks associated with moving, opening, connecting, remotely operating, and locking machine tools, which highlights ICT requirements for safely controlling machine tools. For example, the two patents in this node can be used as technology solutions, or considered opportunities to address the five incident cases.

The second cluster, marked by the A151241 node, includes linking keywords such as lift, load, step, truck, and ladder, all of which relate to high-place work and the increased risk of falls. Most patents in this cluster focus on fall-protection technologies or products for high-place work safety.

The third case involves the P150298 node, which is associated with ICT-based vehicle safety technologies, including trucks, control, signal, sensor, electrical systems, and fences. Sensors, used to monitor trucks, fences, and operational lines are key technologies used to control vehicle operations on construction sites. The matching node labeled P150578 is more focused on hydraulic machines, which are operated by pressure and involve parts like valves, cranes, and lifts. Control issues in these machines are also addressed using ICT patents within this matching node.

Finally, the P140746 node addresses electric hazards such as shocks and burns, indicating the need for power control and circuit design to mitigate these risks. These representative cases demonstrate how the linked keywords in matching nodes provide technological opportunities to address safety needs in both R&D and industrial sites.

4.4. Technology Planning

4.4.1. Data Distribution Analysis

This study proposes technology planning by comparing the matching nodes that include both incident and patent data. Specifically, we can uncover the situation and possible technology strategy by analyzing the data distribution of clusters. Based on the results of mapping information from the two types of documents, incident and patent portfolios can be organized in terms of time and vacancy, as shown in

Table 5. Technology strategy derived from data distribution.

Cluster	Data Distribution	Possible Situation	Possible Technology Strategy
Ⅰ	Balanced documents	Direct applications	Technology application Technology development Feasibility test
Ⅱ	Almost all incident documents	No solutions yet Need for preventive technology	Technology opportunities Technology forecasting
Ⅲ, Ⅳ, Ⅴ	Almost all patent documents	No needs yet Useless technologies	Technology application Licensing Abandonment

. Regarding time, we compare the years of accident incidents with the years of technology development to uncover trends in reactive or preventive technologies. Regarding vacancy, we predict the likelihood of related technological development or accident occurrence by analyzing the distributions of accident and patent cases.

First, as previously mentioned, Cluster I (machine tool work) exhibits a balanced distribution of incident and patent documents. It is noteworthy that many matching cases between incidents and patents are identified, suggesting a direct application of technology to prevent related incidents. This result establishes a pool of relevant technologies and incidents that can address problems in the workplace. In this case, technology planning needs to be formulated, including technology development, application, and feasibility testing. For technology application, patents related to incident data in the matching cells should be explored. As indicated in a large body of the literature, safety managers in construction sites can identify ICT solutions, such as sensors, CCTV, and communications networks, to solve problems in construction environments.

In contrast, other clusters with a bias towards one document type reveal a gap in information between technology and incidents. Most of Cluster II (high-place work) consists of incident documents, signaling that technology matching the incidents is underdeveloped. Although many incidents occur in construction sites, preventive technology for high-place work is still lacking. The final three clusters, which are predominantly patent documents, indicate that technologies in Clusters III, IV, and V are less directly relevant to the incidents.

4.4.2. Technology Roadmapping Using Gap Analysis

In the context of technology roadmapping (TRM), data requirements and examples for extracting appropriate keywords are presented for each layer of the TRM: market, service, product, and technology [60]. In this framework, the incident document is placed in the market layer, while the patent corresponds to the technology layer. The interaction between incident and technology is facilitated by linking keywords present in both documents.

In terms of the time difference between incidents and patents in a given matching node, two types of technology–incident relationships can be identified: technology lag and technology lead. As outlined in

Table 6. Technology gap analysis: lead and lag.

Type of Technology Gap	Criteria	Implication of Technology Strategy
Technology lag	Year of Patent > Year of Incident	Technology development to reduce risks of incident after occurring incidents
Technology lead	Year of Patent < Year of Incident	Late applications of technology despite early technology proposals

, when the year of the incident precedes the year of the patent, it is classified as a technology lag. Conversely, when the patent year precedes the incident year, it is classified as a technology lead. These two types carry distinct implications for technology strategy. A technology lag indicates the need for technology development to reduce risks associated with incidents that have already occurred. In contrast, a technology lead suggests that technology applications are delayed despite early proposals or developments.

We can formulate strategy types through the illustration of possible scenarios, as represented in

Table 7. Technology strategy derived from technology gap analysis: lead and lag.

Gap Strategy	Strategy Type	Related Documents		Linking Keywords	Possible Technology Roadmap
		Incident	Patent
Technology lag (Demand–pull)	Technology development	A140659	P150745	machine, member, operate, move, open, position, part, connect, switch, control, remote, lock, guard
	Incident–pull forecasting	A140659	P150034 P150745
Technology lead (Technology–push)	Technology applications	A150019	P140708	power, electrical, vehicle, burns, line, shock, switch, signal, circuit, voltage, load
	Technology–push forecasting	A150019 A150454	P160084

. Based on the gap strategy, various technology strategies can be adopted. First, technology development is a market–pull strategy, which shows that incident data can be related to new technology information based on linking keywords. Building on incident–pull forecasting, new technologies can be predicted using the keywords associated with future accidents, added to the linking keywords. On the other hand, for technology applications, the roadmap follows a technology–push strategy, from patent to incident. When technology solutions are already available, patents should be applied to address incidents. Additionally, new types of technology can be forecasted to prevent future incidents related to the same issues.

5. Discussion

5.1. Implications

5.1.1. Interest in Stakeholders of Safety Technology

Previously, many R&D activities focused on technology innovation aimed at increasing productivity, but recently, preventive safety technologies have also become a central focus of R&D projects in industrial sites. In South Korea, in particular, occupational safety policies emphasize the adoption of technology solutions to address unsafe conditions and behaviors, highlighting the importance of technology in structuring safety systems. Policy-driven technology planning is fostering technology innovation in the workplace. Among various attempts to reduce risk, technological measures are especially effective in ensuring inherent safety. The results derived from matching information between two types of documents provide valuable insights for CEOs of companies aiming to develop safety technologies in response to occupational safety and health regulations.

Beyond CEOs, interest in safety technology is growing among policymakers, researchers, and workers. (a) Policymakers can gain insights into the status of safety technology and create a technology roadmap based on the matching information between incidents and patents. (b) Researchers can forecast new technologies and products by understanding the gaps between incidents and patents, ensuring that they address practical needs to improve workplace safety. By analyzing incidents, researchers can also identify and address the lack of safety technology in industries. (c) Workers can benefit from the adoption of safety technologies that reduce hazards, mitigate risks, and improve unsafe conditions and behaviors. This research categorizes safety technologies into four main categories, extracted as machine tools, high-place work, vehicles, and hydraulic machines, helping policymakers, researchers, and workers recognize key risks and understand relevant work procedures in construction sites.

5.1.2. Identification of Incident–Technology Linkage

Safety management traditionally relies on experts rather than data, but recent studies have begun recognizing incident reports as valuable data sources for understanding the characteristics and mechanisms of industrial hazards. From the perspective of technology planning, incident reports can be seen as targets that technology needs to address in order to prevent incidents and enhance safety in industrial settings. This study has leveraged the potential use of incident reports as data sources for identifying market demand for safety technology. The effectiveness of utilizing incident reports could be further enhanced by integrating additional information, such as the frequency and severity of hazards, into the technology planning process.

Furthermore, despite the importance of linking demand-related and technology-related information, this connection has rarely been explored in the context of technology planning for safety management. To fill this gap in the literature, this study used incident reports and patent documents as demand- and technology-related data sources, respectively, and proposed analytical methods to capture the relationships between incidents and technologies. Additionally, the study highlighted how the analysis results and linkages between incidents and technology could be applied to technology planning in safety management, offering both reactive and preventive approaches to improving safety in industrial environments.

5.1.3. Technology Planning for Safety Management

While previous studies have primarily focused on the application of new technologies, such as sensors and image processing, fundamental analyses for technology planning have received less attention. The Technology Roadmapping (TRM) approach proposed in this study can assist companies in formulating more concrete technology plans for safety management. By leveraging the valuable data sources of incident and patent documents, the study examines the matching of information between incident factors and technology solutions to guide technology planning. The time analysis also provides a strategy for efficiently developing or applying technology. Through this process, gaps in preventive technology are identified by extracting incident clusters using Self-Organizing Maps (SOM).

The findings of this study offer useful insights for safety managers seeking to implement preventive technology at construction sites and R&D researchers focused on developing new safety technologies. For instance, safety managers can propose the implementation of existing technologies to reduce risks on construction sites, while researchers can gain inspiration for new technology or product development directly linked to incidents.

5.2. Limitations and Future Research

Although this study represents the first attempt to match incident and patent data in the field of safety management, the results of SOM clustering tend to be biased toward one of the document categories due to the differing uses of text in incident and patent documents. Aside from in Cluster I, it is difficult to find matching information between incidents and technologies, which stems from the varied sources and levels of keywords used in the incident and patent data. Incident documents are written in a more informal tone, while patent documents are structured with technical and specialized terms, making it challenging to identify linking keywords in matching nodes between the two types of data.

Additionally, while this study uses SOM for grouping the two types of data based on text information, other machine learning methods, such as Latent Dirichlet Allocation (LDA) and Hidden Markov Chain (HMM), could be considered for clustering and matching elements. As machine learning algorithms continue to evolve, different methods could provide alternative formats and results. By comparing various techniques, the reliability of the findings could be improved.

Furthermore, the issue of validation must be addressed to determine whether the technology solutions align with the incident cases in practice. While this study presents the similarity between incidents and patents, it does not guarantee that the technology identified as a match will offer an appropriate solution for addressing a real-world incident. Regulations can be a barrier to developing new technologies. In Korea, for example, safety certification is required for certain types of safety devices. For instance, a safety helmet with a camera is illegal because the additional device is not regulated in the safety certification specifications. Some patents may also be prohibited in Korea. Therefore, firms initially review the safety certification requirements, and if necessary, the government may amend regulations to facilitate the adoption of new and innovative safety technologies.

6. Conclusions

This research offers a new approach to utilizing incident reports, which have been continuously accumulating, and provides more accessible and understandable information to safety managers. By combining incident data with technology-oriented patent analysis, it expands the scope of research on technology opportunity discovery and technology forecasting, incorporating both the demand-driven factors from industrial incidents and the supply-driven characteristics of safety technologies.

A key practical contribution of this study is its ability to connect safety managers in the field with technology developers in R&D projects by aligning the causes of industrial incidents with the features of safety technologies. To effectively develop safety technologies, the needs must be identified through a comprehensive analysis of incident statistics. This research contributes to the data-driven extraction of safety technology needs using the narrative texts found in incident and patent documents. By matching information between these two types of data, four key areas for technology development and support in construction sites are identified, as follows: machine tool work, high-place work, vehicle-related work, and hydraulic machines. These four areas should be prioritized for safety technology innovation.

Ultimately, this research is expected to help safety managers identify preventive technologies, and offer valuable insights that will help technology developers to anticipate the future needs of safety management systems.