From signals to control: how core technologies shape the evolutionary trajectory of Korean railway systems

The Fourth Industrial Revolution has accelerated technological innovation across industries, making it increasingly important for nations, research institutions and enterprises to enhance their technological competitiveness (Berawi, 2018). The railway industry, as a critical component of national infrastructure, is undergoing rapid transformation driven by advancements in automation, artificial intelligence, energy efficiency and smart transportation systems (Berawi, 2018). To maintain global competitiveness, railway-related enterprises and policymakers must anticipate emerging technologies and strategically incorporate them into long-term development plans.

One of the key strategies for achieving this is the ability to accurately forecast promising technologies and effectively integrate them into business and R&D strategies. The railway industry, characterized by its complex integration of mechanical, electrical and digital systems, heavily relies on continuous innovation to enhance safety, efficiency and sustainability. To strengthen technological competitiveness, companies and research institutions continuously develop and commercialize new technologies, securing intellectual property rights (IPR) through patent applications and registrations. Patents contain a wealth of structured and unstructured information that provides insights into technological coverage, innovation impact and competitive positioning, making them valuable resources for technology analytics and strategic decision-making in railway technology management (Lee et al., 2022; Salmi and Torkkeli, 2009).

However, as shown in our analysis, domestic patenting activity in the Korean railway sector, after peaking in the mid-2010s, has seen a decline in volume. This trend does not necessarily indicate a decline in innovation but rather suggests a maturation of industry and a strategic shift. The focus is moving away from foundational, high-volume patenting toward more complex, specialized and high-value innovations, particularly in digital domains like control systems, sensors and data-driven signaling. This transition makes strategic foresight more critical than ever, as the challenge shifts from broad-based development to identifying and investing in the specific, high-impact technologies that will define the next era of railway transportation.

Despite the growing importance of technology foresight in strategic management, traditional technology forecasting methods have primarily relied on expert-driven qualitative assessments (Nascimento et al., 2021). While expert opinions play a crucial role, qualitative evaluations are often influenced by subjective biases, lack clear evaluation criteria, and require significant time and financial resources. These limitations highlight the need for data-driven, quantitative approaches that can systematically identify and predict promising technologies with greater accuracy and objectivity.

To address these challenges, this study proposes a quantitative patent analytics framework that leverages structured patent data to identify promising technologies and support technology management strategies in the railway industry. Our contribution is not the invention of new analytical techniques but the creation of a novel, synergistic framework that integrates autoregressive integrated moving average (ARIMA) time-series forecasting and social network analysis (SNA). This hybrid methodology allows us to simultaneously evaluate a technology’s temporal growth potential (via ARIMA) and its structural importance within the broader innovation ecosystem (via SNA). By applying this approach to railway-related patent data from 1990 to 2023, this study aims to uncover emerging technological trends, assess their impact, and provide actionable insights for industry leaders and policymakers. The proposed framework enhances technology road mapping by reducing reliance on subjective expert opinions and providing an objective, data-driven methodology for technology forecasting in the railway sector.

Patent data has been widely used as a valuable resource for analyzing technological trends and forecasting promising innovations (Kim and Bae, 2017). Several studies have employed patent analytics to identify emerging technologies by examining country-specific, time-series and domain-specific trends (Garzaniti et al., 2021). These studies typically focus on extracting key technological insights from patent filings, analyzing patent citations and assessing patent intensity to determine technological impact (Alessandri, 2023). However, many existing studies primarily offer a descriptive analysis of patent trends rather than predictive analytics for technology foresight. Time-series forecasting models, particularly the autoregressive integrated moving average (ARIMA) model, have been used in various domains to predict future trends based on historical data (Devi et al., 2013). In the field of technology forecasting, ARIMA has been applied to analyze the temporal growth of patent filings, predict technology adoption rates and estimate future research trends (Lee et al., 2024). While more complex models like long short-term memory (LSTM) networks or Prophet exist, ARIMA remains a powerful and widely used tool for several reasons. For long-term patent data, which often exhibits clear linear trends and seasonality, ARIMA models are highly effective and computationally efficient. Their interpretability is a significant advantage in strategic contexts, as the model’s parameters (p, d, q) directly correspond to understandable concepts like autoregression, differencing and moving averages. In contrast, deep learning models like LSTM can be “black boxes,” and while powerful for complex, nonlinear patterns, they require vast amounts of data and may overfit on simpler time series. Prophet is excellent for series with multiple seasonalities but may not offer a significant advantage over a well-specified ARIMA model for annual patent count data.

Given our goal of creating a transparent and reproducible framework, ARIMA provides a robust and justifiable foundation for trend analysis. While more complex models like LSTM networks or Prophet exist, ARIMA remains a powerful and widely used tool for several reasons. For long-term patent data, which often exhibits clear linear trends and seasonality, ARIMA models are highly effective and computationally efficient. Their interpretability is a significant advantage in strategic contexts, as the model’s parameters (p, d, q) directly correspond to understandable concepts like autoregression, differencing, and moving averages. In contrast, deep learning models like LSTM can be “black boxes,” and while powerful for complex, non-linear patterns, they require vast amounts of data and may overfit on simpler time series. Prophet is excellent for series with multiple seasonalities but may not offer a significant advantage over a well-specified ARIMA model for annual patent count data. Given our goal of creating a transparent and reproducible framework, ARIMA provides a robust and justifiable foundation for trend analysis.

SNA has been increasingly adopted to study the structural relationships among technological concepts, research collaborations and innovation ecosystems (Xu and Yuan, 2024). Several studies have leveraged SNA techniques to analyze co-patenting networks, inventor collaboration networks and keyword co-occurrence networks to identify influential technologies and emerging research areas (Lee, 2024). Through centrality measures such as betweenness, degree and closeness centrality, SNA enables researchers to detect core technologies and assess their influence in a given domain (Kalantari et al., 2021). Common centrality measures include degree, betweenness and closeness, each offering a unique perspective. While other measures like Eigenvector Centrality or PageRank are useful for identifying influence within highly connected clusters, our chosen metrics are particularly suited for technology road mapping. Degree Centrality identifies popular technologies, Betweenness Centrality excels at finding “bridge” technologies that connect disparate fields (critical for identifying interdisciplinary innovation), and closeness centrality measures how quickly technology can influence others. This multifaceted view is essential for a comprehensive strategic assessment. Despite the advantages of network-based technology analysis, existing studies often focus on static network structures rather than dynamic technological evolution over time (Cao et al., 2024).

While existing research has explored patent analytics, SNA and ARIMA-based forecasting separately, there is a lack of studies that integrate these methodologies for a more holistic approach to technology road mapping (Kwon et al., 2022). The combination of time-series forecasting and network analysis can provide a more systematic and data-driven framework for identifying promising technologies, overcoming the limitations of qualitative expert-driven assessments (Maçaira et al., 2018). This study addresses this research gap by using a hybrid methodology that integrates patent analytics, SNA and ARIMA to forecast emerging technologies in Korea’s railway industry. This approach enables policymakers and industry stakeholders to make informed strategic decisions based on objective, data-driven insights rather than relying solely on expert opinions. By bridging these research gaps, this study contributes to the advancement of technology forecasting methodologies and provides a novel framework for data-driven decision-making in the railway sector.

Our research methodology is a multi-stage process designed to systematically extract, analyze and forecast technological trends from patent data. The entire workflow is depicted in Figure 1.

To facilitate a structured analysis of patent data, this study collected 4,352 railway-related patents filed between 1990 and 2023 from the Korean Intellectual Property Office (KIPO) database. The preprocessing workflow, implemented in R, began with data cleaning and tokenization, where patent titles and abstracts were parsed into individual words. Following this, stop word removal was conducted using both standard and domain-specific stop word lists to filter out non-informative words. To identify the most significant technology-related terms, we calculated the term frequency-inverse document frequency (TF-IDF) score for each term. TF-IDF prioritizes words that are frequent in a specific document but rare across the entire corpus, making it effective for identifying characteristic keywords. The top 20 keywords with the highest average TF-IDF scores were selected for analysis. These keywords were then structured into a document-term matrix (DTM), which forms the basis for the subsequent analyses.

Once the technology keywords are structured within the DTM, the next step involves analyzing their temporal significance. ARIMA modeling was applied to forecast future occurrences of each keyword based on historical trends (Carta et al., 2018). An ARIMA model is defined by three parameters: (p, d, q), where “p” represents the order of the autoregressive (AR) part, “d” is the degree of differencing, and “q” is the order of the moving average (MA) part. The model specification process was conducted rigorously for each keyword’s time series. Stationarity was first checked using the Augmented Dickey–Fuller (ADF) test, and nonstationary series were differenced (d = 1, 2,…) until stationarity was achieved. Initial estimates for the p and q parameters were obtained by examining the autocorrelation function (ACF) and partial autocorrelation function (PACF) plots. Multiple candidate ARIMA (p,d,q) models were then fitted, and the optimal model was selected based on the lowest akaike information criterion (AIC) and Bayesian information criterion (BIC) values, ensuring a good balance between model fit and complexity. Based on the selected ARIMA models, forecasts were generated along with 80 and 95% confidence intervals. Keywords were then classified into three groups according to their projected trends (Figure 2): “Hot Fields,” where the lower bound of the 95% confidence interval indicates positive growth; “Active Fields,” where forecasts are stable or growing but do not meet the “Hot” criteria; and “Cold Fields,” where the upper bound of the 95% confidence interval shows a declining trend.

To further refine the analysis, SNA is used to examine the relationships among keywords. A co-occurrence network is constructed based on how frequently keywords appear together in patents, revealing the structural significance of each term. A weighted, undirected co-occurrence network is constructed, where nodes are the top 20 keywords and the weight of an edge between two nodes represents their co-occurrence frequency within the same patent abstracts. To reduce noise, only edges with a co-occurrence frequency above a certain threshold (top 20% of all edge weights) were included in the final visualization. The core metric used in this study is Betweenness Centrality, which measures how often a keyword serves as a bridge between other keywords in the network (Zhang and Luo, 2017). A high Betweenness Centrality score indicates that a keyword plays a pivotal role in connecting different technological domains, suggesting its importance in fostering interdisciplinary advancements. Betweenness Centrality (⁠$CB(v)$⁠)for a node $v$ is calculated as equation (1). In addition, Degree Centrality (⁠$DI$⁠), Closeness Centrality (⁠$Cc$⁠), are calculated to assess the importance and influence of specific technologies within the network in equations (2) and (3):

By justifying our choice of these specific metrics, we gain a multifaceted understanding of each technology’s role: Degree for popularity, Betweenness for bridging/interdisciplinary importance, and Closeness for diffusion potential. This integrated methodology provides a robust foundation for strategic technology road mapping.

To analyze technological trends in the railway industry, we collected 4,352 railway-related patents from KIPO from 1990 to 2023. Figure 3 illustrates the annual frequency of patent applications. The number of patents increased rapidly from the 1990s, peaking in the mid-2010s, before showing a gradual decline. This trend suggests a maturation of foundational technologies and a potential shift toward more specialized innovations. Table 1 presents the top five patent applicants, who collectively account for approximately 38% of all patent filings. The Korea Railroad Research Institute led with 818 patents (18.8%), followed by Hyundai-Rotem Co. (466 patents, 10.7%), Daewoo Heavy Industries (205 patents, 4.7%), Korea Railroad Corporation (104 patents, 2.4%) and Hyundai Mobis (65 patents, 1.5%).

Using TF-IDF analysis on the patent corpus, we identified the top 20 most significant technology keywords. Figure 4 provides a bar chart of their frequencies, offering a clearer quantitative comparison than a word cloud. The analysis shows that foundational terms like “structure” and “rail” are highly frequent, but emerging digital technologies like “control” and “signal” also feature prominently. We then conducted ARIMA-based time-series forecasting for each keyword. Figure 5 shows the historical trends and future forecasts for three representative keywords: “signal” (Hot), “track” (Active) and “test” (Cold). Crucially, the plots now include 80 and 95% confidence intervals, quantifying the forecast uncertainty.

Based on these forecasts, we classified the keywords as shown in Table 2. The criteria are now explicitly defined based on the confidence intervals of the forecasts. Table 3 presents the specific ARIMA model parameters for the representative keywords, ensuring reproducibility. To understand the structural importance of these keywords, we performed SNA. Table 4 summarizes the normalized centrality scores. Keywords like “device”, “control” and “sensor” show high scores across all three metrics, indicating they are not only popular but also serve as crucial bridges in the technology network. Figure 6 visualizes this network, filtered to show only the strongest connections, making the core structure interpretable and avoiding the “hairball” effect.

The integration of ARIMA forecasting and SNA enables a systematic approach to identifying and managing promising railway technologies. Figure 7, the technology classification matrix, synthesizes our findings by plotting each keyword’s network influence against its growth potential. This provides a clear, data-driven basis for prioritization. The priority keywords—device, circuit, speed, sensor and signal – emerged as central elements shaping the industry’s technological trajectory. A comprehensive analysis of patent activity reveals that these key technologies exhibit high influence within the patent network and demonstrate strong growth patterns in time-series forecasting. As illustrated in Table 5, these keywords rank highest in both network influence and temporal expansion, making them prime candidates for strategic R&D investments.

The integrated analysis conclusively identifies a definitive cluster of high-priority technologies driving innovation in Korea’s railway sector. Keywords such as “control,” “signal,” “sensor,” “device” and “speed” consistently ranked highest in both forecasting and network metrics. Their classification as “Hot” fields is supported by ARIMA models showing strong positive growth trends, with lower confidence intervals remaining above zero. Simultaneously, their high normalized scores in degree, betweenness and closeness centrality confirm they are not only prevalent but also serve as critical hubs and bridges within the technology co-occurrence network. This dual validation indicates that the industry’s trajectory is firmly oriented toward cyber-physical systems, where digital control, real-time communication, and sensing are paramount. This finding signals a clear strategic shift from the historical focus on heavy engineering and structural components (“rail”, “structure”) toward intelligent system integration. The data suggests that future competitiveness and resilience will be built on the seamless integration of these core digital technologies. For instance, the centrality of “sensor” and “signal” underscores the importance of data acquisition and transmission as the foundational layer for automation, while the prominence of “control” and “device” highlights the subsequent layer of processing and actuation. This interconnected cluster forms the essential technological backbone for developing autonomous trains, predictive maintenance, and integrated traffic management systems.

This study validates the proposed ARIMA-SNA integrated framework as a powerful tool for objective technology foresight. The framework successfully overcomes the subjectivity of expert-only approaches by providing quantitative, reproducible metrics for assessment. The synergy of the two methods is crucial: ARIMA forecasting alone might highlight momentarily trendy but isolated technologies, while SNA alone might identify historically important but stagnant ones. Their combination, especially using the forecast confidence interval as a classification threshold and multi-faceted centrality for influence mapping, effectively filters and prioritizes technologies that are both growing and structurally significant. The practical implication for R&D managers and policymakers is a data-driven prioritization matrix. Resources should be strategically funneled into the high-growth, high-influence quadrant populated by the identified core cluster. This approach minimizes investment risk in fading technologies and maximizes the impact of funding by strengthening interconnected technological capabilities. The framework offers a replicable model for periodic monitoring, allowing organizations to dynamically update their roadmaps as new patent data emerges, thereby maintaining strategic agility in a rapidly evolving industrial landscape in Table 6.

The interpretation of our results is inherently framed by the domestic scope of the KIPO patent data. The identified core cluster accurately reflects Korea’s strategic emphasis on digitalization and automation within its national innovation system. This focus likely responds to domestic pressures such as labor demographics, urban congestion, and government-led “smart city” initiatives. The decline in overall patent volume after the mid-2010s, juxtaposed with the rise of digital keywords, can thus be interpreted as a strategic consolidation – a transition from broad-based invention to targeted, deep-tech innovation in specific high-value domains. However, this national lens constitutes the study’s primary limitation. A siloed view risks overlooking convergent or disruptive technologies being developed in other leading railway markets like the EU (focusing on hyperloop and cross-border interoperability) or Japan (advancing ultra-high-speed maglev). Our “Hot” technologies might be global standards, or they could be local specializations with limited export potential. Consequently, the derived roadmap is primarily prescriptive for domestic strategic alignment but requires external validation for global competitiveness. The framework itself is robust, but the input data defines its perspective.

To transcend the national limitation, the foremost research direction is a comparative international analysis. Applying the identical ARIMA-SNA framework to patents from the USPTO, EPO and JPO would allow for a benchmarked, global technology landscape map. This would answer critical questions: Is Korea a leader or follower in the “control-sensor-signal” cluster? Are there emerging technology nuclei abroad that are absent domestically? Such an analysis would transform the roadmap from a national strategy into a tool for global positioning and opportunity identification. Further research should also focus on temporal network dynamics. This study used a static network aggregated over decades. Future work could construct dynamic, time-sliced networks to observe how the centrality of keywords like “sensor” has evolved and how the connections between “control” and “signal” have strengthened in Table 7. This would add a crucial evolutionary perspective, reveal the pathways of technology convergence and potentially forecast the emergence of new super-connectors. Integrating these dynamic insights with forecast models could lead to a truly predictive framework for technological convergence.

No funding has been received.