Method for enhancing safety outcomes at a worksite

Abstract

A method to identify which safety leading indicators are not capturing hazards during safety inspections that later contribute to safety incidents and to identify blind spots in safety inspections. The method includes inputting a plurality of incident reports from the worksite and a plurality of inspection reports from the worksite into a system that uses machine learning (ML) and artificial intelligence (AI); identifying socio-technical factors influencing safety risks at the worksite using unsupervised clustering for topic modeling; evaluating semantic similarity between the plurality of incident report and the plurality of inspection reports conducted in a preceding time period to detect missing socio-technical factors not captured during safety inspections; and providing the identified socio-technical factors previously not captured during safety inspections to a stakeholder for development of modified inspection checklists tailored to capture the identified socio-technical factors previously not captured during safety inspection.

Description

METHOD FOR ENHANCING SAFETY OUTCOMES AT A WORKSITE RELATED APPLICATIONS

[0001] This application claims priority benefit of U.S. Provisional Application Serial Number 63 / 658,568 filed 11 Jume 2024, the contents of which are hereby incorporated by reference. FIELD OF THE INVENTION

[0002] The present invention relates to a method for enhancing safety outcomes at a worksite in an industry with a large number of workplace injuries or fatalities, including construction, oil and gas, railroads, and mining, and in particularly to a method for enhancing safety outcomes at a worksite that leverages advanced natural language processing (NLP) techniques to identify safety risk factors that are commonly missed under existing inspection protocols. BACKGROUND OF THE INVENTION

[0003] The construction industry has a large number of fatalities when compared to other sectors, such as oil and gas, as workers are constantly exposed to various safety risks due to changing site conditions (Edwin et al., 2021). In 2021, nearly one in five workplace fatalities occurred in the US construction industry (Bureau of Labor Statistics, 2023). In Canada, the construction industry accounted for up to 20% of workplace fatalities and around 10% of lost- time claims in 2021 (Association of Workers’ Compensation Boards of Canada, 2021).

[0004] Workplace safety inspections are a demonstrable effort of organizations to improve safety continuously (Alruqi & Hallowell, 2019). Reese (2011) reported that a lack of periodic safety inspections can increase the rate of accidents by 40%. According to Shrestha et al. (2011), the significant factors contributing to accidents in the construction industry are unsafesite conditions and are closely linked to inadequate inspection. Although safety inspections can prevent incidents, only a few studies have investigated the site safety inspection procedures in detail (Zhang et al., 2017). In addition, inspection results are rarely analyzed more than once to serve as performance indicators for administrative and management users or to reveal unsafe patterns on site (Lin et al., 2014). Tang et al. (2020) reported that inspection frequency is inadequate to identify and eliminate hazards promptly. Beyond frequency and timeliness, most audits / inspections fail to identify deficiencies that lead to a serious incident or fatality (Hutchinson et al., 2024). The quality of inspections must be examined to know if inspections are adequate and how these could be improved.

[0005] One way to enhance inspections is by identifying the leading indicators. Frequently, lagging indicators such as first aid rates, lost time, and total recordable incident rates (TRIR) are adopted to measure safety performance (Karimi et al., 2016). However, such indicators fail to fulfill the need to make long-term continuous improvements because they are retrospective and reactive (Cheung et al., 2020). In addition, recent research has shown that injury rates have severe statistical limitations and are invalid for nearly all practical purposes because injuries are numerically rare (e.g., 1 per 200,000) and are approximately 98% random unless it is derived from several hundred million work hours (Hallowell et al., 2021; Erkal et al., 2023). On the other hand, leading indicators such as safety training participation are proactive and reflect the actionable, current, and ongoing processes that identify hazards and evaluate opportunities for continual improvement (Zwetsloot et al., 2020). Yet, many leading indicators are only operationalized as presence / absence (e.g., training completed for work with heights, electrical exposures, etc.) or frequency (e.g., number of safety audits / 100,000 person-hours) (Salas & Hallowell, 2016). However, additional insights would be gleaned if these were operationalized as quality assurance (e.g., demonstrated competency) or quality control (e.g.,audited efficacy of controls). Validated safety leading quality assurance indicators are needed to help organizations prioritize their efforts and resources effectively (Poh et al., 2018).

[0006] Deep learning (DL) and natural language processing (NLP) have been widely used in text processing to extract and analyze semantic structures from text data. Methods such as sentence bidirectional encoder representations from transformers (SBERT) and N-grams have been used in safety analysis in aviation (Tikayat Ray et al., 2023) and healthcare (Evans et al., 2020). NLP can analyze multiple—word texts across prospective, real-time, and retrospective risk. In addition, given the advancement in methods to collect structured and unstructured inspection data using building information modeling (BIM) (Ma et al., 2018), text mining approaches are needed to uncover critical safety management issues within the construction industry (Lin et al., 2020).

[0007] Accordingly, in the construction industry, where safety is paramount, the frequency and severity of workplace incidents remain critical concerns. Therefore, site safety inspections have become essential in aiding safety management systems. While incident data is frequently used to identify gaps in the safety management system, inspection reports are rarely analyzed to identify unsafe patterns on site and reveal measures for safety enhancement. In addition, the safety performance is typically only operationalized as presence / absence or frequency rather than quality assurance.

[0008] Thus, there exists a need for a method to identify which safety leading indicators in industries with a large number of workplace injuries or fatalities are not capturing hazards during safety inspections that later contribute to safety incidents and to identify blind spots in safety inspections to enhance safety management systems holistically. SUMMARY OF THE INVENTION

[0009] The present invention provides a method to identify which safety leading indicators in industries with a large number of workplace injuries or fatalities are not capturing hazards during safety inspections that later contribute to safety incidents and to identify blind spots in safety inspections to enhance safety management systems holistically. The inventive method for enhancing safety outcomes at a worksite includes inputting a plurality of incident reports from the worksite and a plurality of inspection reports from the worksite into a system that uses machine learning (ML) and artificial intelligence (AI). The method then includes identifying socio-technical factors influencing safety risks at the worksite using unsupervised clustering for topic modeling and evaluating semantic similarity between the plurality of incident report and the plurality of inspection reports conducted in a preceding time period to detect missing socio-technical factors not captured during safety inspections. The method additionally includes providing the identified socio-technical factors previously not captured during safety inspections to a stakeholder for development of modified inspection checklists tailored to capture the identified socio-technical factors previously not captured during safety inspection. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] The present invention is further detailed with respect to the following figures that depict various aspects of the present invention;

[0011] FIG.1. Research methodology adopted in this study;

[0012] FIG.2. Linking incidents to previous months’ inspections;

[0013] FIG.3. Incident rates by class over the project time period;

[0014] FIG.4. Number of hazards identified in incident and inspection reports;

[0015] FIG.5. Sankey diagram showing the most and least captured indicators in incidents and inspections for the Class A cluster;

[0016] FIG.6. Categorization of Class A incidents;

[0017] FIG.7 shows a bowtie diagram for working at heights;

[0018] FIG.8. Sankey diagram showing the most and least captured indicators in incidents and inspections for the Class B cluster;

[0019] FIG.9. Sankey diagram showing the most and least captured indicators in incidents and inspections for the Class C cluster;

[0020] FIG.10. Four-gram analysis results of incident and inspection reports; and

[0021] FIG. 11. Co-occurrence network created by significant words in the incident description. DESCRIPTION OF THE INVENTION

[0022] The present invention has utility as a method for identifying which safety leading indicators in industries with a large number of workplace injuries or fatalities are not capturing hazards during safety inspections that later contribute to safety incidents and to identify blind spots in safety inspections to enhance safety management systems holistically. The present invention applies deep learning (DL) and natural language processing (NLP) techniques to identify which safety leading indicators are not capturing the hazards during safety inspections that later contribute to incidents. Using DL and NLP techniques, the inventive method identifies and allows visualization of the missing leading indicators in inspections by linking them with subsequent incidents. The present invention has applicability in any industry that has a large number of workplace injuries or fatalities including, but not limited to, construction, oil and gas, railroads, and mining.

[0023] After linking incidents to past inspections, the SBERT models (deep learning) generate semantically meaningful embeddings for incident and inspection descriptions. Embeddings are vectors (sequences of numbers arranged in a line or array), and each number within these vectors serves as a numerical representation of a word, crafted to simplify andenhance computer comprehension. Cosine similarity of incident and inspection embeddings (descriptions) are then evaluated to identify the least similar incident-inspection pairs using similarity threshold. A root cause analysis (five why method) is conducted on the least similar incident scores to understand why inspections fail to prevent incidents. Furthermore, a bowtie is used as a visualization technique to identify the different critical controls and is developed through focus group discussions of front-line supervisors and crew members. Bowtie analysis is a visual risk assessment tool that shows the various relationships between an event's causes, consequences, preventative, and mitigative controls.

[0024] N-gram analysis (natural language processing) is conducted to mine the most frequent topics in incident and inspection reports and validate the results obtained from SBERT analysis. N-gram refers to a set of N adjacent words. To further explain, a uni-gram breaks a sentence into one-word tokens. Similarly, a bigram is a token of two words, and a trigram is three words. In this research, four grams (four consecutive words) are used for topic modelling since they preserve the context and captured the core meaning of the sentences.

[0025] Co-occurrence networks (natural language processing) of incident reports are generated to identify hidden patterns and networks. A co-occurrence network represents the patterns in text data and is commonly used for crucial object extraction. Words in close association are grouped into sub-communities and marked with distinct colors, which helps identify the causal relationships between the words within each community.

[0026] By leveraging DL and NLP techniques, the present invention identifies the leading indicators that are not currently captured or adequately emphasized in safety inspections and later contribute to incidents. Sankey diagrams developed according to embodiments illustrate how the leading indicators are distributed in incident and inspection reports, showing the organization what indicators are being captured and, most importantly, which are not.

[0027] The present invention also offers insights into allocating proactive and effective resources during inspections to prevent incidents. By understanding which indicators predict potential incidents, strategic resource allocation based on data-informed insights can lead to an effective incident prevention strategy and enhance safety outcomes.

[0028] Embodiments of the present invention enhance an organization’s existing safety inspection and PowerBI system: by analyzing inspection and incident reports, the present invention identifies the ‘missing’ indicators that are overlooked during inspections. The results can be used to develop new inspection reports or checklists tailored to capture potential incidents proactively. In addition, by fusing data across systems of prospective data (i.e., HR / training, project planning, toolbox talks, field level risk assessments (FLHAs), job hazard assessments (JHAs)) and retrospective texts (incident investigations), the model can be used to identify additional leading indicators and controls, allowing for real-time hazard identification.

[0029] The present invention will now be described with reference to the following embodiments. As is apparent by these descriptions, this invention can be embodied in different forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art. For example, features illustrated with respect to one embodiment can be incorporated into other embodiments, and features illustrated with respect to a particular embodiment may be deleted from the embodiment. In addition, numerous variations and additions to the embodiments suggested herein will be apparent to those skilled in the art in light of the instant disclosure, which do not depart from the instant invention. Hence, the following specification is intended to illustrate some particular embodiments of the invention, and not to exhaustively specify all permutations, combinations, and variations thereof.

[0030] It is to be understood that in instances where a range of values are provided that the range is intended to encompass not only the end point values of the range but also intermediate values of the range as explicitly being included within the range and varying by the last significant figure of the range. By way of example, a recited range of from 1 to 4 is intended to include 1-2, 1-3, 2-4, 3-4, and 1-4.

[0031] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. The terminology used in the description of the invention herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0032] Unless indicated otherwise, explicitly or by context, the following terms are used herein as set forth below.

[0033] As used in the description of the invention and the appended claims, the singular forms “a,” “an” and “the” are intended to include the plural forms as well, unless the context clearly indicates otherwise.

[0034] Also as used herein, “and / or” refers to and encompasses any and all possible combinations of one or more of the associated listed items, as well as the lack of combinations when interpreted in the alternative (“or”).

[0035] According to embodiments, the present invention uses DL and NLP methods to link incidents with historical inspection reports and identify recurring patterns. SBERT and N-gram techniques are applied to determine which safety leading indicators are missing hazards during safety inspections in an injury prone project, such as a construction project. Using a human and organizational factors framework, a root cause analysis (RCA) is conducted on selected inspection-incidents pairings with low similarity scores to understand why inspections fail to prevent incidents.

[0036] Materials and methods

[0037] Data description

[0038] The initial dataset contains 821 incidents and 12,024 inspection reports from a construction project in Canada from 2015 to 2018. Each incident record includes the date and time of the incident, incident description, and incident classification (Table 1). Each inspection record has information about the date and time of inspection, inspection classification, inspection description, and corrective action target date, as shown in Table 1.

[0039] Table 1. Sample incident and inspection data. This construction company classifies incidents and inspections into three categories: A, B, and C. Class A incidents are severe with a potential threat to life, property, and environment; Class B and C refer to an event that can result in serious and minor injury, illness, or damage to property, respectively. The overall methodology adopted according to embodiments of the present invention is shown in FIG.1.

[0040] Data pre-processing

[0041] Data pre-processing is done on the incident and inspection descriptions since it has been found that text classification and the occurrences of noisy data in a dataset can decrease classification accuracy and result in poor prediction (Gupta & Gupta, 2019). Each dataset is checked for duplicates and missing columns; no duplicates and missing columns were identified in either of the datasets. Stop word removal, punctuation removal, lowercasing, and lemmatization are conducted using the natural language toolkit (NLTK) Python library. Stop words such as ‘a’, ‘the’ are removed. However, certain stop words — such as ‘not’, ‘didn’t’ or ‘wasn’t’ — are not removed to avoid misinterpretation of the incident or inspection description. For punctuation, all characters, such as periods, numbers, colons, and commas, are removed. Finally, lemmatization is conducted to reduce words down to their base forms or “lemmas”, as it simplifies the vocabulary and facilitates semantic work association (Madeira et al., 2021).

[0042] Linking incidents to previous inspections

[0043] After completing the data preprocessing steps, incidents are linked to their previous one- month’s inspections based on their hazard class. Therefore, three clusters are created, where class A, B, and C incidents are linked to past Class A, B, and C inspections, respectively. The reasoning for linking incidents to past inspections is that site conditions usually are improved after an incident. Therefore, an inspection before the occurrence of the incident is deemed representative of the site conditions (Poh et al., 2018). FIG. 2 illustrates a potential scenario where the first incident occurred on 9 January 2015. This incident is then linked to all the past one-month inspections from 9 December 2014 to 9 January 2015. A one-month timeline is chosen to exclude the effect of weather fluctuations as a variable. Some incidents could not be linked to inspections because the first incident is before the first inspection and vice versa. Therefore, overall, 731 incidents are linked to 10,627 inspection reports, and this dataset is used for further analysis.

[0044] SBERT and cosine similarity score generation

[0045] Sentence bidirectional encoder representations from transformers (SBERT) are used according to embodiments of the present invention to derive semantically meaningful sentence embeddings. SBERT uses bigram and trigram network structures, and when fed with a corpus of sentences, the relevant features are encoded into vectors of multi-dimensional space and semantically related words are mapped to nearby points (Reimers & Gurevych, 2019). SBERT is used in the present invention due to its high computational efficiency demonstrated in generating embeddings compared to single BERT (Lombardo et al., 2022). Pretrained SBERT models have been found to satisfactorily perform semantic textual similarity tasks without finetuning using the domain dataset (Reimers & Gurevych 2019). According to emdboiments, the incident and inspection descriptions are transformed into numerical embeddings using SBERT (model: Sentence Transformer model paraphrase- MiniLM-L3-v2), and the similarity between the descriptions is evaluated using cosine similarity.

[0046] A similarity threshold is identified to classify whether the incident and inspection descriptions are similar or not, using the methodology proposed by Orkphol and Yang (2019) since they achieved high accuracy, thereby demonstrating their method's effectiveness. In this approach, a random sample of incident and inspection pairs is chosen, and a binary score (1 if relevant or 0 if not relevant) is assigned to the pairs by a human without knowing the cosine similarity scores. Then, using a logistic regression model by input cosine similarity and binary scores as true labels, the model is fit by minimizing the difference between true and predicted labels. Using this trained model, all the incident-inspection pairs with their cosine similarity scores are used to predict the probability of relevance. For all three incident classes, the similarity score threshold is identified when the probability of relevance is at 0.5. Evaluation metrics are calculated to assess the performance of the model.

[0047] An RCA is conducted on the selected incidents below the threshold score to identify the leading indicators that missed capturing hazards. RCA is effective since the process isiterative and retrospective, therefore allowing continuous improvement (Kum & Sahin, 2015). To identify the indicators, accident analysis models such as human factors analysis and classification system (HFACS) and safety management system (SMS) have been widely used during RCA of incidents (Wong et al., 2016; Esmaeeli et al., 2022). HFACS is reliable because it has the advantages of diagnosis, reliability, and comprehensiveness, particularly in large- scale and complex accidents (Wiegmann & Shappell, 2003). SMS can provide a reliable systemic approach to identifying hazards in complex systems (Wu et al., 2020).

[0048] The qualitative analysis of incidents and inspections identified that the causes of incidents are related to certain elements of the HFACS (e.g., worker violation of procedures and rules) and SMS frameworks (e.g., the process of managing knowledge). Therefore, the present invention developed a human and organizational framework based on some elements from HFACS and SMS frameworks (Table 2). In the framework, human factors analyze the role of human actions, behaviors, and decisions that cause incidents. On the other hand, organizational factors focus on broader systemic and structural factors of organizations that can contribute to incidents.

[0049] Table 2. Human and organizational factors framework.

[0050] In this framework, human factors analyze the role of human actions, behaviors, and decisions that cause incidents. On the other hand, organizational factors focus on broader systemic and structural factors of organizations that can contribute to incidents. Human factors are categorized into ergonomics, communication, and the effect of weather conditions on human performance. The impact of weather conditions is considered a human factor since it influences humans' physical and cognitive capabilities, thus affecting their ability to perform tasks safely. Organizational factors are divided into three subcategories – equipment integrity, training, safe work procedures, and unsafe supervision. While unsafe supervision can be a human factor, it can also be a combination of managerial decisions, resource constraints, and organizational culture. Therefore, according to embodiments of the the present invention classifies this element as an organizational factor.

[0051] Bowtie and Sankey's diagrams are used to illustrate what factors are captured in incident and inspection reports. The bowtie analysis is a visual risk assessment tool that shows the various relationships between the causes and consequences of an event (Sigmann, 2018). Bowties are efficient since they provide a holistic view of any threats, consequences, and controls, facilitating risk conversations and organizational learning (Mulcahy et al., 2017; Baker et al., 2020). In this study, bowtie is developed through focus group discussions of front- line supervisors and crew members. Sankey diagrams are another visualization tool that is usedaccording to embodiments. The Sankey diagram shows flow indicators, and the link's width indicates the flow's magnitude and offers a clear visualization of information (Hernandez et al., 2018). Sankey diagrams are created using an open-source online tool SankeyMATIC (sankeymatic.com).

[0052] N-gram and co-occurrence network analysis

[0053] N-gram analysis is conducted to validate the results obtained from SBERT analysis. N-gram refers to a set of N-adjacent words (Bird et al., 2009). A uni-gram tokenization breaks a sentence into one-word tokens. Similarly, a bigram is a token of two words, and a trigram is three words. This research generated one gram, bigram, trigrams and four grams to mine the most frequent incident and inspection report topics. Unigrams are faster and more efficient since they are single words; however, using a higher value for N preserves more information and is expected to capture contextual information (Wang et al., 2012). On reading through the individual N-grams it is found that four grams are particularly effective since they preserved the context and captured the core meaning of the sentences. Therefore, four grams (four consecutive words) of the incident and inspection descriptions are used for further analysis. Furthermore, co-occurrence networks are generated to identify patterns and networks. A co- occurrence network represents the patterns in text data and is commonly used for crucial object extraction (Grames et al., 2019). Words that are in close association are grouped into sub- communities and marked with distinct colors, which helps identify the causal relationships between the words within each community (Ebrahimi et al., 2023).

[0054] According to embodiments, all the coding is done using the newest Python version, 3.11.3, in the Anaconda environment.

[0055] Results and discussion

[0056] Exploratory data analysis

[0057] An analysis of incident rates over the project timeline (2015-2018) is conducted to assess the occurrence of incident classes in the project, as shown in FIG.3. The incident rates are calculated by dividing the total number of incidents by the total number of work hours. It is observed that the rates for class A incidents have remained steady (0.26-0.61), indicating potential oversight in lowering high consequence low, probability incidents or severe injuries and fatalities (SIFs). Previous research has also reported that the reductions in rates of SIFs have remained relative constant, hence, raising concerns about the safety of workers (Ivensky, 2016; Bureau of Labor Statistics US, 2018; Cooper, 2019; Siever, 2022).

[0058] The next step of the analysis looks at the distribution of incidents and hazards in the dataset. There are 731 incidents, with 30, 127, and 593 class A, B, and C incidents, respectively. On the other hand, the dataset had 10,625 inspection reports with 808, 851, and 9422 Class A, B and C hazards, respectively, as shown in FIG. 4. It can be inferred from the figure that inspections capture Class C hazards (high-frequency, low consequence events) extensively which may explain why there are less Class C incidents occurring. However, Class A and B incident types (low- frequency, high-consequence events) are less identified in inspections, but also frequently occurring.

[0059] To overcome these issues, several clauses that can be included in inspections are developed in this study (Table 3). For example, Table 3 describes two Class A incidents not captured during inspections. The first incident represents a violation in working operations conducted outside the permit boundary. To better capture such hazards, inspections could include a clause that identifies if all the work being performed on-site has received appropriate permits and is conducted in proper locations.

[0060] Table 3. Example of Class A incidents that were not captured in inspections.Cla “O pla Ho it acti the “A wit iro pos ere to r ere wh via stru han to f Altworkers were cleaning the area to end their shifts close to the incident location.”

[0061] SBERT, root cause analysis, and Sankey diagram results

[0062] SBERT is used in this study to derive semantically meaningful sentence embeddings, and a cosine score is used to identify dissimilarities between incident and inspection descriptions. Logistic regression and probability relevance determined the similarity threshold for all classes. The model's accuracy for Class A, B and C clusters were 67%, 93%, and 68%, respectively. For the class A cluster, the total number of incidents and inspections below the similarity threshold score (0.322672) is 13 and 504, respectively. Different topics are mined from incident and inspection reports and are presented in FIG.5. The figure demonstrates that although 81% of the inspections focused on working at height, 100% of the incidents are related to this type of hazard. The other topics that are reported in the inspections are equipment handling / storage (15%), personal protective equipment (PPE) (2%), confined space (1%) and excavation (1%).

[0063] Categorization of Class A incident types is done to identify the most prevalent fault event (FIG.6). The figure shows that “Falling objects” incidents are the most frequent (55%), followed by incidents resulting from “Inadequate / absent fall protection” (36%). Note that the root causes of the incidents are unsafe work procedures (46%), violations (46%), and lack of training (8%).

[0064] Next, bowtie diagrams — a visualization technique – are used to illustrate a hazard, loss of control / containment scenarios, causes, consequences, and controls (de Ruijter & Guldenmund, 2016). FIG.7 shows a bowtie diagram for working at heights. The diagram has a central “knot” that represents the top event (working at height) and loss of control scenarios, where the left side outlines the potential threats (causes), and the right side lists the consequences. The preventative controls on the left side, when present and functional, stop the hazard from leading to a loss of control event, and the mitigative controls on the right side lower the severity of the potential consequence.

[0065] According to embodiments of the present invention, this and other bowties are developed through focus group discussions that facilitate active participation from 120 front- line workers, including supervisors and crew members who have firsthand experience and insights when working at heights. This participatory approach allows for the identification of practical controls applicable to workers' daily operations. In addition, this method allows the workers to visualize how different controls can interact to prevent incidents, making the abstract concept of risk management tangible. Furthermore, FIG. 7 clearly illustrates a profound emphasis on the left side of the bowtie (preventative controls). During the discussion, it is noticed that the participants found it more challenging to conceptualize the controls on the right side of the bowtie (mitigative controls), underscoring the critical need for a balanced focus. Although prioritizing preventative controls is intuitive, it is also essential thatattention is given to mitigative controls. This approach ensures that if an incident occurs, the system is designed to fail safely by relying on robust safety measures rather than “failing lucky”.

[0066] Working at heights is recognized as one of the major hazards in the construction industry (Jahangiri et al., 2019; Bussier & Chong, 2022). Several studies have reported that unsafe work procedures (Guo et al., 2020) and violations (Wong et al., 2016) are a significant cause of incidents when working at heights. Therefore, a multifaceted approach is required to better strategize against fall incidents. Hazard recognition of workers must be improved because poor hazard recognition is the proximal cause of risk misperception and injury in the construction industry (Fass et al., 2017; Albert et al., 2020). Periodic involvement of supervisors is essential during the development of job procedures and planning activities, such as job hazard assessments (JHAs) or field-level risk assessments (FLRAs), to identify hazards and reduce failing controls, such as unsafe work procedures. During the work planning, all crew members must be involved in risk assessment and problem-solving, ensuring that safety priorities are communicated to everyone on the team. In addition, defined supervisory activities are recommended to ensure work activities follow the accepted procedures and proactively prevent safety violations.

[0067] In the analysis of the Class A inspections, it is also observed that some corrective action target dates are assigned for a later date (see Table 1). Given the severity of Class A SIF- potential hazards, delaying corrective action target dates could exacerbate the site conditions and increase the likelihood of incidents. Ensuring effective corrective and timely measures are implemented lowers potential incidents. Immediate action addresses the problem and demonstrates a proactive approach to risk management.

[0068] For the class B cluster, the total number of incidents and inspections below the similarity threshold score (0.172728) is 39 and 352, respectively. The results of the SBERT analysis show that 17% of inspections focused on equipment handling / storage; however, 54%of the incidents are related to this indicator (FIG.8). In addition, 66% of inspections identified instances of working at heights, 41% of incidents occurred in this hazard category. Note that the causes of incidents in this cluster include – unsafe work procedures (76%), violations (14%), equipment integrity (4%), weather (3%), and lack of communication (3%).

[0069] To overcome these challenges, it is recommended that the company establish clear equipment handling and storage protocols. The documents should include instructions on safely using equipment (e.g., lifting and moving heavy machinery) and the appropriate storage conditions. Periodic preventive equipment maintenance is crucial to ensure that the equipment is working in safe conditions. One way to do this is by incorporating equipment management software (Indhu & Ajay, 2023) for efficient equipment maintenance that can mitigate site incidents. Such management software includes the entire lifecycle of equipment, such as repairs, transfers, or disposal, thereby closely tracking to achieve dynamic and closed-loop management of equipment (Chenjie et al., 2021). For the class C cluster, the total number of incidents and inspections below the similarity threshold score (0.4148) was 581 and 8825, respectively. The results show that 73% of the incidents were attributed to ergonomics. However, only 0.4% of inspections focused on this indicator (FIG. 9). Housekeeping is the predominant topic discussed during inspections, accounting for about 50% of the reports. Other issues of inspection include fall protection (11%) and PPE (7%).

[0070] Ergonomics is defined by Fernandez (1995) as the design of the workplace by considering the human’s physical, physiological, and psychological capabilities to optimize the effectiveness and productivity of work while also assuring the safety, health, and well-being of the workers. The construction industry is physically demanding, which includes tasks such as manual material handling and prolonged awkward and static postures (Parida et al., 2016) and long hours often outside that can cause psychological and physiological fatigue (Xing et al., 2020). Across all industries in 2018, work-related musculoskeletal disorders accounted for31.4% of all nonfatal occupational injuries and illnesses that resulted in work absenteeism in the US, with a median of 11 days away from work in the construction industry (Bureau of Labor Statistics. Musculoskeletal injuries are one of the most common occupational injuries and illnesses for workers, accounting for 24% of total injury claims across all sectors in Alberta, Canada (Labour and Immigration, 2022).

[0071] Boatman et al. (2015) conducted focused interviews with workers in the US construction industry. They reported that limited availability and accessibility of tested and practical tools to reduce physical demands, lack of awareness about ergonomics-related injuries, and lack of ergonomics assessment in the job hazard assessment were the significant obstacles that inhibited the implementation of ergonomics in the organization. Therefore, a recommendation is that organizations should ensure that ergonomics-related knowledge is shared among workers create better awareness and require that workers to use trolleys or conveyor belts for tasks such as material transportation to minimize physical strain on workers. Workers can also benefit from adequate training that recognizes the role of lifting weights and fatigue during static repetitive lifting tasks (Antwi-Afari et al., 2017).

[0072] N-gram and Co-occurrence network results

[0073] Four-gram analysis is also conducted on Class C incidents and inspection descriptions to mine topics that are frequently observed, and the results are presented in FIG.10. Housekeeping (68%), safe work procedure (13%), and equipment integrity (9%) are some topics that are most frequently discussed in inspections. The other column (4%) includes miscellaneous topics such as confined space, environment, and fire prevention. An example of a four-gram classified in housekeeping is a ‘cable lying ground walkway.’ However, in incident reports, ergonomics (48%), safe work procedures (24%), and equipment integrity (16%) are the most reported. This shows that specific topics are generally given more attention during inspections (i.e., housekeeping) because these are easier to see. It is recommended, therefore, that the organization allocate more resources to indicators such as ergonomics that are more prone to incidents yet currently under-identified during inspections. An example of a four-gram classified in ergonomics is ‘discomfort lower back repositioning’. These findings validate the results obtained from the SBERT and RCA (FIG.9).

[0074] For all incident classes, the most frequently occurring words (one gram) are plotted to visualize the semantic networks using co-occurrence networks (FIG.11). In a co-occurrence network, several nodes (words) create sub-communities and are represented in different colors. The number of connections determines the size of each node, and the co-occurrence between nodes is indicated by edges (Ebrahimi et al., 2023). The dashed lines demonstrated co- occurrence between the nodes of each sub-community. Different scenarios can be described by the sub-communities with the number of edges and nodes (Libis et al., 2019).

[0075] In sub-community 1 (pink color), the following incidents can be described: “worker” “felt” “discomfort” in the “left” “forearm” and “reported” to the “supervisor.” Nodes from sub- community 2 are connected to sub-community 3, and the incident can be described as “truck” “operator” in “cwa” “area” “made” “contact” “causing” an incident. Similarly, nodes from sub-community 1 are connected to sub-community 2, where the incident happened when a “worker” was “back” when they “noticed” “discomfort” and “reported” “first” in the “morning” and taken for “medical” “assessment”. Sub-community 1 is connected to sub- community 2 and 3 and can be described as an incident that occurred “at approximately” 9:00 while the “worker” was walking in his “work” “area,” he “felt” dust had flown into his “left” “eye.” The “worker” reported this case at “approximately” break “time” and the following “morning” “notified” his “foreman” who then in turn “notified” the general “foreman” superintendent, and “company” health and safety executive. The incident descriptions found in the dataset validate these scenarios generated by the co-occurrence networks (Table 4). These results validate the findings from the RCA analysis that human factors (ergonomics, equipment handling / storage) are critical contributors to incidents in the organization.

[0076] Table 4 A Sample of Incident Descriptions

[0077] Conclusion

[0078] The construction industry continues to have a high frequency of injuries and fatalities due to its dynamic nature. Consequently, it is critical to identify leading indicators in inspections and audits that enhance safety. The present invention leverages DL and NLP techniques to analyze a number of incidents (731) and inspection reports (10,625) to identify the missing safety leading indicators in inspections of an industrial construction project from 2015 to 2018 in Canada. Exploratory data analysis, such as incident rates, show that the rates for class A incidents have remained steady (0.26-0.61), indicating potential oversight in lowering SIFs. This suggests that existing measures may be inadequate, pointing to a need to reinforce leading indicators of critical controls to prevent and mitigate incidents.

[0079] The present invention uses DL and NLP to connect prospective risk management (inspections) with retrospective risk management (incidents) to determine what the organization could ‘foresee’.

[0080] The present invention also identifies how an organization is blind to specific hazards, which subsequently can lead to incidents. Bow-tie diagrams are used to visualize loss of control scenarios for SIF-potential hazards, causes, consequences, and controls. Sankey diagrams areused to visualize how the leading indicators are distributed in incident and inspection reports. Four- gram and co-occurrence network analysis is used to identify patterns in the incident descriptions, and the results validated the findings of RCA.

[0081] Such visualizations quickly show workers what precursory conditions are being captured and, more importantly, which are not. The analysis identifies that working at heights (100%), equipment handling / storage (54%), and ergonomics (73%) are the primary hazards identified in Class A, B, and C incidents, respectively. However, inspections missed these (81%, 17%, and 0.4%, respectively). By identifying these blind spots, the organization enhances its hazard identification, ensures critical controls are in place, and prevents future incidents. By implementing some strategies suggested in this study and using DL techniques, industries can improve their safety performance on sites by assuring the quality of inspections.

[0082] The incidents and inspection data used herein are obtained from the same project site, but are not linked the specific location of the incident and inspection report. Some embodiments include incident location since a construction site can be large, and if an incident and inspection occurred at different locations within the site, then it is challenging to deduce if the inspection effectively captured the hazard. While NLP is used to analyze incident and inspection reports, it can also be used to analyze real- time conversations like job briefings, toolbox talks, job hazard assessments, safety audits, and organizational culture. By comparing these real-time conversations with prospective (audits / inspections) and retrospective texts (incident investigations), organizations can identify blind spots and improve their real-time hazard identification and control.

[0083] Patent documents and publications mentioned in the specification are indicative of the levels of those skilled in the art to which the invention pertains. These documents and publications are incorporated herein by reference to the same extent as if each individual document or publication was specifically and individually incorporated herein by reference.

[0084] While at least one exemplary embodiment has been presented in the foregoing description and attached appendix, it should be appreciated that a vast number of variations exist. It should also be appreciated that the exemplary embodiment or exemplary embodiments are only examples, and are not intended to limit the scope, applicability, or configuration of the described embodiments in any way. Rather, the foregoing description and incorporated references will provide those skilled in the art with a convenient roadmap for implementing the exemplary embodiment or exemplary embodiments. It should be understood that various changes may be made in the function and arrangement of elements without departing from the scope as set forth in the appended claims and the legal equivalents thereof.

[0085] References

[0086] Albert, A., Hallowell, M. R., Kleiner, B., Chen, A., & Golparvar-Fard, M. (2014). Enhancing construction hazard recognition with high-fidelity augmented virtuality. Journal of Construction Engineering and Management, 140(7), 04014024. https: / / doi.org / https: / / doi.org / 10.1061 / (ASCE)CO.1943-7862.0000860

[0087] Albert, A., Pandit, B., & Patil, Y. (2020). Focus on the fatal four: Implications for construction hazard recognition. Safety Science, 128, 104774. https: / / doi.org / https: / / doi.org / 10.1016 / j.ssci.2020.104774

[0088] Alruqi, W. M., & Hallowell, M. R. (2019). Critical success factors for construction safety: Review and meta-analysis of safety leading indicators. Journal of Construction Engineering and Management, 145(3), 04019005. https: / / doi.org / https: / / doi.org / 10.1061 / (ASCE)CO.1943-7862.0001626

[0089] Antwi-Afari, M. F., Li, H., Edwards, D. J., Pärn, E. A., Seo, J., & Wong, A. (2017). Effects of different weights and lifting postures on balance control following repetitive lifting tasks in construction workers. International Journal of Building Pathology and Adaptation, 35(3), 247–263. https: / / doi.org / 10.1108 / IJBPA-05-2017-0025

[0090] Association of Workers’ Compensation Boards of Canada. (2021). National work injury, disease, and fatality statistics. https: / / awcbc.org / en / statistics / national-work-injury- disease-and-fatality-statistics-nwisp-year-at-a-glance /

[0091] Baker, K. E., Macciotta Pulisci, R., Hendry, M. T., & Lefsrud, L. M. (2020). Combining safety approaches to bring hazards into focus: An oil sands tailings case study. The Canadian Journal of Chemical Engineering, 98(11), 2330–2341. https: / / doi.org / 10.1002 / cjce.23813

[0092] Bird, S., Klein, E., & Loper, E. (2009). Natural language processing with Python: analyzing text with the natural language toolkit. O’Reilly Media, Inc.

[0093] Boatman, L., Chaplan, D., Teran, S., & Welch, L. S. (2015). Creating a climate for ergonomic changes in the construction industry. American Journal of Industrial Medicine, 58(8), 858–869. https: / / doi.org / 10.1002 / ajim.22499

[0094] Bureau of Labor Statistics US. (2018). Total injury and fatal injury rates in the US, 1992- 2018. https: / / www.bls.gov / iif / home.htm

[0095] Bureau of Labor Statistics US. (2019). Transportation and material moving workers experienced 184,470 injuries and illnesses in 2018. https: / / www.bls.gov / opub / ted / 2019 / transportation-and-material-moving-workers- experienced-184470-injuries-and-illnesses-in-2018.htm

[0096] Bureau of Labor Statistics. (2023). Construction deaths due to falls, slips, and trips increased 5.9 percent in 2021. https: / / www.bls.gov / opub / ted / 2023 / construction-deaths-due-to- falls-slips-and-trips-increased-5-9-percent-in-2021.htm

[0097] Bussier, M. J. P., & Chong, H. Y. (2022). Relationship between safety measures and human error in the construction industry: working at heights. International Journal of Occupational Safety and Ergonomics, 28(1), 162–173. https: / / doi.org / 10.1080 / 10803548.2020.1760559

[0098] Chenjie, Z., Gang, C., Lingchao, W., Zeyi, Y., Linhao, Z., & Jiajie, W. (2021). Construction of Equipment Management System Based on AI and IoT. 16th International Conference on Computer Science & Education (ICCSE), 197–201. https: / / doi.org / 10.1109 / ICCSE51940.2021.9569679.

[0099] Cheung, C., Xu, J., Manu, P., Ejohwomu, O., & Freitas, A. (2020). Implementing safety leading indicators in construction: Insights on the relative importance of indicators. CIB WO99 & TG59 Int. Web-Conf.2020: Good Health, Wellbeing & Decent Work, 26. de Ruijter, A., & Guldenmund, F. (2016). The bowtie method: A review. Safety science, 88, 211-218.

[0100] Ebrahimi, H., Sattari, F., Lefsrud, L., & Macciotta, R. (2023). A machine learning and data analytics approach for predicting evacuation and identifying contributing factors during hazardous materials incidents on railways. Safety Science, 164, 106180. https: / / doi.org / 10.1016 / j.jlp.2021.104517

[0101] Edwin, K. W., Nilsen, M., & Albrechtsen, E. (2021). Why Is the Construction Industry Killing More Workers Than the Offshore Petroleum Industry in Occupational Accidents? Sustainability, 13(14), 7592. https: / / doi.org / 10.3390 / su13147592

[0102] Erkal, E. D. O., Hallowell, M. R., & Bhandari, S. (2023). Formal evaluation of construction safety performance metrics and a case for a balanced approach. Journal of Safety Research, 85, 380–390. https: / / doi.org / 10.1016 / j.jsr.2023.04.005

[0103] Esmaeeli, N., Sattari, F., Lefsrud, L., & Macciotta, R. (2022). Critical analysis of train derailments in Canada through process safety techniques and insights into enhanced safety management systems. Transportation Research Record, 2676(4), 603–625. https: / / doi.org / 10.1177 / 03611981211062893

[0104] Evans, H. P., Anastasiou, A., Edwards, A., Hibbert, P., Makeham, M., Luz, S., Sheikh, A., Donaldson, L., & Carson-Stevens, A. (2020). Automated classification of primary care patient safety incident report content and severity using supervised machine learning (ML)approaches. Health Informatics Journal, 26(4), 3123–3139. https: / / doi.org / 10.1177 / 1460458219833102

[0105] Fernandez, J. E. (1995). Ergonomics in the workplace. Facilities, 13(4), 20–27. https: / / doi.org / 10.1108 / 02632779510083359

[0106] Gholamy, A., Kreinovich, V., & Kosheleva, O. (2018). Why 70 / 30 or 80 / 20 relation between training and testing sets: a pedagogical explanation.

[0107] Goh, Y. M., & Chua, D. (2013). Neural network analysis of construction safety management systems: a case study in Singapore. Construction Management and Economics, 31(5), 460–470. https: / / doi.org / 10.1080 / 01446193.2013.797095

[0108] Grames, E. M., Stillman, A. N., Tingley, M. W., & Elphick, C. S. (2019). An automated approach to identifying search terms for systematic reviews using keyword co‐ occurrence networks. Methods in Ecology and Evolution, 101(10), 1645–1654. https: / / doi.org / 10.1111 / 2041-210X.13268

[0109] Guo, S., Zhou, X., Tang, B., & Gong, P. (2020). Exploring the behavioral risk chains of accidents using complex network theory in the construction industry. Physica A: Statistical Mechanics and Its Applications, 560, 125012. https: / / doi.org / 10.1016 / j.physa.2020.125012

[0110] Gupta, S., & Gupta, A. (2019). Dealing with noise problem in machine learning data- sets: A systematic review. Procedia Computer Science, 161, 466–474. https: / / doi.org / 10.1016 / j.procs.2019.11.146

[0111] Hallowell, M., Quashne, M., Salas, R., MacLean, B., & Quinn, E. (2021). The statistical invalidity of TRIR as a measure of safety performance. Professional Safety, 66(04), 28– 30.

[0112] Hardy, C., Maguire, S., Power, M., & Tsoukas, H. (2020). Organizing risk: Organization and management theory for the risk society. Academy of Management Annals, 14(2), 1032– 1066. https: / / doi.org / 10.5465 / annals.2018.0110

[0113] Hernandez, A. G., Lupton, R. C., Williams, C., & Cullen, J. M. (2018). Control data, Sankey diagrams, and exergy: Assessing the resource efficiency of industrial plants. Applied Energy, 218, 232–245. https: / / doi.org / 10.1016 / j.apenergy.2018.02.181

[0114] Hinze, J., Thurman, S., & Wehle, A. (2013). Leading indicators of construction safety performance. Safety Science, 51(1), 23–28. https: / / doi.org / 10.1016 / j.ssci.2012.05.016

[0115] Hutchinson, B., Dekker, S., & Rae, A. (2024). Audit masquerade: how audits provide comfort rather than treatment for serious safety problems. Safety science, 169, 106348.

[0116] Indhu, B., & Ajay, J. (2023). Developing Management Information System for Construction Equipment Maintenance. International Conference on Civil Engineering Innovative Development in Engineering Advances, 563–573. https: / / doi.org / 10.1007 / 978-981- 99-6175-7_50

[0117] Jahangiri, M., Solukloei, H. R. J., & Kamalinia, M. (2019). A neuro-fuzzy risk prediction methodology for falling from a scaffold. Safety Science, 117, 88–99. https: / / doi.org / 10.1016 / j.ssci.2019.04.009

[0118] Karimi, H., Taylor, T. R., Goodrum, P. M., & Srinivasan, C. (2016). Quantitative analysis of the impact of craft worker availability on construction project safety performance. Construction Innovation, 16(3), 307–322. https: / / doi.org / 10.1108 / CI-10-2015-0050

[0119] Kum, S., & Sahin, B. (2015). A root cause analysis for Arctic Marine accidents from 1993 to 2011. Safety Science, 74, 206–220. https: / / doi.org / 10.1016 / j.ssci.2014.12.010 Labour and Immigration, G. of A. (2022). Workplace injury, illness and fatality statistics - Provincial summary 2020. https: / / open.alberta.ca / dataset / 4fb90b70-7660-4553-98fc- a59574f8fd5a / resource / 8d783536-2136-4666-8cef-636c26ada677 / download / lbr- workplace- injury-illness-and-fatality-statistics-provincial-summary-2020.pdf

[0120] Levine, D. I., Toffel, M. W., & Johnson, M. S. (2012). Randomized government safety inspections reduce worker injuries with no detectable job loss. Science, 336(6083), 907– 911. https: / / doi.org / 10.1126 / science.1215191

[0121] Libis, V., Antonovsky, N., Zhang, M., Shang, Z., Montiel, D., Maniko, J., Ternei, M. A., Calle, P. Y., Lemetre, C., Owen, J. G., & Brady, S. F. (2019). Uncovering the biosynthetic potential of rare metagenomic DNA using co-occurrence network analysis of targeted sequences. Nature Communications, 10(1), 3848. https: / / doi.org / 10.1038 / s41467- 019-11658-z

[0122] Lin, J. R., Hu, Z. Z., Li, J. L., & Chen, L. M. (2020). Understanding on-site inspection of construction projects based on keyword extraction and topic modeling. IEEE Access, 8, 198503–198517. https: / / doi.org / 10.1109 / ACCESS.2020.3035214

[0123] Lin, K. Y., Tsai, M. H., Gatti, U. C., Lin, J. J. C., Lee, C. H., & Kang, S. C. (2014). A user-centered information and communication technology (ICT) tool to improve safety inspections. Automation in Construction, 48, 53-63. https: / / doi.org / 10.1016 / j.autcon.2014.08.012

[0124] Lombardo, G., Tomaiuolo, M., Mordonini, M., Codeluppi, G., & Poggi, A. (2022). Mobility in unsupervised word embeddings for knowledge extraction—the scholars’ trajectories across research topics. Future Internet, 14(1), 25. https: / / doi.org / 10.3390 / fi14010025

[0125] Ma, Z., Cai, S., Mao, N., Yang, Q., Feng, J., & Wang, P. (2018). Construction quality management is based on a collaborative system using BIM and indoor positioning. Automation in Construction, 92, 35–45. https: / / doi.org / 10.1016 / j.autcon.2018.03.027

[0126] Madeira, T., Melício, R., Valério, D., & Santos, L. (2021). Machine learning and natural language processing for prediction of human factors in aviation incident reports. Aerospace, 8(2), 47. https: / / doi.org / 10.3390 / aerospace8020047

[0127] Mneymneh, B. E., Abbas, M., & Khoury, H. (2019). Vision-based framework for intelligent monitoring of hardhat wearing on construction sites. Journal of Computing in Civil Engineering, 33(2), 04018066. https: / / doi.org / 10.1061 / (ASCE)CP.1943-5487.0000813

[0128] Mulcahy, M. B., Boylan, C., Sigmann, S., & Stuart, R. (2017). Using bowtie methodology to support laboratory hazard identification, risk management, and incident analysis.Journal of Chemical Health & Safety, 24(3), 14–20. https: / / doi.org / 10.1016 / j.jchas.2016.10.003

[0129] Orkphol, K., & Yang, W. (2019). Word sense disambiguation using cosine similarity collaborates with Word2vec and WordNet. Future Internet, 11(5), 114.10.3390 / fi11050114

[0130] Parida, R., Sarkar, S., & Ray, P. K. (2016). Selection of alternate work systems to improve occupational health of Indian construction workers: A design of experiment-based approach. In Smart Technologies for Smart Nations: Perspectives from the Asia-Pacific Region (pp.155–173). https: / / doi.org / 10.1007 / 978-981-287-585-3_11

[0131] Poh, C. Q. X., Ubeynarayana, C. U., & Goh, Y. M. (2018). Safety leading indicators for construction sites: A machine learning approach. Automation in Construction, 93(October 2017), 375–386. https: / / doi.org / 10.1016 / j.autcon.2018.03.022

[0132] Reese, C. D. (2011). Accident / Incident Prevention Techniques (2nd Editio). CRC Press. https: / / doi.org / 10.1201 / b11390

[0133] Reimers, N., & Gurevych, I. (2019). Sentence-bert: Sentence embeddings using siamese bert- networks. ArXiv. https: / / doi.org / 10.2486 / indhealth.2018-0268 Free PMC article

[0134] Salas, R., & Hallowell, M. (2016). Predictive validity of safety leading indicators: Empirical assessment in the oil and gas sector. Journal of Construction Engineering and Management, 142(10), 04016052. https: / / doi.org / 10.1061 / (ASCE)CO.1943-7862.0001167

[0135] Short, J. C., McKenny, A. F., & Reid, S. . (2018). More Than Words? Computer- Aided Text Analysis in Organizational Behavior and Psychology Research. Annual Review ofOrganizational Psychology and Organizational Behavior, 5, 415–435. https: / / doi.org / 10.1146 / annurev-orgpsych-032117-104622

[0136] Shrestha, P. P., Yfantis, E. A., & Shrestha, K. (2011). Construction safety visualization. Proceedings of the 4th International Multi-Conference on Engineering and Technological Innovation (IMETI’11), 243–248.

[0137] Siever, K. (2022, June 3). Workplace injury-related deaths increased under Alberta NDP. Alberta Worker. https: / / albertaworker.ca / news / workplace-injury-related-deaths- increased-under-alberta-ndp /

[0138] Sigmann, S. B. (2018). Playing with fire: Chemical safety expertise required. Journal of Chemical Education, 95(10), 1736–1746. https: / / doi.org / 10.1021 / acs.jchemed.8b00152

[0139] Tang, S., Roberts, D., & Golparvar-Fard, M. (2020). Human-object interaction recognition for automatic construction site safety inspection. Automation in Construction, 120, 103356. https: / / doi.org / 10.1016 / j.autcon.2020.103356

[0140] Tikayat Ray, A., Bhat, A. P., White, R. T., Nguyen, V. M., Pinon Fischer, O. J., & Mavris, D. N. (2023). Examining the Potential of Generative Language Models for Aviation Safety Analysis: Case Study and Insights Using the Aviation Safety Reporting System (ASRS). Aerospace, 10(9), 770. https: / / doi.org / 10.3390 / aerospace10090770

[0141] Tompa, E., Kalcevich, C., Foley, M., McLeod, C., Hogg‐ Johnson, S., Cullen, K., MacEachen, E., Mahood, Q., & Irvin, E. (2016). A systematic literature review of the effectiveness of occupational health and safety regulatory enforcement. American Journal of Industrial Medicine, 59(11), 919–933. https: / / doi.org / 10.1002 / ajim.22605

[0142] Wanberg, J., Harper, C., Hallowell, M. R., & Rajendran, S. (2013). Relationship between construction safety and quality performance. Journal of Construction Engineeringand Management, 139(10), 04013003. https: / / doi.org / 10.1061 / (ASCE)CO.1943- 7862.0000732

[0143] Wang, W., Chen, L., Thirunarayan, K., & Sheth, A. P. (2012). Harnessing twitter" big data" for automatic emotion identification. 2012 International Conference on Privacy, Security, Risk and Trust and 2012 International Conference on Social Computing, 587– 592.

[0144] Weissbrodt, R., Arial, M., Graf, M., Iff, S., & Giauque, D. (2018). Preventing psychosocial risks at work: An evaluation study of labour inspectorate interventions. Safety Science, 110, 355–362. https: / / doi.org / 10.1016 / j.ssci.2018.08.024

[0145] Wiegmann, D. A., & Shappell, S. A. (2003). A Human Error Approach to Aviation Accident Analysis: The Human Factors Analysis and Classification System. Reference and Research Book News.

[0146] Wong, L., Wang, Y., Law, T., & Lo, C. T. (2016). Association of root causes in fatal fall- from-height construction accidents in Hong Kong. Journal of Construction Engineering and Management, 142(7), 04016018. https: / / doi.org / 10.1061 / (ASCE)CO.1943-7862.0001098

[0147] Woodcock, K. (2014). Model of safety inspection. Safety Science, 62, 145–156. https: / / doi.org / 10.1016 / j.ssci.2013.08.021

[0148] Wu, S., Ge, X., & Luo, Y. (2020). A New Model of Safety Management System for Railway Operation. 20207th International Conference on Information Science and Control Engineering (ICISCE), 1576–1584.

[0149] Wu, X., Yang, J., Ruan, H., Wang, S., Yang, Y., Naeem, I., Wang, L., Lui, L., & Wang, D. (2021). The diversity and co-occurrence network of soil bacterial and fungal communities and their implications for a new indicator of grassland degradation. Ecological Indicators, 129, 107989. https: / / doi.org / 10.1016 / j.ecolind.2021.107989

[0150] Xing, X., Zhong, B., Luo, H., Rose, T., Li, J., & Antwi-Afari, M.F. (2020). Effects of pysical fatigue on the induction of mental factigue of construction workers: A pilot studybased on europhysiological approach. Automation in Construction, 120, 103381. https: / / doi.org / 10.1016 / j.autcon.2020.103381

[0151] Zhang, H., Chi, S., Yang, J., Nepal, M., & Moon, S. (2017). Development of a safety inspection framework on construction sites using mobile computing. Journal of Management in Engineering, 33(3), 04016048. https: / / doi.org / 10.1061 / (ASCE)ME.1943-5479.0000495

[0152] Zwetsloot, G., Leka, S., Kines, P., & Jain, A. (2020). Vision zero: Developing proactive leading indicators for safety, health and wellbeing at work. Safety Science, 130, 104890. https: / / doi.org / 10.1016 / j.ssci.2020.104

Claims

CLAIMS 1. A method for enhancing safety outcomes at a worksite, the method comprising: inputting a plurality of incident reports from the worksite and a plurality of inspection reports from the worksite into a system that uses machine learning (ML) and artificial intelligence (AI); identifying socio-technical factors influencing safety risks at the worksite using unsupervised clustering for topic modeling; evaluating semantic similarity between the plurality of incident report and the plurality of inspection reports conducted in a preceding time period to detect missing socio-technical factors not captured during safety inspections; and providing the identified socio-technical factors previously not captured during safety inspections to a stakeholder for development of modified inspection checklists tailored to capture the identified socio-technical factors previously not captured during safety inspections.

2. The method of claim 1 wherein the unsupervised clustering for topic modeling comprises cleaning and processing textual data from the plurality of incident reports and the plurality of inspection reports to obtain preprocessed text.

3. The method of claim 2 wherein cleaning and processing textual data from the plurality of incident reports and the plurality of inspection reports includes removing stop words, stemming, and lemmatization to reduce words to their base or root form.

4. The method of claim 2 wherein the unsupervised clustering for topic modeling further comprises converting the preprocessed text into numerical representations to capture semanticmeanings and obtain vectorized data.

5. The method of claim 4 wherein converting the preprocessed text into numerical representations includes using term frequency-inverse document frequency.

6. The method of claim 4 wherein converting the preprocessed text into numerical representations includes using word embeddings.

7. The method of claim 2 wherein the unsupervised clustering for topic modeling further comprises applying unsupervised learning algorithms to the vectorized data to identify clusters of words representing topics within the vectorized data.

8. The method of claim 7 wherein the unsupervised learning algorithms include Latent Dirichlet Allocation (LDA) or Non-negative Matrix Factorization (NMF).

9. The method of any one of claims 1 to 8 wherein the unsupervised clustering for topic modeling further comprises Analyzing the topics to identify patterns and themes that represent socio-technical factors influencing safety risks at the worksite.

10. The method of any one of claims 1 to 8 wherein the preceding time period is 30 days.

11. The method of claim 1 wherein evaluating semantic similarity comprises converting text of the plurality of incident reports and the plurality of inspection reports into semanticrepresentations to capture a meaning and context of words and phrases from the plurality of incident reports and the plurality of inspection reports.

12. The method of claim 11 wherein converting text of the plurality of incident reports and the plurality of inspection reports into semantic representations is accomplished using natural language processing (NLP) techniques.

13. The method of claim 11 wherein evaluating semantic similarity further comprises quantifying similarity between each of the plurality of incident reports and a preceding inspection report of the plurality of inspection reports.

14. The method of claim 13 wherein quantifying similarity between each of the plurality of incident reports and the preceding inspection report comprises employing an algorithm on the semantic representations.

15. The method of claim 14 wherein the algorithm applied is a cosine similarity algorithm.

16. The method of claim 11 wherein evaluating semantic similarity further comprises selecting inspection reports that are least similar to subsequent incident reports to identify risk factors that were missed in a proceeding inspection.

17. The method of claim 16 further comprising analyzing the risk factors that were missed in a proceeding inspection to identify common oversight patterns.

18. The method of any one of claims 11 to 17 wherein the worksite is in an industry with a large number of workplace injuries or fatalities, including construction, oil and gas, railroads, and mining.

Images