Construction accident prevention: A systematic review of machine learning approaches

3.2.1 Safety management and ML model goals

Work-related accidents in construction are responsible for thousands of deaths by year worldwide. The major causes of the events are generally related to falls (trips, slips, and falls from high elevations), struck-by, caught in-between, and electrocution [6, 18–21].

Operations involving machines and equipment generate thousands of struck-by and caught in-between accidents annually. Moreover, these operations are considered hazardous when executed on construction sites. In this context, researchers developed ML models to avoid collisions between workers and machines by detecting and monitoring entities and trajectory prediction [22–25].

Regarding the fall accidents, some studies tried to collaborate with the height-work safety by developing models of near-miss events recognition, some with the capacity to identify hazardous elements and vulnerable workers [20, 26]. Other research created tools to detect fall protection equipment (PPE, guardrail, scaffolding, and others), safety normative fall protection compliance checking, and detect workers’ unsafe actions [18, 28].

Furthermore, researchers developed solutions to identify the dependency relations between different variables that originate fall accidents [29, 30] and to predict the severity risks of falls [31] and the scaffolding fall risks [32]. Also, Bonifazi et al. [33] created a platform to identify fall occurrences and establish their causes and severity. The information input could activate warning signs and organize/enable rescue management plans. In addition, Sakhakarmi, Park, and Cho [34] proposed a model to classify scaffolding failure cases and predict safety conditions reliably based on strain data sets from scaffolding columns.

The necessity to guarantee adequate ergonomic conditions for workers in safety activities generated ML models to recognize and monitor corporal posture and workload information. Thus, some studies focused on unsafe actions recognition through human action modeling and classification systems using motion capture data [35–37]. In this respect, Yu et al. [38] developed an integrated model to provide detailed automated ergonomic evaluations of construction workers. The proposed method, combined with an ergonomic risk punctuation module based on the Rapid Entire Body Assessment – REBA risk score, identified the high-risk activities and the adequate rest schedule for each service or worker.

Additionally, Alwasel et al. [39] developed a study to classify work poses between groups of masons during standard concrete wall construction. Afterward, the investigation continued by Ryu et al. [40], which identified with high accuracy that masons more experienced adopt postures ergonomically safer and are more productive, indicating less energy use and lost movements.

Whereas an injury is rarely or never the result of one risk factor but a combination of them, understanding the causal factors involved in an accident is essential to offer guidance on effective accident prevention measures [41, 42]. Thus, by manipulating accident data to deduct and systematize significant variables, ML models can discover hidden patterns related to risks and injuries and rationally approach and predict some construction safety problems [43, 44].

In this regard, many researchers developed models to analyze construction data to classify and identify general factors that affect work-related injury incidence [6, 45–49]. For example, Fan [50] designed a tool to identify the relationships between defects and their occurrence probabilities, establish the hierarchical relation and determine the priority in safety management. Furthermore, Zhong et al. [51] conceived a model to identify implicit topics in construction safety-related textual description registers. The tool demonstrated hazard recurrence patterns automatically and reliably.

Other studies were more specific and developed simulation models to understand, for example, the relation between project safety status and the increase in rent search [52]. Additional research utilized data analysis models to classify construction accident reports [53–56], classify works by the project safety risk [57] and determine the construction site’s risk level [58].

Furthermore, forecasting is a significant issue for OHS because it can provide an opportunity to reveal problems before their occurrence [41]. Regarding this, some studies created tools to predict accident risks [10, 60], workers’ fatigue levels [61], incident outcomes [41, 62], fatal accident risks [63] and injury outcomes [9].

The workplace inspection is essential in safety management promotion to verify if the activities have safe and efficient execution. However, due to the dynamic and temporary character of the construction sites, they can be challenging to be monitored [23, 48]. Thus, there is a need to employ tools that automatically recognize and monitor entities in the workplace, such as workers, materials, and equipment/machines [20].

Regarding this, many studies developed recognition models that used computational techniques to recognize the correct use of PPE [64] and the colors of safety helmets [65–67], the detection and movement of workers and activities [11, 69], and sound detection of accidents/hazards [70]. Additionally, the need to reduce time, cost, and error probability in safety inspection reports promote the development of models to automate the compliance verification process [71] and classify near-misses information presented in safety reports [12].

Also, safety training increases knowledge and awareness of HS in the workplace [72]. Therefore, workers’ qualifications are fundamental to construction safety maintenance. In this context, Mariscal et al. [73] developed a Machine Learning model to understand the influence of employee training and information on the probability of accident rates. The results showed that a worker’s chance of suffering an accident could double if the information about the workplace risks is insufficient. In addition to this, it is necessary to identify if the safety culture is absorbed. Thus, Singh e Misra [74] developed a model to assess the workers‘ perceptions to address the issue.

3.2.2 ML methods and data features

According to Fan [50], the use of ML depends on the type of problems addressed, and method selection is closely related to the data collected. Therefore, this review intent to find the data features, such as types, approaches, and data collection methods, and the ML approaches/methods used in the research.

The analysis identified two learning approaches, the supervised and the unsupervised. The supervised approach uses algorithms to generalize knowledge from labeled data to predict other (unlabeled). In unsupervised learning, algorithms learn the underlying relationships from not labeled data and group cases with similar features [75]. In general, the studies analyzed used classification and regression models (supervised learning) and clustering, generalization, and association models (unsupervised learning).

Among the algorithms observed in the research, the SVM was the most recurrent (35%), presented in 26 studies. The classifier showed high performance or better results when compared to other algorithms in cases of classification of construction accident narratives [53], contractual clauses [71], and levels of risk perceived by workers [76], and also in activities recognition [77] and accident predictions [34]. Other algorithms were also recurrent in research, such as isolated or ensemble decision trees (Decision Tree – DT, Boosted Tree – BT, Random Forest – RF, Classification and Regression Tree – CART), other classifiers such as K-Nearest Kneighbor – KNN and Naive Bayes – NB, and deep learning models such as Convolutional Neural Network – CNN. Figure 6 shows the ML methods found in the review.

OPEN IN VIEWER

Fig. 6

ML methods of the publications.

Although the good results of isolated algorithms, the ML methods combination can obtain promising outcomes, especially related to the precision achieved. Moreover, complex problems and voluminous data require ML technologies combination for problem-solving [50]. In this regard, Ayhan and Tokdemir [62] used a clustering method with an ANN to create an accident prediction model, obtaining high performance.

Regarding the type of data, the record of an accident in the construction industry is usually in text format and reported through standardized questionnaires or free-text documents, which generates a large amount of text data with rich narratives of the events. Therefore, processing these texts can benefit the identification, analysis, and assessment of safety risks in the construction. This issue becomes evident in the literature due to the number of proposed methods for analyzing textual data [14].

The analysis found six types of data used in the models: textual, images, videos, mechanical, biomechanical, and mixed data. About half of the research (52%) used text data in ML models’ development, such as accident narratives [55, 63], questionnaires with the workers [59, 78], hazardous situations narratives [51], inspections narratives [50, 57], and others. Furthermore, 23% of the research used images of workers, materials, machines/equipment, activities, and hazardous situations [22, 79]. Some other studies used biomechanical data (14%) regarding workers’ physiological and motion data, mechanical data (3%) about structure strain data, videos (3%) about activities and safety situations records, and mixed data (5%) regarding more than one type of data.

Moreover, the analysis of the data approach followed the proactive and reactive safety approaches [13]. The reactive approach considered the data obtained after an unsafe event (accident data, injuries incidence data, number of lost workdays). On the other hand, the proactive approach considered the data obtained before an incident, used to prevent them from occurring (inspection reports, images/videos of hazardous situations). Thus, the review found that 60% of the publication used proactive data in models’ development, while 37% used a reactive approach and 3% used both.

The mechanical and biomechanical data are one type of proactive approach identified in the review. One example is seen in Park et al. [80] that developed a method to automatically control the position of hydraulic excavators through mechanical data from the machines and ANNs, which obtained satisfactory tracking results. Also, Lim et al. [81] created a model to identify slip and trip events at construction sites by extracting and analyzing workers’ energy variation data.

An important and often limiting factor for the success of the models is the amount of data used. For example, Sakhakarmi, Park, and Cho [34] showed that the accuracy rate considerably increased when there was an addition in the amount of data used to train the model. To address this problem, other authors focused on making databases available to assist in the algorithms’ training, such as Xiao and Kang [82], that created a database with 10,000 images of machines used in construction environments to train object detection algorithms.

Another relevant factor influencing the models’ results is the data collection method. Thus, the review analyzed the publication and found three groups: archival, research, experiment, and mixed. The findings show that around 45% of the studies used archival data for the models’ development, corresponding to data collected before the analysis in electronic databases such as governmental agencies’ reports. Also, 26% of the research used simulation data on construction sites and laboratories, while 22% used data from surveys usually collected in construction companies. Another 7% utilized more than one method of data collection.

Using primary data can be an option for many applications. However, it can cause inconvenience due to the errors in the collection that take time to standardize. Thus, secondary data may offer an advantage, especially if the data comes from reliable vehicles, such as government official agencies. Regarding this, many studies analyzed and predicted outcomes from historical cases of accidents obtained from official databases of Taiwan [46], the USA [6, 10], and China [9].

Although the significant results found by the ML models, some authors reported problems in the development, such as the lack of structured data and classification labels, incomplete data sets, and others. However, some solutions can contribute to more reliable results, such as uniformity of the data and the use of ontology tools in data pre-processing [44, 56].

One example observed is verified in Zhang et al. [55], where the authors used data mining techniques and NLP to assess construction accident reports. The method presented the benefit of not needing classified data, which reduced plenty of manual work. Also, the model is ideal for extracting patterns from a small number of documents with similar grammatical structures.

Other research also used ontological tools and NLP to analyze unstructured data. For example, Tixier et al. [1] created a tool to scan accident reports fast and automatically with high precision. Chin, Lin, and Hsieh [83] designed an ontological classification model for construction safety to collaborate with work-related risk assessment. In addition, to manage the incomplete datasets problem, Dutta et al. [84] developed a data fusion model.

Additionally, further factors can contribute to achieving higher accuracy in the models: use reactive and proactive data combined to consider accidents and incidents before an event; merge data types, such as images, videos, audios, and textual information; use a combination of ML methods in the models; and others.

Another aspect observed in the research that can pave future works is using technologies integrated into the models, such as Building Information Modeling – BIM, Big Data, and Internet of Things – IoT. These technologies can obtain and assess more data, including real-time data, to generate strong models [7, 54]. Using associated technologies can collaborate to more precise and detailed outcomes, predicting unsafe events and reducing the number of accidents on construction sites.

3.2.4 Limitations

The present study has some limitations, such as the year of publications, which considered only papers up to July 2021, excluding other recent relevant studies that could also be relevant to the research. Another limitation of this review was the language selected, studies in English, which could eliminate other suitable papers published in different languages. Although the search terms strategy adopted and the use of many synonyms, the chosen keywords were another factor that limited the review. Therefore, other relevant studies may have been out of this research. Also, the scientific databases selected and the fact that this review only selected papers published in journals may have been other limiting factors to this review.