A Predictive Analytics Framework for Mobile Crane Configuration Selection in Heavy Industrial Construction Projects

Abstract

Predictive analytics have been used to improve efficiency and productivity in the construction industry by leveraging the insights from historical data with a variety of applications in project management. In the planning process of heavy industrial construction projects, mobile crane selection plays a critical role in the project’s success, and poor choice of mobile crane configurations can lead to unnecessary cost-overrun and delayed schedules. In this research, the authors propose a predictive analytics framework for crane configuration selection using combined heuristic search and artificial neural network (ANN) approaches for heavy industrial construction projects. The heuristic search allows the practitioners to select the crane configurations based on engineering rules, while the ANN model utilizes the historical project data to help select crane configurations. The K-fold cross-validation is conducted to validate the designed ANN model and improve the accuracy of predictions. The results from the cross-validation test set have shown 70% accuracy.

1. Introduction

Cranes are widely used in the off-site construction industry to improve construction efficiency and facilitate the on-site assembly of prefabricated components [1]. Many cranes have been designed and produced by crane manufacturers as a result of the increase and diversification of the lifting needs of the construction industry. Crane selection for construction projects is a crucial and complex task. The construction operations’ time, cost, and safety can be significantly influenced by selecting the appropriate crane [2].

During the assembly process of prefabricated modules, heavy mobile cranes are used as the main lifting equipment. Due to the competitiveness of the construction industry, contractors need to analyze the capacity and capability of critical resources in order to reduce the cost and shorten the duration of construction [1]. A typical industrial site where mobile cranes are used to lift the completed modules is shown in Figure 1. The rental cost and crew fees of mobile cranes make them expensive to use. For example, for industrial projects in Alberta, Canada, a dual-crane lift of a single vessel (Demag CC1800 and Manitowoc 4100) costs CAD 320,000, including the rental cost, mobilization/demobilization, and ground preparation [3]. Therefore, cranes should be utilized efficiently to ensure that the projects are successful and productive and avoid budget overruns, schedule delays, and safety issues [3,4].

In practice, the lifting schedules of cranes are managed according to the demand, urgency, and priority of the tasks. Nevertheless, the capacities and locations of cranes can significantly restrict their coverage of service. Consequently, crane planning has a significant impact on construction sequencing, scheduling, budgeting, and safety. Therefore, a proper and feasible crane arrangement is a crucial prerequisite for the success of a construction project. In the past decades, the selection of cranes and their location on the job site is often made manually and largely dependent on the design of the building. This method, manual planning through trial and error based on the site shape, is inefficient and error-prone in today’s construction projects, where projects become more and more complex [1,5,6].

To select the most suitable mobile cranes, this research presented an artificial neural network model which predicts the ideal crane by considering the most relevant features of the modules and the projects. The model has been trained and tested with historical data from real industrial projects. This paper is structured as follows: Section 2 background and problem statement; Section 3 proposed methodology; Section 4 case study and validation; Section 5 research findings; Section 6 conclusion and future study.

2. Literature Review

2.1. Background

Construction sequence, scheduling, budgeting, and safety are notably impacted by crane planning, a challenging task due to the complex trade-offs between the involved parameters (e.g., crane size, lifting sequence, path planning). Previously published studies related to crane planning are mainly focused on the following areas: (1) crane selection; (2) optimum crane location; (3) crane path planning; (4) the usage and coordination of multiple cranes; and (5) simulation and visualization of crane operation [5,6,7,8,9].

As part of crane planning, crane selection consists of two main phases: crane type selection and crane model selection. The former is concerned with the type of the crane (e.g., lattice boom, hydraulic boom, fixed, mobile), and the latter is for determining the exact model of the crane [2]. The parameters that affect the crane type selection often result from the environmental, organizational, market, and industry conditions rather than decision-making [10]. Research shows that the primary factors affecting crane type selection can be categorized as follows: (1) type of use; (2) duration on the site; (3) construction height; (4) site spaciousness; (5) terrain topography; (6) soil stability; (7) construction aspect ratio; (8) crane relocation on the site; and (9) site accessibility. The second phase involves mainly selecting a crane that can fulfill the physical requirements of the lifts to be carried. Therefore, the factors that affect the model selection are summarized as: (1) building height; (2) load dimension; (3) the maximum load to be lifted; (4) the maximum radius at which a load has to be placed; and (5) the depth of placement [6,9].

In this regard, a fuzzy logic approach was proposed by Hanna (1999) to select the type of crane. In this method, the expert’s vague knowledge of the suitability of crane types in various project conditions is translated into fuzzy sets and fuzzy rules. A fuzzy inference engine can quantitatively identify the best crane type for a project based on the description of the project’s condition that an expert provided in words [11]. This system is not capable of performing the crane model selection for a specific task.

IntelliCRANES, which can assist with both crane type and model, was developed by Sawhney and Mund (2001). The system mainly includes two modules: (1) a neural network-based crane type selection module and (2) a knowledge-based expert system module for model selection. This system has some limitations, such as the number of cranes that can be selected as output is limited to only eight, and simultaneous crane selection for more than one crane for a construction site is not possible [2].

LOCRANE was developed by Warszawski (1990) [12] as a test case for applying the expert system methodology to construction planning tasks. It was limited to the selection of cranes for a given building. This knowledge base system cannot combine substitute transportation methods (e.g., equipment pumps, hoists, carts) with cranes for optimization and capturing interrelationships between crane selection and other construction tasks. CRANE ADVISER has been developed by Al-Hussein (1995) to integrate a knowledgebase and algorithm programs to assist in the crane selection for high-rise building projects. Expert knowledge, heuristics, and rules of thumb related to crane selection are contained in a knowledge-based module.

An expert system, entitled NEXPERT, was combined with geographical information system by Varghese (1992) to optimize the route selection when large objects are moving from the pick to set location to optimize the route selection. Wind speed, rental charges, lift radii, weights, and dimensions of the heaviest load have been considered in SELECTCRANE. This system was developed by Hanna (1994) to recommend the type of crane to the user. IntelliCRANES, which can assist with both crane type and model, was developed by Sawhney and Mund (2001). The system mainly includes two modules: (1) a neural network-based crane type selection module; and (2) a knowledge-based expert system module for model selection. This system has some limitations, such as the number of cranes that can be selected as output is limited to only eight, and simultaneous crane selection for more than one crane for a construction site is not possible [2]. Moselhi et al. (2004) developed a 3D modeling crane selection system that searches in its database for all technically feasible cranes according to the module’s dimension, weight, and location. The cranes retrieved from the databases can satisfy the specified clearance between the crane, the lift, and the adjacent buildings [13]. A 3D computer-aided application was developed by [14] to select mobile cranes for heavy lifts. This application can consider the lifting capacity, clearance between the boom or jib to equipment, and ground bearing pressure. Although this system can find all of the feasible cranes, it does not specify which one is the optimum solution, and the user needs to select the crane manually.

Another system developed by Sohn et al. (2014) [15] for tower crane selection considers all of the costs used in the economic analysis. It converts them into the net present value for an accurate comparison. The minimum cost solution, the lateral support structure, the foundation design, and the individual components required are the final output of this system. HELPS2 is a system that can assist in overall lift planning from the preliminary planning stage to the detailed planning stage, final evaluation, and selection. This system was developed by Hornady et al. (1992), and the outputs increase in detail as the lifting plan evolves. HELPS2 optimizes the lifting plan according to the objectives of cost, reliability, safety, and performance. However, it still depends on the user to choose step by step [16].

Moselhi et al. (2004) developed a 3D modeling crane selection system that searches in its database for all technically feasible cranes according to the module’s dimension, weight, and location. The cranes retrieved from the databases can satisfy the specified clearance between the crane, the lift, and the adjacent buildings [13]. A 3D computer-aided application was developed by [14] to select mobile cranes for heavy lifts. This application can consider the lifting capacity, clearance between the boom or jib to equipment, and ground bearing pressure. Although this system can find all of the feasible cranes, it does not specify which one is the optimum solution, and the user needs to select the crane manually.

A three-dimensional-based crane evaluation system to support the selection of crane and plan the crane lift schedule during the crane lift was developed in 2017. This system can design, verify, and simulate three-dimensional (3D) visualization of mobile crane operation. Moreover, this system can be utilized for better collaboration among project stakeholders [17]. An integrated decision support model was proposed by Han et al. (2018). For the purpose of determining the feasible crane type and model that will lead to the most efficient operation, the authors combined the conventional crane model selection methods with a 3D simulation and a crane selection matrix. Although the application of the proposed methodology is generic, the process is not fully automated and for each project the weights need to be assigned by the participants [18].

In summary, the common techniques that have been used for crane selection can fall into two categories: (1) 3D simulation and visualization and (2) statistic and machine learning. All of the methods or the systems that have been developed have some limitations regarding the input/output data or the criteria that were considered. Thus, the author aimed to improve the mobile crane model selection approach by simultaneously considering module dimensions, weights, and safety in this research. By doing so, the productivity of the system has increased. Moreover, two different approaches (i.e., a heuristic and an artificial neural network) have been combined to improve processing time and accuracy.

2.2. Crane Model Selection

The selection of a mobile crane’s model for industrial projects is not a straightforward process. Many projects exist in which both tower (fixed) and mobile cranes constitute practical solutions. There are, however, construction projects where site and job conditions mandate the model of the crane to be used [19]. Further, the factors affecting the selection of the crane model, such as lifting capacity, working radius, monthly rate, mobilization/demobilization cost, and safety, often do not have a linear relationship, and considering all of these factors is a complex process. Therefore, a heuristic algorithm has been developed to select the mobile crane for each project. At the same time, it considers the essential factors simultaneously (the details of this algorithm can be found in [20].

Although the heuristic approach considers all of the key factors, the processing time is the weakness of this method due to the non-linearity of the relationships and the great number of cranes and modules in the database. Consequently, the results of the heuristic approach, which were the most suitable mobile cranes for the previous industrial projects, were used to train an artificial neural network (ANN). The advantage of the ANN method is that it is trained and tested once, and it can be used for future projects for predicting the results; in other words, it does not need to be trained for each project, so the processing time has shortened significantly. The ANN approach and the implementation are explained in detail in the methodology sections. The methodology section is followed by a case study, in which the employment and performance of the model are illustrated. At the end, a brief conclusion and discussion for future works are presented.

2.3. Neural Network Background

Machine learning systems perform based on a simplified model of the biological neurons are called neural networks [21]. In the past few decades, the popularity and usefulness of neural networks for classification, clustering, pattern recognition, and prediction in many disciplines have increased [22]. The internal parameters of the neural networks can change by themselves in the same way as the biological neural network changes itself to perform some cognitive tasks (such as recognizing faces). The topology of a neural network is predefined and composed of a bunch of neurons. Depending on the task that the network has to learn, some possible topologies are kept constant. However, for some applications such as robotics, the topology itself can change dynamically. Therefore, the topology can be considered as a parameter. The type and intensity of the information exchanged are determined by the weights associated with the connection among neurons [23].

For the neural network model, ANNs are constructed from three types of layers, and each layer is composed of simply interconnected elements called nodes or processing elements that act as microprocessors Figure 2. The input and output layers of the neural network and their corresponding nodes are show in Figure 2. Non-linearity introduced to the model through connection weights and activation functions or thresholds within the nodes. The activation of the nodes results from the sum of the weighted inputs and can be negative, zero, or positive. The yellow and blue circles in the hidden layers represent the activation nodes and are usually noted as $\theta$. Each node is connected to every node from the next layer and each connection (arrows) has a particular weight. Each node’s impact on the node from the next layer can be seen as the weight.

To demonstrate how the weights and activation functions work, we consider the top yellow node, which has been connected to all of the previous nodes (green nodes). The values of all of the previous nodes (green) are multiplied by their associated weights and summed, resulted in the value of the top yellow node. The yellow node has a predefined activation function that defines if the node is “activated” or how “active” it will be, based on the summed value.

2.4. Goal and Objectives

This research aims to develop a neural network algorithm (multilayer perceptron neural network (MLP)) for mobile crane selection in heavy industrial construction projects to eliminate the limitations of the current mobile crane selection methods such as low accuracy, high processing time. The primary objectives of this research are:

• Develop an artificial neural network based on historical data to predict the optimum mobile crane configurations for industrial projects.
• Develop an interface to enable the user to interact with the developed applications.

3. Proposed Methodology

The proposed methodology consists of two components, namely the heuristic and ANN, which is based on a cloud-based crane database system that has been developed over years at PCL Industrial Management Inc., in collaboration with the academia. For those who are interested in the details of the developed crane management systems, please refer to previous research work (e.g., [24,25,26,27,28]). As a background, the crane database consists of two sets of data tables for crane-related and project-related information. The crane-related information includes: (1) the crane’s geometric information; (2) lifting capacity as specified by the crane’s manufacturer; and (3) rigging information; while the project-related information contains: (1) the lifting object physical information (e.g., dimensions, weights, center of gravities, etc.); (2) general project information (e.g., site layouts, boundaries, etc.); and (3) selected crane information (e.g., crane locations and configurations, etc.). The crane database interacts with various in-house developed decision support systems and visualization components to assist crane lifting management. In this research, the authors use the database as a data source to develop and implement the proposed methodology (as shown in Figure 3). It is important to note that there has been a challenge for crane selection that solely depends on historical datasets (e.g., companies that do not have sufficient project data), and in such case, heuristic approaches can be taken (stage 1 in Figure 3). During the stage 1, the company can use heuristic approach to obtain decision support information and select cranes for their tasks, and upon completion of the field tasks, the project information can be accumulated over time that can be used for stage 2, ANN network. At the stage 2 (i.e., the company has sufficient data), the heuristic approach is not efficient anymore (e.g., can be biased due to its ability to take the entire historical datasets into consideration). Therefore, ANN can help validate and improve the crane selection incorporate a large number of data points. This approach helps address the challenge with shortage of historical dataset for applications of ANN. Stage 1 and Stage 2 are elaborated in detail as below.

3.1. Heuristic Approach for Crane Selection (Stage 1)

Figure 4 shows the framework for the heuristic approach for crane selection (stage 1 in Figure 3). The framework includes inputs: (1) the crane and project information (mentioned earlier as the PCL database under the “Proposed Methodology” section); and (2) user inputs (e.g., minimum and maximum capacity requirements and minimum clearances for lifting). For more information and detailed explanation about this methodology, readers are encouraged to referred to this paper Azami et al. 2021.

3.2. Artificial Neural Network (ANN) Approach for Crane Selection (Stage 2)

The multilayer perceptron neural network has been selected in this research due to its high accuracy and low runtime compared to other models for multiclass classification. MPLs are feedforward networks trained with backpropagation, which is able to separate data into a specified number of output categories. The input layer of the MLP contains ten processing elements, which equals the number of the features of the modules such as weight, length, width, and height. Likewise, the number of processing elements in the output layer is equal to the number of the crane configurations in the dataset, which in this case were 58 processing elements. One hidden layer with 60 processing elements was selected for the first trial based on the Ward Systems Group (1996) advice, and then the model was modified by increasing the number of hidden layers and processing elements.

Microsoft Visual Studio 2019, Microsoft Access and Excel 2016, and Python 3.8 were utilized to create a user interface, connect a database to the model, and develop the neural network model. The model, which was developed using Python 3.8, reads the input data from the database that includes the training data Figure 5. Following the training step, the model needs to be tested to verify that the model can generalize everything that has been learned.

4. Case Study

A case study including a heavy industrial project was used to illustrate how the crane configurator works. Crane configuration selection was performed for four projects and validated with the historical data of PCL Industrial Management Inc. The results and details of the case study are described in the following paragraphs.

The total number of modules needed to be lifted was 630, weighing between 6 ton and 250 ton, and the momentum range falls between 1000 kN.m and 100,000 kN.m.

Table 1. Examples of the records in the dataset.

Weight (ton)	Height (m)	Length (m)	Width (m)	Elevation (m)	Gross Load (ton)	Moment (kN.m)	Set Radius (m)	Label
65	5	35	6	5	120	36,000	30	LR1400
20	15	6	4	10	46	6900	15	CK-2500
45	10	37	7	20	100	32,000	32	CC-2400
35	12	32	6	20	100	32,000	30	M-18000

shows some records that are included in the dataset.

The column ‘label’ in Table 1 and

Table 2. Examples of normalized data.

Weight	Height	Length	Width	Elevation	Gross Load	Moment	Set Radius	Label
1.28	−0.86	0.51	0.57	−0.39	1.02	1.17	0.43	LR1400
−1.10	1.70	−1.72	−1.73	−1.46	−1.65	−1.59	−1.72	CK−2500
0.14	−0.41	0.71	0.57	0.92	0.31	0.20	0.64	CC−2400
−0.32	−0.41	0.49	0.57	0.92	0.31	0.20	0.64	M−18000

shows the cranes’ models that have been selected for lifting modules with the features presented in the feature columns. The values in Table 1 are the raw data, and each column has a different characteristic, which is called a feature in data science. Since real-world data generally contain noise, missing values, and maybe unusable format, which cannot be directly used for machine learning models preprocessing helps enhance the quality of the data to promote the extraction of meaningful insights from the data.

Data preprocessing includes four main stages: data quality assessment, data cleaning, data transformation, and data reduction. In this research, after careful consideration based on the quality of the available dataset, data normalization was the only preprocessing that was needed to be carried out before feeding the model with the training dataset. Normalization helps to scale the data within a range to avoid building incorrect machine learning models and executing data analysis. Data normalization can be carried out through different techniques (one of which is called ‘standard score’) that has been implemented in this research. Standard score is a variation of another method called minimum-maximum that represents the number of standard deviations away from the mean.

In this study, a standard score has been implemented to normalize the dataset so that the value in each feature follows the standard deviation.

$${\mathrm{x}}^{\prime }=\frac{\mathrm{x}-\mathrm{u}}{\mathrm{s}}$$

(1)

$$\mathrm{u} = \frac{{{\displaystyle \Sigma }}_{\mathrm{i}=1}^{\mathrm{N}}{\mathrm{x}}_{\mathrm{i}}}{\mathrm{N}}$$

(2)

$$\mathrm{s}=\frac{\sqrt{{{\displaystyle \Sigma }}_{\mathrm{i}=1}^{\mathrm{N}}{{(\mathrm{x}}_{\mathrm{i}}-\mathrm{u})}^2}}{\mathrm{N}}$$

(3)

where ${\mathrm{x}}^{\prime }$ = the normalized feature, u = the average of values, ${\mathrm{x}}_{\mathrm{i}}$ = record in the dataset, N = the total number of the records in the dataset, s = the standard deviation of the features.

Validation Metrics

The cost function plays a key role in adjusting a neural network’s weights to create a better fitting machine learning model. Specifically, during forward propagation, the neural network runs on a training dataset and generates outputs. These results and the target labels are compared, and the loss function calculated a penalty for any deviation between the target label and the neural network’s outputs. During backpropagation, the partial derivative of the loss function is calculated for each trainable weight of the neural network, and the partial derivatives adjust the weights. Backpropagation iteratively adjusts the trainable weights of a neural network to produce a model with lower loss [29].

The loss function is a method of evaluating “how badly the algorithm predicts the results”. The loss function would output a higher number if the predictions were completely wrong. In contrast, if the model predicted correctly, it would output a lower number. The loss function helps to understand the learning rate of the model. It includes various approaches for different models, namely, regression loss function, binary classification loss function, and multi-class classification loss function.

In this research a cross-entropy loss function has been implemented, which is a subcategory of the multi-class classification loss function and a generalization of binary cross-entropy loss. In this approach, the model’s performance, whose output is a probability value between 0 and 1, is measured. As long as the predicted probability converges to the actual label, the cross-entropy decreases. The standard binary cross-entropy loss function is given by:

$${\mathrm{J}}_{\mathrm{bce}} = -\frac{1}{\mathrm{M}}\underset{\mathrm{m}=1}{\overset{\mathrm{M}}{{\displaystyle \Sigma }}}\left[{\mathrm{y}}_{\mathrm{m}}\times \log ({\mathrm{h}}_{\mathsf{\theta }}\left({\mathrm{x}}_{\mathrm{m}}\right)+\left(1-{\mathrm{y}}_{\mathrm{m}}\right)\times \log \left(1-{\mathrm{h}}_{\mathsf{\theta }}\left({\mathrm{x}}_{\mathrm{m}}\right)\right)\right]$$

(4)

where M = the number of training examples, ${\mathrm{y}}_{\mathrm{m}}$ is the target label for training example m, ${\mathrm{x}}_{\mathrm{m}}$ is the input for training example m, ${\mathrm{h}}_{\mathsf{\theta }}$ = the model with neural network weights $\mathsf{\theta }$.

Accuracy is another metric for evaluating classification models. It is defined as follows:

$$\mathrm{Accuracy} = \frac{\mathrm{Number} \mathrm{of} \mathrm{correct} \mathrm{prediction}}{\mathrm{Total} \mathrm{number} \mathrm{of} \mathrm{prediction}}$$

(5)

5. Research Findings

In this case study, a multilayer perceptron neural network (MLP) was implemented. The first step was to define the network parameters such as the number of the elements of the input layers, the number of the hidden layer and its elements, and the number of the output layer’s elements. The number of the input elements must be equal to the number of the features of the modules in the dataset, which in this case, there were ten different features. The number of the hidden layer and its elements were selected as 1 and 60, respectively.

Since the value of K for the K-fold has been set to five in this research, the dataset was divided into five portions. One portion consists of %20 of the dataset was selected randomly for the test, and four other portions consist of 80% of the remaining for training the model. Therefore, the following plots in Figure 6 show the model’s accuracy and loss for each fold of the dataset during the training. The number of iterations needs to be found by experiment, and it is a trade-off between the running time of the algorithm and the accuracy of the predicted results. Although the accuracy of the model may increase due to increasing the number of iterations, high variance or overfitting could happen, which could prevent the model from being able to generalize.

Having trained the model using the training dataset, the model must be tested by the test set so that the accuracy and generalization of the model on unseen data can be evaluated.

Table 3. Accuracy of the model including one hidden layer and 60 nodes tested with the available test set.

Fold	Accuracy of the Test Set (%)
1	61.90
2	72.20
3	68.25
4	62.98
5	68.00
Average accuracy	66.61

shows the test results of the model trained in Figure 6. Likewise, the training and testing results of a model with two hidden layers and 60 processing nodes in each hidden layer are shown in Figure 7 and

Table 4. Accuracy of the model including two hidden layer and 60 nodes tested with the available test set.

Fold	Accuracy of the Test Set (%)
1	58.73
2	65.87
3	68.25
4	70.63
5	72.00
Average accuracy	67.10

.

The same dataset was used to train and test other MLP models with one and two hidden layers and 600 and 1200 processing elements to better understand the model’s performance and find an optimum model. Evaluating these plots and tables show that by increasing the number of hidden layers and the processing nodes, the accuracy of the model on the training set and test set increases. However, the rate of increasing the accuracy, which increases either by increasing the hidden layer or the processing nodes, decreases gradually. Therefore, the optimum model needs to be found by trial and error.

The runtimes were measured over the entire training and testing process to assess the performance in providing a real-time solution and selecting the model with the highest accuracy and shortest runtime.

Table 5. The run time (s) of the proposed model for different numbers of hidden layers and processing nodes.

Number of Hidden Layers	Number of Processing Elements
	60	600	1200
	(s)	(s)	(s)
One	11.75	19.95	42.68
Two	15	72.98	317.87

shows the runtimes of different models with one and two hidden layers and different numbers of nodes.

Finally, the trade-off between the accuracy, runtime, number of hidden layers, and the number of processing nodes is illustrated by plotting them against each other. This comparison is noteworthy since, at some points, the number of hidden layers or processing elements increases while the accuracy decreases. Figure 8a shows that for a fixed number of processing elements (1200 nodes), by increasing the number of hidden layers, the accuracy first increases and then decreases. The model with three hidden layers has an accuracy of 75.36%, which was the highest accuracy. By increasing the number of hidden layers to four and five, the accuracy was kept decreasing. Figure 8b,c demonstrate that by increasing the number of hidden layers or the processing nodes, the runtime increases, which results in delays in receiving the outputs. Likewise, Figure 8d, which shows the model’s accuracy against the runtime, indicates that the accuracy was first increasing to a certain point, and after that, it starts to decrease. In this case, the number of nodes was kept constant, and the number of hidden layers was increased. In order to select the most productive model that provides high accuracy within the minimum runtime, such analysis that was presented here is necessary. In this case, the model with two hidden layers and 1200 nodes provided an accuracy of 74.9%, which is slightly lower than the model with three hidden layers, while the runtime for the model with two hidden layers is 317 s and the runtime for the model with three hidden layers is 476 s. In other words, the model with two hidden layers can provide almost the same accuracy within shorter runtime.

6. Conclusions and Future Work

Crane planning still has some limitations in terms of existing solutions and their efficiency. As part of a more extensive research study on crane planning, this paper proposed a machine learning technique incorporating an artificial neural network to classify and predict the most suitable mobile cranes regarding safety, cost, and scheduling for industrial projects. Various models were created and tested to find the one with higher efficiency; the model provides higher accuracy in a shorter time.

The proposed models were validated using test sets, which were selected randomly for each model. The selected model has two hidden layers in which 1200 processing nodes for each layer exist. The accuracy of this model was measured as 74.90%, which was a high accuracy compared to other proposed models. The low computation cost of these algorithms enables lift engineers and construction managers to make a decision faster with higher confidence regarding the results. Besides, a connection was made between the neural network model in this research and the application that was proposed by [20]. The purpose of this connection was to use the results of the heuristic approach for training and testing the neural network. Moreover, a distinguishing characteristic of this work lies in the fact that it can be used for any type of cranes, such as a tower or mobile cranes. The model can be adjusted and used with a new training set to train the model for other types of cranes. This model can be a fast and accurate method for industrial projects to select mobile cranes. However, it is always recommended to consider human judgement before making the final decision for crane selection. Moreover, this research can help other academic researchers to develop and implement machine learning techniques for automating construction processes.

Future work for this research is proposed in three main areas: to improve the developed model in this research, develop a 3D visualization for the crane selection algorithm, and begin the integration of this algorithm with previously developed algorithms for lift planning such as Automated Crane Planning and Optimization (ACPO) and Crane Motion Planning that have been designed for PCL Industrial Management to automate the process of checking the capacity and clearances of a crane [3,24,30]. (1) Model improvement: Even though this paper investigated automated mobile crane selection, other cranes still need to be appropriately studied, such as tower and derrick cranes. In addition, the model can be modified so that it can be used for project cost prediction; (2) 3D visualization: Developing a 3D visualization can help lift engineers better understand the crane’s coordination regarding its surrounding objects. This application can help to improve the clash detection process or increase the safety of the project; (3) future work needs to study the integration of the crane planning applications that have been developed in the past few years by PCL to develop a comprehensive lift planner which can plan for the entire process automatically.