1. Introduction
Studies have shown that most maritime accidents are caused by human and organizational factors (HOF) [
1,
2]. In addition, [
3] found through statistical analysis of the ship loss database that most of the accident causes were related to human factors, and technical factors only accounted for 26%. The International Maritime Organization (IMO) has created a regulatory framework for the safety of seafarers and ships by requiring a range of rules and activities to reduce human errors onboard [
4]. Human factors on ships are the key issue affecting navigation safety. Therefore, an in-depth discussion of the impact of human factors in the process of seafarers performing complex tasks will help ensure the safety of water transportation and personnel.
The mechanism of ship operation is complicated. It is not operated by one person alone, but by the entire crew working collaboratively to operate multiple modules of control conditions, usually including bridge control, communication, steering gear control, and other posts. Therefore, the captain’s judgment and decision-making will directly affect the performance of the entire ship driving team. For each seafarer, they often need to have information about the environment and the state of the ship in a short time to respond to complex task scenarios in a timely manner.
In addition to the driving operations, the ship’s working mechanism is also very particular. Ships must keep running day and night for navigation and production. Seafarers are often on duty during periods of time when rest is required, which can easily lead to heavy workloads. The high workload will affect the physiological and psychological conditions of the seafarer, causing fluctuations in their cognition, judgment, and responsiveness, which can lead to human error. If the high workload is not detected and dealt with in time, it will become a safety hazard to navigation. Therefore, studying the seafarers’ workload related to complex task scenarios will contribute to the basic research and development of the intelligent monitoring and early warning system for the seafarers’ work status.
Human intentions and implicit psychological traits can be expressed through signals generated spontaneously by the human body, such as electroencephalography (EEG), electrocardiograms (ECG), electromyography (EMG), and blood volume pressure (BVP), and cannot be deliberately concealed by these means. Currently, one of the more interesting areas of emerging research is affective computing [
5,
6]. Intelligent emotion classifiers detect emotional cues from a person’s responses and integrate emotional responses, ultimately predicting instantaneous emotional states [
7]. There have been studies on the use of machine learning to identify emotional states with EEG signals, mostly through algorithmic improvements to develop the accuracy of emotion recognition [
8,
9,
10,
11]. Since EEG comes directly from the brain, it is more objective and reliable in capturing the true emotional state of a human being, as well as stress and mental load. Similarly, the detection and integration of workload cues from the EEG and ECG signals of seafarers to discriminate workload states are one of the focuses of this paper.
Considering that the existing studies on seafarers’ workload are mainly based on subjective evaluation and lack of objective evaluation, the main research purposes of this paper are as follows: (1) explore seafarers’ workload variation under different complex tasks using EEG and ECG; (2) propose an EEG-based method for seafarer workload recognition, combining with ECG and task scenarios, to differentiate seafarers’ workload. The workload classification algorithm proposed in this study will solve the problem that current research mostly relies on subjective evaluation and lacks objective data to improve the accuracy of workload identification [
12,
13,
14]. Meanwhile, the portable dry electrode EEG headset used in the experiment has the advantages of high sensitivity and high resolution, and the wireless equipment does not affect the activity, which is suitable for the working environment of seafarers. It helps to achieve real-time monitoring of seafarers’ workload. It is a powerful tool for the intelligent monitoring and early warning system of seafarers’ work status.
The rest of the paper is organized as follows:
Section 2 introduces the literature review of the relevant studies.
Section 3 introduces the experimental setup and experimental procedure;
Section 4 presents the principles and methods of the workload recognition system for seafarers based on ECG and EEG signals;
Section 5 gives the workload classification results;
Section 6 presents the discussion.
6. Discussion
6.1. Seafarers’ Workload Variation in Task Scenarios
Both EEG and ECG are physiological indicators with strong individual differences. From the analysis of the individual physiological characteristics of the seafarers, when the tasks changed, the changes in the captain’s EEG correlate with EGC, showing significant changes at both key peaks of the complete experimental data. In the identification of multi-person workloads, the personalization of the ECG can offset the large variability of the EEG to a certain extent. Therefore, when seafarers’ EEG signals are recognized, the ECG features can be considered in the classification system to obtain good results. In addition, subjects’ physiological signal fluctuations were affected by complex navigation tasks. ECG features showed significant changes with increasing task difficulty, manifested as decreased RMSSD, HF, and increased LF/HF. The prominence of the above features in workload measurements all validate earlier studies [
25,
26,
27].
In the EEG analysis, the overall trend in the variation of EEG-based power ratio (θ + α)/β features for high and low workloads in the final dataset was generally consistent. However, specific values were still clearly differentiated. During the experiment when complex tasks with temporal stress dimensions such as fire alarms occurred, the subject requires high concentration and high alertness, with a greater increase in α and β activity. Additionally, θ activity was more pronounced when in a calmer state. They were the same as the findings of [
29,
46]. The extracted EEG features showed significant changes during the experiment. The results demonstrate that EEG signals can effectively cope with seafarers’ workload and stress and support the use of EEG signals to monitor seafarers’ brain status. The findings are consistent with Liu et al. suggesting [
33].
6.2. Comparison of the Two Recognition Models
Classified EEG data from 23 subjects with 75% accuracy. This is not a satisfactory result for algorithmic results of binary classification. The main reason is that sensitive EEG devices bring about strong individual differences while extracting subtle physiological fluctuations in subjects. Using priori task scenarios alone as workload calibration in algorithm structure is not restrictive enough for EEG classification. In general, there is a lack of an indicator that restricts individual differences in EEG. Therefore, this approach was abandoned, and an improved workload classification algorithm was subsequently developed to bridge this gap.
Compared to the first workload classification, the improved workload classification algorithm using task scenarios combining ECG and EEG showed better performance. The best classifier was Bagging, achieving an optimal AUC of 96% and recall of 96.7%. The workload classification improved accuracy by 26.0% of the first classification method. A comparison of the evaluation indicators between the two methods can be seen in
Table 7. This result validates the effectiveness of using ECG and EEG metrics to determine seafarer workload. The introduction of the individually discriminative ECG when classifying EEG according to the task scenarios offsets individual differences to a certain extent. That makes the classifier fulfill the need for multi-person workload classification and facilitates rapid recognition of the fitness for the duty of seafarers.
The classification results of the machine learning model proposed in this study validate previous research findings that comprehensive measures based on driver physiological data can better represent workload in complex maneuvers involving varying levels of effort [
16]. The study of [
35] can measure the mental load and stress experienced by seafarers during training based on EEG, and output indicative recommendations such as “pass”, “retrain”, or “fail”.
6.3. Implications and Limitations
The main contribution of this study is to propose the method for workload recognition of ship pilots based on task scenarios, ECG, and EEG. The method evaluated the performance of multiple machine learning algorithms and found that the Bagging algorithms were the classifiers with the best classification accuracy. It showed good applicability with a classification accuracy of up to 92.5%, which improved accuracy by 25.9% to model 1.
There are some implications of this study. The improved model of seafarer workload classification trained in this study (
Figure 9) can develop a multi-dimensional and multi-modal perception of on-the-job status, online intelligent monitoring, dynamic risk identification, and human–computer interaction active early warning technology and equipment. Specifically, it can be used to develop an intelligent monitoring and warning system for seafarers’ on-duty workload. This research will provide a theoretical basis for addressing the challenges of workload recognition and risk behavior modification for seafarers, and also risk early warning trigger mechanism design.
Limitations were also faced in this study. First, the experiment is carried out in a bridge simulator, which lacks the realism brought by the ship. Second, the experiment was conducted while the seafarers were in training, not during exams. There is no real-time evaluation process for the performance of each complex task.
Future studies are recommended to be carried out in the real ship navigation environment. This will further calibrate the environmental perception bias that the bridge simulator brings to the seafarer. Moreover, the experiment should form a complete experimental evaluation system combined with the performance of the pilot tasks.