Development of Artificial Neural Networks Model to Determine Labor Rest Period Based on Environmental Ergonomics

Food SMEs (Small and Medium Enterprises) were examples of labor-intensive industry, which involved laborers in pursuing production activities. Food SMEs require complex processes in production activities. Support to increase work productivity and reduce ergonomic risks of the activities was needed. The study was conducted at Tofu SMEs. The determination of the rest period could be developed to give some recovery times to laborers. WBGT (Wet Bulb Globe Temperature) was estimated to determine the rest period. The rest period was determined by the workstation environment and workload labor. ANN (Artificial Neural Networks) model was carried out due to a nonlinear relationship. ANN was used to process the information from the data set and predict the amount of rest period and WBGT. ANN was trained using backpropagation. The backpropagation algorithm used the error value to change the weight with forward and backward propagation. The result showed that dry bulb temperature, heart rate, wet bulb temperature, and gender significantly impacted the rest period and WBGT. A total of 180 data sets from tofu SMEs were divided into training data (80%) and validation data (20%). The optimal ANN structure was determined by four input, four hidden, and two output neurons. The activation function was sigmoid for both layers. SSE (Sum of Squared Errors) was used to obtain the best structure. The value of R² was equal to above 0.900, which indicated that ANN could model the labor rest period based on environmental ergonomics.

Artificial neural networks; Labor; Rest period; Wet bulb globe temperature

Table 1 Environment temperature in SMEs (Ushada et al., 2017)

      SMEs' work system was influenced by various factors and mainly by workplace environmental ergonomics (Ushada and Okayama, 2018). Determining the rest period could be adapted to reduce ergonomic risk at work (Tiacci, 2018). The determination of the rest period was pursued to provide recovery time to laborers. Recovery time gave laborers time to rest their bodies and restore the energy that comes out while working. Every activity in production had specific characteristics; therefore, rest period models, times, and frequencies were diverse and challenging to standardize (Negreiros et al., 2019).
        A Murell formula has been developed to determine the rest periods model (Iridiastadi and Yassierli, 2017). The Murell formula used work time and metabolic rate to determine rest time. Also, in a previous study, Yi and Chan (2013) optimized the rest time schedule for construction laborers in hot and humid working conditions. Ushada et al. (2017) developed an ANN model to determine food-based SMEs' set initial temperature values. Batubara and Dharmastiti (2017) stated that ergonomics intervention was used to improve work systems to reduce workload. Ushada et al. (2017) indicated the nonlinear relationship between worker ergonomics status and workplace environment. ANN (Artificial Neural Networks), as one of the Artificial Intelligence approaches could frame nonlinear structures based on the human brain system and were known as an estimator and could detect nonlinear relationships between variables (Vinoth et al., 2016). Previous studies did not demonstrate a link between the WBGT and the rest period. Thus, an advanced rest period model using ANN was developed in this study, considering the environmental ergonomics, labor circumstances, and production times factor.

Figure 1 Conceptual modelling of the ergonomic parameters of the labor and workstation environment

The previous research by Li et al. (2016) showed that 14:00 to 15:00 was the high-risk time for laborers, while 07:00 to 08:00 was the non-hazardous time for laborers. Based on Li et al. (2016), the data were collected three times: initial, intermediate, and final labor shift.

2.2. Rest Period and WBGT

Metabolic rate (y) was determined by heart rate labor (x) in tofu SMEs. The metabolic rate was calculated using a formula from Astuti in Ushada and Okayama (2018), as shown in Equation 1.

Rest period (R) was determined by labor work time (w), metabolic rate (b), and the threshold value of labor work time (s). The rest period was calculated using a Murell formula in Iridiastadi and Yassierli (2017), as shown in Equation 2.

WBGT was determined by effective temperature (ET) and wet bulb temperature. ET was obtained from Effective Temperature Chart, and wet bulb temperature from Psychrometric Chart. WBGT was calculated using Equation 3 (OSHA, 2012).

Table 2 indicates the categories of metabolic rate and threshold for WBGT value based on the Indonesian government (Minister of Health Indonesia, 2016). Table 2 is a standard by the Minister of Health Indonesia (2016) compared with the actual condition in tofu SMEs.

Table 2 Work cycle and rest period in an hour

2.3. Development of ANN Model

Based on Figure 1, the inputs of ANN were ergonomic parameters of labor and workstation environment. The predicted outputs were the rest period and WBGT. Data sets were divided into training data (n=144) and testing data (n=36). SSE (Sum of Squared Errors) was used as an error value to determine the optimal combination on the ANN model. SSE was minimized between the system output and the neural network model output (Janczak, 2005). The R² correlation was used to express the reliability of the ANN model. ANN was used to predict continuous variables, and a valuable measure of goodness of fit for each output was the coefficient of multiple determination (R²) (Lingireddy and Brion, 2005).

ANN was trained based on the backpropagation algorithm. Backpropagation had a layered feed-forward neural network structure in which the nonlinear neurons were arranged in consecutive layers; afterward, the information passed from the input layer to the output layer through hidden layers (Vinoth et al., 2016). Ushada et al. (2017) found a nonlinear relationship between environmental ergonomics in Food SMEs. Thus, ANN could be a powerful computing approach for complex computations (Cavalieri, Maccarrone, and Pinto, 2004).

The training process was pursued to evaluate weights and biases in ANN (Argatov and Chai, 2019). The training process required a data set of experimentally measured input-output (Argatov and Chai, 2019). The training process was carried out in several stages as follows.

1. Training with variations in the number of neurons and hidden layers

              2. Training with variations in the activation function
        3. Training with variations in the value of the learning rate

3.1. Respondents and Sample Location

       This study obtained sixty (60) respondents from laborers with three times different sampling times. A total sample of one hundred and eighty (180) data sets was obtained. The data samples were obtained from each district in the Special Region of Yogyakarta, such as Yogyakarta, Bantul, Kulon Progo, Gunungkidul, and Sleman. Table 3 indicates the demographic information of the respondents. The average age of the laborers in Tofu SMEs was 42.78 ± 13.64 years old, and the majority ranged from 31 to 40 years old and 41 to 50 years old. The majority gender of the laborers were men (n = 39), and the rest were women (n = 21). Based on the correlation test, age, and weight parameters had a value below 0.200, which indicated a low degree of correlation between the rest period and WBGT.

3.2. Workload and Workstation Environment
       Various tofu-making processes were generally carried out in several SMEs. The processes were soaking, milling, boiling, filtering, processing, solidifying, cutting, and frying. These processes were pursued in a batch continuously. The activities in the tofu production process can be categorized into potentials that cause physical workload (Widyanti et al., 2017). Thus, the activities affected the labor heart rate. Figure 2a shows a graph of labor heart rate, indicating an increase in labor heart rate from the initial to the final shift of the laborer.

        Metabolic rate was counted from labor heart rate. The result indicated three (3) categories such as light, moderate, and heavy (Table 4). Figure 2b indicates an increase in metabolic rate from the initial to final labor shift.

Figure 2 Labor condition: (a) Heart rate; and (b) Metabolic rate

Table 4 Metabolic categories of labor

    The workstation environment was where the laborers did some activities to process the tofu. The environment was observed during working time. The result showed dry bulb temperature increased during work time (Figure 3a).
    The average WBGT value in tofu SME workstations was 26.39 °C. The value was still at the threshold for WBGT in the industry (Minister of Health Indonesia, 2016). Although it met the threshold, the value had a level of discomfort. WBGT value was included in the extremely hot environment category (Chowdhury, Hamada, and Ahmed, 2017). Based on Chowdhury, Hamada, and Ahmed (2017), it had the risk of discomfort, but there was no health risk. Work inconvenience at SMEs arose from 25.82 to 26.86 °C due to the extremely hot environment at the tofu SMEs workstation (Figure 3b). Tofu SMEs use heat sources in most of the process, which could impact increasing the temperature at the workstation.

Figure 3 Workstation environment: (a) Dry bulb temperature; and (b) WBGT

       Table 5 indicates one hundred and sixty-three (163) laborers with a positive WBGT gap. The positive value of the WBGT gap meant that WBGT in laborers did not exceed the threshold value. Laborers were still comfortable with the WBGT values and did well in their activities. On another side, the negative WBGT gap value was found in ten (10) laborers. If the WBGT gap was negative, the WBGT value exceeded the specified threshold. If the WBGT value exceeds the standard value, it could indicate high heat stress on laborers (Bolghanabadi, Ganjali, and Ghalehaskar 2019). Thus, during working hours, the activities increased, and performance did rise after the activities began.

Table 5 Gap of WBGT

3.3. Rest Period
      The rest period was a scheduled break during a working day, and the laborers stopped their activities to rest, eat, and any other needs (ILO, 2019). The rest period was determined by metabolic rate labor during work time. Table 6 indicates the rest period of labor categorized by workload. Work breaks had a range from -44.3 to 11.3 minutes for an hour. The negative number said that laborers did not need an additional rest period because the workload was not too heavy. It could be interpreted that the labor had an extra rest period of 0 minutes to calculate negative numbers. In light and moderate workloads, the average additional rest period was obtained by a negative number, so laborers in tofu SMEs did not need extra work breaks. Labors in the heavy workload category had an average break time of 11.33 minutes. The results show the need for additional breaks of one hour for each work with heavy categories carried out in production activities. The work preferences related to the rest period length were not equal for every labor (Di-Pasquale et al., 2017). Research by Dababneh, Swanson, and Shell (2001) showed that frequent rest periods with short time would not decrease productivity.

Table 6 Rest period based on workload categories

    The rest period depended on the situation in tofu SMEs. Production activities would affect the labor rest period. Activities increased during working hours, and performance increased 2 hours after starting work (Fahed, Ozkaymak, and Ahmed 2018). These conditions were similar to the workstation at tofu SMEs. Labors took a rest during the working day to take toilet breaks, prayer breaks or breaks to address other personal needs (ILO, 2019).

3.4. ANN Model

     Data normalization was performed by the min-max normalization method. The method changed the data into a smaller range. The data changed to a 0 to 1 range. T-test, F-test, and correlation tests were performed. The result showed that dry bulb temperature (i₁), heart rate (i₂), gender (i₃), and wet bulb temperature (i₄) had a significant impact on the rest period (o₁) and WBGT (o₂). These parameters had a high correlation value toward dependent variables so that they could be used as an ANN input.

       The result of the training process was an optimum structure for WBGT and the rest period, as shown in Figure 4. To improve the performance, ANN changed the weights between neurons in each layer to modify the structure of the training data (Darvishi et al., 2017). Training data with the lowest error value was at 4-4-2. The activation function for the between layers was sigmoid. The optimum learning rate was 0.01. That combination got SSE training 0.113 and SSE testing 0.037. The SSE indicated a low error.

Figure 4 ANN structure for the rest period

       The value of R² between the calculation and prediction of resting time was 0.989 (Figure 5a), while the value of R² between the calculation and prediction of WBGT was equal to 0.968 (Figure 5b). An R² value of 1.0 represented a perfect prediction (Lingireddy and Brion, 2005). The results indicated that the calculation of the rest period in explaining the variance of the predicted rest period value was 98.9%. The value showed that the prediction of the rest period could be explained well by calculating the rest period. The WBGT value calculation explains that the WBGT prediction value variance was 96.8%. The predicted WBGT value could be explained well by the WBGT value calculation. An appropriate rest period could reduce the risk caused by workload