Page 176 - Read Online
P. 176
Li et al. Intell Robot 2024;4(3):230-43 I http://dx.doi.org/10.20517/ir.2024.15 Page 236
Data processing Attention-GRU model
section section WOA section
Decode the parameters
Experimental data on passed in by WOA to Input WOA encoding of initial
driving behavior obtain the corresponding values
number of iterations,
batch_size, the number of
neurons working in each Calculation of fitness
Abnormal data handling values for whale
GRU layer, and Dropout
rate population initialization
Operation behavior Establishing Attention- Surrounding Prey
feature extraction GRU fatigue driving Bubble netting for prey
detection model based on Search for prey
incoming parameters
Correlation analysis and
preference of fatigue Calculate the fitness
driving discriminators
Training Attention-GRU value and update the
global optimal solution
Determine the input
variables for Attention-
GRU Y/N
Testing Attention-GRU
Output
Y
Divide the training set
and test set according to Output Attention-GRU
the ratio of 8:2 Evaluation Model optimal hyperparameters
Figure 4. WOA-Attention-GRU algorithm flow (adapted from Li et al., 2023 [24] ). WOA: Whale optimization algorithm; GRU: gated recur-
rent unit.
number of neurons working in the GRU layer, and dropout rate. Fitness value calculation: the fitness values
of the initialized whale population are calculated, and the global optimum is updated. Iterative optimization:
based on the fitness values, the positions of the whale individuals are updated, gradually approaching the
global optimum, and finally outputting the optimal hyperparameters of the Attention-GRU model. In the
Attention-GRU model part, the model is trained and tested using the hyperparameters optimized by WOA.
The main steps include model training: training the model using the training set provided by the data process-
ing part. The attention mechanism focuses on important features in the sequences of driving behavior data,
assigning greater weight to important information and reducing information loss. Model testing: testing the
trained model using the test set to evaluate the model’s performance. The model’s predictive performance is
then assessed by calculating the MSE. Through these steps, we can effectively detect the fatigue state of drivers,
providing more accurate and reliable detection results. “Optimal hyperparameters” refer to the best set of pa-
rameters that can minimize the MSE between the predicted fatigue level and the actual fatigue level. These
hyperparameters include the number of iterations, batch size, the number of neurons in each GRU layer, and
the dropout rate. The network model structure of Attention-GRU is presented in Figure 5.
2.2.5 Fatigue state recognition based on Transformer
Transformer excels at handling long-range dependencies in sequence data, which is particularly beneficial
for time series analysis. We use the standard Transformer architecture, including self-attention mechanism.
The number of layers, number of attention heads, and hidden layer dimensions were adjusted to optimize
performance on our dataset. The model is trained using the same driving behavior data set and adopts the
same preprocessing method as the WOA-Attention-GRU model. The results show that although Transformer
performs well in capturing long-range dependencies, in the context of driving behavior analysis, the method
combining WOA and Attention mechanisms in the GRU model provides more targeted feature extraction and
optimization.