Page 172 - Read Online

Li et al. Intell Robot 2024;4(3):230-43 I http://dx.doi.org/10.20517/ir.2024.15 Page 232

Driving operation behavior

Throttle Vehicle Steering Vehicle Vehicle
opening Vehicle longitudinal wheel traverse lateral
speed
degree acceleration cornering angle acceleration
feature extraction
Driving behavior
Dimensional Features Dimensionless Features

Wavelet Wavelet Wavelet Peak-to- Absolute Root- Waveform
energy scale singular mean Variance mean- Skewness Kurtosis
entropy entropy entropy peak value square factor

Figure 1. Block diagram of driving behavior feature extraction.

including wavelet energy entropy, wavelet scale entropy, and wavelet singular entropy. The chosen wavelet
basis function is db6 within the dbN series, and a three-layer wavelet packet decomposition is employed. The
computational steps are outlined as follows: (1) Utilize the db6 wavelet basis function to decompose driving
behavior data into three layers of wavelet packets, yielding eight subbands. Reconstruct the wavelet subband
components to ensure the new driving behavior data’s length matches that of the original data; (2) Determine
the two-parameter number for each node, square it to yield the node’s energy value, and then sum the energy
across nodes to compute the total wavelet energy. Subsequently, derive the wavelet energy entropy based on
the total wavelet energy; (3) Compute the wavelet scale entropy for each subband; (4) Extract singular values,
construct a vector, generate the singular value spectrum, and perform singular value decomposition to obtain
the wavelet singular entropy.

2.1.2 Driving behavior feature extraction
The recursive feature elimination with cross-validation (RFECV) method involves iterative training of data
using a base model, eliminating features with low weights based on weight coefficients in each round until
the candidate subset meets termination conditions. Given challenges such as fluctuations in real car driving
behaviordata, significantnoise, andsampleimbalance, thepaperadoptstherandomforestasthebasemodelto
address these issues. The algorithm consists of the following steps: (a) Train models using all driving behavior
features, calculate feature importance, and rank them. Extract the top most important features for each
subset , where i ranges from 1 to S; (b) Split the training set into a new training set and a validation set. Train
the model using the new training set and all features, and then evaluate the model with the validation set; (c)
Input the filtered features into the random forest as the initial feature subset and calculate feature importance.
Remove features with the lowest importance from the current subset to obtain a new feature subset. Repeat
this process, inputting the new subset into the random forest, calculating the importance of each feature, and
determining the classification accuracy using cross-validation; (d) Recursively repeat Step 3 until the feature
subset is empty. Ultimately, obtain k feature subsets with varying numbers of features, selecting the subset with
the highest classification accuracy as the optimal feature combination.

2.2 Attention-GRU identification of fatigue levels
2.2.1 GRU neural network
The driving behavior data is inherently sequential, and the connections among the hidden layers of recurrent
neural networks involve integrating the output of the hidden layers from the previous moment into the current
network state [13] . This sequential network structure is effective in preserving dependencies within the data.
Notably, the GRU is particularly skilled at capturing long-range dependencies and efficiently mitigating the
challenges of gradient explosion and gradient disappearance observed in basic recurrent neural networks. The