Feature Comparison of Different Operator Data of a Hydraulic Excavator using LSTM

In construction industry, productivity needs to be improved because of decreasing working population. Hydraulic excavators are essential for construction sites. Operating hydraulic excavators require a high level of skills. Since the productivity at construction sites is affected by the skills of operators, quantitative evaluation of operator skills can help improve the productivity. In this study, a classification model of hydraulic excavator operations is created to classify five common digging motions using Long Short-Term Memory (LSTM). Using time series of three types of operators with different operating skills, we compare the features of the predicted classes and the predicted probabilities of each motion. Skilled and general operators are classified with high accuracy. We find that the accuracy of the classification of unskilled operator is not so good and the predicted probabilities of “idle” is high for changing to the next operation or in the motion. These reveal that the operation of the unskilled operator is not stable.


Introduction
In recent years, construction industry needs to improve its productivity due to the decrease of the number of skilled workers (1) . The Ministry of Land, Infrastructure, Transport and Tourism of Japan is promoting "i-Construction" and is using Information and Communication Technology (ICT) to improve the productivity and safety at construction sites (2)(3) . Hydraulic excavators are essential construction machines which are used to dig and load soil. Operating a hydraulic excavator requires a high level of skills, and in general, skilled operators are more productive. In other words, the productivity of a construction site is affected by the operators' skills of the hydraulic excavator. Assessing the quantitative features of operators' skills will lead to a better understanding of the productivity and help us to improve it at the construction sites.
In previous studies, deep learning is used to classify the motions of a hydraulic excavator (4)(5) , but the evaluation is based on data from only one single operator and do not extend to the comparison of different operators. To compare different operators, motion classification using Random Forest has been proposed (6)(7) , but there is no comparison with other methods such as deep learning. In addition, a research has been conducted to build a system to evaluate operation skills and to compare the skills of different operators for swing operation (8) , however it only evaluates a specific motion. In the medical field, machine learning-based assessment of surgical skills has been used to train novices and surgical robots (9) .
In this study, we create a model to classify the motions of a hydraulic excavator based on data obtained by multiple sensors. We utilize the time series of a hydraulic excavator by different operators to classify digging motions and compare the features of the operations.
We may use a recurrent neural network that can account for time dependency a time series. However, a recurrent neural network is often not able to learn sufficiently due to the vanishing gradient in the parameters (10) . Thus, we adopt a Long Short-Term Memory (LSTM), which is one of the deep learning methods, since it is said that LSTM can learn the long-term time dependency (11) .
In this experiment, we use time series of a 20-ton class hydraulic excavator. The general digging operations of the hydraulic excavator include "digging", "lifting and swing", "dumping", and "repositioning". In addition to these motions, "idle", which corresponds to no operation, is added. Our model classifies these five motions. Operators having different operating skills are classified as "skilled operator", "general operator", and "unskilled operator". We compare the features of the predicted class and the predicted probabilities of each motion.
The proposed model can accurately classify time series of skilled operator and general operator into five digging motions. Their predicted probabilities are also high for each motion. There is no significant difference between the features of skilled operator and general operator. On the other hand, the classification of unskilled operator is not accurate because it is aperiodic and is often classified as "idle" state in the motions. In addition, the predicted probability of "idle" is high even in the middle of motions or changes to the other motions. Since these features are not shown by skilled and general operators, they are unique to unskilled operator. These results show that the operation of unskilled operator is unstable and their skills are low. Fig. 1 shows that Long Short-Term Memory (LSTM) blocks consisting of three types of gates (input gate, output gate, and forget gate) and cells called memory cells are introduced.

LSTM model
Given input ( ) at time t, hidden layer ( − 1) at time t-1, weights W, U, V, and bias b and with the subscripts corresponding to the input gate, output gate, and forget gate as i, o, and f, respectively. The output ( ) of input gate, the output ( ) of output gate, and the output ( ) of forget gate are represented as where ( ) is the output of memory cell and σ represents the activation function, where the ReLU (Rectified Linear Unit) function is used. When ( ) is represented the value of the input ( ) activated and ⨀ represents the element product of the elements of the vector multiplied by each other, the output ( ) of the memory cell is represented as When ′( ) is the value of the memory cell output ( ) activated by the tanh function, the output ( ) in the hidden layer is represented as Next, when the number of classes is n, the output layer consists of the fully connected layers that output n values. Also, when transformed by the softmax function, these values sum to 1. Given the output ( ) of class k (k = 1, 2, …, n) at time t, the predicted probability ( ) of class k is represented as equation (6).
Then, the class corresponding to the output with the highest prediction probability is used as predicted class.

Experiment overview
We utilize 20-ton class hydraulic excavator (SK200-9 made by Kobelco Construction Machinery Co., Ltd.) shown in Fig. 2. Then, the digging operation is classified for three types of time series whose operating skills are classified as "skilled operator", "general operator", and "unskilled

h(t) x(t)
input gate forget gate output gate memory cell

i(t) o(t) f(t) c(t)
operator", respectively. The time series consists of 33 types of sensors data which is sampled at 0.1 [s]. These sensors data include pump pressure and operating pilot pressure.
As shown in Fig. 3, the digging operation consists of four operations "digging", "lifting and swing", "dumping", and "repositioning" as one cycle. There are five classification classes, including "idle" state, which corresponds to no operation.
Since the digging operation of skilled operator is used as a reference, the time series of skilled operator is labeled with the operation classification. The classification model is created using the latter 500 samples of skilled operator's data as test data and the remaining data as training data. Similarly, the last 500 samples of general operator and unskilled operator are extracted as test data, and their digging operations are classified by the model trained on the skilled operator's data. An overview of the digging operations data is shown in Table 1.
At the time of training, the training data is divided into multiple batches, and each batch is input to the model. The model is trained to minimize the loss between the prediction and the actual data at that time. The model is repeated for 30 epochs, with one epoch being the time when all training data has been input. At each epoch, 30% of the data is randomly extracted from the training data, and these data are used as validation data. Fig. 4 shows the training results. Table 2 presents the percentage of correct answers and losses at 30 epochs. As shown in Fig. 4, the loss decreases and the correct classification increases with each iteration of training. Table 2 demonstrates that the loss at 30 epochs is sufficiently small, and the model can classify the validation data with high accuracy. Therefore, the model has a high generalization performance even for unknown data.

Results
Trained model was used to classify the digging motion of each operator for test data. Fig. 5 illustrates the results of predicted classes for each operator. Fig. 5 shows that the classification for skilled operator and general operator is highly accurate. It also presents that digging operations are    On the other hand, Fig. 5 shows that the classification for unskilled operator is not highly accurate. The predicted classes for unskilled operator demonstrate that, unlike the other operators, the motions are not classified as equally spaced and are aperiodic in each motion. The results of unskilled operator present that they are classified as "idle" state which corresponds to no operation during digging operation. We found that it is the feature of unskilled operator that their predicted classes are classified as "idle" state despite in the digging motions.
As shown in Figs. 6-8, the predicted probabilities for the five motions of each operator are compared. Fig. 6 presents the predicted probabilities for each motion of skilled operator. These graphs illustrate the high prediction probabilities for each motion and the small number of false positives. Fig. 7 presents the predicted probabilities for each motion of general operator. As in the case of skilled operator, high prediction probabilities are shown for each motion and their motions are correctly classified. Then, they can be seen that the motions of both skilled and general operators have similar features. Fig. 8 illustrates the predicted probabilities for each motion of unskilled operator. As shown in Fig. 8, the unskilled operator data tend to be different from the other operators. When the motions change to the other motions, the predicted probabilities of "idle" state is higher. For example, when the motion changes from "repositioning" to "digging" at 250 timesteps, the predicted probabilities of "idle" is high. Also, the probabilities of "idle" is high in the middle of motions, for example, during "lifting and swing" motion at 150 timesteps. Although "idle" state corresponds to no operation, the data of unskilled operator presents the features of no operation even during digging operations. Since these features were not shown by skilled and general operators, they were unique to unskilled operator. These results were influenced by low operation skills and unstable operation of unskilled operator.

Conclusion
In this study, a classification model of digging motion was created using LSTM, and the digging motion of a hydraulic excavator was classified. The results of motion classification using time series of three operators, "skilled operator", "general operator", and "unskilled operator" were compared. It was found that the results of skilled and general operators classified periodic motions. On the other hand, the data of unskilled operator had features of low accuracy in motion classification and that classified several motions as "idle" state during digging operation.