### 1. Introduction

### 2. Research Methodology

### 2.1 Recurrent Neural Network

*t*, the LSTM structure consists of a forget gate, an input gate, and output gate. In Fig. 2, F, I, and O boxes represent the forget, input, and output gates, respectively.

*h*

_{t}_{− 1}represents the hidden state at

*t*− 1,

*x*represents the input data at

_{t}*t*, and

*W*and

_{f}*b*represent the weight and bias of the forget gate, respectively. Using sigmoid function

_{f}*σ*, unnecessary information is assigned a weight close to 0 to forget the information, and important information is assigned a weight close to 1 to learn to completely preserve the information. The calculated

*f*is transferred to the cell state for the next operation.

_{t}*x*and previous hidden state

_{t}*h*

_{t}_{− 1},

*i*indicating the degree to remember the current input value and the current local state

_{t}*C̄*are obtained. Global cell state

_{t}*C*to be transmitted to the next hidden state is determined by reflecting the received previous cell state

_{t}*C*

_{t}_{− 1}. In this case,

*W*and

_{i}*b*denote the weight and bias of the input gate, respectively, and

_{i}*W*and

_{C}*b*denote the weight and bias of the cell state, respectively.

_{C}*h*that will be delivered to the hidden layer reflecting current global cell state

_{t}*C*obtained with

_{t}*o*, which expresses the extent of memory of the current output value calculated using previous hidden layer

_{t}*h*

_{t}_{− 1}, and current input value

*x*. In this case,

_{t}*W*and

_{O}*b*denote the weight and bias of the output gate, respectively.

_{O}### 2.2 Multiple Classification Model Performance Indicators

### 3. Data and Learning

### 3.1 Damaged Ship Motion Data

*Max*, owing to collision is expressed as Eq. (11), and the maximum damage area at the hull bottom,

_{HS}*Max*, owing to stranding is expressed as Eq. (12).

_{HB}*L*indicates the overall length,

*D*indicates the depth, and

*B*indicates the moulded breadth.