Autonomous ship sea condition estimation method based on noisy labeled ship motion data

Abstract

The invention discloses an autonomous ship sea condition estimation method based on noisy labeled ship motion data, and belongs to the field of data mining and machine learning. An end-to-end neural network is used as an infrastructure, and targeted improvement is provided for problems existing in the field of sea condition estimation. In order to solve the problems of multi-scale characteristics, noise labels and the like of collected ship motion data, a multi-scale characteristic learning module with excellent representation learning ability is firstly provided, and the need for extra noise processing steps and hyper-parameter tuning is reduced by focusing on extraction of key characteristics. Besides, in order to prevent an overfitting problem possibly caused by strong feature extraction capability, a loss function for inhibiting an abnormal value is combined, and the design further ensures the stability of the model. The method is not only suitable for sea condition estimation, but also has good applicability to time series data classification problems in other fields.

Description

Technical Field

[0001] The invention relates to an autonomous ship sea condition estimation method based on noisy annotated ship motion data, and belongs to the field of data mining and machine learning. Background Art

[0002] Accurately estimating sea conditions to adjust navigation strategies for safe and efficient operation is a key challenge in the development of autonomous ships. Autonomous ships are ships designed to navigate, operate, and make decisions. These ships are usually equipped with complex sensors and communication equipment, enabling them to navigate independently, optimize routes, and manage operations in different sea conditions. As autonomous ship technology continues to develop, sea state estimation is essential to ensure the safe and efficient operation of autonomous ships.

[0003] Traditional sea condition assessment usually relies on means such as weather forecasts or wave radars, rather than ship motion data. However, these methods are limited by human subjectivity, low cloud cover, and the need for frequent calibration. In recent years, researchers have proposed sea condition estimation methods based on ship motion data, using ships as large sensors to directly infer sea conditions through their motion data. These methods are divided into physical models and data-driven models. Traditional physical-based models usually derive the ship response amplitude operator (RAO) through complex ocean mechanics theory, and use expert knowledge to associate the ship's motion with ocean conditions. In contrast, data-driven methods focus more on automatically revealing the intrinsic correlation between ship motion data and sea conditions. Shallow machine learning methods usually rely on manual feature extraction, but this often limits the model's expressiveness and generalization capabilities. Deep learning methods can mine more representative and task-related features by combining feature extraction with classifier training, and show stronger adaptability and robustness in sea condition estimation.

[0004] However, in practical applications, ship motion data usually exhibit multi-scale characteristics due to the interaction between complex environmental forces and ship dynamics. Changes in environmental conditions and human labeling errors can cause sensor data to be interfered with by noise, resulting in inaccurate labeling, which in turn affects the reliability of prediction results. Early methods mainly relied on fixed rules and statistical models. Although they can handle some noise, they have limited performance in complex sea conditions. With the development of deep learning and data-driven methods, sea condition estimation based on noisy annotated ship motion data has become possible. By building a robust neural network model, this field is expected to significantly improve the perception and adaptability of autonomous ships to sea conditions.

[0005] To address the above challenges, this study proposes a multi-scale feature learning module with superior representation learning capabilities. Unlike traditional noise-resistant methods, our approach focuses on enhancing the model's feature extraction capabilities. By enhancing feature extraction capabilities, we reduce the need for hyperparameter tuning that relies on noise processing, and improve computational efficiency and model robustness. In addition, a robust loss function that can suppress outliers is used to further prevent overfitting and ensure model stability. Summary of the invention

[0006] The present invention aims to provide an autonomous ship sea state estimation network model based on ship motion data with noise annotation, so as to achieve accurate estimation of sea conditions in complex environments.

[0007] In order to achieve the above objectives, the solution of the present invention is: in order to avoid the introduction of additional error problems, we focus on the feature extraction of data, and combine a more flexible loss function that can suppress outliers. The model involved has better performance and stronger generalization ability. The specific steps are as follows:

[0008] (1) Obtain the ship motion data collected by the autonomous ship sensor and perform preprocessing operations such as analysis, division, and interpolation on the collected data. The training set and test set are divided into a 4:1 ratio and randomly divided five times to more stably evaluate its performance.

[0009] (2) Design a network model for autonomous ship sea state estimation using noisy labeled ship motion data. The processed sea state dataset is input into the network and first passes through a multi-scale feature learning module to obtain a more comprehensive feature representation. Here, the emphasis is on using the intrinsic structure of the data to obtain highly distinguishable features. In order to further reduce the impact of noisy labels, we design a robust loss function that can suppress outliers to calculate the training loss, and then perform backpropagation to update the model parameters.

[0010] (3) Conduct multiple experiments to explore the optimal model parameters, and finally use the F1 score as an evaluation metric to evaluate the model.

[0011] (4) Design ablation experiments to explore the effectiveness of each proposed module and the specific problems it solves.

[0012] The beneficial effects of the present invention are: the method can make accurate judgments on ocean conditions in complex environments. Some challenging problems have been solved: ship motion data usually has multi-scale characteristics and is prone to labeling errors. Early methods mainly relied on fixed rules and statistical models. Although they can handle some noise, their performance is limited in complex sea conditions. Although the mainstream methods for solving the noise labeling problem can achieve a certain degree of robustness, they are very likely to introduce additional erroneous information during the training process, and most methods are designed based on certain assumptions or preset conditions. When faced with different application scenarios, they may not be flexibly adjusted and optimized. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 This is the entire algorithm flow chart of the present invention DETAILED DESCRIPTION

[0015] Obtain the ship motion data collected by the autonomous ship sensor, and perform pre-processing operations such as analysis, division, and interpolation on the collected data. Divide the training set and test set into a 4:1 ratio and randomly divide them five times. Input the processed data into the network for training.

[0016] We first introduced a multi-scale feature learning module that combines a multi-scale mechanism with a large-kernel convolution structure. The core of this module uses a one-dimensional convolutional neural network (1D-CNN) as the main feature extractor. It consists of N layers of convolution, each of which contains K convolution kernels of different scales. By using convolution kernels of different sizes in parallel, the model can capture both local and global features of the time series, thereby obtaining a more comprehensive feature representation. In addition, the module combines large-kernel convolution with dilated convolution. Dilated convolution expands the receptive field to obtain feature correlations in different ranges, improving the efficiency of the model while keeping the computational complexity low.

[0017] In addition, time series data not only has the problem of noisy labels, but the data collection process may also be disturbed by factors such as changes in sampling frequency and inaccurate sensors, resulting in data distortion. This distortion increases the difficulty of model learning patterns. Large kernel convolution structures have significant advantages in dealing with these data distortions and can enhance the robustness of the model to noise and uncertainty. Therefore, this multi-scale feature learning module improves the classification performance of the model on time series data with noisy labels by effectively extracting features.

[0018] The specific operation of each convolution layer is as follows: Given an input X, the operation of the mth layer can be expressed by the formula:

[0019]

[0020] X″ m，k =BN(X′ m,k ) (2)

[0021] X″′ m,k =ReLU(X″ m,k ) (3)

[0022] X m =Cat[X″′ m，1 , X″′ m，2 , ..., X″′ m,k ] (4)

[0023] here represents the size of the kth convolution kernel in the mth layer, d represents the dilation size. BN stands for batch normalization, X′ m，k , X″ m，k , X″′ m，k They are the outputs of the kth kernel Cnovld, BN, and ReLU layers respectively. Finally, the outputs of each kernel are concatenated to obtain the output X of the convolution layer. i .

[0024] The size of the convolution kernel is defined as follows:

[0025] F={2k+1},1≤k≤K (5)

[0026] Here, F represents the set of kernel sizes for each layer of convolution, and K represents the number of convolution kernels. In terms of the selection of convolution kernel sizes, we start from 3 and gradually increase. First, use smaller conventional convolution kernel sizes (such as 3, 5, and 7) to extract detail features, and then gradually introduce larger convolution kernels (such as 9 and 11) to expand the receptive field. This setting can effectively combine the extraction of local and global information, thereby enhancing the expressiveness of the model and adapting to the learning of complex patterns. Compared with the combination of convolution kernels with larger spans (such as 4) and the combination of the two, this gradual expansion method can balance the extraction of local and global features without losing information. By selecting the appropriate convolution kernel size, the model can cover multiple receptive fields and promote the learning of multi-scale features.

[0027] Finally, the features output by the last layer of the network are passed to the classifier through the average pooling layer and the fully connected layer, and the output vector z = [z1, z2, ..., z n ], where z i =[z i1 , z i2 , ..., z iq ] represents the score of the i-th sample in each category.

[0028] Powerful feature extraction capabilities will inevitably lead to overfitting of the model, so a robust loss function is used when calculating the loss during training. By suppressing possible outliers in the network output, overfitting can be further prevented and the generalization ability of the model can be improved. The main design of this loss function is as follows:

[0029] First, calculate the training loss of each sample. The cross entropy loss function is used here: Let L = {L1, L2, ..., L n} represents the loss value set of all samples in a batch (n is the number of samples), where L i is the loss of the ith sample. Let C = {c1, c2, ..., c q} represents the set of all categories (q is the number of categories), where c j is the jth class. Let L cj ={L j |y i =c j} means it belongs to category c j The loss set of all samples in .

[0030] Next, the obtained losses are normalized class by class. j For example, the loss L for each sample i belonging to this category i Standardization: The formula is as follows:

[0031]

[0032] in, is the mean, is the standard deviation, Represents the normalized loss value.

[0033] This process is performed once for each category, resulting in a set of losses that are standardized for each category. At this time, the losses of different categories have the same distribution (mean 0 and variance 1), which avoids the selection preference problem when selecting outliers in the next step.

[0034] For the normalized loss, we use a dynamic threshold to select possible noise samples, and the selected index set is:

[0035]

[0036] Here, we use the normalized loss average as the threshold τ.

[0037] Next, we trim the logits of the selected high-loss samples to prevent abnormal samples from having too much impact on model training. The formula is as follows:

[0038]

[0039] Among them, z i is the original logits vector of the ith sample, ||·|| p represents the p-norm, is the clipped logits. Adjusting with the norm also keeps the logits vector in the same direction as the original logits vector.

[0040] Finally, recalculate the cross entropy loss using the adjusted Logits:

[0041]

[0042] By using the pruned Logits to calculate the final loss, we can weaken the negative impact of abnormal samples during training, thereby improving the robustness of the model.

Claims

1. An autonomous ship sea state estimation method based on noisy annotated ship motion data, characterized in that: The steps include: (1) Obtain ship motion data collected by autonomous ship sensors and perform preprocessing operations such as analysis, partitioning, and interpolation on the data. The data is divided into training and test sets at a ratio of 4:1 and randomly divided five times to more stably evaluate the model performance. (2) Design an autonomous ship sea state estimation network model based on noisy annotated ship motion data. The preprocessed sea state data is input into the network. First, a comprehensive feature representation is obtained through a multi-scale feature learning module, emphasizing the use of the intrinsic structure of the data to obtain highly distinguishable features. In order to reduce the impact of noisy labels, a robust loss function is designed to suppress outliers, and the training loss is calculated based on it, and back-propagation is performed to update the model parameters. (3) The optimal model parameters were explored through multiple experiments, and the F1 score was finally used as the evaluation indicator to evaluate the model performance. (4) Design ablation experiments to explore the effectiveness of each module and the specific problems it solves.

2. The method for autonomous ship sea state estimation based on noisy annotated ship motion data according to claim 1 is characterized in that: In step (1), after the collected ship motion data is preprocessed, five cross-validations are used to verify the effectiveness of the model and the average value is calculated.

3. The method for autonomous ship sea state estimation based on noisy annotated ship motion data according to claim 1, characterized in that: In step (2), the processed data is input into the feature map extraction network (with a one-dimensional convolutional neural network as the core feature extractor), and a more comprehensive feature representation is obtained through a multi-scale mechanism combining large kernel convolution and dilated convolution. Then, the feature map output by the feature extraction module is input into the classifier to obtain the category probability distribution of each sample.

4. The method for autonomous ship sea state estimation based on noisy annotated ship motion data according to claim 1, characterized in that: In step (2), the loss is not calculated directly using the probability distribution of the classifier output, but is calculated after adjustment according to the designed loss function. The specific steps are: first, the cross entropy loss of each sample is calculated through the probability distribution, and the obtained loss is normalized class by class, so that the loss of each category follows the same distribution and avoids the preference of noise sample selection; then, through the adaptively adjusted threshold, the samples with larger loss are selected from the normalized loss as abnormal samples; the abnormal samples are adjusted, that is, the original probability distribution vector is divided by its bi-norm to obtain the adjusted vector, ensuring that the direction of the vector remains unchanged.

5. The method for autonomous ship sea state estimation based on noisy annotated ship motion data according to claim 1, characterized in that: In step (2), the final loss is calculated using the adjusted probability distribution vector, and backpropagation is performed to update the model parameters.

6. The method for autonomous ship sea state estimation based on noisy annotated ship motion data according to claim 1, characterized in that: In step (3), the F1 score is used as an indicator to evaluate the network performance.

7. The method for autonomous ship sea state estimation based on noisy annotated ship motion data according to claim 1, characterized in that: In step (4), the rationality and effectiveness of the designed loss function are evaluated through ablation experiments.

Images