Water surface unmanned system state automatic evaluation method based on multi-dimensional feature parallel extraction and fusion modeling

Abstract

The invention relates to the technical field of unmanned ships, in particular to a water surface unmanned system state automatic evaluation method based on multi-dimensional feature parallel extraction and fusion modeling. The method mainly faces multi-source time series data of a water surface unmanned system, and constructs structured and unstructured dual-channel feature representation in combination with statistical feature extraction with physical interpretability and deep convolution modeling. And a gradient boosting Catboost model based on classification enhancement and a gated recurrent neural network GRU model are further fused, and the static classification capability and the time sequence modeling capability are considered. And meanwhile, a classification enhancement fusion strategy and a Bayesian optimization mechanism are introduced, so that the classification performance and the model adaptive capacity of the model are improved. The method has the advantages of high accuracy, high robustness and good generalization ability, is suitable for various complex operation scenes, and can significantly improve the operation efficiency and safety of the water surface unmanned system.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned ships, and in particular to a method for automatically evaluating the state of a surface unmanned system based on multi-dimensional feature parallel extraction and fusion modeling. Background Art

[0002] With the widespread application of unmanned surface systems in intelligent inspection, water quality monitoring, water rescue, and other fields, the issue of state assessment during their operation is receiving increasing attention. The operational status assessment of unmanned systems is a core link in ensuring their safe and efficient operation. During navigation, unmanned surface systems must cope with variable environmental factors such as waves, current velocity, and wind speed, which will affect the system's operational stability and navigation safety. Traditional assessment methods often rely on static rules, empirical thresholds, or simple statistical features, making it difficult to effectively analyze the system's operational behavior in a dynamic environment. This is especially true in complex scenarios such as wind and wave disturbances and sudden state changes, where the assessment results are prone to misjudgments or omissions.

[0003] With the widespread application of unmanned surface systems in intelligent inspection, water quality monitoring, water rescue, and other fields, the issue of state assessment during their operation is receiving increasing attention. The operational status assessment of unmanned systems is a core link in ensuring their safe and efficient operation. During navigation, unmanned surface systems must cope with variable environmental factors such as waves, current velocity, and wind speed, which will affect the system's operational stability and navigation safety. Traditional assessment methods often rely on static rules, empirical thresholds, or simple statistical features, making it difficult to effectively analyze the system's operational behavior in a dynamic environment. This is especially true in complex scenarios such as wind and wave disturbances and sudden state changes, where the assessment results are prone to misjudgments or omissions. Summary of the Invention

[0004] The present invention discloses a method for automatically evaluating the state of an unmanned surface system based on parallel extraction and fusion modeling of multi-dimensional features. The specific method is as follows:

[0005] Obtain the status information of the unmanned ship and perform preprocessing;

[0006] For the preprocessed information, structural feature information and convolution feature information are extracted respectively;

[0007] The structural feature information is combined with the CatBoost model to obtain the first state prediction probability distribution, and the convolutional feature information is combined with the GRU model to obtain the second state prediction probability distribution;

[0008] The covariance cross fusion algorithm is used to fuse the first state prediction probability distribution and the second state prediction probability distribution to obtain the joint decision result.

[0009] Furthermore, the unmanned ship status information includes: position data, attitude data, hull data, environmental data and operating status level.

[0010] Further, preprocessing is performed, and the specific method is as follows:

[0011] The moving window mechanism is used to segment the continuous time series data, and a fixed-length time window and a set step size are set to slide to extract several overlapping time segments.

[0012] All numerical features are normalized using the standardization method. The specific formula is as follows:

[0013]

[0014] Where x i represents the original value of the i-th feature, μ i and σ i are the mean and standard deviation of the feature in the sample set, x i ′ represents the normalized eigenvalue.

[0015] Furthermore, structural feature information is extracted. The specific content and calculation formula are as follows:

[0016]

[0017] Where, X t Indicates the horizontal position value at the tth moment, Y t The vertical position value at the tth moment, Represents the average increment of longitudinal position, Z t The vertical position value at the tth moment, Roll t The roll angle at the tth moment, Represents the average value of the roll angle sequence in the sliding window, Pitch t Indicates the pitch angle at the tth moment, Yaw t Indicates the heading angle at the tth moment, RPM t The engine speed value at the tth moment, Represents the mean value of the engine speed within the sliding window, VB t represents the vibration frequency value at the t-th moment, Represents the mean value of the vibration frequency in the sliding window, TE t represents the cabin temperature at time t, WS t represents the wind speed at the tth moment, Represents the average value of the wind speed sequence in the sliding window, Rain t represents the rainfall at the tth moment, AT t represents the weather temperature at the tth moment, Indicates the average increment of weather temperature, CS trepresents the water velocity at the tth moment, t represents the time step index, N represents the sliding window size, X_DM represents the mean of the first-order difference of the lateral position, Y_DSTD represents the standard deviation of the first-order difference of the longitudinal position, Z_RNG represents the vertical height range, RL_STD represents the standard deviation of the roll angle, PT_SLP represents the linear trend slope of the pitch angle, YW_DMAX represents the maximum value of the heading angle difference, RPM_STD represents the standard deviation of the engine speed, VB_KURT represents the vibration frequency kurtosis, TE_SLP represents the linear trend slope of the cabin temperature, WS_STD represents the standard deviation of the wind speed, RAIN_MEAN represents the mean of the rainfall, AT_DSTD represents the standard deviation of the first-order difference of the weather temperature, and CS_SLP represents the linear fitting slope of the water velocity.

[0018] Furthermore, the CatBoost model takes minimizing the multi-classification cross entropy loss function as its objective function. The specific formula is as follows:

[0019]

[0020] in An indicator variable indicating whether sample i belongs to category k, represents the k-th class probability predicted by the model, M represents the total number of samples, K represents the total number of categories, Represents the multi-class cross entropy loss function.

[0021] Furthermore, the convolution feature information is extracted. The specific method is as follows:

[0022] Assume that the time series X={x0,x1,x2,...,x t-1 ,x t}, the dilated convolution operation formula of its element xt is:

[0023]

[0024] Where, F={f1,f2,...,f K-1 ,f K} represents the convolution kernel of size K, f k represents the convolution kernel, x t represents the sample value at time t in the input time series, d represents the expansion factor, C(F,x t ) represents the output result of the sequence after dilation convolution.

[0025] Furthermore, the state update formula of the GRU model is as follows:

[0026] z t =σ(W z ×[h t-1 ,x t ])

[0027] r t =σ(W r ×[h t-1 ,x t ])

[0028]

[0029]

[0030] Where x t Indicates the current input, h t represents the current hidden state, h t-1 represents the hidden state at the previous moment, z t represents the update gate vector, r t Represents the reset gate vector, W z ,W r ,W represent the weight parameter matrices of the update gate, reset gate and main network respectively, σ(·) represents the Sigmoid activation function, and tanh(·) represents the hyperbolic tangent function.

[0031] Furthermore, the hyperparameters of the CatBoost model and the GRU model are optimized using the Bayesian optimization algorithm. The specific method is as follows:

[0032] Modeling target performance by constructing a surrogate function;

[0033] The most promising parameter combination is selected based on the acquisition function to achieve automatic hyperparameter tuning.

[0034] Furthermore, the covariance cross fusion algorithm is used to fuse the first state prediction probability distribution and the second state prediction probability distribution. The specific formula of the fusion algorithm is as follows:

[0035]

[0036] Among them, P a is the covariance matrix of model A, P b is the covariance matrix of model B, ω∈[0,1] is the weighting coefficient, P CI is the uncertainty matrix of the fused output, represents the output of model A, represents the output of model B, Represents the final estimated output result after fusion.

[0037] Furthermore, the CatBoost model and GRU model use accuracy, precision, F1 score, and recall as evaluation indicators. The specific formulas are as follows:

[0038]

[0039] Where TP represents the number of positive samples correctly identified by the model; TN represents the number of negative samples correctly identified by the model; FP represents the number of negative samples incorrectly predicted as positive samples by the model; and FN represents the number of positive samples incorrectly predicted as negative samples by the model.

[0040] Due to the adoption of the above technical solution, the present invention has the following beneficial effects:

[0041] The present invention is oriented towards multi-source time series data of surface unmanned systems, combining statistical feature extraction with physical interpretability and deep convolutional modeling to construct structured and unstructured dual-channel feature representation. It further integrates the gradient boosting Catboost model based on classification enhancement and the gated recurrent neural network GRU, taking into account both static classification capabilities and time series modeling capabilities. At the same time, the classification enhancement fusion strategy and Bayesian optimization mechanism are introduced to improve the classification performance and model adaptability of the model. The present invention has high accuracy, high robustness and good generalization ability, is applicable to a variety of complex operation scenarios, and can significantly improve the operating efficiency and safety of surface unmanned systems.

[0042] Other advantages, objects, and features of the present invention will be described in part in the following description and, in part, will be apparent to those skilled in the art upon examination of the following description or may be learned from practice of the present invention. The objects and other advantages of the present invention may be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] The accompanying drawings of the present invention are described below.

[0044] Figure 1 Schematic diagram of the physical feature extraction module.

[0045] Figure 2 Schematic diagram of TCN convolution structure.

[0046] Figure 3 This is the GRU model structure diagram.

[0047] Figure 4 Catboost model structure diagram

[0048] Figure 5 A schematic diagram of the overall framework.

[0049] Figure 6 This is the accuracy distribution diagram of the simulation experiment model.

[0050] Figure 7 This is the ROC curve of the simulation experiment. DETAILED DESCRIPTION

[0051] The present invention will be further described below with reference to the accompanying drawings and examples.

[0052] An automatic assessment method for the status of unmanned surface systems based on parallel extraction and fusion modeling of multi-dimensional features, such as Figure 5 The specific steps are as follows:

[0053] S1. Obtain the status information of the unmanned ship and perform preprocessing.

[0054] In this embodiment, when evaluating the operational status of an unmanned surface system, multiple types of key feature information should be comprehensively considered to fully reflect the system's dynamic behavior and external environmental influences. Position information includes X, Y, and Z positions, which are used to describe the system's motion trajectory in space; attitude information includes pitch, roll, and yaw angles, which can be used to characterize the system's attitude changes and stability during navigation; hull status data includes engine speed, vibration frequency, and cabin temperature, reflecting the operating status of internal mechanical systems; environmental information includes wind speed, rainfall, air temperature, and water flow velocity, revealing the interference and challenges of the external natural environment on system operation; and the operational status level, as an indicator of the system's comprehensive performance, is the final basis for the status assessment model. By integrating these multi-source heterogeneous features, accurate assessment of the operational status of unmanned surface systems and intelligent decision-making can be effectively supported.

[0055] In order to enhance the model's sensitivity to time series changes, a sliding window mechanism is used to segment continuous time series data. Specifically, a fixed-length time window and a set step size are set. This embodiment sets a fixed length to select 10 consecutive time points, sets a step size of 1 for sliding, and extracts multiple overlapping time segments, each of which serves as a basic unit for subsequent processing. In view of the fact that the original features of each dimension may have inconsistent dimensions and significant differences in value ranges, all numerical features are normalized using a standardized method in the preprocessing stage. The normalization method uses standard deviation normalization, and the processing formula is shown in Formula 1.

[0056]

[0057] Among them, x i represents the original value of the i-th feature, μ i and σ i are the mean and standard deviation of the feature in the sample set, x i ′ represents the normalized eigenvalue.

[0058] The above processing process can ensure that the constructed time series samples not only retain the continuity and dynamic change characteristics of the data, but also have good model input consistency and stability, providing a data basis for subsequent dual-channel feature extraction and modeling.

[0059] S2. For the preprocessed information, extract the structural feature information and the convolution feature information respectively.

[0060] In step S2, the multidimensional observation data in each time window are input into two parallel feature extraction channels respectively to generate a fused multi-level feature representation.

[0061] S21, the first channel is used to extract structured features with physical interpretability.

[0062] This channel, based on modeling the physical behavior of key state variables, constructs statistical indicators reflecting system operational trends, fluctuation characteristics, and rates of change. This approach aims to characterize the macroscopic patterns of system change within a sliding window. To enhance the adaptability and explanatory power of the channel, a structured physical feature system, as shown in the table below, was constructed based on the typical operational characteristics of unmanned surface systems.

[0063]

[0064]

[0065] The calculation of each feature is based on standard statistics and mathematical transformation methods, involving first-order difference, range, standard deviation, linear fitting slope, kurtosis and other operations. Figure 1 The figure below illustrates the extraction process for X, Y, Z, Roll, Pitch, Yaw, engine speed, vibration frequency, cabin temperature, wind speed, rainfall, weather temperature, and water velocity. The specific mathematical expressions are shown in Formulas 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, and 14. These can be used to quantitatively characterize the evolution of different types of variables within a sliding window and serve as structured input vectors in subsequent model building.

[0066]

[0067]

[0068] Among them, X t Indicates the horizontal position value at the tth moment, Y t The vertical position value at the tth moment, Represents the average increment of longitudinal position, Z t The vertical position value at the tth moment, Roll t The roll angle at the tth moment, Represents the average value of the roll angle sequence in the sliding window, Pitch t Indicates the pitch angle at the tth moment, Yaw t Indicates the heading angle at the tth moment, RPMt The engine speed value at the tth moment, Represents the mean value of the engine speed within the sliding window, VB t represents the vibration frequency value at the t-th moment, Represents the mean value of the vibration frequency in the sliding window, TE t represents the cabin temperature at time t, WS t represents the wind speed at the tth moment, Represents the average value of the wind speed sequence in the sliding window, Rain t represents the rainfall at the tth moment, AT t represents the weather temperature at the tth moment, Indicates the average increment of weather temperature, CS t represents the water velocity at the tth moment, t represents the time step index, N represents the sliding window size, X_DM represents the mean of the first-order difference of the lateral position, Y_DSTD represents the standard deviation of the first-order difference of the longitudinal position, Z_RNG represents the vertical height range, RL_STD represents the standard deviation of the roll angle, PT_SLP represents the linear trend slope of the pitch angle, YW_DMAX represents the maximum value of the heading angle difference, RPM_STD represents the standard deviation of the engine speed, VB_KURT represents the vibration frequency kurtosis, TE_SLP represents the linear trend slope of the cabin temperature, WS_STD represents the standard deviation of the wind speed, RAIN_MEAN represents the mean of the rainfall, AT_DSTD represents the standard deviation of the first-order difference of the weather temperature, and CS_SLP represents the linear fitting slope of the water velocity.

[0069] S22, the second channel uses a temporal convolutional network to perform dynamic feature extraction on the multi-dimensional temporal data within the window.

[0070] The network combines the causal convolution and dilated convolution structures. On the one hand, it ensures that the convolution process does not introduce future information, which meets the causal requirements of time series modeling. On the other hand, it effectively captures long-range dependencies through the exponential expansion of the receptive field. The TCN convolution structure is shown in the attached figure. Figure 2 shown.

[0071] In long sequence modeling, traditional methods usually rely on extremely deep network structures or large-size convolution kernels to expand the receptive field. However, the former has high computational overhead, and the latter is prone to loss of local features. TCN introduces dilated causal convolution, which effectively improves the model's ability to express time series change patterns while maintaining computational efficiency. For a time series X = {x0, x1, x2, ..., x t-1 ,x t}, all its elements x t The dilated convolution operation is shown in Formula 15.

[0072]

[0073] Where F={f1,f2,...,f K-1 ,f K} represents the convolution kernel of size K, f k represents the convolution kernel, x t represents the sample value at time t in the input time series, d represents the expansion factor, C(F,x t ) represents the output result of the sequence after dilation convolution.

[0074] After processing through the structured physical feature channel, the raw time series data is compressed into low-dimensional vectors, focusing on depicting the overall trend and statistical characteristics of the variables within the window. Simultaneously, the data is mapped to higher dimensions through the time series convolutional network channel to extract local dynamic details and nonlinear evolution patterns. The feature vectors output by the two channels are concatenated and fused at the feature layer to form a unified fused feature representation, which serves as input to subsequent modeling modules to support the accurate classification and assessment of the operational status of unmanned surface systems.

[0075] S3. The structural feature information is combined with the CatBoost model to obtain the first state prediction probability distribution, and the convolutional feature information is combined with the GRU model to obtain the second state prediction probability distribution.

[0076] In the time series modeling channel, GRU is used to capture the implicit time dependency structure in the fusion features. GRU introduces a gating mechanism based on the traditional recurrent neural network (RNN). t and update gate z t Control the flow of information and reset the gate to control how much information of the previous state is written into the current candidate set The smaller the reset gate, the less information of the previous state is written. This allows the model to have the ability to model long-term and short-term dependencies while maintaining computational efficiency. Figure 3 As shown, its state update process is shown in Formula 16, Formula 17, Formula 18, and Formula 19.

[0077] z t =σ(W z ×[h t-1 ,x t ]) (16)

[0078] r t =σ(W r ×[h t-1 ,x t ]) (17)

[0079]

[0080] where x t Indicates the current input, h t represents the current hidden state, h t-1 represents the hidden state at the previous moment, z t represents the update gate vector, r t Represents the reset gate vector, W z ,W r ,W represent the weight parameter matrices of the update gate, reset gate and main network respectively, σ(·) represents the Sigmoid activation function, and tanh(·) represents the hyperbolic tangent function.

[0081] In the structural feature modeling channel, the CatBoost classification model based on gradient boosting is used to discriminate the fusion vector. The structure diagram is shown in the attached Figure 4 As shown in . This model integrates multiple symmetrical decision trees and uses an iterative optimization strategy to gradually approximate the objective function, thereby improving the accuracy and generalization of multi-category classification. A major advantage of CatBoost lies in its efficient encoding mechanism for categorical features: by introducing sequential statistical encoding and category prior information, it effectively alleviates target leakage and overfitting problems, significantly improving the model's performance in scenarios with small samples and imbalanced categories. During training, CatBoost aims to minimize the multi-category cross-entropy loss function, as shown in Equation 20.

[0082]

[0083] in An indicator variable indicating whether sample i belongs to category k, represents the k-th class probability predicted by the model, M represents the total number of samples, K represents the total number of categories, Represents the multi-class cross entropy loss function.

[0084] To further enhance model performance and stability, this paper introduces a Bayesian optimization mechanism to jointly search for key hyperparameters of the two sub-models, including the GRU hidden state dimension, learning rate, and number of training rounds, as well as the CatBoost tree depth, learning rate, and regularization term coefficient. Bayesian optimization models the target performance by constructing a proxy function and selects the most promising parameter combinations based on the acquisition function, enabling efficient automatic parameter tuning and avoiding the performance bottlenecks associated with traditional empirical parameter tuning.

[0085] S4. Use the covariance cross fusion algorithm to fuse the first state prediction probability distribution and the second state prediction probability distribution to obtain a joint decision result.

[0086] After the parallel modeling phase is complete, the system obtains the predicted probability distributions for each category state from the two submodels. Traditional weighted fusion methods are often influenced by the dominance of the majority class in tasks with uneven category distributions, resulting in insufficient recognition accuracy for minority class samples. This invention, however, employs a covariance cross-CI fusion method to jointly determine multiple prediction results, effectively enhancing the ability to discriminate against minority classes and mitigating the bias in the final prediction results caused by category skew. The formulas for the CI fusion algorithm are shown in Equations 21 and 22.

[0087]

[0088] Among them, P a is the covariance matrix of model A, P b is the covariance matrix of model B, ω∈[0,1] is the weighting coefficient, P CI is the uncertainty matrix (covariance) of the fused output, represents the output of model A, represents the output of model B, Represents the final estimated output result after fusion.

[0089] In order to quantitatively evaluate the performance of the model, the system finally uses classification indicators such as accuracy, precision, F1 score and recall rate to test the classification effect as core evaluation indicators, as shown in Formula 23, Formula 24, Formula 25 and Publication 26.

[0090]

[0091] Among them, TP represents the number of positive samples correctly identified by the model; TN represents the number of negative samples correctly identified by the model; FP represents the number of negative samples incorrectly predicted as positive samples by the model; FN represents the number of positive samples incorrectly predicted as negative samples by the model.

[0092] Simulation experiment:

[0093] During cruising, the unmanned system continuously collects multi-source data, including its position and attitude, hull state, and environmental disturbances, and uses this data as time-series input to determine whether it is currently in a stable and safe operating state. The system consists of three components: a perception module, a feature processing module, and a classification and evaluation module. The perception module is responsible for collecting real-time information such as GPS position, attitude angle, wind speed, water flow, rainfall, and cabin temperature to form a complete time-series data sequence. The feature processing module uses a dual-channel structure based on physical modeling and deep time-series modeling to extract structured statistical features and unstructured dynamic features from the input data and fuse them into a representation. The classification and evaluation module, consisting of a parallel CatBoost and GRU algorithm, performs state discrimination on the fused feature vectors. Finally, a fusion strategy is used to output the final system safety level and its confidence level.

[0094] This experiment selected 2923 original samples, with a total of 13 features, including position X, position Y, position Z, pitch angle Pitch, roll angle Roll, yaw angle Yaw, hull engine speed, vibration frequency, cabin temperature, wind speed, rainfall, temperature, water flow velocity, and 5 operating status levels, of which level 1 is the highest safety and level 5 is the lowest safety. 2914 groups of time series samples were generated through the sliding window construction method, with a window size of 10 and a step size of 1. 80% was used as the training set, totaling 2331 groups, and 20% was used as the test set, totaling 583 groups. The classification result evaluation indicators of the data set are shown in the following table, and the accuracy distribution diagram of the model in the classification task is shown in the attached figure. Figure 6 As shown in the attached figure, the classification ROC curve of the model is shown in the attached figure. Figure 7 shown.

[0095]

[0096] As shown in the table above, the proposed model achieved an overall performance of 86.11% accuracy, 87.00% precision, 86.11% recall, and 86.29% F1 score on the test set, demonstrating high accuracy and stability in multi-category state discrimination. The high precision and recall reflect the model's excellent ability to distinguish and cover different categories, while the balanced F1 score further demonstrates the robustness of the fusion model in handling inter-class imbalances, demonstrating the stability and practicality of the proposed model in state assessment tasks for unmanned surface systems.

[0097] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered by the scope of protection of the claims of the present invention.

Claims

1. A method for automatic assessment of the state of unmanned surface systems based on parallel extraction and fusion modeling of multi-dimensional features, characterized in that: The specific method is as follows: Obtain the status information of the unmanned ship and perform preprocessing; For the preprocessed information, structural feature information and convolution feature information are extracted respectively; The structural feature information is combined with the CatBoost model to obtain the first state prediction probability distribution, and the convolutional feature information is combined with the GRU model to obtain the second state prediction probability distribution; The covariance cross fusion algorithm is used to fuse the first state prediction probability distribution and the second state prediction probability distribution to obtain the joint decision result.

2. The method for automatically assessing the state of an unmanned surface system based on parallel extraction and fusion modeling of multi-dimensional features according to claim 1, characterized in that: The status information of the unmanned ship includes: position data, attitude data, hull data, environmental data and operating status level.

3. The method for automatically assessing the state of an unmanned surface system based on parallel extraction and fusion modeling of multi-dimensional features according to claim 2, characterized in that: Perform preprocessing, the specific method is as follows: The moving window mechanism is used to segment the continuous time series data, and a fixed-length time window and a set step size are set to slide to extract several overlapping time segments. All numerical features are normalized using the standardization method. The specific formula is as follows: Where x i represents the original value of the i-th feature, μ i and σ i are the mean and standard deviation of the feature in the sample set, x′ i represents the normalized eigenvalue.

4. The method for automatically assessing the state of an unmanned surface system based on parallel extraction and fusion modeling of multi-dimensional features according to claim 1, characterized in that: Extract structural feature information. The specific content and calculation formula are as follows: Where, X t Indicates the horizontal position value at the tth moment, Y t The vertical position value at the tth moment, Represents the average increment of longitudinal position, Z t The vertical position value at the tth moment, Roll t The roll angle at the tth moment, Represents the average value of the roll angle sequence in the sliding window, Pitch t Indicates the pitch angle at the tth moment, Yaw t Indicates the heading angle at the tth moment, RPM t The engine speed value at the tth moment, Represents the mean value of the engine speed within the sliding window, VB t represents the vibration frequency value at the t-th moment, Represents the mean value of the vibration frequency in the sliding window, TE t represents the cabin temperature at time t, WS t represents the wind speed at the tth moment, Represents the average value of the wind speed sequence in the sliding window, Rain t represents the rainfall at the tth moment, AT t represents the weather temperature at the tth moment, Indicates the average increment of weather temperature, CS t represents the water velocity at the tth moment, t represents the time step index, N represents the sliding window size, X_DM represents the mean of the first-order difference of the lateral position, Y_DSTD represents the standard deviation of the first-order difference of the longitudinal position, Z_RNG represents the vertical height range, RL_STD represents the standard deviation of the roll angle, PT_SLP represents the linear trend slope of the pitch angle, YW_DMAX represents the maximum value of the heading angle difference, RPM_STD represents the standard deviation of the engine speed, VB_KURT represents the vibration frequency kurtosis, TE_SLP represents the linear trend slope of the cabin temperature, WS_STD represents the standard deviation of the wind speed, RAIN_MEAN represents the mean of the rainfall, AT_DSTD represents the standard deviation of the first-order difference of the weather temperature, and CS_SLP represents the linear fitting slope of the water velocity.

5. The method for automatically assessing the state of an unmanned surface system based on parallel extraction and fusion modeling of multi-dimensional features according to claim 4 is characterized in that: The CatBoost model takes minimizing the multi-classification cross entropy loss function as its objective function. The specific formula is as follows: in An indicator variable indicating whether sample i belongs to category k, represents the k-th class probability predicted by the model, M represents the total number of samples, K represents the total number of categories, Represents the multi-class cross entropy loss function.

6. The method for automatically evaluating the state of an unmanned surface system based on parallel extraction and fusion modeling of multi-dimensional features according to claim 1, characterized in that: Extract convolution feature information. The specific method is as follows: Assume that the time series X={x0,x1,x2,...,x t-1 ,x t }, the dilated convolution operation formula of its element xt is: Where, F={f1,f2,...,f K-1 ,f K } represents the convolution kernel of size K, f k represents the convolution kernel, x t represents the sample value at time t in the input time series, d represents the expansion factor, C(F,x t ) represents the output result of the sequence after dilation convolution.

7. The method for automatically assessing the state of an unmanned surface system based on parallel extraction and fusion modeling of multi-dimensional features according to claim 6, characterized in that: The state update formula of the GRU model is as follows: z t =σ(W z ×[h t-1 ,x t ]) r t =σ(W r ×[h t-1 ,x t ]) Where x t Indicates the current input, h t represents the current hidden state, h t-1 represents the hidden state at the previous moment, z t represents the update gate vector, r t Represents the reset gate vector, W z ,W r ,W represent the weight parameter matrices of the update gate, reset gate and main network respectively, σ(·) represents the Sigmoid activation function, and tanh(·) represents the hyperbolic tangent function.

8. The method for automatically assessing the state of an unmanned surface system based on parallel extraction and fusion modeling of multi-dimensional features as described in claims 5 and 7 is characterized in that: The hyperparameters of the CatBoost model and the GRU model are optimized using the Bayesian optimization algorithm. The specific method is as follows: Modeling target performance by constructing a surrogate function; The most promising parameter combination is selected based on the acquisition function to achieve automatic hyperparameter tuning.

9. The method for automatically assessing the state of an unmanned surface system based on parallel extraction and fusion modeling of multi-dimensional features as described in claim 1 is characterized in that: The covariance cross fusion algorithm is used to fuse the first state prediction probability distribution and the second state prediction probability distribution. The specific formula of the fusion algorithm is as follows: Among them, P a is the covariance matrix of model A, P b is the covariance matrix of model B, ω∈[0,1] is the weighting coefficient, P CI is the uncertainty matrix of the fused output, represents the output of model A, represents the output of model B, Represents the final estimated output result after fusion.

10. The method for automatically evaluating the state of an unmanned surface system based on parallel extraction and fusion modeling of multi-dimensional features as claimed in claim 1, characterized in that: The CatBoost model and GRU model use accuracy, precision, F1 score, and recall as evaluation indicators. The specific formulas are as follows: Where TP represents the number of positive samples correctly identified by the model; TN represents the number of negative samples correctly identified by the model; FP represents the number of negative samples incorrectly predicted as positive samples by the model; and FN represents the number of positive samples incorrectly predicted as negative samples by the model.

Images