Neural network training and application method for forecasting workability of engineering ship

Abstract

The invention relates to the technical field of ship operation forecasting, and discloses a neural network training and application method for engineering ship workability forecasting. According to the method, a position sensor device and an environment sensor device are arranged on an engineering ship, and the current position and surrounding environment monitoring data of the ship are obtained in real time; an operation plan acquisition device is arranged in the ship control system to acquire current operation plan data; generating an input feature set from the data, inputting the trained neural network model, and calculating to generate an workability forecasting result; and comparing the forecast result with a preset judgment standard, and judging the operation state. According to the method, multi-source data are integrated, complex association among the data is mined by using the neural network model, accurate forecasting of the workability of the engineering ship is realized, a scientific basis is provided for operation decision making, the method is adaptive to a complex operation environment and diversified tasks, and engineering ship operation is assisted to be carried out more efficiently and safely.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of ship operation prediction, in particular to a neural network training and application method for predicting the operability of engineering ships. BACKGROUND

[0002] Engineering ships play a crucial role in port construction, ocean engineering, and channel dredging, and their operation efficiency and safety directly affect the progress and cost of engineering projects. However, the operation process of engineering ships is influenced by a variety of complex factors, and the uncertainty of the operating environment poses great challenges to the efficient and safe operation of the ships.

[0003] Currently, the operation arrangement of engineering ships relies on experience or simple environmental parameter evaluation, which cannot fully consider the correlation and influence between various factors. For example, when performing dredging operations, relying solely on real-time wind speed or water flow speed to determine whether it is suitable for operation may overlook the matching degree of the ship's own position and the terrain of the operating area, or the special requirements of different processes in the operation plan for environmental conditions. This one-sided judgment method not only may lead to low operation efficiency and missed optimal operation opportunities, but also may force operation under unsuitable environmental conditions, causing equipment damage, personnel safety risks, etc.

[0004] With the continuous development of ocean engineering, the operating environment of engineering ships is becoming increasingly complex, and the operation tasks are becoming more diversified. The traditional operation judgment method has been difficult to meet the actual needs. Although data collection and analysis methods have been introduced in some scenarios, data processing is mostly limited to simple statistics and comparison, and cannot deeply explore the internal relationship between different parameters, making it difficult to achieve accurate and timely prediction of the operability of engineering ships. Therefore, how to use advanced technical means, integrate multi-source data, and build an efficient prediction model to realize scientific evaluation of the operability of engineering ships has become a problem to be solved in the field of engineering ship operation. SUMMARY

[0005] The present application aims to provide a neural network training and application method for predicting the operability of engineering ships to solve the problems raised in the background.

[0006] To achieve the above-mentioned purpose, the present application provides a neural network training and application method for predicting the operability of engineering ships, which comprises:

[0007] A position sensor device and an environmental sensor device are arranged on the engineering ship, the position sensor device is used to obtain the current position of the engineering ship in real time, and the environmental sensor device is used to obtain environmental monitoring data around the engineering ship in real time;

[0008] An operation plan acquisition device is installed in the control system of the engineering vessel, which is used to acquire the current operation plan data of the engineering vessel;

[0009] An input feature set is generated using the current location of the engineering vessel, the environmental monitoring data, and the current work plan data.

[0010] The input feature set is input into the trained neural network model to calculate and generate the workability prediction results for engineering vessels;

[0011] The operational status of the engineering vessel is determined by comparing the predicted operational status with the preset judgment criteria.

[0012] Preferably, the step of generating an input feature set using the current position of the engineering vessel, the environmental monitoring data, and the current work plan data includes:

[0013] A position feature sequence is generated based on the current position of the engineering vessel;

[0014] An environmental feature sequence is generated based on the environmental monitoring data.

[0015] A plan feature sequence is generated based on the current work plan data;

[0016] The location feature sequence, the environmental feature sequence, and the plan feature sequence are merged into the input feature set.

[0017] Preferably, the step of calculating and generating a position feature sequence based on the current position of the engineering vessel includes:

[0018] Obtain historical location data of engineering vessels;

[0019] The movement trajectory features are calculated and generated using the current position and historical position data of the engineering vessel.

[0020] The movement trajectory features are normalized into the position feature sequence.

[0021] Preferably, the step of calculating and generating an environmental feature sequence based on the environmental monitoring data includes:

[0022] Acquire real-time wind speed and wave height data for engineering vessels;

[0023] Environmental fluctuation characteristics are calculated and generated using the real-time wind speed data and the real-time wave height data.

[0024] The environmental fluctuation characteristics are standardized into the environmental characteristic sequence.

[0025] Preferably, the step of calculating and generating a plan feature sequence based on the current work plan data includes:

[0026] Acquire data on the target operating area and expected operating time of engineering vessels;

[0027] The operation requirement characteristics are calculated and generated using the target operation area data and the expected operation time data.

[0028] The job requirement features are encoded into the plan feature sequence.

[0029] Preferably, the training of the completed neural network model includes the following steps:

[0030] Acquire multiple sets of historical operational data of engineering vessels during normal operation. Each set of historical operational data includes historical location data, historical environmental monitoring data, and historical operation plan data of the engineering vessels.

[0031] Training feature vectors are generated by calculating historical location data, historical environmental monitoring data, and historical work plan data from each set of historical operational data.

[0032] The initial neural network model is trained using all training feature vectors to generate the trained neural network model.

[0033] Preferably, the step of calculating and generating training feature vectors from historical location data, historical environmental monitoring data, and historical work plan data in each set of historical operational data includes:

[0034] Historical location features are calculated and generated based on the historical location data.

[0035] Historical environmental characteristics are calculated and generated based on the historical environmental monitoring data.

[0036] Historical plan features are calculated and generated based on the historical work plan data;

[0037] The historical location features, historical environment features, and historical plan features are combined into the training feature vector.

[0038] Preferably, training the initial neural network model using all training feature vectors includes:

[0039] The input layer of the neural network model is configured to receive the training feature vector;

[0040] The training feature vectors are processed through the hidden layers of the neural network model to generate hidden features;

[0041] The prediction result is generated based on the hidden features through the output layer of the neural network model;

[0042] Adjust the weight parameters of the neural network model to optimize the prediction results.

[0043] Preferably, the application of a trained neural network model for predicting the workability of engineering vessels includes:

[0044] The input feature set is input into the input layer of the trained neural network model;

[0045] The input feature set is processed by the hidden layer of the trained neural network model to generate intermediate features;

[0046] The output layer of the trained neural network model generates the workability forecast result of the engineering vessel based on the intermediate features.

[0047] Preferably, comparing the workability forecast result of the engineering vessel with the preset judgment criteria includes:

[0048] Obtain the preset job probability threshold;

[0049] Compare the workability forecast results of the engineering vessel with the work probability threshold;

[0050] When the predicted workability of the engineering vessel meets the work probability threshold, the work status of the engineering vessel is determined to be feasible; otherwise, the work status of the engineering vessel is determined to be infeasible.

[0051] Compared with the prior art, the beneficial effects of the present invention are:

[0052] By installing position and environmental sensors on engineering vessels, real-time and comprehensive data on the vessel's current position and surrounding environment can be acquired, providing rich and accurate foundational information for subsequent workability forecasts. Real-time position data acquisition clearly reveals the relative position of the vessel to the work area, preventing misjudgments due to positional errors. Real-time environmental monitoring data collection covers various key factors affecting operations, such as wind speed, water flow, and waves, ensuring comprehensive awareness of the work environment.

[0053] The setup of the work plan acquisition device incorporates the current work plan data of engineering vessels, making workability forecasts no longer limited to environmental and location factors, but closely integrated with specific operational tasks. Different work plans have different requirements for environmental conditions and vessel positions. Integrating work plan data into the input feature set allows the neural network model to better match actual operational needs, resulting in more targeted and practical forecasts.

[0054] By generating an input feature set and feeding it into a trained neural network model, intelligent forecasting of the workability of engineering vessels was achieved. The neural network model possesses powerful nonlinear fitting capabilities, enabling it to deeply explore the complex correlations between multi-dimensional data such as location, environment, and operational plans. Compared to traditional experience-based judgments or simple data comparisons, its forecast results are more objective and accurate. This intelligent forecasting method effectively reduces judgment errors caused by human factors and improves the scientific rigor of workability assessments.

[0055] By comparing forecast results with preset judgment criteria to determine the operational status, a clear basis for operational decisions is provided. Based on the judgment results, operators can rationally arrange operation time and methods, making full use of favorable conditions to improve operational efficiency when operations are suitable, and adjusting plans in a timely manner when operations are unsuitable to avoid unnecessary resource consumption and risks. This data- and model-based operational status determination enables engineering vessel operations to be more rational and orderly, adapting to complex and ever-changing operating environments and diverse operational tasks, and promoting the development of engineering vessel operations towards a more efficient and safer direction. Attached Figure Description

[0056] Figure 1 This is a schematic diagram illustrating the working principle of the neural network training and application method for predicting the workability of engineering vessels as described in this invention.

[0057] Figure 2 A flowchart for generating the input feature set;

[0058] Figure 3 A flowchart for generating positional feature sequences;

[0059] Figure 4 A flowchart for training a neural network model;

[0060] Figure 5 A flowchart for generating training feature vectors. Detailed Implementation

[0061] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0062] Please see Figure 1 This invention provides a neural network training and application method for predicting the workability of engineering vessels, the method comprising:

[0063] Position sensors and environmental sensors are installed on the engineering vessel. The position sensors continuously collect the vessel's current position information, while the environmental sensors continuously collect environmental monitoring data around the vessel. A work plan acquisition device is configured in the vessel's control system to acquire the vessel's current work plan data. Subsequently, based on the vessel's current position, environmental monitoring data, and current work plan data, a comprehensive input feature set is generated. This input feature set is fed into a pre-trained neural network model. The neural network model processes the input feature set and outputs a prediction of the vessel's workability. Finally, this prediction result is compared and analyzed with preset judgment criteria to determine the vessel's current operational status.

[0064] Example 1: See Figure 2 The first step in this process is generating a sequence of positional features based on the current position information of the engineering vessel. Position sensor devices, such as GPS receivers, continuously collect the vessel's latitude and longitude coordinates, heading angle, and speed information. These real-time data streams are transmitted to the data processing unit. The calculations not only target the instantaneous position but also incorporate recent historical position data. The data processing unit maintains a cache queue containing position information from several past time points. By analyzing the sequence relationship between the current and historical positions, a set of derived parameters characterizing the vessel's motion state can be calculated. These parameters include, but are not limited to: the vessel's lateral and longitudinal deviations relative to the planned route, the average speed over a past period, the rate of change of heading, and the curvature of the position's trajectory. The calculated raw position parameters typically have different dimensions and numerical ranges; for example, latitude and longitude coordinates and speed values ​​differ significantly in magnitude. Therefore, normalization must be performed. Normalization linearly maps the value of each position parameter to a preset common range, such as zero to one. This process eliminates dimensional differences and prevents certain features from dominating subsequent model calculations due to excessively large values. All normalized position parameters are arranged in a predefined order, forming a one-dimensional numerical array, i.e., a position feature sequence. This sequence encapsulates the ship's current dynamic positioning status and motion trend in a standardized form.

[0065] Generating environmental feature sequences based on environmental monitoring data is a parallel processing flow. Environmental sensor devices, including anemometers, wave sensors, and current meters, continuously collect real-time environmental elements of the sea area where ships operate. This raw monitoring data, such as instantaneous wind speed, gust wind speed, average wave height, maximum wave height, wave period, and current speed and direction, is transmitted to the same or collaboratively working data processing unit. The computational operations focus on extracting key features from the raw data that effectively characterize the impact of environmental conditions on ship operations. For example, combining wind speed and wave height data can calculate environmental disturbance energy; analyzing wave spectrum data can yield parameters characterizing wave concentration; and statistically analyzing the standard deviation of environmental parameters over a past period can reflect environmental stability. These calculated environmental fluctuation characteristics also suffer from inconsistencies in units, such as wind speed being measured in meters per second (m / s) and wave height in meters (m). Therefore, standardization is necessary. Standardization typically employs the Z-score method, which calculates the average and standard deviation of each environmental feature parameter over a relatively long historical time window, then subtracts the average from the current value and divides by the standard deviation. This transforms each environmental characteristic parameter into a value with a mean of zero and a standard deviation of one. All standardized environmental parameters are arranged in a predefined order, forming another one-dimensional numerical array, namely the environmental characteristic sequence. This sequence quantifies the severity of the current external environmental conditions and their fluctuation characteristics.

[0066] The generation of a plan feature sequence based on current work plan data constitutes the third independent processing branch. The work plan acquisition device reads current task instructions or plan information from the engineering vessel's control system. This data typically includes the latitude and longitude coordinates of the target work area's center point, the scope of the work area (e.g., radius or boundary coordinates), the planned start and end times of the work, the expected work content (e.g., dredging, lifting, installation, etc.), and the work accuracy requirements. The calculation operation aims to convert this typically structured plan information into numerical features that the model can process. For example, it calculates the distance and bearing between the vessel's current position and the center point of the target work area; calculates the remaining time interval from the current time to the planned work start time; and maps these intervals to different numerical codes based on the work content type and accuracy requirements. These calculated work requirement features require further encoding processing. For numerical features, such as distance and time interval, appropriate scaling may be applied. For categorical features, such as work type, one-hot encoding is typically used, converting them into a binary vector containing only 0s and 1s, where each bit represents a work type. All encoded plan parameters are arranged in a predefined order, forming the third one-dimensional numerical array, i.e., the plan feature sequence. This sequence expresses the specific requirements and constraints that the task places on ship operation in numerical form.

[0067] After the position feature sequence, environmental feature sequence, and planning feature sequence are generated, a merging operation is performed to create the final input feature set. This merging operation is implemented in the data processing unit using vector concatenation. Specifically, all elements of the position feature sequence, environmental feature sequence, and planning feature sequence are concatenated end-to-end in a strictly predefined order to form a longer one-dimensional numerical vector. This merged one-dimensional vector is the input feature set, which integrates all information from the ship's real-time status, external environmental conditions, and mission plan requirements. The consistency of the order during the merging process is crucial; it must perfectly match the input feature arrangement expected by the neural network model during training. Otherwise, the model will interpret the data incorrectly. The generated input feature set is then fed into the subsequent neural network model for computation to produce the final operational forecast result. The entire process from raw data acquisition to feature set generation is automated and continuous, ensuring the real-time performance of the forecast system.

[0068] Example 2: See Figure 3The generation of position feature sequences begins with the acquisition and utilization of historical position data. Position sensor devices continuously record the latitude and longitude coordinates, ground speed, and heading angle of the engineering vessel. This data is stored in the vessel's navigation data recording unit in the form of a time series. When workability forecasting is required, the data processing unit retrieves historical position data from the current moment and a specific time window prior to it. The length of this time window is pre-set according to the vessel's motion characteristics and forecasting requirements, and may cover navigation data from the past tens of minutes to several hours. The core of the computational operation lies in extracting features from these raw position time series that can effectively characterize the vessel's motion state and trends. For example, by analyzing continuous position points, the displacement components of the vessel in the east-west and north-south directions can be calculated, thereby obtaining the average speed. By analyzing changes in the heading angle, the vessel's turning rate can be calculated. By comparing the deviation between the actual navigation trajectory and the planned route, lateral and longitudinal deviations can be calculated. The radius of curvature of the trajectory can also be calculated to reflect the flexibility or constraint of vessel operation. These calculated trajectory characteristic parameters differ significantly in numerical range and units. Speed ​​may be measured in knots, latitude and longitude deviations in degrees or minutes, and turning rates in degrees per second. This inconsistency in units can cause difficulties for subsequent neural network model processing. Therefore, normalization must be performed. Normalization linearly transforms the value of each trajectory feature parameter to a common numerical range, such as between zero and one. Specifically, for each feature type, a possible maximum and minimum value are set, and then the actual value is subtracted from the minimum value and divided by the difference between the maximum and minimum values. After this processing, all trajectory features are transformed into dimensionless, scale-uniform values. These normalized values ​​are arranged in a predefined fixed order, forming a one-dimensional numerical array, which constitutes the position feature sequence. This sequence encodes the ship's recent motion state and trend in a standardized, dimension-uniform manner.

[0069] The generation process of environmental feature sequences focuses on the processing and analysis of real-time environmental monitoring data. Environmental sensor devices, including anemometers, wave radars, and current meters installed on the ship, continuously collect environmental data around the vessel. Anemometers record instantaneous wind speed, average wind speed, and wind direction; wave radars measure significant wave height, maximum wave height, and wave period; and current meters provide information on current speed and direction. This real-time environmental monitoring data is transmitted to the data processing unit. The computational operations aim to extract features from this raw environmental data that characterize the degree to which environmental conditions affect ship operations. For example, by combining wind speed and direction data, relative wind speed (i.e., the vector difference between the real wind and the ship's wind) can be calculated; wave energy can be calculated by analyzing wave height and wave period; and the fluctuation characteristics of environmental conditions can be obtained by statistically analyzing changes in environmental parameters over a period of time. These calculated environmental fluctuation characteristics also suffer from a lack of dimensional consistency. Wind speed is measured in meters per second, wave height in meters, and wave period in seconds, while the calculated wave energy may have more complex units. This dimensional difference affects the processing performance of the neural network model. Therefore, standardization is necessary. Standardization typically employs the Z-score method, which calculates the mean and standard deviation for each environmental characteristic parameter based on long-term historical observation data. The current value is then subtracted from the mean and divided by the standard deviation. This process transforms each environmental characteristic parameter into a value with a mean of zero and a standard deviation of one. All standardized environmental parameters are arranged in a predefined, fixed order, forming another one-dimensional numerical array, thus creating the environmental characteristic sequence. This sequence quantifies the severity and fluctuation characteristics of the current environmental conditions in a standardized and dimensionally consistent manner.

[0070] Although the generation processes of the location feature sequence and the environmental feature sequence involve different data types, they both follow similar data processing logic: meaningful features are extracted from the raw data, then dimensional differences are eliminated through standardization, and finally, they are arranged in a fixed order to form a feature sequence. These two sequences provide key input information for subsequent operational forecasting from two dimensions: the ship's own motion state and external environmental conditions, respectively. The entire processing is automated, ensuring the real-time performance and reliability of the forecasting system.

[0071] Example 3: See Figure 4The generation of the planned feature sequence begins with the parsing and transformation of the current work plan data. The work plan acquisition device reads the current task instructions from the engineering vessel's control system. This data is typically stored in a structured format, containing metadata such as the geometric description of the target work area (e.g., the latitude and longitude coordinates of the center point and the area radius, or a sequence of coordinates of multiple boundary points), the planned work start time and duration, the work type identifier, and the work accuracy level. The computational operations aim to transform this heterogeneous planning information into a set of numerical work requirement features. For example, calculating the spherical distance between the vessel's current position and the center point of the target work area, which needs to consider the influence of the Earth's curvature; calculating the difference between the current UTC time and the planned work start time to obtain the remaining preparation time in seconds; and classifying and encoding these features according to the work type and accuracy requirements. These calculated work requirement features have different properties and require encoding processing. For continuous numerical features, such as distance and time intervals, min-max scaling is typically used to map them to a range of zero to one. For categorical features, such as work type, one-hot encoding is used to create a binary vector where each bit corresponds to a possible work type, with the corresponding bit set to 1 and the rest to 0. All coded planning parameters are arranged in a predefined, fixed order and combined into a one-dimensional numerical array, namely the planning feature sequence. This sequence fully expresses the spatial, temporal, and functional requirements of the operational task for the ship in a machine-readable numerical form.

[0072] The acquisition of the trained neural network model is a data-driven process relying on a large amount of historical operational data. Multiple sets of historical operational data recorded by engineering vessels during their previous normal operations are collected. Each set of data constitutes a complete data sample, containing historical location data (latitude, longitude, speed, and heading) recorded at a specific historical point in time, historical environmental monitoring data (wind speed, wave height, cycle, etc.), and historical operational plan data being executed at that time. For each set of historical operational data, a corresponding training feature vector needs to be generated. This process is completely consistent with the process of generating the input feature set in the application phase: based on historical location data, historical location features are generated by calculating and normalizing movement trajectory features; based on historical environmental monitoring data, historical environmental features are generated by calculating and standardizing environmental fluctuation features; based on historical operational plan data, historical operational requirement features are generated by calculating and encoding operational requirements features. These three features are then combined into a unified training feature vector. The training feature vectors generated from all historical samples constitute the training dataset, with each vector corresponding to a known actual operational status label.

[0073] Training the initial neural network model using all training feature vectors involves defining the model structure and optimizing parameters. The initial neural network model typically employs a multilayer perceptron structure, including an input layer, hidden layers, and an output layer. The number of neurons in the input layer is exactly the same as the dimension of the training feature vectors. The hidden layer contains several neurons, the number of which is determined by hyperparameter tuning. Each neuron performs a linear weighted sum of its inputs and then applies a non-linear activation function. The design of the output layer depends on the nature of the prediction task; for binary classification problems, a single neuron with a sigmoid activation function is typically used, outputting a task probability value between zero and one.

[0074] The training process optimizes model parameters by minimizing the difference between the predicted output and the true label. A loss function is defined to measure this difference; for binary classification problems, binary cross-entropy loss is commonly used. for:

[0075]

[0076] Where: N represents the total number of training samples, y i This represents the true label of the i-th sample (0 indicates no job, 1 indicates job). Let represent the predicted output probability of the model for the i-th sample, and log represent the natural logarithm function. The gradient of the loss function with respect to the parameters of each layer of the model is calculated using the backpropagation algorithm, and the weights and bias parameters are iteratively updated using the gradient descent optimization algorithm, gradually reducing the loss function value and improving the model's prediction accuracy. This process continues until the model performance stabilizes on the validation set, ultimately resulting in a trained neural network model with generalization capabilities, which can be used for operational prediction of new real-time data.

[0077] Example 4: See Figure 5The generation of training feature vectors requires systematic processing of each set of historical operational data. Based on historical position data, calculations are performed to generate historical position features. This process calls upon historical position data records, including the ship's latitude and longitude coordinates, speed, and heading information within a specific time window. By analyzing this time-series data, a set of parameters characterizing the ship's historical motion state is calculated, such as average speed, heading stability, and position deviation. These parameters are normalized to form standardized historical position feature vectors. Based on historical environmental monitoring data from the historical operational data, calculations are performed to generate historical environmental features. This process processes environmental parameters in the records, including measurements of wind speed, wave height, and wave period. Features reflecting historical environmental conditions, such as average wind speed, maximum wave height, and environmental parameter volatility, are extracted from these raw data. These features are standardized to eliminate dimensional differences, forming historical environmental feature vectors. Based on historical operational plan data from the historical operational data, calculations are performed to generate historical plan features. This process parses historical operational plan records, including information such as target operational area, planned operational time, and operational type. A set of parameters characterizing task requirements is calculated, such as planned task distance, time urgency, and task complexity. These parameters are appropriately encoded and converted into numerical features, forming historical plan feature vectors. After generating historical location features, historical environment features, and historical plan features separately, a combination operation is performed. This operation concatenates these three feature vectors in a predefined order into a longer one-dimensional vector, the training feature vector. Each training feature vector corresponds to a complete state description at a historical moment and is associated with the actual task state label at that moment.

[0078] Training the initial neural network model using all generated training feature vectors is a systematic process. The input layer structure of the neural network model is configured such that the number of neurons perfectly matches the dimension of the training feature vectors, ensuring that all feature values ​​of each training feature vector are fully received. The training feature vectors pass through the input layer to the hidden layers of the neural network model. The hidden layers typically consist of multiple fully connected layers, where each neuron performs a weighted summation of its received input and applies a non-linear activation function. The hidden layers perform multi-level non-linear transformations and feature extraction operations on the input training feature vectors, gradually abstracting deep patterns and regularities from the input data to generate hidden features. These hidden features are then passed to the output layer of the neural network model. The output layer performs computational operations based on the received hidden features to generate prediction results. For binary classification problems, the output layer typically uses the sigmoid activation function, mapping the output value to between zero and one, representing the predicted job probability. The loss value is calculated based on the difference between the predicted result and the actual label; the loss function measures the degree of error in the model's prediction. The gradient of the loss function with respect to all model parameters is calculated using the backpropagation algorithm, indicating the direction and magnitude of parameter adjustments. Based on the calculated gradient, an optimization algorithm is used to adjust the weight parameters of the neural network model, gradually reducing the loss function value so that the model's predictions become increasingly closer to the true labels. This process is repeated multiple times until the model's performance on the training data reaches a satisfactory level (see Table 1).

[0079] Table 1: Training Feature Vector Table.

[0080] Feature Type Feature Name Numerical Treatment Location Feature Normalized Average Speed 0.75 Min-Max Normalization Location Feature Normalized Heading Stability 0.82 Min-Max Normalization Location Feature Normalized Lateral Deviation 0.15 Min-Max Normalization Environmental Feature Normalized Wind Speed 1.25 Z-score Standardization Environmental Feature Normalized Wave Height 0.87 Z-score Standardization Environmental Feature Normalized Wave Volatility -0.32 Z-score Standardization Planning Feature Encoded Job Type [0,1,0] One-Hot Encoding Planning Feature Scaled Job Distance 0.68 Min-Max Scaling Planning Feature Scaled Time Urgency 0.45 Min-Max Scaling

[0081] During training, each training feature vector is input into the network as a complete sample. The model calculates predictions through forward propagation and adjusts parameters through backpropagation. The sufficiency and representativeness of the training data have a crucial impact on model performance, requiring coverage of various possible job scenarios and environmental conditions. Hyperparameters such as the number of iterations and learning rate during training need to be adjusted using validation set performance to avoid overfitting or underfitting. The final trained model is able to learn the complex mapping relationship from input features to job states.

[0082] Example 5: This describes the practical application process of a trained neural network model and the specific operational procedures for determining the operational status based on its output. The process begins with the input of a real-time generated input feature set. This input feature set is composed of a sequence of positional features, an environmental feature sequence, and a planned feature sequence, containing information about the current status, environmental conditions, and task requirements of the engineering vessel. This input feature set is transmitted to the input layer of the trained neural network model. The input layer, acting as the model's interface, has a number of neurons that perfectly match the dimension of the input feature set, ensuring that all feature values ​​are fully received. After receiving the input feature set, the input layer passes its internal data to the next layer, namely the model's hidden layers.

[0083] The hidden layers of a trained neural network model process the received input feature set. Hidden layers typically contain one or more computational layers, each consisting of multiple neurons. Each neuron performs a linear weighted summation of its received input data and applies a pre-defined nonlinear activation function to the sum. This transformation introduces the model's nonlinear expressive power. Through its multi-layered structure, the hidden layers extract and abstract the input feature set layer by layer. Each neuron, based on its weight parameters and activation function, performs nonlinear combinations and transformations of the input data, progressively extracting higher-level, more abstract feature representations. After all layers of the hidden layers have processed the data, the original input features are transformed into a set of internal representations—intermediate features—that more fundamentally characterize the relationship between the current ship's operational status and environmental conditions. These intermediate features, as the output of the hidden layers, are passed to the next layer of the neural network model—the output layer.

[0084] The output layer of the trained neural network model performs computational operations based on the received intermediate features. The structural design of the output layer depends on the specific requirements of the forecasting task. In the scenario of predicting the workability of engineering vessels, a binary classification output structure is typically used. The output layer receives the intermediate feature vectors from the hidden layers as its input. A neuron in the output layer (usually one) performs a linear weighted summation of its inputs. Then, an activation function is applied to the summation result. For probability outputs, the sigmoid activation function is a common choice, which maps any real number to a value between zero and one. Through its computational operations, the output layer maps the intermediate features to a specific numerical output, namely the workability forecast result of the engineering vessel. This forecast result is a value between zero and one, and is usually interpreted as a probability estimate that the engineering vessel can safely and effectively perform the predetermined operational task under the conditions described by the current input features. This forecast result is output for use in subsequent decision-making processes.

[0085] Comparing the predicted workability of engineering vessels with preset judgment criteria is a crucial step in determining the operational status. First, the preset operational probability threshold is obtained. This threshold is a pre-defined, fixed numerical constant stored in the configuration file or database of the engineering vessel's control system. It represents the minimum probability requirement for the system to determine that the operational status is feasible; its specific value is usually determined based on domain knowledge, safety regulations, or historical operational experience. For example, this threshold might be set to 0.85, meaning that the system considers the operational status feasible only when the predicted operational probability reaches or exceeds 85%. After obtaining the operational probability threshold, a comparison operation is performed. This operation directly compares the predicted workability result (a specific probability value) output by the neural network model with the obtained operational probability threshold. This comparison is a simple arithmetic comparison, determining whether the predicted result is greater than or equal to the threshold.

[0086] Based on the results of the comparison operation, the system determines the operational status of the engineering vessel. The determination logic is a clear binary decision: if the comparison operation indicates that the predicted workability of the engineering vessel is greater than or equal to a preset operational probability threshold, then the operational status of the engineering vessel is determined to be feasible. This means that under the current vessel position, environmental conditions, and operational plan, the system predicts that the vessel can safely and effectively perform the operational task. Conversely, if the comparison operation indicates that the predicted workability of the engineering vessel is less than the preset operational probability threshold, then the operational status of the engineering vessel is determined to be infeasible. This means that the system predicts that performing the operational task under the current conditions carries a high risk or difficulty and is not suitable for operation. The determined operational status result (feasible or infeasible) is output, usually displayed on the monitoring interface in the vessel's control room, or directly transmitted to the vessel operators or automatic control system, providing a direct, model-based predictive reference for operational decisions. The entire process from inputting the feature set to the final status determination is automated and continuous, providing real-time assessment support for the operational safety of engineering vessels.

[0087] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0088] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for training and applying a neural network for predicting the workability of engineering vessels, characterized in that, The method includes: A position sensor device and an environmental sensor device are installed on the engineering vessel. The position sensor device is used to obtain the current position of the engineering vessel in real time, and the environmental sensor device is used to obtain environmental monitoring data around the engineering vessel in real time. An operation plan acquisition device is installed in the control system of the engineering vessel, which is used to acquire the current operation plan data of the engineering vessel; An input feature set is generated using the current location of the engineering vessel, the environmental monitoring data, and the current work plan data. The input feature set is input into the trained neural network model to calculate and generate the workability prediction results for engineering vessels; The operational status of the engineering vessel is determined by comparing the predicted operational status with the preset judgment criteria.

2. The neural network training and application method for predicting the workability of engineering vessels as described in claim 1, characterized in that, The process of generating an input feature set using the current position of the engineering vessel, the environmental monitoring data, and the current work plan data includes: A position feature sequence is generated based on the current position of the engineering vessel; An environmental feature sequence is generated based on the environmental monitoring data. A plan feature sequence is generated based on the current work plan data; The location feature sequence, the environmental feature sequence, and the plan feature sequence are merged into the input feature set.

3. The neural network training and application method for predicting the workability of engineering vessels as described in claim 2, characterized in that, The step of calculating and generating a position feature sequence based on the current position of the engineering vessel includes: Obtain historical location data of engineering vessels; The movement trajectory features are calculated and generated using the current position and historical position data of the engineering vessel. The movement trajectory features are normalized into the position feature sequence.

4. The neural network training and application method for predicting the workability of engineering vessels as described in claim 3, characterized in that, The calculation and generation of environmental feature sequences based on the environmental monitoring data includes: Acquire real-time wind speed and wave height data for engineering vessels; Environmental fluctuation characteristics are calculated and generated using the real-time wind speed data and the real-time wave height data. The environmental fluctuation characteristics are standardized into the environmental characteristic sequence.

5. The neural network training and application method for predicting the workability of engineering vessels as described in claim 4, characterized in that, The step of calculating and generating a plan feature sequence based on the current work plan data includes: Acquire data on the target operating area and expected operating time of engineering vessels; The operation requirement characteristics are calculated and generated using the target operation area data and the expected operation time data. The job requirement features are encoded into the plan feature sequence.

6. The neural network training and application method for predicting the workability of engineering vessels as described in claim 5, characterized in that, The training of the completed neural network model includes the following steps: Acquire multiple sets of historical operational data of engineering vessels during normal operation. Each set of historical operational data includes historical location data, historical environmental monitoring data, and historical operation plan data of the engineering vessels. Training feature vectors are generated by calculating historical location data, historical environmental monitoring data, and historical work plan data from each set of historical operational data. The initial neural network model is trained using all training feature vectors to generate the trained neural network model.

7. The neural network training and application method for predicting the workability of engineering vessels as described in claim 6, characterized in that, The process of generating training feature vectors by calculating historical location data, historical environmental monitoring data, and historical work plan data from each set of historical operational data includes: Historical location features are calculated and generated based on the historical location data. Historical environmental characteristics are calculated and generated based on the historical environmental monitoring data. Historical plan features are calculated and generated based on the historical work plan data; The historical location features, historical environment features, and historical plan features are combined into the training feature vector.

8. The neural network training and application method for predicting the workability of engineering vessels as described in claim 7, characterized in that, The process of training the initial neural network model using all training feature vectors includes: The input layer of the neural network model is configured to receive the training feature vector; The training feature vectors are processed through the hidden layers of the neural network model to generate hidden features; The prediction result is generated based on the hidden features through the output layer of the neural network model; Adjust the weight parameters of the neural network model to optimize the prediction results.

9. The neural network training and application method for predicting the workability of engineering vessels as described in claim 8, characterized in that, The application of trained neural network models for predicting the workability of engineering vessels includes: The input feature set is input into the input layer of the trained neural network model; The input feature set is processed by the hidden layer of the trained neural network model to generate intermediate features; The output layer of the trained neural network model generates the workability forecast result of the engineering vessel based on the intermediate features.

10. The neural network training and application method for predicting the workability of engineering vessels as described in claim 9, characterized in that, The comparison of the workability forecast results of the engineering vessel with the preset judgment criteria includes: Obtain the preset job probability threshold; Compare the workability forecast results of the engineering vessel with the work probability threshold; When the predicted workability of the engineering vessel meets the work probability threshold, the work status of the engineering vessel is determined to be feasible; otherwise, the work status of the engineering vessel is determined to be infeasible.

Images