A ship propeller full-condition multi-target intelligent regulation and control method based on transformer time series prediction

Abstract

The application discloses a ship propeller full-condition multi-target intelligent regulation and control method based on a Transformer time series prediction, prediction results are taken as inputs of rolling time domain model prediction control based on future working condition and propelling response prediction results output by a Ship-LInformer model, optimal control amounts of propeller pitch angle and main engine rotating speed are solved jointly under constraint conditions of pitch angle, main engine rotating speed, power, torque, air bubble safety margin and actuator response capacity and the like, and a closed-loop mechanism of "prediction-optimization-execution-feedback-off-line iteration" is used to realize full-condition multi-target intelligent regulation and control of the ship propeller.

Description

Technical Field

[0001] This invention relates to the field of ship and marine engineering equipment technology, and in particular to a multi-objective intelligent control method for ship propellers under all operating conditions based on Transformer time-series prediction. Background Technology

[0002] As a core functional component of the ship's propulsion system, the operating conditions of the propeller directly determine the ship's economic performance, operational safety, and passenger comfort. Controllable pitch propellers (CPPs) can adjust the propeller blade pitch angle in real time while maintaining a constant main engine output speed, flexibly adapting to varying navigation conditions. They are now widely used in various mainstream ship types, including coastal transport merchant ships, marine engineering vessels, and ocean-going cargo ships.

[0003] Throughout their actual navigation, ships are continuously disturbed by multiple complex and unsteady factors, including coupled disturbances from wind, waves, and currents; dynamic changes in ship draft and cargo load; marine biofouling on the hull; and distortion of the propeller disk's wake field. Existing propeller control methods struggle to achieve optimal coordinated control across all sea states and load conditions, resulting in long-standing industry-wide technical challenges such as severe propulsion efficiency losses, high fuel consumption, susceptibility to propeller cavitation erosion, and excessive stern vibration and underwater noise. Against the backdrop of the industry's focus on intelligent, green, and low-carbon ship development, leveraging artificial intelligence algorithms to achieve intelligent optimization and control of controllable pitch propellers has become a key research and development direction in the field of shipbuilding and marine engineering equipment.

[0004] Currently, there are four main existing technologies in the field of intelligent control of ship propellers, each with obvious technical shortcomings: The first category is traditional control technology, which is currently the mainstream technology in commercial shipboard applications. It is based on the propeller's open-water characteristic curve to construct a fixed MAP diagram for open-loop control, or uses PID single-variable closed-loop control, which can only adapt to preset ideal design conditions. This technology cannot simultaneously achieve multiple objectives such as efficiency and cavitation suppression; at the same time, the control has a strong lag, which is a post-event feedback control, and it has no ability to predict operating conditions. Under extreme sea conditions, it cannot avoid risks such as propeller cavitation and main engine overload in advance.

[0005] The second category is traditional machine learning-based control schemes, which use parameter optimization methods based on SVM, random forest, and BP neural networks. These methods train models using historical navigation data to achieve static optimization of propeller pitch angle and main engine speed. This technology cannot capture the temporal variation patterns of navigation conditions and can only achieve parameter optimization under static conditions. Furthermore, its multi-source data fusion capability is weak, failing to effectively integrate heterogeneous data from the marine environment, propeller status, and different sampling frequencies.

[0006] The third category is recurrent neural network (RNN) control technology. This technology uses RNNs to build time-series prediction models to capture the temporal variation patterns of navigation conditions and achieve dynamic propeller control. However, this technology has low accuracy in predicting long-term navigation conditions and cannot capture long-term sea state variation patterns. Furthermore, the model is difficult to train, has poor stability, and is prone to overfitting.

[0007] The fourth category is Transformer-related application technologies. Existing research applies Transformer to single tasks such as ship trajectory prediction and equipment fault diagnosis. This technology has not yet formed a complete closed-loop engineering solution for real-time propeller control and is only used for single speed or operating condition prediction. At the same time, the original Transformer model has a large number of parameters and high computational complexity. Without lightweight processing, it cannot achieve real-time inference on shipborne edge devices and can only remain in the laboratory simulation stage.

[0008] In summary, existing propeller control technologies suffer from multiple defects, including lack of multi-objective collaborative optimization capabilities, lack of long-term operating condition prediction capabilities, poor heterogeneous data fusion effects, insufficient long-term prediction accuracy, and inability to deploy models onboard in real time. The industry urgently needs a smart closed-loop control solution for ship pitch propellers that can achieve multi-objective collaborative optimization under all operating conditions and is compatible with shipboard edge hardware. Summary of the Invention

[0009] This invention provides a multi-objective intelligent control method for ship propellers under all operating conditions based on Transformer time-series prediction. It utilizes the future operating conditions and propulsion response prediction results output by the lightweight Informer time-series prediction model Ship-LInformer, and uses these prediction results as input to the rolling time-domain model predictive control. Under constraints such as pitch angle, main engine speed, power, torque, cavitation safety margin, and actuator response capability, the propeller pitch angle is jointly solved. With the main unit speed The optimal control quantity is determined, and intelligent multi-objective control of ship propellers under all operating conditions is achieved through a closed-loop mechanism of "prediction-optimization-execution-feedback-offline iteration". This solves the problems of insufficient long-term dependency modeling capability, low long-cycle prediction accuracy, slow serial calculation and inference speed, and inability to adapt to multi-source heterogeneous ship data when only used for single speed prediction in existing technologies.

[0010] The technical solution adopted by this invention to solve its technical problem is a multi-objective intelligent control method for ship propellers under all operating conditions based on Transformer time-series prediction, comprising: S1: Construct and train the Ship-Linformer model, a lightweight encoder-based multi-step temporal prediction model based on the ProbSparse attention mechanism in Informer. The Ship-Linformer model includes an input feature mapping layer, a temporal sequence position encoding layer, several layers of ProbSparse multi-head self-attention encoder layers, and a multi-step prediction output layer. S2: Collect ship navigation data in all dimensions at a fixed frequency, construct the original time series dataset, and perform outlier removal, time series alignment, normalization, and sliding window construction. S3: The input features are mapped to the model feature dimension through the input feature mapping layer to obtain the feature embedding sequence, and a position encoding matrix is ​​generated through the temporal sequence position encoding layer. The position encoding matrix is ​​added element by element to the feature embedding sequence to obtain the input sequence. S4: Determine several candidate control quantities, generate candidate control sequences, normalize them, map them to the model feature space, and input them together with the input sequence into the ProbSparse sparse multi-head self-attention encoder layer to predict the non-control response prediction results other than the candidate control quantities in multiple future time steps. The non-control response prediction results are inversely normalized and combined with the candidate control sequences to form a future comprehensive prediction sequence. S5: Based on the future comprehensive prediction sequence, rolling time-domain model predictive control is adopted. With multiple objectives such as recommendation efficiency, cavitation risk, vibration noise and host operating stability, candidate control sequences are solved and the optimal control sequence is obtained in each control cycle. S6: Send the first control quantity in the optimal control sequence as the execution command for the current control cycle to the ship's execution control unit to complete the control action and return to step S2 to form a closed loop.

[0011] As a preferred option of the above scheme, 24-hour real ship operation data, control command data, and regulation effect data are collected daily to construct an incremental training dataset. The Ship-LInformer model is incrementally trained in an offline computing environment. After the offline incremental training is completed, validation set testing, historical navigation data playback testing, typical sea state simulation testing, and safety boundary verification are carried out in sequence. When the incremental model meets the preset prediction error threshold, control stability threshold, and safety constraints, the updated model weights are deployed to the shipborne edge computing module.

[0012] As a preferred embodiment of the above scheme, in step S1, the ProbSparse sparse multi-head self-attention encoder layer includes a ProbSparse multi-head self-attention sub-layer and a feedforward neural network sub-layer, and residual connections and layer normalization operations are respectively introduced after the ProbSparse multi-head self-attention sub-layer and the feedforward neural network sub-layer; The ProbSparse sparse multi-head self-attention encoder layer described in the previous layer is represented as follows: in, Indicates the first Intermediate features of the layer after ProbSparse multi-head self-attention, residual connections and layer normalization; Indicates the first Layer encoder output features; Presentation layer normalization operation; For the first The output of the ProbSparse multi-head self-attention sublayer; For the first The output of a sublayer of a feedforward neural network.

[0013] As a preferred embodiment of the above scheme, in step S2, the ship navigation full-dimensional data includes: Ship navigation status data: speed , longitudinal tilt Horizontal tilt Draft at the bow and stern , ; Marine environmental data: wind speed ,wind direction Significant wave height Wave cycle Flow direction Flow rate ; Host operating data: Rotation speed ,power Torque Fuel consumption rate Cooling water temperature ; Propeller operating data: Rotation speed Pitch angle ,thrust Torque ; Safety and comfort data: Stern triaxial vibration acceleration underwater noise level Propeller disk pressure pulsation .

[0014] As a preferred embodiment of the above scheme, in step S3, the input sequence is represented as: Where E is the mapped feature embedding sequence, , As input features, The weight matrix of the input feature mapping layer, The bias vector of the input feature mapping layer; This is the position encoding matrix.

[0015] As a preferred embodiment of the above scheme, in step S4, the candidate control quantity includes the future pitch angle. With the main unit speed At the current time step Under these conditions, the candidate control variable is defined as: Candidate control sequences are defined as follows: in, Indicates the length of the control time domain; when At that time, the candidate control variable remains at the first... The values ​​of each control step are: After generating the candidate control sequence, the candidate control sequence is expanded to a length of The prediction uses the control sequence: in:

[0016] As a preferred embodiment of the above scheme, the non-control response prediction result is expressed as follows: In the formula, This represents the Ship-LInformer time series prediction model; This represents all trainable parameters of the model; This represents the predicted results of multidimensional future non-control responses, excluding candidate control variables, under normalized scaling. For a given history input window, Control sequences for prediction The normalized embedded control sequence after being mapped by the input feature mapping layer. , The control sequence for prediction after normalization. The input feature mapping layer weight matrix, The bias vector of the input feature mapping layer; The inversely normalized result of the non-control response prediction is expressed as follows: in, This represents element-wise multiplication; These represent the minimum and maximum values ​​of the multidimensional non-control response variables (excluding candidate control variables) in the training set, respectively. This represents the prediction results of multidimensional future non-control responses, excluding candidate control quantities, at the physical quantity scale. The future comprehensive prediction sequence is represented as follows: in, This means that the multidimensional future non-control response prediction results (excluding candidate control quantities) and candidate control quantities are merged into a multidimensional future comprehensive prediction sequence according to a preset feature order.

[0017] As a preferred embodiment of the above scheme, in step S5, the optimization problem of the rolling time-domain model is defined as: The cumulative objective function for future prediction in the time domain is: In the formula, Indicates the current time step The rolling time-domain cumulative objective function; Represents the time discount factor, satisfying ; , , , This represents the weight coefficient of each sub-objective; Indicates the first The efficiency sub-goal for advancing each future time step; Indicates the first A sub-target for cavitation risk at a future time step; Indicates the first Vibration noise sub-target at a future time step; Indicates the first The target for the stability of host operating conditions at a future time step.

[0018] As a preferred embodiment of the above scheme, the sub-objective functions are defined as follows: in, For the smoothing penalty function, ; To predict the real-time propulsion efficiency of the propeller under operating conditions; This is the design rated propulsion efficiency of the propeller under this operating condition. This represents the minimum number of cavitation bubbles at the propeller disk surface; The critical cavitation number of the propeller; To predict the probability of cavitation occurring under operating conditions; The rated vibration acceleration threshold designed for ships; To predict underwater noise levels under operating conditions; The rated noise threshold designed for ships; To predict the main engine speed under operating conditions; This represents the actual rotational speed of the main unit at the current time step. Predict the host power under operating conditions; The actual power of the host at the current time step; Propeller real-time propulsion efficiency Calculated based on the propeller open water characteristic curve: In the formula, Indicates the first The rate of advance for each future time step; This indicates the advance velocity at the propeller disk surface; Indicates the propeller speed; Indicates the propeller diameter; Indicates the thrust coefficient; Indicates the torque coefficient; Indicates the first Propeller propulsion efficiency in a future time step; Probability of cavitation From the cavitation risk estimation function Calculations were made based on future predicted pressure pulsations, underwater noise, and stern triaxial vibration acceleration: In the formula, Indicates the first Predicted values ​​of propeller disk pressure pulsation at one future time step; This represents the predicted underwater noise level.

[0019] As a preferred embodiment of the above scheme, in step S6, the ship execution control unit performs amplitude limiting and smoothing processing on the execution command before execution: in, This represents the amplitude limiting function.

[0020] The advantages of this invention are: The Ship-Linformer model, a lightweight encoder-based multi-step temporal prediction model based on the ProbSparse sparse attention mechanism in Informer, effectively captures the dynamic coupling changes between ship navigation status, marine environment, main engine operation, and propeller operation over long periods, improving the accuracy of operational condition prediction; the Ship-LInformer model employs a parallelized sparse attention computation structure, significantly reducing long-sequence inference latency and increasing inference speed by more than 40% compared to LSTM, meeting the requirements of real-time rolling control on actual ships; the Ship-LInformer model designs a lightweight input mapping layer, a ProbSparse encoder structure, and a multi-step prediction output layer for 25-dimensional multi-source heterogeneous ship navigation data, enabling stable operation on shipborne edge computing devices and solving the problems of ordinary T... The Ransformer model suffers from large parameter count, high computational complexity, and difficulty in real-time deployment on actual ships. A comprehensive dataset covering 25 dimensions across five categories—ship navigation status, marine environment, main engine operation, propeller operation, and safety and comfort—was constructed, encompassing the main factors affecting propeller operation. This dataset effectively integrates multi-source data, providing comprehensive input information for model prediction and optimization, and significantly improving the model's generalization ability. A closed-loop control strategy of "prediction-optimization-regulation-feedback" is proposed, enabling proactive regulation based on predictions of future operating conditions. This addresses the lag problem of traditional control and can proactively avoid risks such as propeller cavitation and overload under extreme sea conditions. A configurable weighted multi-objective optimization function is designed, simultaneously considering multiple core optimization objectives, resolving the pain points of traditional control and achieving multi-objective collaborative optimal control of the propeller under all operating conditions. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the multi-objective intelligent control method for ship propellers under all operating conditions in Example 1. Detailed Implementation

[0022] The technical solution of the present invention will be further described below through embodiments and in conjunction with the accompanying drawings.

[0023] Example 1: This embodiment presents a multi-objective intelligent control method for ship propellers under all operating conditions based on Transformer time-series prediction, such as... Figure 1 As shown, it includes the following steps: S1: Construct and train the Ship-Linformer model, a lightweight encoder-based multi-step temporal prediction model based on the ProbSparse sparse attention mechanism in Informer.

[0024] The Ship-Linformer model comprises an input feature mapping layer, a temporal sequence position encoding layer, several layers of ProbSparse multi-head self-attention encoders, and a multi-step prediction output layer. Compared to the baseline Transformer model with a 6-layer encoder, this embodiment sets the number of ProbSparse multi-head self-attention encoder layers to [missing information]. The model feature dimension is set to The standard full self-attention mechanism is replaced with a ProbSparse sparse attention mechanism. The complex decoder structure is removed, and an encoder-style multi-step prediction output head is adopted. During the shipboard edge deployment phase, the trained Ship-LInformer model is quantized using FP16 half-precision and deployed to the Jetson Orin NX edge computing module to reduce model memory usage and inference latency. The Ship-LInformer model receives 25-dimensional ship operation characteristic data from consecutive historical windows and combines it with candidate control sequences to predict the ship's navigation conditions and propulsion response results within the next 15 minutes, providing a predictive basis for subsequent rolling time-domain multi-objective optimization control of propeller pitch angle and main engine speed.

[0025] The Ship-LInformer model employs the ProbSparse sparse multi-head self-attention mechanism from the Informer model to reduce computational complexity in long-term time-series prediction. Unlike the standard Transformer, which requires full attention calculations across all time steps, the ProbSparse attention mechanism selects the top-contributing keys based on the sparsity of the correlation between the query vector and the key vector. Each query vector participates in the main attention calculation, thereby reducing the computational cost while maintaining prediction accuracy.

[0026] The sparsity evaluation function is as follows: The number of query vectors selected for attention calculation is: .

[0027] The ProbSparse attention calculation formula is as follows: In the above formula, For the first The query vector corresponding to each time step; The key vector matrix; For the first The key vector corresponding to each time step; For the first Sparsity evaluation value of each query vector; Enter the length of the history input window; The number of query vectors selected to participate in the main attention calculation; For sparse sampling coefficients, in this embodiment, we take... ; The top values ​​obtained based on the sparsity evaluation function A matrix composed of query vectors; In this embodiment, to query the dimensions of the vector and key vector, ; It is a value vector matrix; This is the output of the ProbSparse sparse attention mechanism.

[0028] In this embodiment, the ProbSparse sparse multi-head self-attention encoder layer includes a ProbSparse multi-head self-attention sub-layer and a feedforward neural network sub-layer. For the first... The ProbSparse multi-head self-attention sublayer and the feedforward neural network sublayer, the first The formula for calculating the attention points is as follows: No. The multi-head self-attention output of the ProbSparse layer is: In the formula, Indicates the first Layer encoder output features; Indicates the first Layer The output of a sparse attention head; , , They represent the first Layer The query vector, key vector, and value vector of each attention head are linearly transformed into a weight matrix; Indicates the first Multi-head attention output linear transformation weight matrix; In this embodiment, the number of attention heads is indicated. .

[0029] The feedforward neural network sublayer uses a two-layer fully connected network with a GELU activation function added in between, as shown in the following formula: in, This is the output of the ProbSparse multi-head self-attention sublayer; is the weight matrix in the feedforward neural network, and are the trainable parameters of the model; represents the bias term in the feedforward neural network, and represents the trainable parameters of the model. is the activation function for the Gaussian error linear unit.

[0030] In this implementation, residual connections and layer normalization operations are introduced after the ProbSparse multi-head self-attention sublayer and the feedforward neural network sublayer, respectively. The computation process of the ProbSparse sparse multi-head self-attention encoder layer is as follows: in: In the formula, Indicates the first Intermediate features of the layer after ProbSparse multi-head self-attention, residual connections and layer normalization; Indicates the first Layer encoder output features; Presentation layer normalization operation.

[0031] S2: Collect full-dimensional data of ship navigation at a fixed frequency, construct the original time-series dataset, and perform outlier removal, time-series alignment, normalization, and sliding window construction.

[0032] In this embodiment, ship navigation data in all dimensions are collected at a fixed sampling frequency to construct the original time-series dataset. The data sources and feature dimensions collected are as follows: Ship navigation status data: speed , longitudinal tilt Horizontal tilt Draft at the bow and stern , ; Marine environmental data: wind speed ,wind direction Significant wave height Wave cycle Flow direction Flow rate ; Host operating data: Rotation speed ,power Torque Fuel consumption rate Cooling water temperature ; Propeller operating data: Rotation speed Pitch angle ,thrust Torque ; Safety and comfort data: Stern triaxial vibration acceleration underwater noise level Propeller disk pressure pulsation .

[0033] Original time series dataset ∈ ,in 25 represents the length of the time series sequence and the total feature dimension.

[0034] Original time series dataset The processing flow is as follows: Outlier removal: using 3 The criterion is to remove outliers from the original data, specifically those exceeding the mean. Values ​​three times the standard deviation are completed using linear interpolation.

[0035] Timing alignment: Data sources with different sampling frequencies are aligned to a time step of 1Hz to ensure that all features are synchronized at each time step.

[0036] Normalization: Min-max normalization is used to map all features to the [0,1] interval. The normalization formula is as follows: in, These are the normalized feature data and the standardized data input to the model. These are the original feature data and the original time-series data collected. This is the minimum value of the feature in the training set, and it is a statistical value of a fixed training set. This represents the maximum value of the feature in the training set, and is a statistical value for a fixed training set.

[0037] Sliding window construction: The normalized 25-dimensional feature data is processed according to the length of the historical input window. and prediction step size Construct the model input samples and training labels.

[0038] in, This represents the total time length of the original complete time series dataset; This indicates the length of the historical window for each input to the model; This represents the future time step predicted by the model. In this embodiment, the sampling frequency... Length of historical input window This corresponds to historical data from the past 15 minutes; predict step size. This corresponds to the prediction result for the next 15 minutes.

[0039] At the current time step The model input window is defined as follows: in: The future true label during the training phase is defined as: in: In the formula, Indicates the first The 25-dimensional normalized feature vectors corresponding to each time step.

[0040] S3: The input features are mapped to the model feature dimension through the input feature mapping layer to obtain the feature embedding sequence, and a position encoding matrix is ​​generated through the temporal sequence position encoding layer. The position encoding matrix is ​​added element by element to the feature embedding sequence to obtain the input sequence.

[0041] The Ship-LInformer model itself cannot directly identify the positional information of a time series, therefore a time series positional encoding layer is used to embed the input features into the sequence. Injection position encoding. In this embodiment, the temporal sequence position encoding layer uses sine-cosine position encoding, as shown in the following formula: In the formula: For the position encoding matrix, the first The time step, the first Encoded values ​​for an even number of dimensions; For the position encoding matrix, the first The time step, the first Encoded values ​​for an odd number of dimensions; This represents the time step position in the time series, with a value range of [value missing]. ; This is the feature dimension index, with a value range of [value range missing]. ; represents the feature dimensions of the Ship-LInformer model.

[0042] The Ship-LInformer model maps the 25-dimensional original input features to a unified model feature dimension through a linear mapping layer, as shown in the following formula: In the formula, This is a sequence of multi-source heterogeneous features of ships within a historical input time window. ; The length of the history input window. This represents historical data from the past 900 seconds. The weight matrix of the input feature mapping layer, , which are the trainable parameters of the model; is the bias vector of the input feature mapping layer, are the trainable parameters of the model; E is the mapped feature embedding sequence. .

[0043] Next, the position encoding matrix With input feature embedding sequence By adding elements one by one, we obtain the final input sequence containing the injected position information: In the formula, Let represent the input sequence of the Ship-LInformer encoder, satisfying: In the above, the model input and labels are constructed as follows: S4: Determine several candidate control variables, generate candidate control sequences, normalize them, map them to the model feature space, and input them together with the input sequence into the ProbSparse sparse multi-head self-attention encoder layer to predict the non-control response prediction results other than the candidate control variables in multiple future time steps. The non-control response prediction results are inversely normalized and combined with the candidate control sequences to form a future comprehensive prediction sequence.

[0044] In this embodiment, the future pitch angle With the main unit speed As candidate control variables given by the optimizer. Current time step. Under these conditions, the candidate control variable is defined as: Candidate control sequences are defined as follows: in, Indicates the length of the control time domain. When At that time, the candidate control variable remains at the first... The values ​​of each control step are: To ensure that the candidate control sequence is consistent with the prediction time domain Align, with a length of The candidate control sequence is expanded to a length of The prediction uses the control sequence: in: Subsequent model predictions, normalization, and control input embedding all employ... Instead of directly adopting .

[0045] The Ship-LInformer model, given a historical input window and candidate control sequences Predicting the future under these conditions Within each time step, except for the control quantity and In addition to the 23 non-control response variables, the candidate control sequences are first normalized to allow them to participate in the prediction. and This represents the minimum and maximum values ​​used for control variable normalization. Then, the candidate control sequences are mapped to the model feature space through a control variable embedding layer. In the formula, This represents the weight matrix of the control embedding layer. This represents the bias vector of the control embedding layer. The non-control response prediction result of the Ship-LInformer model is expressed as: in: In the formula, This represents the Ship-LInformer time series prediction model; This represents all trainable parameters of the model; This represents the 23-dimensional future uncontrolled response prediction result at a normalized scale. Since the model output is a normalized result, it needs to be denormalized back to physical quantities before performing multi-objective optimization. in, This represents element-wise multiplication; These represent the minimum and maximum values ​​of the 23-dimensional uncontrolled response variable in the training set, respectively. This represents the 23-dimensional future uncontrolled response prediction results at the physical quantity scale. The final 25-dimensional comprehensive prediction sequence used for optimization calculations is defined as follows: in: In the formula, This means that the 23-dimensional non-control response prediction results and the 2-dimensional candidate control quantities are merged into a 25-dimensional future comprehensive prediction sequence according to the preset feature order.

[0046] S5: Based on the future comprehensive prediction sequence, rolling time-domain model predictive control is adopted. With multiple objectives such as recommended efficiency, cavitation risk, vibration noise and host operating stability, candidate control sequences are solved in each control cycle and the optimal control sequence is obtained.

[0047] In this embodiment, the rolling time-domain optimization problem is defined as: The cumulative objective function for future prediction in the time domain is: In the formula, Indicates the current time step The rolling time-domain cumulative objective function; Represents the time discount factor, satisfying ; , , , This represents the weight coefficient of each sub-objective. ; Indicates the first The efficiency sub-goal for advancing each future time step; Indicates the first A sub-target for cavitation risk at a future time step; Indicates the first Vibration noise sub-target at a future time step; Indicates the first The target for the stability of host operating conditions at a future time step.

[0048] The sub-objective functions are defined as follows: in, This is a smoothing penalty function, used to replace non-smoothness. function: In the formula, The smoothing coefficient is , and satisfies ; To predict the real-time propulsion efficiency of the propeller under operating conditions, the prediction results are used... Calculated using the open-water characteristic curve of the propeller; The rated propulsive efficiency of the propeller under this operating condition is a fixed value obtained from propeller model tests. The minimum cavitation number at the propeller disk is calculated from the predicted operating conditions, including speed, draft, and pitch angle. The critical cavitation number of the propeller is a fixed value determined by the propeller design parameters. To predict the probability of cavitation occurrence under operating conditions, based on Obtain; To predict the triaxial vibration acceleration of the stern under operating conditions, from ; The rated vibration acceleration threshold for ships is a fixed value. To predict the underwater noise level under operating conditions, from ; The rated noise threshold designed for ships is a fixed value. To predict the main engine speed under operating conditions, from ; This represents the actual rotational speed of the main unit at the current time step. To predict the host power under operating conditions, from ; This represents the actual power of the host at the current time step.

[0049] Among them, the real-time propulsion efficiency of the propeller Calculated based on the propeller open water characteristic curve: In the formula, Indicates the first The rate of advance for each future time step; This indicates the advance velocity at the propeller disk surface; Indicates the propeller speed; Indicates the propeller diameter; Indicates the thrust coefficient; Indicates the torque coefficient; Indicates the first Propeller propulsion efficiency in a future time step.

[0050] in, It is not a feature dimension output solely by the Ship-LInformer model, but rather derived from the cavitation risk estimation function. Calculated based on future predicted pressure pulsations, underwater noise, and stern triaxial vibration acceleration. In the formula, Indicates the first Predicted values ​​of propeller disk pressure pulsation at a future time step. This represents the predicted underwater noise level. , , This represents the predicted value of the triaxial vibration acceleration at the stern.

[0051] The constraints for rolling time-domain optimization are as follows: The mechanical constraint on the pitch angle is: The host speed constraint is: The constraint on the rate of change of pitch angle is: The constraint on the rate of change of the main engine speed is: The host power constraint is: The main engine torque constraint is: The cavitation safety margin constraint is: In the formula, and These represent the lower and upper limits of the allowable pitch angle of the controllable propeller, respectively. and These represent the lower and upper limits of the host's permissible rotational speed, respectively. This indicates the maximum allowable pitch angle change within a unit control cycle; This indicates the maximum allowable change in main engine speed within a unit control cycle; Indicates the maximum allowed power of the host; Indicates the maximum allowable torque of the main unit; This represents the minimum allowable number of cavitation bubbles corresponding to the cavitation safety margin.

[0052] To meet To meet the real-time requirements of the control cycle, this embodiment will control the time domain. Set to less than the prediction time domain In the short time domain, this embodiment sets , That is, each control cycle optimizes only the candidate control sequences for the next 30 control steps, but uses the prediction results for the next 900 time steps for risk assessment and cumulative target calculation. Simultaneously, a maximum number of iterations is set for the sequence quadratic programming algorithm. and maximum solution time This ensures that the total time consumed within a single control cycle satisfies: in, This indicates the prediction time taken by the Ship-LInformer model; This indicates the time taken by the optimization algorithm to solve the problem; Indicates the time taken for safety constraint verification; Indicates the time consumed in sending control command communications; Indicates the control cycle, in this embodiment .

[0053] In this embodiment, when using a constrained sequential quadratic programming algorithm to solve the above-mentioned rolling time-domain optimization problem, since the Ship-LInformer model, the inverse normalization process, and the objective function are all in a differentiable form, and the non-smooth penalty term is derived from the smoothing function... Therefore, the gradient of the objective function with respect to the candidate control sequence can be obtained through automatic differentiation: In the formula, This represents the gradient of the rolling time-domain cumulative objective function with respect to the candidate control sequence.

[0054] S6: Send the first control quantity in the optimal control sequence as the execution command for the current control cycle to the ship's execution control unit to complete the control action and return to step S2 to form a closed loop.

[0055] In this embodiment, the execution instructions for the current control cycle are: Before execution, the control command undergoes amplitude limiting and smoothing processing: in, This represents the limiting function. After completing the adjustment action, it reverts to step 2 to complete one closed-loop control cycle with a control period of 1 second, achieving real-time rolling adjustment. When the model output exceeds the safety boundary, CAN communication malfunctions, or the edge computing unit fails, the system automatically switches to the ship's original PID control mode to ensure the safe operation of the ship's propulsion system.

[0056] In addition, in this embodiment, 24-hour real-ship operation data, control command data, and regulation effect data are collected daily to construct an incremental training dataset. The incremental training process does not directly update the model weights during the actual online control process; instead, the Ship-LInformer model is incrementally trained in an offline computing environment. After offline incremental training is completed, validation set testing, historical navigation data playback testing, typical sea state simulation testing, and safety boundary verification are performed sequentially. Only when the incremental model meets the preset prediction error threshold, control stability threshold, and safety constraints are the updated model weights deployed to the shipborne edge computing module. Through the closed-loop iterative approach of offline training, simulation verification, edge deployment, and real-ship feedback, the Ship-LInformer model's ability to model the long-term temporal coupling relationships between ship navigation status, marine environment, main engine operation status, propeller operation status, and safety and comfort data is continuously optimized, improving the prediction accuracy of the operating conditions for the next 15 minutes and the multi-objective regulation effect of the propeller.

[0057] Example 2: This embodiment compares the multi-objective intelligent control method for ship propellers under all operating conditions based on Transformer time-series prediction in Embodiment 1, the mainstream PID control method in actual ship trials, and the mainstream LSTM control method in existing technologies through simulation experiments and actual ship trials. Specific quantitative results are as follows: Mean absolute error (MAE) of the predicted operating conditions in the next 5 minutes: 2.12% for the method in Example 1, and 3.05% for the LSTM method, representing a 30.5% improvement in accuracy; Mean absolute error (MAE) of the predicted operating conditions in the next 15 minutes: 3.87% for the method in Example 1, and 5.21% for the LSTM method, representing an accuracy improvement of 25.7%. Under normal navigation conditions: Compared with PID control, the method in Example 1 improves propeller propulsion efficiency by an average of 9.2% and reduces ship fuel consumption by an average of 8.5%; compared with LSTM control, propulsion efficiency is improved by 4.6% and fuel consumption is reduced by 4.1%. Extreme sea conditions: Compared with PID control, the method in Example 1 improves propulsion efficiency by an average of 13.5% and reduces fuel consumption by an average of 12.2%; compared with LSTM control, it improves propulsion efficiency by 7.8% and reduces fuel consumption by 7.1%. Ship entry and exit from port and changing operating conditions: Compared with PID control, the method in Example 1 improves propulsion efficiency by an average of 10.8% and reduces fuel consumption by an average of 9.7%.

[0058] Cavitation suppression effect: The method in Example 1 can effectively reduce the probability of propeller tip vortex cavitation. The duration of cavitation occurrence is reduced by 68.3% compared with PID control and by 42.7% compared with LSTM method, which greatly reduces the risk of propeller erosion and extends the service life of propeller. Vibration and noise control: Compared with PID control, the method in Example 1 reduced the average stern vibration acceleration by 18.6% and the average underwater noise level by 12.3 dB; compared with LSTM method, the vibration acceleration was reduced by 9.2% and the underwater noise level was reduced by 5.7 dB. Main unit operating stability: Compared with PID control, the method in Example 1 reduces the main unit speed fluctuation by 72.5% and the torque fluctuation by 65.8%, which significantly reduces main unit wear and extends the main unit overhaul cycle.

[0059] In this embodiment, the operating condition prediction error is calculated using the multivariate average relative error. The characterization, and its calculation formula are as follows: In the formula, Indicates the number of test samples; Indicates the first Predicted values ​​for each test sample; Indicates the first The true value of each test sample; This represents the smallest positive number that prevents the denominator from being zero.

[0060] When expressed as prediction accuracy, prediction accuracy is defined as: in, In calculation Use decimal form.

[0061] Furthermore, the Ship-LInformer model of the method in Example 1 has a single prediction inference time of [time missing] on the shipboard Jetson Orin NX edgebox. In the control time domain Prediction time domain Under the specified settings, the system imposes maximum iteration count and maximum solution time constraints on the optimization algorithm to ensure that the total time consumed by predictive inference, rolling optimization, security verification, and control command issuance meets the following requirements: Therefore, it meets the requirements of the actual ship. The system meets the requirements for real-time rolling control during the control cycle. It can be directly adapted to the existing ship's controllable pitch propeller and main engine control systems without requiring large-scale hardware modifications, resulting in low modification costs, short installation cycles, and easy implementation on actual ships.

[0062] The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to substitute them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

Claims

1. A multi-objective intelligent control method for ship propellers under all operating conditions based on Transformer time-series prediction, characterized by: include: S1: Construct and train the Ship-Linformer model, a lightweight encoder-based multi-step temporal prediction model based on the ProbSparse attention mechanism in Informer. The Ship-Linformer model includes an input feature mapping layer, a temporal sequence position encoding layer, several layers of ProbSparse multi-head self-attention encoder layers, and a multi-step prediction output layer. S2: Collect ship navigation data in all dimensions at a fixed frequency, construct the original time series dataset, and perform outlier removal, time series alignment, normalization, and sliding window construction. S3: The input features are mapped to the model feature dimension through the input feature mapping layer to obtain the feature embedding sequence, and a position encoding matrix is ​​generated through the temporal sequence position encoding layer. The position encoding matrix is ​​added element by element to the feature embedding sequence to obtain the input sequence. S4: Determine several candidate control quantities, generate candidate control sequences, normalize them, map them to the model feature space, and input them together with the input sequence into the ProbSparse sparse multi-head self-attention encoder layer to predict the non-control response prediction results other than the candidate control quantities in multiple future time steps. The non-control response prediction results are inversely normalized and combined with the candidate control sequences to form a future comprehensive prediction sequence. S5: Based on the future comprehensive prediction sequence, rolling time-domain model predictive control is adopted. With multiple objectives such as recommendation efficiency, cavitation risk, vibration noise and host operating stability, candidate control sequences are solved and the optimal control sequence is obtained in each control cycle. S6: Send the first control quantity in the optimal control sequence as the execution command for the current control cycle to the ship's execution control unit to complete the control action and return to step S2 to form a closed loop.

2. The intelligent multi-objective control method for ship propellers under all operating conditions based on Transformer time-series prediction as described in claim 1, characterized in that: Daily 24-hour real-time ship operation data, control command data, and regulation effect data are collected to construct an incremental training dataset. The Ship-LInformer model is incrementally trained in an offline computing environment. After the offline incremental training is completed, validation set testing, historical navigation data playback testing, typical sea state simulation testing, and safety boundary verification are carried out in sequence. When the incremental model meets the preset prediction error threshold, control stability threshold, and safety constraints, the updated model weights are deployed to the shipborne edge computing module.

3. The intelligent multi-objective control method for ship propellers under all operating conditions based on Transformer time-series prediction as described in claim 1, characterized in that: In step S1, the ProbSparse sparse multi-head self-attention encoder layer includes a ProbSparse multi-head self-attention sub-layer and a feedforward neural network sub-layer. Residual connections and layer normalization operations are introduced after the ProbSparse multi-head self-attention sub-layer and the feedforward neural network sub-layer, respectively. The ProbSparse sparse multi-head self-attention encoder layer described in the previous layer is represented as follows: in, Indicates the first Intermediate features of the layer after ProbSparse multi-head self-attention, residual connections and layer normalization; Indicates the first Layer encoder output features; Presentation layer normalization operation; For the first The output of the ProbSparse multi-head self-attention sublayer; For the first The output of a sublayer of a feedforward neural network.

4. The intelligent multi-objective control method for ship propellers under all operating conditions based on Transformer time-series prediction as described in claim 1, characterized in that: In step S2, the full-dimensional data of ship navigation includes: Ship navigation status data: speed , longitudinal tilt Horizontal tilt Draft at the bow and stern , ; Marine environmental data: wind speed ,wind direction Significant wave height Wave cycle Flow direction Flow rate ; Host operating data: Rotation speed ,power Torque Fuel consumption rate Cooling water temperature ; Propeller operating data: Rotation speed Pitch angle ,thrust Torque ; Safety and comfort data: Stern triaxial vibration acceleration underwater noise level Propeller disk pressure pulsation .

5. The intelligent multi-objective control method for ship propellers under all operating conditions based on Transformer time-series prediction as described in claim 1, characterized in that: In step S3, the input sequence is represented as: Where E is the mapped feature embedding sequence, , As input features, The weight matrix of the input feature mapping layer, The bias vector of the input feature mapping layer; This is the position encoding matrix.

6. The intelligent multi-objective control method for ship propellers under all operating conditions based on Transformer time-series prediction as described in claim 1, characterized in that: In step S4, the candidate control quantity includes the future pitch angle. With the main unit speed At the current time step Under these conditions, the candidate control variable is defined as: Candidate control sequences are defined as follows: in, Indicates the length of the control time domain; when At that time, the candidate control variable remains at the first... The values ​​of each control step are: After generating the candidate control sequence, the candidate control sequence is expanded to a length of The prediction uses the control sequence: in: 。 7. The intelligent control method for multi-objective control of ship propellers under all operating conditions based on Transformer time-series prediction as described in claim 6, characterized in that: The non-control response prediction result is expressed as follows: In the formula, This represents the Ship-LInformer time series prediction model; This represents all trainable parameters of the model; This represents the predicted results of multidimensional future non-control responses, excluding candidate control variables, under normalized scaling. For a given history input window, Control sequences for prediction The normalized embedded control sequence after being mapped by the input feature mapping layer. , The control sequence for prediction after normalization. The input feature mapping layer weight matrix, The bias vector of the input feature mapping layer; The inversely normalized result of the non-control response prediction is expressed as follows: in, This represents element-wise multiplication; These represent the minimum and maximum values ​​of the multidimensional non-control response variables (excluding candidate control variables) in the training set, respectively. This represents the prediction results of multidimensional future non-control responses, excluding candidate control quantities, at the physical quantity scale. The future comprehensive prediction sequence is represented as follows: in, This means that the multidimensional future non-control response prediction results (excluding candidate control quantities) and candidate control quantities are merged into a multidimensional future comprehensive prediction sequence according to a preset feature order.

8. The intelligent control method for multi-objective control of ship propellers under all operating conditions based on Transformer time-series prediction as described in claim 1, characterized in that: In step S5, the optimization problem of the rolling time-domain model is defined as: The cumulative objective function for future prediction in the time domain is: In the formula, Indicates the current time step The rolling time-domain cumulative objective function; Represents the time discount factor, satisfying ; , , , This represents the weight coefficient of each sub-objective; Indicates the first The efficiency sub-goal for advancing each future time step; Indicates the first A sub-target for cavitation risk at a future time step; Indicates the first Vibration noise sub-target at a future time step; Indicates the first The target for the stability of host operating conditions at a future time step.

9. The intelligent control method for multi-objective control of ship propellers under all operating conditions based on Transformer time-series prediction as described in claim 8, characterized in that: The sub-objective functions are defined as follows: in, For the smoothing penalty function, ; To predict the real-time propulsion efficiency of the propeller under operating conditions; This is the design rated propulsion efficiency of the propeller under this operating condition. This represents the minimum number of cavitation bubbles at the propeller disk surface; The critical cavitation number of the propeller; To predict the probability of cavitation occurring under operating conditions; The rated vibration acceleration threshold designed for ships; To predict underwater noise levels under operating conditions; The rated noise threshold designed for ships; To predict the main engine speed under operating conditions; This represents the actual rotational speed of the main unit at the current time step. Predict the host power under operating conditions; The actual power of the host at the current time step; Propeller real-time propulsion efficiency Calculated based on the propeller open water characteristic curve: In the formula, Indicates the first The rate of advance for each future time step; This indicates the advance velocity at the propeller disk surface; Indicates the propeller speed; Indicates the propeller diameter; Indicates the thrust coefficient; Indicates the torque coefficient; Indicates the first Propeller propulsion efficiency in a future time step; Probability of cavitation From the cavitation risk estimation function Calculations were made based on future predicted pressure pulsations, underwater noise, and stern triaxial vibration acceleration: In the formula, Indicates the first Predicted values ​​of propeller disk pressure pulsation at one future time step; This represents the predicted underwater noise level.

10. The intelligent multi-objective control method for ship propellers under all operating conditions based on Transformer time-series prediction as described in claim 1, characterized in that: In step S6, the ship's execution control unit performs amplitude limiting and smoothing processing on the execution command before execution: in, This represents the amplitude limiting function.

Images