A ship oil throttle self-adaptive control system for complex working conditions

Abstract

The application discloses a ship throttle self-adaptive control system for complex working conditions, which comprises a data sensing module, a working condition evaluation module, a trajectory modeling module, a self-adaptive parameter updating module and a throttle control module. The application relates to the technical field of throttle self-adaptive control and utilizes a working condition evaluation model combining load feedforward estimation and lightweight time series convolution to realize comprehensive determination of working condition types, health index and trends. A trajectory prediction model combining multi-time scale physical constraints is adopted to constrain and solve the throttle opening change trajectory, thereby generating a dynamically optimized throttle control reference trajectory. Through a PID parameter adjustment method fusing working condition information and physical constraint optimization, adaptive correction of control parameters is realized under a reinforcement learning framework, the throttle control can realize front prediction and dynamic optimization under complex speed and sea condition conditions, and the response speed, control precision and operation safety of the ship under complex working conditions of rapid maneuvering and wave excitation are improved.

Description

Technical Field

[0001] This invention belongs to the field of throttle adaptive control technology, specifically a ship throttle adaptive control system for complex operating conditions. Background Technology

[0002] The ship's adaptive throttle control system is an intelligent control device designed for ship propulsion. It collects multi-source operational data, intelligently analyzes complex operating conditions, predicts throttle opening changes, and optimizes control parameters in real time. This allows throttle control to maintain dynamic response and stability under different speeds, loads, and sea states. Its function is to improve the ship's maneuverability, shorten the response time of throttle adjustments, and avoid the risks of overheating or insufficient stability margin due to over-adjustment, thereby ensuring the ship's power safety and operational efficiency in complex combat and navigation environments.

[0003] However, existing ship throttle adaptive control systems suffer from several technical problems: reliance on single sensor feedback makes it impossible to predict throttle control in advance under complex speed and sea state conditions, leading to lag in throttle opening adjustment and slow system response; reliance on static threshold judgments makes it difficult to accurately estimate transient load torque, resulting in insufficient accuracy in condition identification and one-sided health assessment; reliance on empirical prediction models without considering physical constraints leads to a lack of engineering feasibility and safety assurance in trajectory optimization results; and reliance on fixed PID parameters or manual experience in updating control system parameters makes it difficult to achieve adaptive parameter adjustment under different speeds, sudden load changes, or cabin environmental disturbances. Summary of the Invention

[0004] To address the above issues and overcome the shortcomings of existing technologies, this invention provides a ship throttle adaptive control system for complex operating conditions. Addressing the technical problems of existing ship throttle adaptive control systems that rely on single sensor feedback, failing to achieve pre-prediction of throttle control under complex speed and sea state conditions, resulting in lag in throttle opening adjustment and slow system response, this solution creatively employs pre-emptive intelligent data analysis combining operating condition assessment and throttle control trajectory modeling as the foundation of intelligent control. By improving the reward-based control system parameter update method, it achieves proactive and pre-emptive throttle opening adjustment, improving the ship's dynamic response capability and stability under complex operating conditions such as wave excitation and rapid maneuvering. Furthermore, addressing the technical problems of existing operating condition status assessment processes that rely on static threshold judgments, making it difficult to accurately estimate transient load torque, leading to insufficient accuracy in operating condition identification and one-sided health assessment, this solution creatively adopts a lightweight temporal convolutional model based on load feedforward estimation to enhance operating condition observability for operating condition status assessment. By constructing a quasi-linear reduced-order observer, it performs negative... The system dynamically estimates the load torque and introduces an attention mechanism to weightedly fuse operating condition characteristics and physical load characteristics. Finally, it utilizes multi-task learning to achieve operating condition type identification, health index calculation, and trend prediction, significantly improving the observability and diagnostic accuracy of the operating condition. It can provide early warning and trend prediction for early anomalies such as shaft vibration abnormalities and uneven temperature fields. Addressing the technical problem in existing throttle control trajectory modeling that relies solely on empirical prediction models without considering physical constraints, resulting in a lack of engineering feasibility and safety assurance in trajectory optimization results, this solution creatively adopts an improved time-series prediction model that combines multi-timescale physical constraints for throttle control trajectory modeling. It constructs a physical constraint objective function and couples the dynamic model of the actuator process at a fast timescale with the dynamic model of the load output at a slow timescale for prediction, performing real-time parameter correction. Simultaneously, it embeds soft constraints such as upper temperature threshold, stability margin, and valve rate into the optimization objective function, forming a trajectory optimization with safety boundaries. This achieves a balance between the dynamism and safety of throttle trajectory prediction, avoiding problems such as overheating, insufficient stability margin, or actuator overload caused by excessive control commands.To address the technical challenge of existing control system parameter updates relying on fixed PID parameters or manual experience, which makes adaptive parameter adjustment difficult under varying speeds, load changes, or cabin environmental disturbances, this solution creatively employs a PID parameter adjustment method that combines operating condition information and physical constraint optimization. This method adjusts control system parameters by loading the operating conditions and predicted trajectory into a state vector using a reinforcement learning framework. Candidate PID parameters are generated using scaling factors, and a reward function incorporating error progression, phase consistency, safety margin barriers, and actuator penalties is used for strategy updates. This yields reference data for parameter correction, ensuring the dynamic adaptability of control parameters and achieving stability in throttle control under complex conditions such as rapid load increases and drastic sea state fluctuations. This significantly reduces the risks of overshoot and actuator saturation.

[0005] The technical solution adopted by the present invention is as follows: The present invention provides a ship throttle adaptive control system for complex working conditions, including a data sensing module, a working condition evaluation module, a trajectory modeling module, an adaptive parameter update module, and a throttle control module;

[0006] The data sensing module is used for collecting operating status data. Through collecting operating status data, it obtains raw operating status data and sends the raw operating status data to the operating condition evaluation module.

[0007] The operating condition assessment module is used to assess the operating condition status, obtain operating condition assessment information through the operating condition status assessment, and send the operating condition assessment information to the trajectory modeling module and the adaptive parameter update module.

[0008] The trajectory modeling module is used for throttle control trajectory modeling. Through throttle control trajectory modeling, throttle opening change prediction data is obtained, and the throttle opening change prediction data is sent to the adaptive parameter update module.

[0009] The adaptive parameter update module is used to control system parameter adjustment, obtain real-time correction reference data of control parameters by controlling system parameter adjustment, and send the real-time correction reference data of control parameters to the throttle control module.

[0010] The throttle control module is used to execute throttle opening commands. By executing throttle opening commands, it obtains throttle opening commands and controls the opening of the throttle valves of the power supply system.

[0011] Furthermore, the operation status data acquisition is used to collect operation status data in real time, specifically by acquiring raw operation status data through multi-sensor data acquisition;

[0012] The multi-sensor data acquisition includes the following steps: sensor layout acquisition, signal standardization, data preprocessing, and data frame storage.

[0013] Furthermore, the operational condition status assessment is used to assess the specific operational condition at present. Specifically, based on the original operational condition data, an improved lightweight temporal convolutional model with enhanced operational condition observability based on load feedforward estimation is used to assess the operational condition status and obtain operational condition assessment information. This includes the following steps: load feedforward estimation, operational condition feature extraction, observability enhancement fusion, and operational condition classification output.

[0014] The load feedforward estimation specifically involves using energy conversion unit parameter data and rotating component data to construct a reduced-order state observer based on a quasi-linear model to dynamically estimate the current load torque and obtain estimated load torque value data.

[0015] The operating condition feature extraction specifically involves using the estimated load torque value data and the original operating status data to construct an improved lightweight temporal convolutional network to extract features with multi-scale and long-term dependencies, thereby obtaining a high-dimensional feature vector of operating condition information.

[0016] The observability enhancement fusion specifically involves dynamically calculating the fusion weights through the attention layer based on the high-dimensional feature vector of the operating condition information and the estimated load torque value data, and performing adaptive weighted fusion of the operating condition features and the load feedforward features of the physical model to obtain the observability enhancement fusion feature vector.

[0017] The operational condition classification output is specifically based on the observability enhancement fusion feature vector, and through a multi-task learning network branch, operational condition type identification, health assessment, and short-term trend prediction are performed to obtain operational condition assessment information.

[0018] The operational condition assessment information specifically includes operational condition labels, health index, and predicted trends.

[0019] Furthermore, the throttle control trajectory modeling is used to dynamically generate optimized throttle control trajectories. Specifically, based on the operating condition evaluation information, an improved time-series prediction model combining multi-time-scale physical constraints is used to model the throttle control trajectory and obtain throttle opening change prediction data. This includes the following steps: target constraint construction, multi-time-scale prediction model construction, physical constraint embedding objective function construction, constraint trajectory solving, and throttle control trajectory modeling.

[0020] The construction of the target constraint specifically involves constructing a basic objective function containing an error term and a control smoothing term based on the operating condition evaluation information.

[0021] The error term is used to measure the deviation between the predicted throttle trajectory and the target trajectory;

[0022] The control smoothing term is used to limit the fluctuation of the throttle opening change rate;

[0023] The construction of the multi-timescale prediction model specifically involves establishing a fast-timescale actuator process dynamic prediction model and a slow-timescale load output dynamic prediction model based on the operating condition evaluation information and the basic objective function. The model parameters are then corrected in real time through a parameter recursive update method to obtain the multi-timescale prediction output.

[0024] The construction of the physical constraint embedding objective function specifically involves introducing physical constraint terms and weighting and combining the physical constraint terms with the terms in the basic objective function to obtain the physical constraint embedding optimization objective function, which is used for constraint trajectory solving.

[0025] The physical constraints specifically combine the health status and trend boundaries in the operating condition assessment information to apply soft constraints to the upper temperature threshold, stability margin, and valve rate.

[0026] The formula for calculating the physical constraint term is:

[0027] ;

[0028] In the formula, J phy These are physical constraints. It is the soft constraint weight of the upper temperature threshold. It is a soft constraint on the upper temperature threshold, T EGT (k) is the predicted upper temperature threshold value, where k is the discrete time step index. It is the safe upper limit threshold for the temperature upper limit. It is a stability margin soft constraint weight. It is a soft constraint with stability margin. It is the lower bound threshold of the stability margin, and SM(k) is the predicted value of the stability margin. It is the valve rate soft constraint weight. It is a valve rate soft constraint. It is the predicted valve rate of throttle opening, r max This is the maximum valve rate of the throttle opening. + It is a positive function; the soft constraint weights of the upper temperature threshold, stability margin, and valve rate are specifically adaptively adjusted based on the health index and the boundary of the predicted trend in the operating condition assessment information.

[0029] The formula for calculating the physical constraint embedding optimization objective function is as follows:

[0030] ;

[0031] In the formula, J is the physical constraint embedding optimization objective function. It is the weight of the error term, J err It is an error term constraint. It controls the weights of the smoothing term, J smooth It controls the smoothing term constraint;

[0032] The constraint trajectory solution is specifically performed by embedding the physical constraint into the optimization objective function and using a rolling optimization method to solve the constraint on the throttle opening change trajectory, thereby obtaining a throttle opening change prediction sequence that satisfies the constraint conditions.

[0033] The throttle control trajectory modeling specifically involves generating a time-stamped throttle control reference trajectory based on the throttle opening change prediction sequence, thereby obtaining throttle opening change prediction data.

[0034] Furthermore, the control system parameter adjustment is used to adjust the parameters of the control system in real time. Specifically, based on the operating condition evaluation information and the throttle opening change prediction data, a PID parameter adjustment method that combines operating condition information and physical constraints is used to adjust the control system parameters to obtain real-time correction reference data for the control parameters. This includes the following steps: loading state parameters, generating action parameters, improving control rewards, and updating control system parameters.

[0035] The loading of state parameters specifically involves constructing a reinforcement learning state vector based on the operating condition evaluation information and the throttle opening change prediction data output by the throttle control trajectory modeling module.

[0036] The action parameter generation specifically involves constructing a three-dimensional scaling factor and scaling the baseline PID parameters according to the three-dimensional scaling factor to obtain candidate parameters, which are then used as reinforcement learning action vectors.

[0037] The improved control reward specifically involves constructing an improved control reward function that includes an error progression term, a phase consistency term, a control smoothing term, a safety margin barrier term, and an actuator saturation penalty term, which serves as the reinforcement learning reward function.

[0038] The formula for calculating the improved control reward function is as follows:

[0039] ;

[0040] In the formula, r t It is an improved control reward function, where t is the time step index. It is the weight of the error progression term. It is the error progression term, where e t It is the control error at time step t, e t-1It is the control error at time step t-1. It is the weight of the phase consistency term. It is the phase consistency term, where, It is the angle between the rate of change vector of the error vector and the error vector itself. It controls the weights of the smoothing terms. It is the control smoothing term, where, It is the change in throttle opening control at time step t. It is the weight of the safety margin barrier item. It is a safety margin barrier term, where m t It refers to the safety margin, specifically calculated based on the upper temperature threshold, stability margin, and physical constraints on valve speed. It is a nonlinear barrier function. It is the weight of the executor saturation penalty term, Sc t It is the actuator saturation penalty term, specifically represented by the physical limit indicator function;

[0041] The control system parameter update is specifically achieved by using the reinforcement learning state vector, the reinforcement learning action vector, and the reinforcement learning reward function, employing a deterministic policy gradient method and constructing a policy commentator network, training a small-step update reinforcement learning model to obtain a control system parameter update model, and then using the control system parameter update model to update the PID parameters of the ship's throttle control system in real time to obtain real-time correction reference data for the control parameters.

[0042] Furthermore, the throttle opening command execution is used to execute the final throttle opening command. Specifically, it involves real-time correction of reference data based on the control parameters, calculation of the PID regulator output of the ship's throttle controller, and generation of the corresponding throttle opening control quantity. By converting the throttle opening control quantity into an actuator drive signal, the throttle valve opening of the power supply system is controlled to obtain the throttle opening command.

[0043] The beneficial effects achieved by the present invention using the above solution are as follows:

[0044] (1) In view of the technical problems in the existing ship throttle adaptive control system, which relies on feedback from a single sensor and cannot achieve advance prediction of throttle control under complex speed and sea conditions, resulting in lag in throttle opening adjustment and slow system response, this solution creatively adopts advance intelligent data analysis combining working condition assessment and throttle control trajectory modeling as the basis of intelligent control. By improving the reward-based control system parameter update method, the advance and initiative of throttle opening adjustment is realized, and the dynamic response capability and stability of the ship under complex working conditions such as wave excitation and rapid maneuvering are improved.

[0045] (2) In view of the technical problems in the existing operation condition status assessment process, which rely on static threshold judgment, make it difficult to accurately estimate transient load torque, resulting in insufficient accuracy of operation condition identification and one-sided health assessment, this solution creatively adopts a lightweight temporal convolution model based on load feedforward estimation to improve the observability of operation condition and conduct operation condition status assessment. By constructing a quasi-linear reduced-order observer, the load torque of the energy conversion unit and rotating parts is dynamically estimated. An attention mechanism is introduced to weightedly fuse the operation condition features and physical load features. Finally, multi-task learning is used to realize operation condition type identification, health index calculation and trend prediction, which significantly improves the observability and diagnostic accuracy of operation condition status and can realize early warning and trend prediction under early anomalies such as shaft vibration abnormality and temperature field inhomogeneity.

[0046] (3) In the existing throttle control trajectory modeling process, there is a technical problem that relies solely on empirical prediction models without considering physical constraints, resulting in a lack of engineering feasibility and safety assurance in trajectory optimization results. This solution creatively adopts an improved time-series prediction model that combines physical constraints at multiple time scales to model the throttle control trajectory. It constructs a physical constraint objective function, and performs prediction by coupling the dynamic model of the actuator process at a fast time scale with the dynamic model of the load output at a slow time scale, and performs real-time parameter correction. At the same time, it embeds soft constraints such as the upper limit threshold of temperature, stability margin and valve rate in the optimization objective function to form a trajectory optimization with a safety boundary. This achieves the unity of dynamism and safety in throttle trajectory prediction and avoids problems such as overheating, insufficient stability margin or actuator overload caused by excessive control commands.

[0047] (4) In view of the technical problem that the existing control system parameter update process relies on fixed PID parameters or manual experience adjustment, and it is difficult to achieve adaptive adjustment of parameters under different speeds, load changes or cabin environment disturbances, this solution creatively adopts a PID parameter adjustment method that combines operating condition information and physical constraints to adjust the control system parameters. The state vector is constructed by loading the operating condition state and predicted trajectory through the reinforcement learning framework, and candidate PID parameters are generated by using scaling factors. The strategy is updated by combining the reward function containing error progression, phase consistency, safety margin barrier and actuator penalty, thereby obtaining parameter correction reference data, ensuring the dynamic adaptability of control parameters, realizing the stability of throttle control under complex conditions such as rapid load increase and violent sea state fluctuations, and significantly reducing the risk of regulation overshoot and actuator saturation. Attached Figure Description

[0048] Figure 1 A schematic diagram of the structure of a ship throttle adaptive control system for complex operating conditions provided by the present invention;

[0049] Figure 2A flowchart illustrating the steps performed by the system provided for this invention;

[0050] Figure 3 A flowchart illustrating the steps performed by the operating condition assessment module;

[0051] Figure 4 A flowchart illustrating the steps performed by the trajectory modeling module;

[0052] Figure 5 A flowchart illustrating the steps performed by the adaptive parameter update module.

[0053] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0054] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0055] Example 1, see Figure 1 The technical solution adopted by the present invention is as follows: The present invention provides a ship throttle adaptive control system for complex working conditions, including a data sensing module, a working condition evaluation module, a trajectory modeling module, an adaptive parameter update module and a throttle control module;

[0056] The data sensing module is used for collecting operating status data. Through collecting operating status data, it obtains raw operating status data and sends the raw operating status data to the operating condition evaluation module.

[0057] The operating condition assessment module is used to assess the operating condition status, obtain operating condition assessment information through the operating condition status assessment, and send the operating condition assessment information to the trajectory modeling module and the adaptive parameter update module.

[0058] The trajectory modeling module is used for throttle control trajectory modeling. Through throttle control trajectory modeling, throttle opening change prediction data is obtained, and the throttle opening change prediction data is sent to the adaptive parameter update module.

[0059] The adaptive parameter update module is used to control system parameter adjustment, obtain real-time correction reference data of control parameters by controlling system parameter adjustment, and send the real-time correction reference data of control parameters to the throttle control module.

[0060] The throttle control module is used to execute throttle opening commands. By executing throttle opening commands, it obtains throttle opening commands and controls the opening of the throttle valves of the power supply system.

[0061] By performing the above operations, this solution addresses the technical problem in existing ship throttle adaptive control systems that rely on feedback from a single sensor, making it impossible to predict throttle control in advance under complex speed and sea state conditions, resulting in lag in throttle opening adjustment and slow system response. This solution creatively adopts advance intelligent data analysis combining operating condition assessment and throttle control trajectory modeling as the basis for intelligent control. By improving the reward-based control system parameter update method, it achieves the advance and proactive nature of throttle opening adjustment, thereby improving the ship's dynamic response capability and stability under complex operating conditions such as wave excitation and rapid maneuvering.

[0062] Example 2, this example is based on the above example, see reference. Figure 1 , Figure 2 The aforementioned operational status data acquisition is used to collect operational status data in real time, specifically by acquiring raw operational status data through multi-sensor data acquisition;

[0063] The multi-sensor data acquisition includes the following steps: sensor layout acquisition, signal standardization, data preprocessing, and data frame storage;

[0064] The sensor layout acquisition specifically involves deploying multi-source sensors at key measuring points to obtain raw physical sensing data. The key measuring points include the air intake channel, the inlet and outlet of the compression unit, the energy conversion unit, the inlet and outlet of the power transmission unit, the bearing housing, and the output shaft. The multi-source sensors include a temperature sensor, a pressure sensor, a speed sensor, and a vibration accelerometer.

[0065] The signal standardization specifically involves conditioning the non-standard signals in the original physical sensing data using amplifiers and filters to convert them into standard acquisition signals, thereby obtaining standardized sensing signal data.

[0066] The data preprocessing specifically involves performing preliminary processing on the standardized sensor signal data, including engineering unit conversion, invalid data removal, and moving average filtering operations, to obtain preprocessed optimized data.

[0067] The data frame storage specifically involves packaging the preprocessed and optimized data into complete data frames according to timestamps to obtain the original running status data;

[0068] Preferably, Table 1 is an example table of data content of the original operating status data. As shown in the table, the complete data content of the original operating status data includes: intake channel parameter data, compression unit parameter data, energy conversion unit parameter data, power transmission unit parameter data, rotating component data, output shaft and power data, and environmental auxiliary data.

[0069] The intake channel parameter data includes total intake pressure, intake temperature, intake moisture content, and intake filter pressure difference; the compression unit parameter data includes inlet and outlet pressure and temperature, pressure ratio, rotational speed, and power input signal; the energy conversion unit parameter data includes energy input flow rate, input temperature, energy conversion unit outlet temperature distribution, and combustion state detection signal; the power transmission unit parameter data includes inlet and outlet pressure and temperature, rotational speed, and blade vibration acceleration; the rotating component data includes the temperature of each bearing, bearing vibration acceleration, rotor eccentricity, and lubricating oil pressure and temperature; the output shaft and power data include output shaft rotational speed, output power, torque signal, and dynamic load change rate; the environmental auxiliary data includes cabin ambient temperature and humidity, speed, propeller pitch, roll and pitch angular velocities, altitude and air pressure correction values, sensor health status labels, and timestamps;

[0070] Table 1. Example of data content for raw running status data

[0071]

[0072] As a further optimization of this embodiment, after the above-mentioned data are standardized by signals and preprocessed by data, they are packaged into data frames of a unified format in chronological order. Each data frame corresponds to the complete operating state of the ship at a certain moment and serves as the basic input for operating condition assessment, trajectory modeling and adaptive parameter update.

[0073] Example 3, this example is based on the above examples, see reference. Figure 1 , Figure 2 and Figure 3 The operational condition status assessment is used to assess the current specific operational condition. Specifically, based on the original operational condition data, a lightweight temporal convolutional model with improved operational condition observability based on load feedforward estimation is used to assess the operational condition status and obtain operational condition assessment information. The assessment includes the following steps: load feedforward estimation, operational condition feature extraction, observability enhancement fusion, and operational condition classification output.

[0074] The load feedforward estimation specifically involves using energy conversion unit parameter data and rotating component data to construct a reduced-order state observer based on a quasi-linear model to dynamically estimate the current load torque, obtaining the estimated load torque value data. The calculation formula is as follows:

[0075] ;

[0076] In the formula, This is the estimated load torque value data, where L is the observer index, k is the discrete time step index, and E... k It is a linear transformation matrix, A k It is the state transition matrix. B is the reduced-order state estimate vector from the previous discrete time step. k This is the input action matrix, used to represent the degree of influence of the ship's operational inputs on the reduced-order state estimation vector, u k-1 L is the control operation applied at the previous discrete moment. k It is the observer gain matrix, y k It is the observation vector at the current moment, specifically represented as a combination of energy conversion unit parameter data and rotating component data, C k It is the mapping matrix from state to observation;

[0077] The operating condition feature extraction specifically involves using the estimated load torque value data and the original operating status data to construct an improved lightweight temporal convolutional network to extract features with multi-scale and long-term dependencies, thereby obtaining a high-dimensional feature vector of operating condition information.

[0078] Preferably, Table 2 is an example table of structural parameters for the improved lightweight temporal convolutional network. As shown in the table, in this embodiment, the lightweight temporal convolutional network is improved based on the traditional one-dimensional convolutional structure; the improvements include:

[0079] By introducing depthwise separable convolution to replace conventional convolution, and decomposing temporal convolution operations into channel-wise convolution and point convolution, the number of parameters and computational load are significantly reduced, ensuring the feasibility of network deployment in the real-time ship condition monitoring environment.

[0080] By introducing convolutional kernels with different dilation rates into each convolutional block, corresponding to the feature extraction of short-term, medium-term, and long-term dependencies respectively, the receptive field is expanded, and the temporal modeling capability of working condition information is effectively improved without increasing the number of layers.

[0081] Setting residual connections between convolutional blocks ensures the stability of feature transfer between layers and avoids the gradient vanishing problem in deep networks.

[0082] Based on multi-scale output, a fusion layer of feature concatenation and one-dimensional convolution aggregation was designed to compress and unify temporal features of different scales, resulting in a high-dimensional feature vector of working condition information with consistent dimensions.

[0083] Table 2. Examples of structural parameters for the improved lightweight temporal convolutional networks.

[0084]

[0085] The observability enhancement fusion specifically involves dynamically calculating the fusion weights through the attention layer based on the high-dimensional feature vector of the operating condition information and the estimated load torque value data, and performing adaptive weighted fusion of the operating condition features and the load feedforward features of the physical model to obtain the observability enhancement fusion feature vector.

[0086] The formula for calculating the observability enhancement fusion feature vector is as follows:

[0087] ;

[0088] In the formula, f k It is an observability-enhanced fusion feature vector. It is the dynamic fusion weight of the attention layer, h k It is a high-dimensional feature vector of working condition information, q k It is the load feedforward feature of the physical model, specifically the feature vector obtained by linear projection of the estimated load torque value data;

[0089] The operational condition classification output is specifically based on the observability enhancement fusion feature vector, and through a multi-task learning network branch, operational condition type identification, health assessment, and short-term trend prediction are performed to obtain operational condition assessment information.

[0090] Preferably, the operating condition type identification branch is used to perform classification calculations based on the observability enhancement fusion feature vector to obtain an operating condition label, which is used to characterize the operating condition status of the ship.

[0091] Preferably, the health assessment branch is used to perform regression calculations based on the observability enhancement fusion feature vector to obtain a health index, which reflects the health level of key components or the overall system.

[0092] Preferably, the short-term trend prediction branch is used to perform prediction calculations based on the observability enhancement fusion feature vector to obtain the predicted trend, which helps to identify operational fluctuations or potential risks in advance.

[0093] The operational condition assessment information specifically includes operational condition labels, health index, and predicted trends.

[0094] By performing the above operations, this solution addresses the technical problems in existing operational condition assessment processes, which rely on static threshold judgments, making it difficult to accurately estimate transient load torque, resulting in insufficient accuracy in condition identification and one-sided health assessment. It creatively employs a lightweight temporal convolutional model based on load feedforward estimation to enhance operational condition observability. This model assesses operational conditions by constructing a quasi-linear reduced-order observer to dynamically estimate the load torque of energy conversion units and rotating components. An attention mechanism is introduced to weightedly fuse operational condition features and physical load features. Finally, multi-task learning is used to achieve operational condition type identification, health index calculation, and trend prediction, significantly improving the observability and diagnostic accuracy of operational conditions. This enables early warning and trend prediction for early anomalies such as shaft vibration abnormalities and uneven temperature fields.

[0095] Example 4, this example is based on the above examples, see below. Figure 1 and Figure 4 The throttle control trajectory modeling is used to dynamically generate optimized throttle control trajectories. Specifically, based on the operating condition evaluation information, an improved time-series prediction model combining multiple time-scale physical constraints is used to model the throttle control trajectory and obtain throttle opening change prediction data. The steps include: target constraint construction, multi-time-scale prediction model construction, physical constraint embedding objective function construction, constraint trajectory solving, and throttle control trajectory modeling.

[0096] The construction of the target constraint specifically involves constructing a basic objective function containing an error term and a control smoothing term based on the operating condition evaluation information.

[0097] The error term is used to measure the deviation between the predicted throttle trajectory and the target trajectory, and the calculation formula is as follows:

[0098] ;

[0099] In the formula, J err This is an error term constraint, where H is the prediction time domain length. It is the predicted throttle opening value corresponding to the kth discrete time. It is a reference throttle opening value;

[0100] The control smoothing term, used to limit fluctuations in the rate of change of throttle opening, is calculated using the following formula:

[0101] ;

[0102] In the formula, J smooth It controls the smoothing term constraint. It is the predicted throttle opening value corresponding to the (k+1)th discrete time.

[0103] The construction of the multi-timescale prediction model specifically involves establishing a fast-timescale actuator process dynamic prediction model and a slow-timescale load output dynamic prediction model based on the operating condition evaluation information and the basic objective function. The model parameters are then corrected in real time through a parameter recursive update method to obtain the multi-timescale prediction output.

[0104] Preferably, the fast timescale actuator process dynamic prediction model is implemented based on the linear state-space modeling method. Specifically, it is modeled and predicted by calling the modeling and simulation functions ss and lsim of the Control System Toolbox v10.14 in MATLAB R2022b, which are used to capture the rapid response relationship between valve opening and energy input and energy conversion unit outlet temperature.

[0105] Preferably, the slow-time-scale load output dynamic prediction model is based on an autoregressive model with exogenous input. Specifically, it uses the arx function provided in the System Identification Toolbox v9.15 of MATLAB R2022b, and combines speed and power data to perform reduced-order modeling to achieve slow dynamic prediction under propulsion load.

[0106] Preferably, the parameter recursive update method is based on the recursive least squares algorithm, specifically calling the recursiveLS interface provided in the System Identification Toolbox v9.15 of MATLAB R2022b to complete the real-time correction of model parameters, so that the fast time scale and slow time scale prediction models can adapt to different operating conditions and maintain prediction accuracy.

[0107] The construction of the physical constraint embedding objective function specifically involves introducing physical constraint terms and weighting and combining the physical constraint terms with the terms in the basic objective function to obtain the physical constraint embedding optimization objective function, which is used for constraint trajectory solving.

[0108] The physical constraints specifically combine the health status and trend boundaries in the operating condition assessment information to apply soft constraints to the upper temperature threshold, stability margin, and valve rate.

[0109] The formula for calculating the physical constraint term is:

[0110] ;

[0111] In the formula, J phy These are physical constraints. It is the soft constraint weight of the upper temperature threshold. It is a soft constraint on the upper temperature threshold, T EGT(k) is the predicted upper temperature threshold value, where k is the discrete time step index. It is the safe upper limit threshold for the temperature upper limit. It is a soft constraint weight for stability margin. It is a soft constraint with stability margin. It is the lower bound threshold of the stability margin, and SM(k) is the predicted value of the stability margin. It is the valve rate soft constraint weight. It is a valve rate soft constraint. It is the predicted valve rate of throttle opening, r max This is the maximum valve rate of the throttle opening. + It is a positive function; the soft constraint weights of the upper temperature threshold, stability margin, and valve rate are specifically adaptively adjusted based on the health index and the boundary of the predicted trend in the operating condition assessment information.

[0112] Preferably, the formula for calculating the soft constraint weights for adaptive boundary adjustment is as follows:

[0113] ;

[0114] In the formula, These are soft constraint weights, used to represent the upper temperature threshold soft constraint weight, the stability margin soft constraint weight, and the valve rate soft constraint weight. These are baseline weights, used to represent the initial set values ​​under normal health conditions and safety boundary conditions. H is the health adjustment coefficient, representing the sensitivity of the health index to the weights. idx It is a health index, derived from the health index in the operational condition assessment information, used to reflect the current or predicted health level. Z is the boundary proximity adjustment coefficient, representing the sensitivity of the predicted trend to the weights in relation to the safety boundary. bound,i Z is the safety boundary value of a physical quantity. current,i These are the predicted values ​​corresponding to the physical quantities;

[0115] In this embodiment, the absolute value operator is used to represent the deviation from the boundary value; however, it should be noted that the positive part function can also be used instead of the absolute value, and a penalty contribution is only generated when the predicted value exceeds the safety upper limit (or falls below the safety lower limit), thereby avoiding misjudgment as deviation on the safe side; in this way, the adaptive weight adjustment can better meet the actual needs of physical constraints.

[0116] The formula for calculating the physical constraint embedding optimization objective function is as follows:

[0117] ;

[0118] In the formula, J is the physical constraint embedding optimization objective function. It is the weight of the error term, J err It is an error term constraint. It controls the weights of the smoothing term, J smooth It controls the smoothing term constraint;

[0119] The constraint trajectory solution is specifically performed by embedding the physical constraint into the optimization objective function and using a rolling optimization method to solve the constraint on the throttle opening change trajectory, thereby obtaining a throttle opening change prediction sequence that satisfies the constraint conditions.

[0120] The throttle control trajectory modeling specifically involves generating a time-stamped throttle control reference trajectory based on the throttle opening change prediction sequence, thereby obtaining throttle opening change prediction data.

[0121] By performing the above operations, this solution addresses the technical problem in existing throttle control trajectory modeling processes that rely solely on empirical prediction models without considering physical constraints, resulting in a lack of engineering feasibility and safety assurance in trajectory optimization results. This solution creatively employs an improved time-series prediction model that combines multi-timescale physical constraints for throttle control trajectory modeling. It constructs a physical constraint objective function and couples the dynamic model of the actuator process at a fast timescale with the dynamic model of the load output at a slow timescale for prediction, performing real-time parameter correction. Simultaneously, it embeds soft constraints such as upper temperature threshold, stability margin, and valve rate into the optimization objective function, forming a trajectory optimization with safety boundaries. This achieves a balance between the dynamism and safety of throttle trajectory prediction, avoiding problems such as overheating, insufficient stability margin, or actuator overload caused by excessive control commands.

[0122] Example 5, this example is based on the above examples, see below. Figure 1 and Figure 5 The control system parameter adjustment is used to adjust the parameters of the control system in real time. Specifically, based on the operating condition evaluation information and the throttle opening change prediction data, a PID parameter adjustment method that combines operating condition information and physical constraints is used to adjust the control system parameters and obtain real-time correction reference data for the control parameters. The adjustment includes the following steps: loading state parameters, generating action parameters, improving control rewards, and updating control system parameters.

[0123] The loading of state parameters specifically involves constructing a reinforcement learning state vector based on the operating condition evaluation information and the throttle opening change prediction data output by the throttle control trajectory modeling module.

[0124] The formula for calculating the reinforcement learning state vector is:

[0125] ;

[0126] In the formula, s t It is the reinforcement learning state vector, e t This is the current error. It is the rate of change of error. It is the reference valve position increment, C t It is a working condition classification code, HI t It's about health, m t It is a safety margin;

[0127] The action parameter generation specifically involves constructing a three-dimensional scaling factor and scaling the baseline PID parameters according to the three-dimensional scaling factor to obtain candidate parameters, which are then used as reinforcement learning action vectors.

[0128] The formula for calculating the three-dimensional scaling factor is:

[0129] ;

[0130] In the formula, a t It is the 3D scaling factor, s p It is the scaling factor of the proportional element, s i It is the scaling factor of the integral element, s d It is the scaling factor of the differential element;

[0131] The formula for calculating the candidate parameters is as follows:

[0132] ;

[0133] In the formula, K p It is the candidate proportional gain. It is the baseline scaling gain, K i It is the candidate integral gain. It is the baseline integral gain, K d It is the candidate differential gain. It is the baseline differential gain;

[0134] The improved control reward specifically involves constructing an improved control reward function that includes an error progression term, a phase consistency term, a control smoothing term, a safety margin barrier term, and an actuator saturation penalty term, which serves as the reinforcement learning reward function.

[0135] The formula for calculating the improved control reward function is as follows:

[0136] ;

[0137] In the formula, r t It is an improved control reward function, where t is the time step index. It is the weight of the error progression term. It is the error progression term, where e tIt is the control error at time step t, e t-1 It is the control error at time step t-1. It is the weight of the phase consistency term. It is the phase consistency term, where, It is the angle between the rate of change vector of the error vector and the error vector itself. It controls the weights of the smoothing terms. It is the control smoothing term, where, It is the change in throttle opening control at time step t. It is the weight of the safety margin barrier item. It is a safety margin barrier term, where m t It refers to the safety margin, specifically calculated based on the upper temperature threshold, stability margin, and physical constraints on valve speed. It is a nonlinear barrier function. It is the weight of the executor saturation penalty term, Sc t It is the actuator saturation penalty term, specifically represented by the physical limit indicator function;

[0138] The formula for calculating the phase consistency term is:

[0139] ;

[0140] In the formula, It is the change in throttle opening control at time step t. It is the reference valve position increment. It is to prevent the parameter from being removed from zero;

[0141] The calculation formula for the safety margin barrier term is as follows:

[0142] ;

[0143] In the formula, This is the safety margin threshold, with a default value of 0.15. This is a penalty constant for violations, with a default value of 10;

[0144] Preferably, the default value of the error progression term weight is 1.0, the default value of the phase consistency term weight is 0.5, the default value of the control smoothing term weight is 0.2, the default value of the safety margin barrier term weight is 0.7, and the default value of the actuator saturation penalty term weight is 0.5.

[0145] The control system parameter update is specifically achieved by using the reinforcement learning state vector, the reinforcement learning action vector, and the reinforcement learning reward function, employing a deterministic policy gradient method and constructing a policy commentator network, training a small-step update reinforcement learning model to obtain a control system parameter update model, and then using the control system parameter update model to update the PID parameters of the ship's throttle control system in real time to obtain real-time correction reference data for the control parameters.

[0146] By performing the above operations, this solution addresses the technical problem that existing control system parameter updates rely on fixed PID parameters or manual experience, making it difficult to achieve adaptive parameter adjustment under different speeds, sudden load changes, or cabin environmental disturbances. It creatively employs a PID parameter adjustment method that combines operating condition information and physical constraint optimization to adjust control system parameters. A state vector is constructed by loading the operating condition and predicted trajectory using a reinforcement learning framework. Candidate PID parameters are generated using scaling factors, and a reward function incorporating error progression, phase consistency, safety margin barriers, and actuator penalties is used for strategy updates. This yields reference data for parameter correction, ensuring the dynamic adaptability of control parameters and achieving stability of throttle control under complex conditions such as rapid load increases and drastic sea state fluctuations. This significantly reduces the risks of regulation overshoot and actuator saturation.

[0147] Example 6, this example is based on the above examples, see below. Figure 1 The throttle opening command execution is used to execute the final throttle opening command. Specifically, it involves real-time correction of reference data based on the control parameters, calculation of the PID regulator output of the ship's throttle controller, and generation of the corresponding throttle opening control quantity. By converting the throttle opening control quantity into an actuator drive signal, the throttle valve opening of the power supply system is controlled to obtain the throttle opening command.

[0148] It should be noted that, in this document, relational terms such as "first" and "second" are used merely 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 or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process or method.

[0149] 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.

[0150] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A ship throttle adaptive control system for complex operating conditions, characterized in that: It includes a data perception module, a working condition assessment module, a trajectory modeling module, an adaptive parameter update module, and a throttle control module; The data sensing module is used for collecting operating status data. Through collecting operating status data, it obtains raw operating status data and sends the raw operating status data to the operating condition evaluation module. The operating condition assessment module uses a lightweight temporal convolutional model based on load feedforward estimation to enhance operating condition observability and assess the operating condition status, obtaining operating condition assessment information. This includes the following steps: load feedforward estimation, operating condition feature extraction, observability enhancement fusion, and operating condition classification output. The operating condition assessment information is then sent to the trajectory modeling module and the adaptive parameter update module. The load feedforward estimation specifically involves using energy conversion unit parameter data and rotating component parameter data to construct a reduced-order state observer based on a quasi-linear model to dynamically estimate the current load torque and obtain estimated load torque value data. The observability enhancement fusion specifically involves dynamically calculating the fusion weights through the attention layer based on the high-dimensional feature vector of the operating condition information and the estimated load torque value data, and performing adaptive weighted fusion of the operating condition features and the load feedforward features of the physical model to obtain the observability enhancement fusion feature vector. The trajectory modeling module employs an improved time-series prediction model incorporating multi-timescale physical constraints to model the throttle control trajectory and obtain throttle opening change prediction data. This includes the following steps: target constraint construction, multi-timescale prediction model construction, physical constraint embedding objective function construction, constraint trajectory solving, and throttle control trajectory modeling; and then sends the throttle opening change prediction data to the adaptive parameter update module. The throttle control trajectory modeling is used to dynamically generate optimized throttle control trajectories; The construction of the target constraint specifically involves constructing a basic objective function containing an error term and a control smoothing term based on the operating condition evaluation information. The construction of the multi-timescale prediction model specifically involves establishing a fast-timescale actuator process dynamic prediction model and a slow-timescale load output dynamic prediction model based on the operating condition evaluation information and the basic objective function. The model parameters are then corrected in real time through a parameter recursive update method to obtain the multi-timescale prediction output. The construction of the physical constraint embedding objective function specifically involves introducing physical constraint terms and weighting and combining the physical constraint terms with the terms in the basic objective function to obtain the physical constraint embedding optimization objective function, which is used for constraint trajectory solving. The constraint trajectory solution is specifically performed by embedding the physical constraint into the optimization objective function and using a rolling optimization method to solve the constraint on the throttle opening change trajectory, thereby obtaining a throttle opening change prediction sequence that satisfies the constraint conditions. The throttle control trajectory modeling specifically involves generating a time-stamped throttle control reference trajectory based on the throttle opening change prediction sequence, thereby obtaining throttle opening change prediction data. The adaptive parameter update module uses a PID parameter adjustment method that combines operating condition information and physical constraints to adjust the control system parameters and obtain real-time correction reference data for the control parameters. The steps include: loading state parameters, generating action parameters, improving control rewards, and updating control system parameters; and sending the real-time correction reference data for the control parameters to the throttle control module. The throttle control module is used to execute throttle opening commands. By executing throttle opening commands, it obtains throttle opening commands and controls the opening of the throttle valves of the power supply system.

2. The ship throttle adaptive control system for complex operating conditions according to claim 1, characterized in that: The operation status data acquisition is used to collect operation status data in real time, specifically by acquiring raw operation status data through multi-sensor data acquisition; The multi-sensor data acquisition includes the following steps: sensor layout acquisition, signal standardization, data preprocessing, and data frame storage; The raw operating status data includes: intake channel parameter data, compression unit parameter data, energy conversion unit parameter data, power transmission unit parameter data, rotating component data, output shaft and power data, and environmental auxiliary data.

3. The ship throttle adaptive control system for complex operating conditions according to claim 2, characterized in that: The operating condition status assessment is used to assess the specific operating condition currently in which the system is operating; The operating condition feature extraction specifically involves using the estimated load torque value data and the original operating status data to construct an improved lightweight temporal convolutional network to extract features with multi-scale and long-term dependencies, thereby obtaining a high-dimensional feature vector of operating condition information. The operational condition classification output is specifically based on the observability enhancement fusion feature vector, and through a multi-task learning network branch, operational condition type identification, health assessment, and short-term trend prediction are performed to obtain operational condition assessment information. The operational condition assessment information specifically includes operational condition labels, health index, and predicted trends.

4. The ship throttle adaptive control system for complex operating conditions according to claim 3, characterized in that: The physical constraints specifically combine the health status and trend boundaries in the operating condition assessment information to apply soft constraints to the upper temperature threshold, stability margin, and valve rate. The formula for calculating the physical constraint term is: ； In the formula, J phy These are physical constraints. It is the soft constraint weight of the upper temperature threshold. It is a soft constraint on the upper limit of temperature threshold, T EGT (k) is the predicted upper temperature threshold value, where k is the discrete time step index. It is the safe upper limit threshold for the temperature upper limit. It is a stability margin soft constraint weight. It is a soft constraint with stability margin. It is the lower bound threshold of the stability margin, and SM(k) is the predicted value of the stability margin. It is the valve rate soft constraint weight. It is a valve rate soft constraint. It is the predicted valve rate of throttle opening, r max This is the maximum valve rate of the throttle opening. + It is a positive function; the soft constraint weights of the upper temperature threshold, stability margin, and valve rate are specifically adaptively adjusted based on the health index and the boundary of the predicted trend in the operating condition assessment information. The formula for calculating the physical constraint embedding optimization objective function is as follows: ； In the formula, J is the physical constraint embedding optimization objective function. It is the weight of the error term, J err It is an error term constraint. It controls the weights of the smoothing term, J smooth It controls the smoothing term constraint.

5. The ship throttle adaptive control system for complex operating conditions according to claim 4, characterized in that: The control system parameter adjustment is used to adjust the parameters of the control system in real time; The loading of state parameters specifically involves constructing a reinforcement learning state vector based on the operating condition evaluation information and the throttle opening change prediction data output by the throttle control trajectory modeling module. The action parameter generation specifically involves constructing a three-dimensional scaling factor and scaling the baseline PID parameters according to the three-dimensional scaling factor to obtain candidate parameters, which are then used as reinforcement learning action vectors. The improved control reward specifically involves constructing an improved control reward function that includes an error progression term, a phase consistency term, a control smoothing term, a safety margin barrier term, and an actuator saturation penalty term, which serves as the reinforcement learning reward function. The control system parameter update is specifically achieved by using the reinforcement learning state vector, the reinforcement learning action vector, and the reinforcement learning reward function, employing a deterministic policy gradient method and constructing a policy commentator network, training a small-step update reinforcement learning model to obtain a control system parameter update model, and then using the control system parameter update model to update the PID parameters of the ship's throttle control system in real time to obtain real-time correction reference data for the control parameters.

6. The ship throttle adaptive control system for complex operating conditions according to claim 5, characterized in that: The formula for calculating the improved control reward function is as follows: ； In the formula, r t It is an improved control reward function, where t is the time step index. It is the weight of the error progression term. It is the error progression term, where e t It is the control error at time step t, e t-1 It is the control error at time step t-1. It is the weight of the phase consistency term. It is the phase consistency term, where, It is the angle between the rate of change vector of the error vector and the error vector itself. It controls the weights of the smoothing terms. It is the control smoothing term, where, It is the change in throttle opening control at time step t. It is the weight of the safety margin barrier item. It is a safety margin barrier term, where m t It refers to the safety margin, specifically calculated based on the upper temperature threshold, stability margin, and physical constraints on valve speed. It is a nonlinear barrier function. It is the weight of the actuator saturation penalty term, Sc t It is the actuator saturation penalty term, specifically represented by the physical limit indicator function.

7. The ship throttle adaptive control system for complex operating conditions according to claim 6, characterized in that: The throttle opening command execution is used to execute the final throttle opening command. Specifically, it involves real-time correction of reference data based on the control parameters, calculation of the PID regulator output of the ship's throttle controller, and generation of the corresponding throttle opening control quantity. By converting the throttle opening control quantity into an actuator drive signal, the throttle valve opening of the power supply system is controlled to obtain the throttle opening command.

Images