A deep learning and data-driven based online cooperative control system for rotational speed of series hybrid engine

Abstract

A kind of online collaborative control system of series hybrid engine speed based on deep learning and data driving, including parameter acquisition and preprocessing system, control decision system, execution system, feedback system and hybrid engine model;Hybrid engine model is responsible for state generation, parameter acquisition and preprocessing system is responsible for obtaining and processing data;Hybrid engine model as the virtual physical environment and data source of system;Parameter acquisition and preprocessing system real-time capture hybrid engine multidimensional simulation running data, carry out data cleaning through internal filtering and abnormal detection mechanism, and according to the comprehensive perception accuracy calculated, data quality is judged, abnormal signal is intercepted, to ensure that the information passed to the back stage has high precision and reliability;Control decision system is responsible for decision support, provides command issued;Execution system is responsible for physical action conversion.The present application establishes multidimensional parameter collaborative perception and closed-loop feedback system, accurately compensates uncertainty interference, and improves speed control precision.

Description

Technical Field

[0001] This invention belongs to the field of online coordinated control technology of hybrid engine speed, and specifically relates to an online coordinated control system for series hybrid engine speed based on deep learning and data-driven principles. Background Technology

[0002] Limited by current technological bottlenecks in battery energy density and electric motor performance, existing battery technology cannot yet support the mission requirements of all-electric flight for high-thrust aircraft. Currently, the energy density of commercial lithium batteries can reach 200Wh / kg, with some laboratory levels reaching 500Wh / kg, while the energy density of fuel is 13kWh / kg, dozens of times higher than that of batteries. Therefore, all-electric propulsion systems cannot completely replace traditional fuel propulsion systems and are only applied to some small general aviation electric aircraft, failing to achieve a leapfrog innovation in propulsion systems. Consequently, pure electric aircraft struggle to overcome technological barriers in real-world scenarios such as long-distance, long-range flights, and cannot meet the diverse requirements of aircraft for range, payload, and reliability. Against this backdrop, hybrid electric propulsion systems (HEPS) have emerged as an innovative and practical aviation power solution.

[0003] HEPS typically consists of a power supply system, a propulsion system, and an energy storage system. This coupled design of "traditional aircraft engine + electric system" inevitably increases the complexity of the original power system, thus placing stringent requirements on the rationality and accuracy of the control method.

[0004] The speed control accuracy of hybrid engines directly affects stability and safety. Currently, existing methods utilize adaptive neural networks for trajectory tracking in rehabilitation robots, exhibiting some nonlinear adaptability; others predict the current of dual active bridge transformers using neural networks, improving measurement accuracy but lacking real-time performance. Some researchers have proposed a control method combining fractional-order proportional-integral-derivative (PID) circuits with neural networks, enhancing robustness, but its implementation is complex. The "fuel + electric" coupling design of HEPS increases control complexity, and speed control accuracy directly impacts safety and stability. The current mainstream control scheme is "fractional-order PID + single neural network compensation," which, while performing well in some scenarios, suffers from strong modeling dependence, insufficient coordination, single feedback, and an imbalance between real-time performance and accuracy. It cannot adapt to complex operating conditions such as hybrid mode switching and sudden load changes, thus failing to accurately respond to dynamic speed changes and correct speed deviations, affecting control accuracy and stability.

[0005] Existing speed control for hybrid aircraft engines must simultaneously meet three core requirements: nonlinear coupling adaptation, multi-condition response, and dual-actuator collaboration. However, current mainstream control schemes have failed to fully adapt to these requirements, resulting in problems such as insufficient control accuracy and poor stability during actual flight.

[0006] Therefore, there is an urgent need for a deep learning-based online collaborative adjustment system for the speed control of a series hybrid engine, which can dynamically correct the fuel supply and motor output power to achieve dynamic speed adjustment. This system can ensure high precision in speed control and improve stability under complex operating conditions such as hybrid mode switching and sudden load changes, ultimately achieving precision and stability in intelligent speed control of the hybrid engine.

[0007] The current mainstream solution for speed control of hybrid engines is the "fractional PID + single neural network compensation" control scheme, which is implemented as follows:

[0008] 1. It adopts separate control of "fuel regulation unit + motor power regulation unit" and has no collaborative decision-making module;

[0009] 2. Parameters such as engine speed and fuel flow rate are processed only by simple mean filtering, without anomaly detection and coupling interference suppression;

[0010] 3. Based on the mechanism modeling of the hybrid engine (fixed parameters such as combustion efficiency and motor response characteristics need to be preset), the fuel supply is adjusted by fractional-order PID, and a single neural network (e.g., a single neural network) is used to compensate for steady-state error.

[0011] 4. It adopts a semi-closed-loop feedback, only collecting engine speed as a feedback parameter, without combining dynamic parameters such as battery SOC and load power for real-time correction.

[0012] Disadvantages of existing technology:

[0013] 1. It relies on a precise dynamic model of the hybrid engine, which has poor adaptability to unknown parameters and fluctuations in operating conditions;

[0014] 2. The fuel regulation and electric regulation are controlled separately without a coordinated constraint mechanism, which easily leads to action conflicts and large speed fluctuations under complex operating conditions, affecting stability;

[0015] 3. Based solely on rotational speed feedback correction, without integrating multi-dimensional parameters such as battery SOC and load power, it cannot compensate for coupling uncertainties and has weak interference suppression capabilities;

[0016] 4. Complex algorithms (such as fractional PID + neural network) have cumbersome implementation processes, long control cycles, insufficient real-time performance, and the complex calculations increase the computing power burden on airborne equipment. Summary of the Invention

[0017] To overcome the above technical problems, the present invention aims to provide an online collaborative control system for the speed of a series hybrid engine based on deep learning and data-driven principles. This system improves the system's adaptability to nonlinear coupling characteristics and operating condition fluctuations; it constructs a collaborative control mechanism for fuel-electric dual actuators to avoid action conflicts and enhance stability under complex operating conditions; and it establishes a multi-dimensional parameter collaborative perception and closed-loop feedback system to accurately compensate for uncertain interference and improve speed control accuracy.

[0018] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0019] A series hybrid engine speed online coordinated control system based on deep learning and data-driven principles includes a parameter acquisition and preprocessing system, a control decision system, an execution system, a feedback system, and a hybrid engine model.

[0020] Each module forms a complete closed-loop collaborative working system based on the logic of data acquisition → decision calculation → execution control → closed-loop feedback.

[0021] The hybrid engine model is responsible for state generation, while the parameter acquisition and preprocessing system is responsible for acquiring and processing the data.

[0022] The hybrid engine model serves as the system's virtual physical environment and data source, dynamically generating simulation operation data that includes multiple dimensions such as engine speed, battery SOC, and load power. The parameter acquisition and preprocessing system captures this raw data in real time, cleans the data through internal filtering and anomaly detection mechanisms, and determines the data quality based on the calculated comprehensive sensing accuracy, intercepting abnormal signals to obtain high-quality data vectors. This ensures that the information transmitted to subsequent stages has high accuracy and reliability.

[0023] The parameter acquisition and preprocessing system is responsible for transmitting data to the control and decision system, which provides decision support and issues commands. The control and decision system is responsible for transmitting commands to the execution system, which is responsible for converting physical actions.

[0024] The high-quality data vectors obtained from the parameter acquisition and preprocessing system are input to the control decision system. Based on the high-quality data vectors, the control decision system uses a deep learning model to predict the speed change trend and performs online calculations in conjunction with a sliding mode adaptive control law to generate the optimal fuel and electric power coordinated control command.

[0025] High-quality data vectors specifically refer to a high-quality data sequence composed of turboshaft engine gas turbine speed, power turbine speed, oil-electric power distribution ratio, load demand power, battery SOC, and exhaust temperature.

[0026] The control decision system sends the calculated coordinated control commands to the execution system. After receiving the commands, the execution system drives its internal fuel regulation and electric regulation dual execution units to convert the digital commands into corresponding simulated fuel supply and motor output power.

[0027] The execution system is responsible for controlling the application of oil and electricity, which drives the mathematical model of the engine to rotate and calculate a completely new current speed;

[0028] The actual fuel quantity and motor power output by the execution system directly affect the hybrid engine model; after receiving the synergistic effect of the two power sources, the model's internal dynamic state changes, thereby outputting the updated actual speed.

[0029] Every 10 milliseconds, the difference between the actual speed and the target speed is checked (error closed-loop operation), and then the fuel injection quantity and motor power are immediately fine-tuned (dynamic correction) to make it as close as possible to the target speed;

[0030] The feedback system monitors and collects the actual engine speed after the change in speed in real time, compares it with the target speed (the predicted value of the BP neural network), and calculates the real-time speed deviation signal. The working cycle of this system is strictly consistent with the control cycle, that is, it is executed once every 10 milliseconds (10ms) to ensure real-time dynamic correction.

[0031] After the feedback system detects the deviation, it sends this signal back to the control decision system. The control decision system uses this deviation data to dynamically update the compensation parameters in the sliding mode adaptive control law for the next cycle (10ms), thereby calculating the optimal correction command.

[0032] The feedback system monitors the actual speed output by the hybrid engine model in real time and compares it with the target speed to calculate the real-time speed deviation. Subsequently, the feedback system sends the deviation signal back to the control decision system. The control decision system calculates the correction compensation amount based on this and instructs the system to return to the parameter acquisition stage to start the next sampling cycle, thereby realizing closed-loop adaptive compensation for interference and error.

[0033] A method for online coordinated control of the speed of a series hybrid engine based on deep learning and data-driven approaches, comprising the following steps;

[0034] Step 1: Collect multi-dimensional operational data, perform three levels of preprocessing, determine data quality based on comprehensive sensing accuracy, and output high-quality data vectors;

[0035] The data vector is a high-quality data sequence (i.e., a data vector containing six dimensions, including speed, power distribution, battery SOC, exhaust temperature, etc.) from the current cycle and the previous 99 cycles, which is input into the BP neural network. Based on this data, the neural network predicts the target speed at future times.

[0036] Step 2: The BP neural network predicts the future speed change trend, uses data vector-driven modeling to judge the current dynamic model, and combines the speed deviation to generate coordination commands;

[0037] Step 3: After receiving the coordination command, adjust the fuel injection quantity and the electric motor output power to work together to change the actual engine speed;

[0038] Step 4: The difference between the actual speed and the target speed is calculated and sent back to the control decision system as a feedback signal to correct the control command for the next cycle. Then the system enters the next control cycle, returns to Step 1 to collect real-time data again, and realizes closed-loop control.

[0039] Step one is executed by the parameter acquisition and preprocessing system in conjunction with the hybrid engine model. The engine model generates the initial physical state, and the parameter acquisition and preprocessing system is responsible for capturing these data and performing three-level filtering, purification, and quality judgment to finally output high-quality data.

[0040] Step two is executed independently by the control decision system; the system receives the data from step one, runs a BP neural network and a sliding mode adaptive algorithm, and calculates the optimal fuel and electric coordination command.

[0041] Step three is completed by the execution system acting on the hybrid engine model; the execution system receives digital instructions, converts them into actual physical control variables (such as fuel injection quantity and motor power), and drives the engine model to change its current operating speed.

[0042] Step four is executed by the feedback system. The feedback system compares the actual speed after the model change with the target speed in real time, transmits the error data back to the control decision system for parameter correction, and instructs the system to return to execute step one to start the next complete cycle of iteration, so that the actual speed of the engine smoothly approaches the target speed.

[0043] Step one involves collecting multi-dimensional operational data, and after three levels of preprocessing, outputting high-quality data vectors based on the data quality assessment results.

[0044] The preprocessing process involves three levels of preprocessing: arithmetic mean filtering, adaptive Kalman filtering, and isolated forest anomaly detection (in the parameter acquisition and preprocessing system). The data quality is quantified by calculating the comprehensive sensing accuracy. When the comprehensive sensing accuracy is ≥95%, the data is deemed valid and a high-quality data vector is output. Otherwise, abnormal signals are intercepted to ensure the stability and reliability of the data input to the control and decision-making system.

[0045] Cruise operating points at specific altitudes and Mach numbers are selected, and the collected real-time parameters are substituted into the control system for filtering and anomaly detection. The fuel injection quantity and the increase in motor output power are selected as control variables, and data are continuously collected during engine operation.

[0046] At flight altitude , Initial steady-state speed As a basic operating point, and in A sudden 20% increase in load demand was set to simulate a motorized climb or encounter with gusts of wind; to verify the robustness of the system, high-frequency pulse noise with an amplitude of 5% was artificially introduced into the acquisition channel.

[0047] For series hybrid engines, as flight missions become more complex, it is necessary to more precisely coordinate and adjust the fuel and motor power to avoid conflicts in the actions of the two actuators.

[0048] The essence of online coordinated speed control of a series hybrid engine is a control process aimed at rapidly eliminating speed deviations while ensuring the safety of the battery and motor. To ensure the safe operation of the aircraft during missions requiring rapid power response, such as sudden load changes, the controlled independent variable is the change in fuel flow rate. and motor power , The rated power constraints for the motor are as follows;

[0049] Constraints:

[0050]

[0051] This represents the battery's current state of charge. , These are the minimum and maximum allowable states of charge for the battery, respectively.

[0052] This is the fuel injection quantity adjustment value within a single control cycle. , These are its lower and upper limits for adjustment, respectively;

[0053] This represents the current output power of the motor. This refers to the rated output power of the motor.

[0054] In step two, the BP neural network predicts the future speed change trend and judges the current dynamic model through data-driven modeling. In each control cycle, the control decision system predicts the target speed at the next moment based on the high-quality data sequence of the previous 100 cycles, and updates the pseudo-partial derivative parameters online. This overcomes the mismatch error of traditional fixed-parameter mechanism models;

[0055] Subsequently, a sliding surface with cooperative constraints is established, and cooperative instructions are generated by combining the prediction results; the optimal conflict-free instructions in the current iteration process are solved and output only if the calculated control increments satisfy the constraints.

[0056] .

[0057] In the formula, To obtain the optimal control command vector, the fuel injection quantity adjustment is used. With motor power adjustment composition, This represents the transpose of a vector.

[0058] In step three, after receiving the instruction, the fuel injection quantity and the motor output power are adjusted to work together to change the engine speed.

[0059] The execution system (including the fuel conditioning unit and the electric module control unit) receives the coordinated control command issued in step two. The fuel injection supply in the combustion chamber of the simulated engine and the output power of the three-phase permanent magnet synchronous motor are precisely adjusted respectively. The two power sources work together without conflict on the hybrid engine model, forcing it to change its current operating speed.

[0060] Step four uses the difference between the actual rotational speed and the target rotational speed as a feedback signal to achieve closed-loop correction.

[0061] After executing the system response, the feedback system collects the actual engine speed in real time and compares it with the target speed to obtain a deviation signal. This signal is sent back to the control decision system to dynamically update the compensation parameters in the sliding mode adaptive control law for the next cycle (10ms). Subsequently, the system returns to step one to start a new round of sampling to ensure that the control command continues to evolve.

[0062] Step one, from data perception to execution, forms the foundation of the entire control process, providing high-quality input for step two. Step two aims to generate commands based on prediction and modeling. Step three involves the implementation of actions, translating commands into changes in engine power. Step four verifies the effectiveness of the results from the first three steps and is crucial for providing parameter correction for the next control cycle. By feeding back deviations to step two through step four and triggering the system to return to step one to begin a new iteration, a self-optimizing dynamic closed loop is achieved, ensuring high-precision speed control under complex operating conditions.

[0063] Step one is executed by the parameter acquisition and preprocessing system in conjunction with the hybrid engine model. The engine model generates the initial physical state, and the parameter acquisition and preprocessing system is responsible for capturing this data and performing three-level filtering and quality assessment, ultimately outputting high-quality data.

[0064] Step two is executed independently by the control decision system. This system receives the data from step one, runs a BP neural network and a sliding mode adaptive algorithm, and calculates the optimal fuel and electric power coordination command.

[0065] Step three is accomplished by the execution system acting on the hybrid engine model. The execution system receives digital instructions, converts them into actual physical control variables (such as fuel injection quantity and motor power), and drives the engine model to change its current operating speed.

[0066] Step four is executed by the feedback system. The feedback system compares the actual speed after the model change with the target speed in real time, transmits the error data back to the control decision system for parameter correction, and instructs the system to return to execute step one to start the next complete cycle of iteration, so that the actual engine speed smoothly approaches the target speed.

[0067] The beneficial effects of the present invention.

[0068] This invention employs data-driven modeling (without requiring a mechanistic model) and updates pseudo-partial derivative parameters online through a parameter projection estimation algorithm. Compared to existing schemes that rely on mechanistic models, this improves the speed deviation control accuracy under fluctuating operating conditions.

[0069] The design incorporates sliding surface constraints with co-constraints and quadratic programming constraints to avoid conflicts in the operation of the fuel-electric dual-unit system, thereby reducing speed fluctuations during hybrid mode switching and load change conditions.

[0070] By integrating BP neural network feedforward prediction and sliding mode adaptive feedback compensation, combined with multi-dimensional parameter closed-loop feedback, the speed deviation is small and the control accuracy is high.

[0071] The control algorithm process is simplified (data-driven modeling + sliding mode adaptive control) to avoid complex mechanism calculations and balance real-time performance and robustness. The data preprocessing mechanism (filtering + anomaly detection) quantifies the comprehensive sensing accuracy. Compared with the existing single filtering scheme, the data anomaly rate is reduced, ensuring the reliability of control commands.

[0072] This invention employs data-driven modeling, eliminating the need for precise mechanistic models. It uses a parameter projection estimation algorithm to update pseudo-partial derivative parameters online in real time, enabling the system's equivalent dynamic model to self-correct in response to changes in the external environment and operating conditions. This overcomes the errors generated by traditional fixed-parameter mechanistic models when encountering unknown operating condition fluctuations. Compared to existing solutions, the speed deviation control accuracy is significantly improved.

[0073] This invention designs a sliding mode surface with cooperative constraints and a quadratic programming constraint optimization mechanism. It uses the fuel-electric coupling coefficient and physical boundary conditions such as battery SOC and motor rated power as hard constraint inputs to the optimization solver, fundamentally eliminating the possibility of mutual interference or overshoot in the allocation of dual power sources. This effectively avoids action conflicts between the fuel-electric dual actuators, significantly reducing speed fluctuations under complex operating conditions such as hybrid mode switching and sudden load changes.

[0074] This invention integrates BP neural network feedforward prediction and sliding mode adaptive feedback compensation, combined with closed-loop feedback of multi-dimensional state parameters. This architecture enables the control system to not only predict future speed trends but also quickly and adaptively suppress current nonlinear coupling disturbances, effectively reducing speed deviation and ensuring extremely high overall control accuracy.

[0075] The control system of this invention simplifies the control algorithm process through (data-driven modeling + sliding mode adaptive control), avoiding the complex and time-consuming calculations required by traditional aero-engines to solve massive nonlinear partial differential equations, thus greatly reducing the computing load and latency of airborne equipment. Therefore, while avoiding complex mechanism calculations, the system achieves excellent control real-time performance and robustness.

[0076] This invention constructs a three-level data preprocessing mechanism including arithmetic averaging, Kalman filtering, and isolated forest anomaly detection. It also sets a hard threshold (≥95%) for data acceptance by quantifying the overall sensing accuracy, effectively blocking high-frequency noise and sudden distortion signals from the sensors from flowing into the control decision system. Compared to existing single-filtering schemes, the data anomaly rate is significantly reduced, ensuring the absolute reliability of control command output from the data source. Attached Figure Description

[0077] Figure 1 System architecture diagram.

[0078] Figure 2 Overall process diagram.

[0079] Figure 3 A schematic diagram of a BP neural network structure.

[0080] Figure 4 Turboshaft engine structure.

[0081] Figure 5 Three-phase voltage vector.

[0082] Figure 6 Schematic diagram of energy flow direction in a series hybrid engine system.

[0083] Figure 7 This is an example of a three-level preprocessing and data quality assessment process.

[0084] Figure 8 This is pseudocode for a collaborative control algorithm based on deep learning and data-driven approaches.

[0085] Figure 9 Comparison of optimization results under load mutation conditions. Detailed Implementation

[0086] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0087] like Figure 1 , Figure 2 As shown, the intelligent speed control online adjustment system based on hybrid engine of the present invention includes a parameter acquisition and preprocessing system, a control decision system, an execution system, and a feedback system.

[0088] The parameter acquisition and preprocessing system is responsible for importing the full-dimensional simulation data source of the hybrid engine, covering parameters such as turboshaft engine gas turbine speed (rpm), power turbine speed (rpm), electric power distribution ratio (%), load demand power (kW), battery SOC (%), and exhaust temperature (K). The simulation data source is generated programmatically through the hybrid engine mechanism model and undergoes filtering and anomaly detection by the preprocessing module to ensure the reliability of the input information.

[0089] Based on the collected multi-dimensional data, the control decision system uses a deep learning prediction model and an adaptive control algorithm to first estimate the trend of speed change, and then combine the fuel-electric coupling characteristics to generate a coordinated control command and transmit it to the execution system.

[0090] The execution system adopts a dual execution unit design of "fuel regulation + electric regulation". On the one hand, the fuel injection control simulation unit calculates and outputs the corresponding simulated fuel supply according to the fuel regulation command to adapt to the fuel demand of the engine under different loads. On the other hand, the electric module control simulation unit dynamically calculates and adjusts the output power of the simulated motor according to the electric power regulation command. Through the collaborative simulation of the two units, the system achieves precise matching of the target speed.

[0091] The feedback system adopts a closed-loop feedback structure to monitor the deviation between the actual speed and the target speed in real time, dynamically correct the fuel supply and motor output power, and realize dynamic simulation adjustment of speed. This not only ensures the high accuracy of speed control simulation results, but also improves the stability under complex operating conditions such as hybrid mode switching and load changes, ultimately achieving the accuracy and stability of intelligent speed control for hybrid engines.

[0092] Parameter acquisition and preprocessing system:

[0093] After the parameter acquisition system collects the data, the data first enters the preprocessing module, where digital filtering algorithms, including arithmetic mean filtering and adaptive Kalman filtering, are used to remove high-frequency noise and sudden abnormal signals.

[0094] Arithmetic mean filtering is to find a This makes it consistent with the measured values ​​within a sampling period. Error between The sum of squares is minimized, where:

[0095]

[0096] Sum of squares for:

[0097]

[0098] right Find the extreme value:

[0099]

[0100] get:

[0101]

[0102] The higher the value, the better the smoothness of the signal.

[0103] The Kalman filter update process can be expressed by the following formula:

[0104]

[0105] This is the filtered state estimate at the current moment; This is the state estimate obtained from the previous time step; The Kalman gain matrix; The current observation (sensor measurement value); This is the observation matrix.

[0106] In addition, the overall system perception accuracy It can be represented as

[0107]

[0108] Through comprehensive sensing accuracy Quantify the overall error level (data reliability is considered to be ≥95% of the overall perception accuracy) to ensure the reliability of the preprocessed data.

[0109] The preprocessed data is first subjected to digital filtering (arithmetic mean filtering and adaptive Kalman filter) to achieve preliminary noise reduction. Then, outliers are analyzed and removed based on the isolated forest algorithm. Finally, the overall data quality is quantified by calculating the comprehensive sensing accuracy. When the comprehensive sensing accuracy is ≥95%, the data is considered reliable, ensuring that the data input to the control decision system has a high degree of confidence.

[0110] Control Decision System:

[0111] 1. Control algorithm implementation:

[0112] The control algorithm of this invention adopts a hybrid architecture of "deep learning predictive feedforward + sliding mode adaptive control + fuel-electric collaborative control". It fundamentally solves the control challenges of hybrid engines, including "nonlinear coupling, dynamic changes in operating conditions, and collaborative operation of dual actuators", achieving high-precision, fast-response, and highly stable speed control. The specific logic is "data-driven modeling + sliding mode surface design + adaptive reaching law + constraint optimization". The adaptive control method is an extension and supplement to sliding mode adaptive control, implemented as follows:

[0113] Specific implementation of sliding mode adaptive control

[0114] Sliding mode adaptive control serves as the feedback core. It adapts to the nonlinear characteristics of the hybrid engine through data-driven modeling, and compensates for fuel-electric coupling uncertainties in real time using adaptive laws. Finally, it outputs the optimal control command through constraint optimization. The specific implementation is as follows:

[0115] (1) Data-driven modeling

[0116] Without relying on a precise dynamic model of the hybrid engine, an equivalent dynamic model is established solely through system input and output data, adapting to unknown parameters and fluctuating operating conditions.

[0117] Define the input vector:

[0118]

[0119] in These correspond to the turbine speed (rpm), power turbine speed (rpm), electric power distribution ratio (%), load demand power (KW), battery SOC (%), and exhaust temperature (K) of the turboshaft engine, respectively.

[0120] Define the output vector:

[0121] ( (This refers to the actual speed of the hybrid engine).

[0122] Establish a dynamic data model (in discrete time):

[0123]

[0124] By describing the relationship between the current output, the change in engine speed, and the change in input, the dynamic characteristics of the system's input and output are directly described. This dynamic model uses a recursive relationship of discrete time steps to characterize the correlation between the output (engine speed) at the next moment and the current output, the change in engine speed, and the change in input, thus adapting to the dynamic characteristics of complex nonlinear systems.

[0125] In the formula: This represents the change in rotational speed; Input the change amount; , We define the pseudo-partial derivative parameters and design the following parameter projection estimation algorithm, which is used for online updating.

[0126] Parameter projection estimation algorithm:

[0127]

[0128]

[0129] in, , Yes and The estimated values ​​ensure that the model always fits the actual dynamics of the system, thus guaranteeing the model's accuracy. To achieve adaptive gain, the step size of parameter updates is controlled, balancing the stability and speed of parameter convergence. This is a regularization parameter to avoid parameter oscillations caused by data fluctuations. This ensures that the parameter estimation converges.

[0130] This method ensures online adaptive updates of model parameters because operating conditions and characteristics change dynamically during system operation, which traditional fixed-parameter models cannot adapt to. Therefore, this algorithm can achieve online adaptive updates through real-time data acquisition. , The data is continuously being revised.

[0131] (2) Sliding surface design

[0132] The design incorporates a sliding surface that includes speed deviation and dual-actuator coordination terms, enabling explicit quantification of deviation and coordination constraints.

[0133]

[0134] In the formula:

[0135] For real-time speed deviation ( (The target speed is )); the integral term is used to eliminate steady-state deviation. The integral coefficient is used to balance response speed and steady-state accuracy. This refers to the engine output power corresponding to the fuel consumption command. This represents the total power demand on the electrical side. For the collaborative constraint coefficient, The fuel-electric coupling coefficient is set at 0.2-0.9; dynamically adjusted according to the operating condition and speed deviation. For example, it is set at 0.7-0.9 to enhance the cooperative constraint during sudden load changes, and at 0.2-0.4 to reduce the control intensity during cruise conditions, to ensure that there is no conflict in the power distribution between fuel and electric motor.

[0136] The core function of the sliding surface is to enable the system state to converge rapidly to a fixed state. This means the system has reached an ideal control state. At the same time, it can quickly suppress disturbances and ensure speed stability under complex operating conditions such as hybrid mode switching and sudden load changes.

[0137] (3) Adaptive reaching law design

[0138] Design a discrete-time adaptive reaching law to ensure finite-time convergence of the sliding surface, while compensating for coupling uncertainties:

[0139]

[0140] This equation is the dynamic update rule for the sliding surface. Through the recursion of discrete time steps, it forces the sliding surface to converge to 0 in a finite time.

[0141] In the formula: The attenuation coefficient is... To control the period, the attenuation term This means lowering the current value of the sliding surface to accelerate the convergence speed; The disturbance suppression coefficient is... For saturation function (in conjunction with threshold ( (This is a threshold for smooth transition), replacing the sign function to suppress chattering, avoiding frequent fluctuations in fuel injection and motor power, resulting in a smoother output, for example when... When it is large, The output is 1 to ensure convergence speed. When approaching 0, Output a gradual value to avoid frequent fluctuations in fuel injection quantity and power; To address the coupling uncertainty in adaptive law estimation, adaptive compensation is achieved (e.g.) Get bigger It will also increase, and this term is used to offset the effects of interference, allowing the sliding surface to still converge stably. The update rule is:

[0142]

[0143] in For adaptive gain, the estimation error can be kept small.

[0144] By designing an adaptive reaching law, sliding mode control can be adapted to the complex operating conditions of aero-engines, while overcoming the shortcomings of traditional sliding mode control, ultimately achieving precise, stable, and rapid speed control.

[0145] (4) Control command solution and constraint optimization

[0146] Combining sliding mode control law and dual-actuator physical constraints, the optimal control command is solved using quadratic programming:

[0147] First, derive the control increment. :

[0148] Substituting the sliding surface convergence condition into the dynamic data model, we obtain:

[0149]

[0150] In the formula , This model predicts the target rotational speed at the next moment using deep learning; it can be used to calculate the control amount required to achieve the target rotational speed from the current speed. The extent of the adjustment required.

[0151] Secondly, physical constraints are added:

[0152]

[0153] This represents the change in fuel flow rate; The battery is in its state of charge. Motor power, This refers to the rated power of the motor. Physical constraints are used to prevent commands from exceeding the safe range of the hybrid engine.

[0154] Finally, the optimal control command is solved using quadratic programming:

[0155]

[0156] In the formula The nominal control instructions are used to ensure that the control instructions are optimal and do not exceed the limits of the execution unit.

[0157] 2. Deep learning prediction models ( Figure 3 ):

[0158] The engine speed prediction model is built on a backpropagation (BP) neural network. By learning the nonlinear mapping relationship between the input parameters and engine speed of the hybrid engine, it predicts the engine speed trend at several future moments, enhancing the controller's forward-looking adjustment capability and improving the response speed and accuracy of speed control. The specific implementation is as follows:

[0159] The BP neural network model consists of an input layer, hidden layers, and an output layer, with a purely linear transfer function between the hidden and output layers. The model input is the simulation data sequence of the current cycle and the previous 99 cycles. The turbine speed (rpm), power turbine speed (rpm), electric power distribution ratio (%), load demand power (KW), battery SOC (%), and exhaust temperature (K) of the turboshaft engine.

[0160] There are 6 channels in total, with an input dimension of 100×6.

[0161] (1) Data preparation:

[0162] Data sets for training and testing the neural network were prepared. Simulation data were generated programmatically using a turboshaft engine component-level model and a HEPS power switching dynamics model. Specifically, the gas turbine speed and power turbine speed were obtained through rotor dynamics constraint simulation of the engine component-level model; exhaust temperature was output through simulation using the model's built-in combustion efficiency-temperature correlation logic (no additional mechanistic formula derivation required); battery SOC, load demand power, and hybrid power distribution ratio were generated by simulation using the HEPS power switching model after pre-setting boundary conditions for different operating conditions. This invention collected 100 sets of simulation data covering five core operating conditions: takeoff, cruise, descent, sudden load changes, and hybrid mode switching.

[0163] The input parameters, namely the input layer of the BP neural network, include six key state parameters: turboshaft engine gas turbine speed (rpm), power turbine speed (rpm), oil-electric power distribution ratio (%), load demand power (KW), battery SOC (%), and exhaust temperature (K).

[0164] The output parameter, i.e. the output layer of the BP neural network, is the target speed of the hybrid engine.

[0165] One hundred sets of time series data (each set containing 100 periods and six parameters) are divided into 80 training sets and 20 prediction sets. The input matrix of the training set is (80×100×6) (80 sets of time series data, each set containing 100 periods, and each period containing 6 features), and the output matrix is ​​(80×1); the input matrix of the prediction set is (20×100×6), and the output matrix is ​​(20×1).

[0166] Data preprocessing: The reliability of input data is ensured through digital filtering (arithmetic mean + Kalman filtering) and isolated forest anomaly detection in the parameter acquisition and preprocessing system.

[0167] (2) Determine the network architecture:

[0168] The BP neural network constructed in this invention includes an input layer, a hidden layer, and an output layer.

[0169] There are 6 input layer parameters;

[0170] After comparing the network prediction results with different numbers of nodes in the hidden layer, it was finally determined that the best fit to the black box function was achieved when there were 10 hidden layer nodes. Therefore, 10 hidden layer nodes were selected.

[0171] The output layer outputs one parameter, which corresponds to the target speed of the hybrid engine.

[0172] (3) Initialize weights and thresholds:

[0173] In the programming software, use the rand function to randomly initialize the weight matrix V (6×10) between the input layer and the hidden layer, the weight matrix W (10×1) between the hidden layer and the output layer, as well as the hidden layer threshold b1 and the output layer threshold b2, with an initialization range of [-0.3, 0.3].

[0174] (4) Select activation function

[0175] In this invention, the activation function of the neural network is the Sigmoid function. The Sigmoid function maps the linear input of the hidden layer to a nonlinear output, enabling the model to learn complex nonlinear relationships between input parameters and rotational speed (such as the nonlinear impact of changes in the fuel-electric power distribution ratio on rotational speed). The Sigmoid function provides smooth gradients, preventing gradient vanishing during gradient descent and ensuring model convergence stability. Multiple neurons in the neural network can represent complex nonlinear functions through combinations of Sigmoid functions. The smooth gradients across most of the input range are crucial for the stability and convergence of the gradient descent optimization algorithm. Therefore, the Sigmoid function is chosen as the activation function of the neural network, mapping the output range to between 0 and 1. Its function form is:

[0176]

[0177] Output layer transfer function:

[0178]

[0179] The function is to directly output the linear combination result of the hidden layer, preserving the true numerical range of the rotational speed (rpm), and avoiding the rotational speed prediction deviation caused by nonlinear mapping.

[0180] (5) Forward propagation:

[0181] This involves passing input data into the network, calculating the output of each neuron through the neurons in each layer, and the propagation process of the BP neural network.

[0182] The input layer is denoted by L1, the output layer by L2, the weights between L1 and the hidden layer are denoted by V, and the matrix size is 6*10; the weights between L2 and the hidden layer are denoted by W, and the matrix size is 10*1.

[0183]

[0184] Where X is the input parameter vector (6×1) and H is the hidden layer output vector (10×1).

[0185]

[0186] Where Y is the output parameter vector (1×1).

[0187] (6) Back propagation:

[0188] Using only the mean squared error (MSE) of speed prediction as the loss function, the model training is entirely focused on reducing speed prediction bias. The function is in the form of:

[0189]

[0190] Where: N is the number of training samples. The predicted rotational speed is the value predicted by the model. The actual speed value is the actual rotational speed value, and the loss function value directly reflects the accuracy of the speed prediction.

[0191] 3. Adaptive control method:

[0192] The adaptive control scheme of this invention is based on "data-driven parameter identification + sliding mode adaptive reaching law + dual-execution unit cooperative constraints," and its core solution addresses the problems of nonlinear coupling, dynamic changes in operating conditions, and conflicts in the actions of the two execution units in hybrid engines, achieving dynamic parameter optimization and real-time correction of control commands. To address the problem of insufficient convergence under strong noise and rapid operating conditions, this scheme compensates through a three-level preprocessing mechanism, the introduction of a variable step-size forgetting factor, and a parameter limiting strategy. The control law adopts a "sliding surface dynamic adjustment + cooperative constraint gain adaptation" structure, optimizing the sliding mode reaching law parameters and the fuel-electric coupling coefficient in real time based on the parameter identification results, achieving rapid response to speed errors and stable suppression under complex operating conditions.

[0193] 4. Performance optimization mechanism:

[0194] The performance optimization mechanism of this invention is based on "data-driven parameter optimization + dual-unit collaborative constraints + multi-dimensional index dynamic balancing". It solves the problem of comprehensively improving the speed control accuracy and dual-unit collaborative efficiency of hybrid engines under multiple operating conditions. By deeply integrating with deep learning predictive feedforward, sliding mode adaptive control and closed-loop feedback system, it achieves continuous and stable high performance of the control system.

[0195] The optimization mechanism has four core objectives: "precise speed control, conflict-free dual-unit collaboration, optimal overall energy consumption, and adaptive operating conditions." By monitoring multi-dimensional data in real time, dynamically adjusting control parameters, and constraining physical boundaries, it ensures that the system continuously meets the standards under complex operating conditions such as takeoff, sudden load changes, and hybrid mode switching.

[0196] Execution system:

[0197] (1) Engine modeling Figure 4 )

[0198] In the modeling, the cross-sections of the engine components are numbered as follows: 1-engine intake, 2-compressor inlet, 3-compressor outlet, 4-combustion chamber outlet (gas turbine inlet), 4.5-power turbine inlet, 5-power turbine outlet, 8-exhaust nozzle outlet.

[0199] 1. Gas turbine rotor power balance equation:

[0200]

[0201] Indicates the output power of the gas turbine. Indicates the compressor drive power. This represents the moment of inertia of the gas turbine rotor. Indicates the rotational speed of the gas turbine rotor. This represents the angular acceleration of the gas turbine rotor. This represents the deviation term in the equation; this formula describes the dynamic balance of the gas turbine rotor (torque balance determines speed stability), and changes in speed directly reflect changes in power distribution.

[0202] 2. Power balance equation for the power turbine rotor:

[0203]

[0204] Indicates the output power of the power turbine. This indicates the power demand of the load (such as the propulsion power demand of an electric motor driving a propeller). This represents the moment of inertia of the power turbine rotor. Indicates the rotational speed of the power turbine rotor. This represents the angular acceleration of the power turbine rotor. This represents the deviation term in the equation; this formula supports the ratio of turbine speed to torque distribution, clearly demonstrating the nonlinear mapping relationship between speed and torque.

[0205] The two equations together define the dynamic boundary of the engine speed, and the derived real-time speed is obtained. and As a closed-loop feedback parameter, it provides input support for data-driven modeling, avoids speed overshoot or instability, and is also used to calculate speed deviation.

[0206] (2) Electric motor modeling:

[0207] The modeling object selected for this study is a three-phase permanent magnet AC synchronous motor. This motor has two advantages: first, it is energy-saving and highly efficient, with minimal efficiency loss under light loads; second, it has a high torque density, which enables small size and lightweight design, a crucial factor for hybrid power systems.

[0208] In a three-phase permanent magnet AC synchronous motor, voltage can be decomposed into three-phase voltage vectors. Figure 5 )

[0209] 1. The electromagnetic torque equation of an electric motor:

[0210]

[0211] It is a three-phase current. The three-phase winding flux linkage is quantified to represent the relationship between flux linkage and electromagnetic torque, indicating the power output of the motor, so that the torque output on the motor side conforms to the laws of electromagnetism.

[0212] 2. Equation of motion of the electric motor:

[0213]

[0214]

[0215]

[0216]

[0217] In the formula: Let be the mechanical angular velocity of the electric motor. This represents the load torque of the motor. This is the damping coefficient of the motor. Let be the moment of inertia of the motor. Electromagnetic angular velocity, Let represent the motor speed. This equation describes the dynamic balance of the motor rotor (electromagnetic torque is converted into speed changes), clarifying the relationship between motor speed and torque. It supports the coordinated control of motor speed and torque, ensuring that the torque distribution ratio output by the neural network can be converted into actual speed.

[0218] 3. Equations for the input and output power of the electric motor:

[0219]

[0220]

[0221] Quantify the energy conversion efficiency of the motor (converting input electrical power into output mechanical power), including loss calculation.

[0222] The three equations construct a physical mapping of power input into speed output, limiting the motor power to not exceed the rated power and the nonlinear relationship between speed and torque. Among them, the real-time motor speed derived from the motion equations and the energy efficiency calculated from the input / output power equations serve as the basis for collaborative control feedback, avoiding conflicts between the actions of the two execution units.

[0223] (3) Hybrid operating system of electric motor and generator ( Figure 6 )

[0224] 1. Generator speed-torque matching equation:

[0225]

[0226]

[0227] In the formula: The speed at the output of the reducer. The speed at the input end of the reducer. The output torque of the reducer. The input torque of the reducer. The reduction ratio of the reducer. This equation represents the mechanical efficiency of the reducer. The input is connected to the engine's power shaft, and the output is connected to the generator's working shaft. This equation quantifies the conversion relationship between engine speed, generator speed, and motor speed, resolving the speed mismatch problem between the engine and generator. Simultaneously, this equation clarifies the logic of power transmission from the engine to the motor.

[0228] All-electric operating mode

[0229]

[0230] Charging mode

[0231]

[0232] 2. Basic voltage equations for a three-phase motor in synchronous rotational coordinates along the dq axis:

[0233]

[0234] This equation transforms the motor model from a natural coordinate system to a synchronous rotating coordinate system, simplifying control calculations. At the same time, it improves the real-time performance of motor control, adapting to the fast response requirements of neural networks.

[0235] 3. SOC Calculation

[0236]

[0237] Upper limit of charging: SOC ≤ 0.95; Lower limit of discharging: SOC ≥ 0.2.

[0238] This equation quantifies the remaining battery capacity based on the coulomb counting method and clarifies the charge and discharge boundaries.

[0239] The engine speed and generator speed are correlated by the speed-torque matching equation to ensure that the power distribution ratio of oil and electricity meets the mechanical constraints; the dq axis voltage equation simplifies the motor control calculation, improves the response speed of control commands, and adapts to the real-time requirements of sliding mode adaptive control; the charge and discharge constraints are directly applied to the motor power command through the SOC boundary to avoid conflict between the actions of the two execution units.

[0240] Closed-loop feedback system:

[0241] Closed-loop feedback system workflow:

[0242] 1. Feedback Parameter Acquisition: Real-time acquisition of the actual engine speed output by the execution system. Actual motor speed The ratio of electric to oil power (%), combined with the load demand power. Battery SOC (k), exhaust temperature ;

[0243] 2. Deviation Calculation: Calculate the actual rotational speed. With target speed Bias of (BP neural network predictions) ;

[0244] 3. Feedback and correction: Addressing deviations Deviation change rate Input control decision system, adjust sliding surface Adaptive law parameters;

[0245] 4. Update Instructions: Based on the corrected parameters, the control decision system re-solves for the optimal control instructions. This information is transmitted to the execution system to complete the closed-loop adjustment.

[0246] 5. Feedback period: Consistent with the control period (10ms) to ensure real-time dynamic correction.

[0247] Figure 7 The specific example illustrates the three-stage preprocessing and data quality assessment process, which is explained in detail below:

[0248] like Figure 7 The diagram illustrates an example of the three-level preprocessing and data quality assessment process provided in this embodiment of the invention. In this example, the system continuously captures simulation data generated by the hybrid engine model at a 10-millisecond cycle. At sampling times of 4.98s and 4.99s, the engine operates stably, and the calculated comprehensive perception accuracies are 98.2% and 98.5%, respectively, both greater than or equal to the preset 95% threshold. The system performs normal data acquisition and directly outputs high-quality speed data. At sampling time of 5.00s, due to high-frequency noise from the sensor or occasional external interference, the original engine speed signal experiences a high-frequency spike with an instantaneous amplitude of 12650.0 rpm, causing the calculated comprehensive perception accuracy to plummet to 81.4% (below 95%). At this point, the preprocessing module immediately triggers an interception mechanism, rejecting the abnormal signal from flowing into the control decision system and forcibly maintaining the output value of the current cycle at the valid value of the previous cycle (i.e., 12001.8 rpm). At sampling time 5.01s, the engine encountered a sudden load, and the original speed began to physically drop (to 11980.5 rpm). Since this change belongs to the continuous evolution of dynamic conditions rather than an outlier, after the isolation forest algorithm, its overall perception accuracy is still as high as 96.1% (≥95%). The system judges the data to have high confidence and accepts it normally. This example fully demonstrates that the three-level preprocessing mechanism of this invention can accurately intercept high-frequency distortion noise and accurately identify and retain real transient changes in operating conditions, ensuring the absolute reliability of control command output from the source.

[0249] Figure 8 The pseudocode flow of the online coordinated speed control algorithm for a series hybrid engine based on deep learning and data-driven principles is presented to demonstrate the feasibility of the scheme. The algorithm operates in a closed loop with a control cycle of 10 milliseconds.

[0250] like Figure 9 As shown, to verify the effectiveness of the present invention, simulation results were extracted and analyzed, as illustrated in the table below. The results are presented in... Furthermore, under a 20% load surge operating point, the performance optimization of online collaborative speed control based on deep learning and data-driven methods was verified. As shown in the table, compared with the traditional control scheme, the maximum speed drop of the present invention is reduced by approximately 82%, and the recovery time is shortened by approximately 82%. Throughout the adjustment process, both battery SOC and motor power are strictly within safety constraints, avoiding conflicts between the fuel and electric units, and the three-level preprocessing mechanism ensures an extremely low data anomaly rate. This demonstrates the effectiveness of the algorithm of the present invention.

Claims

1. A series hybrid engine speed online coordinated control system based on deep learning and data-driven principles, characterized in that, This includes a parameter acquisition and preprocessing system, a control decision system, an execution system, a feedback system, and a hybrid engine model; The hybrid engine model serves as the virtual physical environment and data source for the system, dynamically generating multi-dimensional simulation operation data including engine speed, battery SOC, and load power. The parameter acquisition and preprocessing system captures multi-dimensional simulation operation data of the hybrid engine in real time, cleans the data through internal filtering and anomaly detection mechanisms, judges the data quality based on the calculated comprehensive perception accuracy, intercepts abnormal signals, and obtains high-quality data vectors. The parameter acquisition and preprocessing system is responsible for transmitting data to the control and decision-making system, which is responsible for decision support and issuing commands; the control and decision-making system is responsible for transmitting commands to the execution system, which is responsible for the physical action conversion. The feedback system monitors and collects the actual engine speed after the change in real time, compares it with the target speed, calculates the real-time speed deviation signal, and sends the speed deviation signal back to the control decision system. The control decision system calculates the correction compensation amount based on this and instructs the system to return to the parameter acquisition stage to start the next sampling cycle, so as to realize closed-loop adaptive compensation for interference and error.

2. The online coordinated control system for the speed of a series hybrid engine based on deep learning and data-driven principles as described in claim 1, characterized in that, The high-quality data vectors obtained from the parameter acquisition and preprocessing system are input to the control decision system. Based on the high-quality data vectors, the control decision system uses a deep learning model to predict the speed change trend and combines it with a sliding mode adaptive control law to perform online calculations and generate the optimal fuel and electric power coordinated control command. The control decision system sends the calculated coordinated control commands to the execution system. After receiving the commands, the execution system drives its internal fuel regulation and electric regulation dual execution units to convert the digital commands into corresponding simulated fuel supply and motor output power. The execution system is responsible for controlling the applied oil and electricity to drive the engine's mathematical model to rotate and calculate a completely new current speed; The actual fuel quantity and motor power output by the execution system directly affect the hybrid engine model; after receiving the synergistic effect of the two power sources, the model's internal dynamic state changes, thereby outputting the updated actual speed. Every 10 milliseconds, the difference between the actual speed and the target speed is checked, and then the fuel injection quantity and motor power are immediately fine-tuned to make it as close as possible to the target speed.

3. A control method for an online coordinated control system of a series hybrid engine speed based on deep learning and data-driven principles, as described in any one of claims 1-2, characterized in that... Includes the following steps; Step 1: Collect multi-dimensional operational data, and after three levels of preprocessing, determine the data quality based on the comprehensive sensing accuracy and output a high-quality data vector; Step 2: The BP neural network predicts the future speed change trend, uses data vector-driven modeling to judge the current dynamic model, and combines the speed deviation to generate coordination commands; Step 3: After receiving the coordination command, adjust the fuel injection quantity and the electric motor output power to work together to change the actual engine speed; Step 4: Calculate the difference between the actual speed and the target speed, use it as a feedback signal, and send it back to the control decision system to correct the control command for the next cycle. The system then enters the next control cycle, returning to step one to re-collect real-time data, thus achieving closed-loop control.

4. The online coordinated speed control method for a series hybrid engine based on deep learning and data-driven principles as described in claim 3, characterized in that... Step one is executed by the parameter acquisition and preprocessing system in conjunction with the hybrid engine model. The hybrid engine model generates the initial physical state, and the parameter acquisition and preprocessing system is responsible for capturing these data and performing three-level filtering purification and quality judgment, and finally outputting high-quality data. Step two is executed independently by the control decision system; the system receives the data from step one, runs a BP neural network and a sliding mode adaptive algorithm, and calculates the optimal fuel and electric coordination command. Step three is completed by the execution system acting on the hybrid engine model; the execution system receives digital instructions, converts them into actual physical control variables, and drives the engine model to change its current operating speed; Step four is executed by the feedback system. The feedback system compares the actual speed after the model change with the target speed in real time, transmits the error data back to the control decision system for parameter correction, and instructs the system to return to execute step one to start the next complete cycle of iteration, so that the actual speed of the engine smoothly approaches the target speed.

5. The online coordinated speed control method for a series hybrid engine based on deep learning and data-driven principles according to claim 4, characterized in that, Step one involves collecting multi-dimensional operational data, and after three levels of preprocessing, outputting high-quality data vectors based on the data quality assessment results. The preprocessing process involves three levels of preprocessing: arithmetic mean filtering, adaptive Kalman filtering, and isolated forest anomaly detection. The data quality is quantified by calculating the comprehensive sensing accuracy. When the comprehensive sensing accuracy is ≥95%, the data is deemed valid and a high-quality data vector is output. Otherwise, abnormal signals are intercepted to ensure the stability and reliability of the data input to the control and decision-making system. Cruise operating points at specific altitudes and Mach numbers are selected, and the collected real-time parameters are substituted into the control system for filtering and anomaly detection. The fuel injection quantity and the increase in motor output power are selected as control variables, and data are continuously collected during engine operation.

6. The online coordinated speed control method for a series hybrid engine based on deep learning and data-driven principles according to claim 5, characterized in that, The online coordinated control of the speed of a series hybrid engine is essentially a control process aimed at quickly eliminating speed deviations while ensuring the safety of the battery and motor; the independent variable is the change in fuel flow rate. and motor power , The rated power constraints for the motor are as follows; Constraints: ； This represents the battery's current state of charge. , These are the minimum and maximum allowable states of charge for the battery, respectively. This is the fuel injection quantity adjustment value within a single control cycle. , These are its lower and upper limits for adjustment, respectively; This represents the current output power of the motor. This refers to the rated output power of the motor.

7. The online coordinated speed control method for a series hybrid engine based on deep learning and data-driven principles as described in claim 5, characterized in that, In step two, the BP neural network predicts the future speed change trend and judges the current dynamic model through data-driven modeling. In each control cycle, the control decision system predicts the target speed at the next moment based on the high-quality data sequence of the previous 100 cycles, and updates the pseudo-partial derivative parameters online. ; Subsequently, a sliding surface with cooperative constraints is established, and cooperative instructions are generated by combining the prediction results; the optimal conflict-free instructions in the current iteration process are solved and output only if the calculated control increments satisfy the constraints. ； In the formula, To obtain the optimal control command vector, the fuel injection quantity adjustment is used. With motor power adjustment composition, This represents the transpose of a vector.

8. The online coordinated speed control method for a series hybrid engine based on deep learning and data-driven principles according to claim 5, characterized in that, In step three, after receiving the instruction, the fuel injection quantity and the motor output power are adjusted to work together to change the engine speed. The execution system receives the collaborative control command issued in step two. The fuel injection supply in the combustion chamber of the simulated engine and the output power of the three-phase permanent magnet synchronous motor are precisely adjusted respectively. The two power sources work together without conflict on the hybrid engine model, forcing it to change its current operating speed.

9. The online coordinated speed control method for a series hybrid engine based on deep learning and data-driven principles as described in claim 5, characterized in that... Step four uses the difference between the actual rotational speed and the target rotational speed as a feedback signal to achieve closed-loop correction. After executing the system response, the feedback system collects the actual engine speed in real time and compares it with the target speed to obtain the deviation signal. This signal is sent back to the control decision system to dynamically update the compensation parameters in the sliding mode adaptive control law for the next cycle (10ms). Subsequently, the system returns to step one to start a new round of sampling to ensure that the control command continues to evolve.

Images