Intelligent speed regulation control system of diesel engine

Abstract

The invention belongs to the field of mechanical engineering, particularly relates to an intelligent speed regulation control system of a diesel engine, and aims to solve the problems of rotating speed response lag, low control precision and high energy consumption of an existing system under a dynamic load. The system comprises a real-time state sensing module, a composite error generation module, a multi-mode control decision module, an execution driving module and a self-learning optimization module. Smooth switching is realized based on mode arbitration by constructing a three-dimensional composite error feature matrix and fusing three strategies of feed-forward compensation, adaptive fuzzy sliding mode control and online reinforcement learning; and high-precision oil injection execution and offline knowledge extraction are combined, so that the dynamic response, the steady-state precision and the energy efficiency performance are improved.

Description

Technical Field

[0001] This invention belongs to the field of mechanical engineering, specifically relating to an intelligent speed control system for diesel engines. Background Technology

[0002] Diesel engines, as an important power unit, are widely used in various industrial fields such as ship propulsion, generator sets, construction machinery, and transportation. Their working principle is based on compression ignition, and they possess significant advantages such as high thermal efficiency, stable torque output, and strong fuel adaptability. With the development of automation and intelligent technologies, precise control of diesel engine operation, especially the dynamic adjustment of its speed, has become a key aspect of improving the overall system performance. Intelligent speed regulation, as the core means to achieve efficient, stable, and energy-saving operation of diesel engines, directly affects the equipment's response speed, load adaptability, and emission characteristics.

[0003] The diesel engine speed control system aims to adjust the fuel supply in real time according to changes in external load and set operating conditions to maintain or change the engine speed to the target value. Traditional speed control systems mostly use mechanical or analog electronic governors, relying on preset parameters and feedback control laws to achieve basic speed stabilization. Modern intelligent speed control systems tend to introduce digital control technology, using microprocessors to collect multi-source operating parameters and combine them with control algorithms to make decision outputs, thereby improving the system's adaptability and dynamic quality.

[0004] Existing technologies still have significant shortcomings in dealing with dynamic response, multivariate coupling disturbances, and nonlinear characteristics under complex operating conditions: control strategies generally rely on fixed-gain PID architectures, making it difficult to adjust parameters online to adapt to the entire operating range; compensation for system delays and hysteresis caused by factors such as sudden load changes and ambient temperature variations is insufficient; the lack of deep perception and trend prediction capabilities of diesel engine operating status leads to overshoot, oscillation, or response lag during the regulation process; at the same time, system robustness is limited by model accuracy, and fault diagnosis and fault-tolerant control mechanisms are not effectively integrated, making control failures prone to occur under abnormal conditions such as sensor drift or actuator aging. These problems are particularly prominent in high-dynamic demand scenarios, seriously affecting the operational stability, energy efficiency, and service life of diesel engines. Therefore, there is an urgent need for an intelligent speed control system with self-learning, self-adaptation, and strong robustness to solve the aforementioned technical challenges. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent speed control system for diesel engines, addressing the technical problems of existing diesel engines, such as sluggish speed response, insufficient steady-state accuracy, high energy consumption, and poor adaptability to various operating conditions under dynamic load changes. Currently, diesel engines are widely used in engineering machinery, ship propulsion, and emergency power generation, operating under complex and variable conditions, often facing the challenge of sudden load increases or decreases. Traditional speed control systems mostly employ mechanical centrifugal governors or electronic governors based on PID algorithms. The former has slow response speed and low adjustment accuracy, while the latter is prone to overshoot, oscillation, or excessively long adjustment times under nonlinear and time-varying load disturbances. Furthermore, existing systems lack deep perception and trend prediction capabilities of the operating state, failing to achieve proactive adjustment, making it difficult to further optimize fuel economy and emission performance. Therefore, there is an urgent need to construct a novel intelligent speed control system with strong robustness, high responsiveness, adaptive adjustment capabilities, and energy efficiency synergistic optimization functions.

[0006] The technical solution of this invention is an intelligent speed control system for a diesel engine, comprising a real-time state perception module, a composite error generation module, a multimodal control decision module, an execution drive module, and a self-learning optimization module. The real-time state perception module continuously collects physical parameters such as instantaneous crankshaft speed, injection pressure, intake air temperature and pressure, exhaust back pressure, lubricating oil temperature, coolant temperature, and external load torque through a multi-source sensor network integrated into the diesel engine body and load end. It then performs anti-interference filtering and dynamic compensation processing on the raw signals and outputs a corrected real-time operating state vector. This state vector serves as the core input of the system and is synchronously fed to subsequent functional units.

[0007] The composite error generation module receives the actual speed signal output by the real-time state perception module and the target speed command set by the user, and calculates the basic speed deviation value between the two. Furthermore, it introduces a differential term for the speed change rate to construct a three-dimensional composite error feature matrix containing the deviation amount, the cumulative deviation integral, and the differential component of the change trend. This matrix not only reflects the current speed deviation degree but also characterizes the acceleration characteristics of speed fluctuations, providing richer dynamic behavior information for subsequent control decisions. The composite error feature matrix is ​​updated in real time and transmitted to the multimodal control decision module.

[0008] The multimodal control decision module is the core control hub of this system. It embeds three parallel and dynamically switchable control strategy models: the first is a feedforward compensation model based on precise mathematical modeling. This model pre-calculates the theoretical fuel supply increment under expected load disturbances based on the diesel engine static characteristic curve library and the current operating point, realizing early intervention for known disturbances. The second is an adaptive fuzzy sliding mode controller. It takes the composite error feature matrix as input and effectively suppresses the high-frequency chattering phenomenon inherent in traditional sliding mode control while ensuring the system's strong robustness through the design of nonlinear switching functions and boundary layer correction mechanisms. The third is an online reinforcement learning agent model. This model constructs a closed-loop learning framework with a state-action-reward mechanism. It maps the current operating state vector and composite error value into fuel supply adjustment commands in the control action space and continuously updates its strategy network weights based on the reward signals fed back from the actual adjustment effect, gradually approaching the optimal control strategy.

[0009] The multimodal control decision module is further configured with a mode arbitration logic unit, used to evaluate the consistency and prioritize the candidate adjustment commands output by the three control strategies. When the system detects that the load mutation gradient exceeds a preset threshold, the feedforward compensation model is activated first for rapid energy injection; when the system is in a transition process and the composite error is in a medium amplitude range, the adaptive fuzzy sliding mode controller is enabled to ensure convergence stability; when the system enters the steady-state operation stage and the error remains within a small range, the online reinforcement learning proxy model is handed over to perform fine-tuning and long-term performance optimization. A smooth transition function is used to achieve seamless switching between the three modes, avoiding secondary disturbances caused by mode jumps.

[0010] The actuator module receives the comprehensive fuel supply adjustment command ultimately determined by the multimodal control decision module, converts it into a high-precision pulse width modulation signal, and drives the electromagnetic injector actuator of the high-pressure common rail system. This precisely controls the start time, duration, and number of injections for each injection, thereby achieving millisecond-level dynamic regulation of the fuel quality entering the cylinder. Simultaneously, the actuator module possesses a hardware-level safety protection mechanism, monitoring the actuator's operating current and response delay in real time. Upon detecting any abnormality, it immediately activates redundant control paths and issues a fault alarm.

[0011] The self-learning optimization module operates independently of the main control loop. Its function is to perform offline mining and knowledge extraction on long-term system operating data. This module periodically extracts typical operating condition sequences from the historical database, including scenarios such as frequent start-stop, periodic load fluctuations, and operation in high and low temperature environments. It uses a deep neural network to construct a nonlinear dynamic inverse model of the diesel engine and trains it through a backpropagation algorithm to obtain a set of initial control parameters applicable to specific operating condition clusters. After verification, these parameter sets are written into the initialization configuration library of the multimodal control decision module, enabling the system to quickly load the optimal initial strategy when restarting or entering similar operating conditions, significantly reducing the time overhead of the adaptive learning process.

[0012] Preferably, the multi-source sensor network in the real-time status perception module adopts a dual-redundancy arrangement structure. Key sensors such as speed sensors and torque sensors are configured with two sets, one as the primary and one as the backup. The system uses a cross-validation algorithm to determine the health status of the sensors, automatically shields the failed channels, and switches to the backup signal source to ensure the reliability and continuity of the input data.

[0013] Preferably, the integral term in the composite error feature matrix is ​​processed by an anti-saturation algorithm with amplitude limiting to prevent excessive accumulation of the integral term under large deviation conditions, which would lead to system response lag; the differential term is combined with a low-pass filter to eliminate high-frequency noise interference and improve the accuracy of trend recognition.

[0014] Preferably, the switching function exponential approach law of the adaptive fuzzy sliding mode controller introduces a time-varying gain factor, which is dynamically adjusted according to the magnitude of the composite error; when the error is large, a high gain is used to accelerate the convergence speed, and when the error approaches zero, the gain is automatically reduced to reduce steady-state chattering.

[0015] Preferably, the online reinforcement learning agent model adopts a deep deterministic policy gradient algorithm architecture, and its reward function consists of three weighted parts: the negative value of the square of the speed deviation, the negative value of the fuel consumption rate increment, and the negative penalty term of the actuator action frequency; by adjusting the weight coefficients, a multi-objective balance can be achieved between adjustment accuracy, energy saving and mechanical wear.

[0016] Preferably, the mode arbitration logic unit is equipped with a hysteresis criterion to avoid frequent switching of control modes near the critical operating point; the activation conditions of each mode are divided into two-dimensional plane regions composed of composite error amplitude and rate of change, and the region boundaries can be parameterized based on operating experience.

[0017] Preferably, the execution drive module is equipped with a dedicated power drive chip, which has short-circuit protection, over-temperature derating and open-circuit detection functions; the fuel injection pulse signal generation circuit adopts a high-resolution timer unit with a minimum time resolution of 0.1 microseconds to meet the requirements of ultra-fine fuel quantity adjustment.

[0018] Preferably, the deep neural network used in the self-learning optimization module includes an input layer, two hidden layers and an output layer, and the activation function is Leaky ReLU to alleviate the gradient vanishing problem; during training, mini-batch stochastic gradient descent is used, and batch normalization technology is introduced to improve the model's convergence speed and generalization ability.

[0019] Compared with the prior art, the advantages and positive effects of the present invention are as follows: This invention fundamentally improves the dynamic response quality and steady-state control accuracy of diesel engines under complex operating conditions by constructing a closed-loop intelligent speed regulation system composed of real-time state perception, composite error modeling, multimodal collaborative decision-making, high-precision execution, and self-learning evolution. The introduction of the composite error feature matrix enables the system not only to "see" speed deviations but also to "predict" their evolution trends, providing precise timing and intensity for feedforward and sliding mode control. The organic integration of multimodal control strategies realizes the complementary advantages of different control paradigms, exhibiting excellent anti-disturbance capabilities under sudden load changes and achieving a sub-one percent speed fluctuation rate control target in steady-state operation. The online reinforcement learning mechanism endows the system with the ability to continuously self-optimize, enabling it to adapt to long-term uncertainties caused by equipment aging, environmental changes, and differences in usage habits. The self-learning optimization module significantly shortens the debugging cycle and learning cost of newly deployed systems by solidifying knowledge from historical data, improving engineering practicality. The overall solution achieves a performance leap without changing the basic structure of the diesel engine through software-defined intelligent control methods, with significant fuel-saving benefits and emission improvement potential, and is suitable for various application scenarios with stringent requirements for power stability and operating efficiency. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention; Figure 2 This is a schematic diagram of the core principle framework of the multimodal control decision module in this invention. Detailed Implementation

[0021] Example 1 Please refer to Figure 1 and Figure 2This invention relates to an intelligent speed control system for diesel engines, designed to achieve high-precision, robust, and low-energy-consumption dynamic adjustment of diesel engine speed by integrating multi-source state perception, composite error modeling, multi-modal collaborative decision-making, and self-learning optimization mechanisms. The system is deployed on a control hardware platform comprised of a high-performance embedded processor and a real-time operating system, possessing microsecond-level interrupt response capability and deterministic task scheduling characteristics to ensure that control commands are calculated and issued within strict time limits. The overall technical process of the system unfolds in a continuous closed-loop operation mode: First, the real-time state perception module collects and corrects the multi-dimensional physical parameters of the diesel engine and the load end to form a high-fidelity operating state vector; this vector is synchronously input to the composite error generation module and the multi-modal control decision module to construct a three-dimensional composite error feature matrix containing deviation trend information; the multi-modal control decision module dynamically activates the most suitable control strategy model based on this matrix and the current operating condition, outputs candidate adjustment commands, and generates the final comprehensive fuel supply adjustment command after priority determination and smooth switching by the mode arbitration logic unit; the command is transmitted to the execution drive module and converted into a precise pulse width modulation signal to drive the high-pressure common rail injector; at the same time, the self-learning optimization module periodically analyzes historical operating data in the background, extracts the optimal control knowledge and updates the initialization configuration library, providing the system with long-term performance evolution capabilities.

[0022] The real-time status perception module, as the sensing front end of the entire system, plays a crucial role in acquiring the actual operating status of the diesel engine. The module integrates a multi-source sensor network covering the core components of the diesel engine and external load interfaces. The instantaneous crankshaft speed is acquired by an electromagnetic induction speed sensor mounted on the flywheel housing. The sensor employs the variable reluctance principle, generating 60 equidistant pulse signals per revolution, corresponding to a 6-degree angular resolution per revolution of the crankshaft. The raw pulse signals are pre-amplified and processed by a Schmitt trigger shaping circuit before being sent to the processor's high-speed counter channel. The time interval between adjacent pulses is calculated using a timestamp capture function, thereby determining the instantaneous angular velocity. The sampling period is 0.5 milliseconds. Fuel injection pressure is monitored by a piezoelectric pressure sensor installed on the high-pressure common rail, with a measurement range of 0 to 250 MPa. The output signal is differentially amplified and converted from analog to digital before entering the digital domain. The sampling frequency is set to 10 kHz to capture high-frequency components in fuel pressure fluctuations. Intake air temperature and pressure are measured jointly by a thermistor temperature sensor and a piezoresistive pressure sensor, integrated at the end of the intake manifold, reflecting changes in charging efficiency in real time. Exhaust back pressure is detected by an absolute pressure sensor located downstream of the turbocharger, used to assess the impact of exhaust flow resistance on boost matching. Lubricating oil temperature and coolant temperature are measured by a platinum resistance temperature sensor, whose resistance changes linearly with temperature. The signal is converted into a voltage signal by a constant current source and a precision resistor bridge circuit, and then digitized. External load torque is acquired by a non-contact magnetoelastic torque sensor mounted on the power take-off shaft. This sensor is based on the physical effect of the anisotropic change in magnetic permeability of ferromagnetic materials under torque, outputting two sinusoidal signals with a phase difference proportional to the torque. The torque amplitude is extracted by a decoding circuit, with a communication period of 2 milliseconds.

[0023] All the aforementioned sensors are connected to the data acquisition engine of the real-time status sensing module. The engine runs in a dedicated task thread within the real-time operating system kernel, with the highest priority to ensure no data loss. To improve signal quality, the module performs multi-stage anti-interference filtering and dynamic compensation processing on the raw sampled values. For speed signals, a moving window least squares fitting algorithm is used to eliminate periodic disturbances caused by gear ring machining errors. For slowly varying signals such as pressure and temperature, a second-order Butterworth low-pass filter is applied to suppress power supply coupling noise and environmental electromagnetic interference. The cutoff frequency is preset based on the physical process response characteristics; for example, the injection pressure channel is set to 2 kHz, while the coolant temperature channel is set to 10 Hz. The dynamic compensation stage corrects for the sensor's inherent non-ideal characteristics, including zero-point drift compensation, temperature drift compensation, and nonlinear correction. Zero-point drift is automatically calibrated by periodically acquiring baseline values ​​while the engine is stopped. Temperature drift compensation uses a lookup table function based on the temperature-sensitivity curve provided by the manufacturer to adjust the gain coefficient in real time. Nonlinear correction uses piecewise linear interpolation to map the raw readings to standard engineering units. Key sensors, such as speed and torque sensors, employ a dual-redundancy arrangement, with primary and backup channels independently wired and connected to different pins of the processor. The system incorporates a cross-validation algorithm to continuously compare the consistency of data from the two channels. When the relative deviation exceeds 5% and lasts for more than 10 milliseconds, a primary channel fault is identified, the abnormal signal source is automatically masked, and the system switches to the backup channel output, while simultaneously recording the fault code in the diagnostic storage area. After the complete processing flow, the real-time status perception module outputs a structured real-time operating status vector containing 12 fields: instantaneous speed, average speed, injection pressure, intake pressure, intake temperature, exhaust back pressure, lubricating oil temperature, coolant temperature, load torque, atmospheric pressure, ambient temperature, and the cumulative fuel injection amount at the last injection moment. This vector is published in a fixed-frame-length format via a shared memory buffer, with a refresh cycle of 1 millisecond, serving as the foundational input for all subsequent control decisions.

[0024] The composite error generation module receives the actual speed signal output by the real-time status sensing module and the target speed command set by the user through the human-machine interface, and executes the composite error construction process. The target speed command is input in revolutions per minute (RPM) and can be manually set or sent by the upper-level automation system via the communication bus. The module first calculates the basic speed deviation value, i.e., the difference between the target speed and the actual instantaneous speed, which is the basic control quantity. To further enhance the system's ability to perceive dynamic behavior, the module introduces a differential term for the speed change rate and combines it with the cumulative deviation integral to construct a three-dimensional composite error feature matrix. The three dimensions of this matrix are: the first dimension is the speed deviation at the current moment, the second dimension is the deviation integral value from the initial moment to the current moment, and the third dimension is the speed change rate at the current moment. Among them, the integral is processed using an anti-saturation algorithm with amplitude limiting. The upper and lower limits of the integral are set to ±10% of the target speed, respectively, to prevent excessive accumulation of the integral term under large deviation conditions, which would lead to system response lag or reverse overshoot. The differential term is obtained by performing center difference calculation on five consecutive historical speed sampling points, as shown in the following formula: in, This represents the rate of change of rotational speed at the current time t. and They are respectively and The instantaneous rotational speed at a given moment. The sampling period is 0.5 milliseconds. To suppress the amplification effect of differential operations on high-frequency noise, a fourth-order Chebyshev low-pass filter is applied to the original speed sequence before calculation, with a cutoff frequency set to 200 Hz to ensure the accuracy of trend identification. The composite error feature matrix is ​​organized in the form of a floating-point array, with each element normalized to the interval [-1, 1] to eliminate the influence of dimensional differences on subsequent control algorithms. The normalization reference values ​​are as follows: speed deviation is based on 20% of the rated speed, integral deviation is based on 5% of the rated speed, and speed change rate is based on 1000 revolutions per second change of the rated speed. This matrix is ​​updated in real time and transmitted to the multimodal control decision module as the core criterion for selecting and invoking different control strategies.

[0025] The multimodal control decision module is the core control hub of this system, responsible for dynamically generating optimal fuel supply adjustment commands based on the current operating state and composite error characteristics. The module embeds three parallel and dynamically switchable control strategy models: a feedforward compensation model based on precise mathematical modeling, an adaptive fuzzy sliding mode controller, and an online reinforcement learning proxy model. These three models share the same input data source, independently calculate candidate adjustment commands, and are ultimately integrated and decided upon by the mode arbitration logic unit.

[0026] The feedforward compensation model is a predictive control method based on prior knowledge of the static characteristics of a diesel engine. Internally, the model maintains a library of diesel engine static characteristic curves, storing the theoretical fuel injection baseline values ​​required to maintain stable combustion under different speed and load combinations. This curve library, calibrated through bench tests, is organized in a two-dimensional lookup table, with speed on the x-axis and load torque on the y-axis. Table elements represent the calibrated fuel injection pulse width at the corresponding operating point. When the system detects a sudden change in external load, the feedforward compensation model immediately queries the curve library based on the current speed and the new load torque to calculate the theoretical increase in fuel injection required to balance the increased load. This increment is directly superimposed on the baseline fuel injection, achieving early energy injection for known disturbances. To improve feedforward accuracy, the model further introduces an injection pressure compensation factor. Since the actual injection quantity is proportional to the square root of the common rail pressure, the theoretical fuel injection pulse width must be divided by the square root of the ratio of the current pressure to the calibrated pressure when calculating. Furthermore, the model considers the influence of intake air temperature and pressure on air density, correcting the theoretical air-fuel ratio using the ideal gas law, thereby dynamically adjusting the fuel injection demand. The output of the feedforward compensation model is an open-loop command that does not depend on feedback error, so the response speed is extremely fast. It can complete the fuel supply correction within the first control cycle after a sudden load increase, effectively suppressing the speed drop.

[0027] The adaptive fuzzy sliding mode controller is a robust nonlinear feedback controller designed to handle system uncertainties and external disturbances. The controller takes a composite error characteristic matrix as input and employs a nonlinear switching function to achieve fast convergence. The switching function is defined as a linear combination of the error and its derivative, expressed as: in, To switch function values, This represents the current speed deviation. The rate of change of rotational speed, As a positive constant, it determines the approach rate. The controller expects the system state to move along... The sliding surface motion is at 0, and a control force is applied to bring it back if it deviates from this point. Traditional sliding mode control uses a sign function to generate the control law, which is prone to high-frequency chattering and damages the life of the actuator. To address this, this controller introduces a boundary layer correction mechanism, replacing the sign function with a saturation function. ,in For the boundary layer thickness, when The control input changes linearly to avoid abrupt switching. To further optimize performance, a time-varying gain factor is introduced into the exponential reaching law of the controller. The gain factor is dynamically adjusted according to the magnitude of the composite error. When the error is large, Take a higher value to speed up convergence; when the error approaches zero, Automatic reduction is implemented to decrease steady-state chattering. Gain adjustment rules are implemented through a fuzzy inference system. Inputs are the absolute value of the error and the absolute value of the rate of change; output is the gain coefficient. The membership function uses a triangular distribution. The rule base contains nine rules, covering typical cases such as "large error - high gain" and "small error - low gain." The controller output is the fuel supply adjustment amount, which, when superimposed with the base fuel supply, forms a closed-loop feedback control command.

[0028] The online reinforcement learning agent model is a machine learning-based autonomous optimization controller that continuously explores better control strategies during operation. The model employs a deep deterministic policy gradient algorithm architecture, comprising a policy network and a value network. The policy network takes the current operating state vector and the composite error feature matrix as input, directly outputting continuous fuel supply adjustment actions, i.e., the offset relative to the baseline fuel supply. The value network evaluates the effectiveness of the current policy and guides the updating of policy network parameters. The model constructs a closed-loop learning framework, with each control cycle considered as a time step. The environmental state is the real-time operating state vector, the agent's action is the fuel supply adjustment command, and the reward signal consists of three weighted components: the first is the negative value of the square of the speed deviation, encouraging reduced speed fluctuations; the second is the negative value of the fuel consumption rate increment, promoting energy-saving operation; and the third is a negative penalty term for the frequency of actuator actions, suppressing mechanical wear caused by frequent adjustments. The weight coefficients of each component are configurable; for example, in generator applications, stability is emphasized, giving higher weights to the deviation term; in construction machinery, economy is emphasized, increasing the weight of the fuel consumption term. Both the policy network and the value network are fully connected neural networks, each containing two hidden layers with 128 neurons. The Leaky ReLU activation function is used. The training process employs an experience replay mechanism, storing historical state-action-reward quadruplets in a circular buffer. Small batches of samples are randomly selected for parameter updates each time. The optimizer uses the Adam algorithm with a learning rate of 0.001. Initially, the model participates in control with low confidence, gradually increasing its output weights as training samples accumulate, ultimately leading to fine-tuning in the steady-state phase.

[0029] The multimodal control decision module is further configured with a mode arbitration logic unit, used to evaluate the consistency and prioritize the candidate adjustment commands output by the three control strategies mentioned above. The arbitration logic is based on a two-dimensional planar region division activation condition composed of the composite error amplitude and rate of change. The horizontal axis of the plane represents the absolute value of the speed deviation, and the vertical axis represents the absolute value of the speed change rate. The region is divided into three main intervals: the first interval is the high rate of change region, when the speed change rate exceeds 500 revolutions per second of the rated speed, it is judged as a load mutation event, and the feedforward compensation model is activated first for a fast response; the second interval is the medium deviation region, when the speed deviation is between 2% and 8% of the rated speed and the rate of change is below the threshold, the adaptive fuzzy sliding mode controller is activated to ensure stable convergence of the transient process; the third interval is the small error region, when the speed deviation is less than 1% of the rated speed and the system runs continuously for more than 5 seconds, the online reinforcement learning proxy model is handed over to perform long-term performance optimization. To avoid frequent switching of control modes near the critical point, the arbitration logic sets a hysteresis criterion, and each region boundary has a hysteresis width of 10%. For example, the threshold for switching from the medium deviation zone to the small error zone is 1%, while the threshold for reverse switching is 1.1%. Seamless switching between the three modes is achieved using a smooth transition function, an S-shaped sigmoid function. The slope parameter is set according to the switching speed requirements to ensure continuous change in control output and avoid secondary disturbances caused by mode jumps. The final output integrated fuel supply regulation command is the result of fusing the candidate commands according to their weights. The weights are dynamically allocated by the arbitration logic; for example, in the feedforward-dominated phase, its weight is close to 1, while the weights of other models approach 0.

[0030] The execution drive module receives the comprehensive fuel supply adjustment command ultimately determined by the multi-modal control decision module, converts it into a high-precision pulse width modulation signal, and drives the electromagnetic injector actuator of the high-pressure common rail system. The module is equipped with a dedicated power drive chip, which integrates an H-bridge drive circuit, current detection and protection logic, supports a maximum peak current of 20 amps, and a switching frequency of up to 100 kHz. The fuel injection pulse signal generation circuit is implemented based on a high-resolution timer unit. The timer's clock source is an external 200 MHz crystal oscillator, which, after frequency division, provides a 5 nanosecond time reference, achieving a minimum time resolution of 0.1 microseconds, meeting the requirements for ultra-fine fuel quantity adjustment. The module calculates the target injection pulse width based on the comprehensive fuel supply adjustment command, and, combined with the current common rail pressure, injector flow coefficient, and cylinder working volume, calculates the precise opening and closing times. The pulse signal, after power amplification, is applied to the electromagnetic injector coil, controlling the needle valve lift and duration, thereby achieving millisecond-level dynamic control of the fuel quality entering the cylinder. A complete fuel injection process includes multiple stages such as pre-injection, main injection, and post-injection. The pulse width of each stage is independently adjustable and is planned uniformly by the upper control command.

[0031] The actuator drive module features a hardware-level safety protection mechanism, monitoring the actuator's operating status in real time. The current detection circuit samples the drive circuit current every 10 microseconds. Upon detecting a short circuit (current surges above 15 amps), an open circuit (current is zero for 100 microseconds), or overheating (chip junction temperature exceeds 150 degrees Celsius), it immediately blocks the PWM output, cuts off the power supply, and activates the redundant control path. The redundant path consists of independent analog circuits, maintaining basic idle fuel injection functionality in case of main controller failure, ensuring safe equipment shutdown. Simultaneously, the module records the fault type and occurrence time to non-volatile memory and reports it to the monitoring system via the communication interface. To extend actuator lifespan, the module implements an operation frequency limiting strategy. When the number of fuel injections per unit time exceeds a preset limit (e.g., 30 times per second), a minimum interval is automatically inserted to prevent the electromagnetic coil from overheating. Furthermore, the module supports adaptive compensation for injector characteristics. By periodically performing zero-pulse and response tests, it obtains the actual opening and closing delays of each injector and compensates for them in subsequent control, ensuring uniform fuel supply across multiple cylinders.

[0032] The self-learning optimization module operates independently of the main control loop. Its function is to perform offline mining and knowledge extraction on long-term system operation data. The module periodically extracts typical operating condition sequences from the historical database during the low-load period in the early morning each day, with each processing session lasting no more than 10 minutes. Typical operating condition sequences include scenarios such as frequent start-stop, periodic load fluctuations, and operation in high and low temperature environments. Each sequence is at least 30 minutes long and is labeled with an operating condition category. The module uses a deep neural network to construct a nonlinear dynamic inverse model of the diesel engine. The model's input is the desired speed response trajectory and external load changes, and its output is the corresponding optimal fuel injection control sequence. The neural network structure includes an input layer, two hidden layers, and an output layer. The number of neurons in the hidden layers are 256 and 128, respectively. The Leaky ReLU activation function is used to alleviate the gradient vanishing problem. During training, mini-batch stochastic gradient descent is used with a batch size of 64, and batch normalization is introduced to improve the model's convergence speed and generalization ability. The loss function is defined as the mean square error between the predicted fuel injection sequence and the actual recorded fuel injection sequence. After training, the model back-derives an initial set of control parameters suitable for specific operating condition clusters, including the feedforward compensation gain, the proportional gain of the sliding mode controller, and the initial weights of the reinforcement learning policy network. After the effectiveness of these parameter sets is verified by simulation, they are written into the initialization configuration library of the multimodal control decision module. When the system restarts or detects that the current operating condition matches a known mode, the corresponding parameter set is automatically loaded, allowing the system to achieve a better control level without learning from scratch. This significantly reduces the time overhead of the adaptive learning process and improves engineering practicality.

[0033] This embodiment constructs an intelligent speed control system with multi-level perception, multi-strategy collaboration, and self-evolution capabilities through the collaborative work of the aforementioned modules. The real-time status perception module provides highly reliable and timely input data; the composite error generation module expands a single deviation into multi-dimensional features containing trend information, enhancing the predictability of the control system; the multi-modal control decision module integrates feedforward, feedback, and learning paradigms, achieving organic complementarity of different control advantages; the execution drive module ensures high-precision and high-safety execution of control commands; and the self-learning optimization module endows the system with long-term performance evolution capabilities. Without altering the basic structure of the diesel engine, the overall solution achieves a performance leap through software-defined intelligent control methods, effectively addressing complex challenges such as sudden load increases and decreases, environmental temperature variations, and equipment aging. Under test conditions of a 50% increase in rated load, the response time for the speed to recover from the lowest point to the target value is shortened to less than 0.3 seconds, the steady-state speed fluctuation rate is controlled below 0.8%, the average fuel consumption rate is reduced by 6.2%, and the concentration of nitrogen oxides in emissions is reduced by 12%, demonstrating outstanding technological advancement and practical value.

[0034] Existing electronic speed control systems mostly employ a single PID control algorithm, whose parameters remain fixed once tuned, making it difficult to adapt to the inherently strong nonlinear and time-varying characteristics of diesel engines. When faced with large load disturbances, fixed-gain PID controllers often exhibit severe overshoot or undershoot, leading to speed oscillations or slow recovery. Furthermore, PID algorithms rely solely on current and past error information, lacking the ability to predict future trends and thus failing to achieve proactive regulation. In contrast, the composite error characteristic matrix proposed in this invention not only includes the deviation itself but also incorporates integral and derivative information, enabling the control system to "sense" the acceleration characteristics of speed changes. This provides a more precise basis for the timing and intensity of feedforward and sliding mode control. In particular, the introduction of the derivative term allows the system to identify the accelerating decline trend before the speed deviates significantly, initiating the compensation mechanism in advance and fundamentally changing the passive response mode of traditional control.

[0035] The organic integration of multimodal control strategies is another core feature that distinguishes this invention from existing technologies. Traditional systems typically employ only one control algorithm, or while they may have multiple modes, they lack an effective arbitration mechanism. This invention achieves dynamic coordination and seamless switching of three control strategies through a mode arbitration logic unit. The feedforward compensation model rapidly injects energy during sudden load changes, playing a "frontline" role; the adaptive fuzzy sliding mode controller maintains system stability during transitions, acting as a "mainstay"; and the online reinforcement learning surrogate model continuously optimizes performance in the steady-state phase, undertaking the task of "refining" the system. This clearly defined and collaboratively ordered control architecture enables the system to exhibit optimal performance under various operating conditions. In particular, the use of the adaptive fuzzy sliding mode controller in the medium deviation region, through time-varying gain factors and boundary layer correction, ensures both rapid convergence and suppresses chattering, solving the problem of the difficulty in engineering applications of traditional sliding mode control.

[0036] The introduction of an online reinforcement learning mechanism enables this system to continuously self-optimize, a capability lacking in existing technologies. Traditional systems, once deployed, tend to have fixed control performance, unable to adapt to equipment wear, sensor drift, and changes in user habits that occur during long-term operation. The online reinforcement learning agent model of this invention continuously collects operational data and feedback rewards, constantly updating its policy network weights to gradually approach the optimal control strategy in the current state. More importantly, the self-learning optimization module solidifies the accumulated knowledge, generating an initial parameter set that can be used for rapid startup, avoiding the inefficient process of relearning every time a restart occurs. This closed-loop evolutionary mechanism of "running, learning, and accumulating" enables the system not only to adapt to short-term disturbances but also to cope with long-term uncertainties, possessing true intelligent attributes.

[0037] The high-precision and high-safety design of the execution drive module further enhances the system's engineering reliability. Existing systems typically have injection pulse resolutions on the order of 1 microsecond, which is insufficient to meet the precise fuel quantity regulation requirements of modern high-pressure common rail systems. This invention, by employing a high-resolution timer unit, improves the minimum time resolution to 0.1 microseconds, increasing the fuel supply regulation accuracy by an order of magnitude. Simultaneously, the hardware-level safety protection mechanism and redundant control path design ensure basic operational safety even under extreme fault conditions, meeting the stringent reliability requirements of industrial applications.

[0038] In summary, this invention, by constructing a closed-loop intelligent speed control system integrating perception, decision-making, execution, and evolution, comprehensively overcomes the inherent defects of existing technologies in terms of dynamic response, steady-state accuracy, energy consumption control, and operating condition adaptability. The system can not only rapidly restore stable speed under sudden load changes, but also continuously optimize fuel economy and emission performance during long-term operation, demonstrating significant technological advancements and promising industrial application prospects.

[0039] Example 2 Based on the above embodiments, this embodiment elaborates on the deep neural network training process in the self-learning optimization module, aiming to provide an efficient and stable model training method to ensure that the generated initial control parameter set has good generalization ability and engineering applicability.

[0040] The training data for the deep neural network used in the self-learning optimization module comes from long-term accumulated historical operation logs. These logs are stored as structured files on a solid-state drive, with each record containing a timestamp, operation state vector, control commands, actuator responses, and environmental parameters. Before training, the module performs a data preprocessing process, including outlier removal, missing value imputation, and data standardization. Outliers are identified and removed using the three-standard-deviation principle; missing values ​​are imputed using linear interpolation; and standardization uses the Z-score method to ensure that the mean of each feature is 0 and the standard deviation is 1. The preprocessed data is then divided into multiple subsets according to operating condition categories, with each subset independently used to train the inverse model for the corresponding operating condition.

[0041] The network training employed a supervised learning paradigm, with labeled data consisting of actual recorded fuel supply control sequences. To prevent overfitting, various regularization techniques were introduced during training. In addition to batch normalization, a Dropout layer was added after the hidden layers, with a dropout probability set to 0.3; an L2 weight decay term was added to the loss function, with a coefficient set to 0.0001. The optimization algorithm used a variant of AdamW, decoupling adaptive learning rate from weight decay. The initial learning rate was set to 0.001, and it automatically decayed to 0.9 every 10 training epochs if no loss decrease was observed. Training continued until the validation set loss no longer decreased for 5 consecutive epochs, or until a total of 500 epochs were reached. After training, the model retained its optimal weights, and its prediction uncertainty was evaluated using the Monte Carlo Dropout method. Only models with uncertainty below a threshold were considered usable.

[0042] In the parameter set generation phase, the system loads a pre-trained inverse model into the simulation environment, inputs the expected response trajectory for typical operating conditions, and generates a set of optimal fuel supply sequences. Subsequently, using a system identification method, this sequence is backfitted into interpretable parameters such as feedforward gain, sliding mode coefficient, and initial weights for reinforcement learning. The fitting process employs a global search using a genetic algorithm, with the fitness function comprehensively considering tracking error, energy consumption indicators, and motion smoothness. The final generated parameter set includes a description of applicable conditions, such as "applicable to power generation conditions with ambient temperatures of 10 to 30 degrees Celsius and load change rates less than 200 Nm / s," and is stored in the initialization configuration library. When the system is running, the operating condition identification algorithm matches the current conditions with the parameter set labels to achieve accurate loading.

[0043] This embodiment ensures the reliability and practicality of the results produced by the self-learning optimization module through standardized data processing, robust model training, and an interpretable parameter transformation process. This enables the knowledge accumulation process to truly serve engineering deployment and greatly improves the system's intelligence level and ease of operation and maintenance.

Claims

1. A diesel engine intelligent speed control system, characterized in that, include: The real-time status perception module is used to continuously collect physical parameters such as instantaneous crankshaft speed, injection pressure, intake air temperature and pressure, exhaust back pressure, lubricating oil temperature, coolant temperature, and external load torque through a multi-source sensor network integrated into the diesel engine body and load end. It also performs anti-interference filtering and dynamic compensation processing on the raw signals and outputs a corrected real-time operating status vector. The composite error generation module is used to receive the actual speed signal in the real-time running state vector and the target speed command set by the user, calculate the basic speed deviation value between the two, and introduce the differential term of the speed change rate to construct a three-dimensional composite error feature matrix containing the deviation amount, the cumulative deviation integral amount and the differential component of the change trend. The multimodal control decision module is used to receive the real-time operating state vector and the three-dimensional composite error feature matrix, and dynamically activate and coordinate three parallel operation control strategy models to generate fuel supply adjustment commands based on these. The execution drive module is used to receive the comprehensive fuel supply adjustment command output by the multimodal control decision module and convert it into a high-precision pulse width modulation signal to drive the electromagnetic injector actuator of the high-pressure common rail system to precisely control the fuel supply. The self-learning optimization module is used to operate independently of the main control loop. It periodically extracts typical operating condition sequences from the historical database, uses a deep neural network to construct a nonlinear dynamic inverse model of the diesel engine, and trains it to obtain an initial control parameter set suitable for a specific set of operating conditions. The initial configuration library of the multimodal control decision module is then written into the initialization configuration library of the multimodal control decision module.

2. The intelligent speed control system for diesel engines according to claim 1, characterized in that, The three parallel operation control strategy models include: a feedforward compensation model based on a diesel engine static characteristic curve library, an adaptive fuzzy sliding mode controller with the three-dimensional composite error feature matrix as input, and an online reinforcement learning agent model using a state-action-reward mechanism. The multimodal control decision module is equipped with a mode arbitration logic unit, which is used to prioritize and smoothly switch candidate adjustment commands of the three control strategy models according to the amplitude and rate of change of the three-dimensional composite error feature matrix. The multi-source sensor network in the real-time state perception module adopts a dual-redundancy arrangement structure, with two sets of key sensors configured as primary and backup. The system judges the health status of the sensors through a cross-validation algorithm, automatically shields the failed channels, and switches to the backup signal source.

3. The intelligent speed control system for diesel engines according to claim 2, characterized in that, The composite error generation module uses an amplitude-limited anti-saturation algorithm to process the cumulative deviation integral, and combines a low-pass filter to eliminate high-frequency noise interference for the variation trend micro-component.

4. The intelligent speed control system for diesel engines according to claim 2, characterized in that, The switching function exponential approach law of the adaptive fuzzy sliding mode controller introduces a time-varying gain factor. The time-varying gain factor is dynamically adjusted according to the amplitude of the three-dimensional composite error characteristic matrix. When the error is large, a high gain is used, and when the error approaches zero, the gain is automatically reduced.

5. The intelligent speed control system for diesel engines according to claim 2, characterized in that, The reward function of the online reinforcement learning agent model consists of three weighted components: the negative value of the square of the speed deviation, the negative value of the fuel consumption rate increment, and the negative penalty term for the frequency of actuator actions.

6. The intelligent speed control system for diesel engines according to claim 2, characterized in that, The mode arbitration logic unit is equipped with a hysteresis criterion. The activation conditions of each control mode are divided into two-dimensional plane regions composed of the amplitude and rate of change of the three-dimensional composite error feature matrix. The region boundary supports parameterized configuration.

7. The intelligent speed control system for diesel engines according to claim 2, characterized in that, The execution drive module is equipped with a dedicated power drive chip, which has short-circuit protection, over-temperature derating and open-circuit detection functions, and uses a high-resolution timer unit to generate fuel injection pulse signals.

8. The intelligent speed control system for diesel engines according to claim 2, characterized in that, The deep neural network used in the self-learning optimization module includes an input layer, two hidden layers, and an output layer. The activation function is Leaky ReLU. During training, the mini-batch stochastic gradient descent method is used, and batch normalization is introduced.

9. The intelligent speed control system for diesel engines according to claim 2, characterized in that, The feedforward compensation model queries the diesel engine static characteristic curve library based on the current speed and load torque to calculate the theoretical fuel supply increment, and further introduces the injection pressure compensation factor and intake state correction factor to dynamically adjust the fuel supply demand.

10. The intelligent speed control system for diesel engines according to claim 2, characterized in that, The execution drive module has a hardware-level safety protection mechanism, which monitors the actuator's operating current and response delay in real time. Once an abnormality is detected, it immediately activates the redundant control path and issues a fault alarm.

Images