A fuel injector performance self-calibration system and method fusing digital twin and neural network

Abstract

The application discloses a fuel injector performance self-calibration system and method fusing digital twin and neural network, comprising: constructing a physical-data hybrid driven fuel injector digital twin model; realizing real-time synchronization of the digital twin model and the state of the physical injector through a state estimation algorithm; performing performance prediction and ideal trajectory planning based on the synchronized digital twin model; constructing and running a control parameter reverse solving neural network model, reversely solving calibration instructions according to the ideal trajectory; and executing the calibration instructions and continuously optimizing the model online to form a closed-loop adaptive control. The application solves the problems that the traditional method cannot adapt to performance aging and model precision deficiency, and through high-precision twin prediction and intelligent reverse decision, a closed-loop adaptive online calibration system is constructed, which can improve the fuel efficiency of the engine in the whole life cycle, reduce emissions and enhance reliability.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and engine control technology, and in particular to a fuel injector performance self-calibration system and method that integrates digital twins and neural networks. Background Technology

[0002] As a key actuator in internal combustion engines, the fuel injector's performance accuracy and stability play a decisive role in the engine's combustion efficiency, power performance, fuel economy, and emission control. An ideal fuel injection process requires precisely metered fuel, atomized in a specific manner, and injected at a specific angle into the cylinder at precise moments. However, in real-world operating environments, the performance of fuel injectors is not static but degrades or drifts with time and changing operating conditions. On one hand, minute tolerances in the manufacturing process result in initial performance differences for each injector at the time of manufacture. On the other hand, during long-term use, moving parts inside the injector, such as the needle valve, experience mechanical wear, and the nozzles accumulate carbon deposits due to impurities and incomplete combustion products in the fuel. These physical changes cause key performance parameters such as the actual fuel injection quantity and injection rate to deviate from their design values. Furthermore, fluctuations in fuel quality, operating temperature, and pressure can also have a momentary impact on its performance.

[0003] The aforementioned performance drift can trigger a series of technical problems. For example, deviations in fuel injection quantity can disrupt the engine's precise air-fuel ratio control, leading to incomplete combustion, which in turn increases fuel consumption and emissions of harmful substances such as hydrocarbons (HC) and nitrogen oxides (NOx). In multi-cylinder engines, inconsistencies in the performance of injectors in each cylinder can disrupt the power balance between cylinders, causing engine vibration, noise, and even misfires or knocking under certain operating conditions, severely affecting the engine's smoothness and reliability.

[0004] To address the performance drift issue of fuel injectors, existing technologies primarily employ two approaches. The first is offline calibration and periodic maintenance. This method involves comprehensive performance testing of the injector on a standard test bench during the product development phase, and the resulting characteristic data is stored in the engine control unit (ECU) in the form of a lookup table. The ECU then queries this static data based on the engine's current operating conditions to determine the injector's drive commands. The drawback of this method is that the calibration data is static and cannot reflect the performance aging process of the injector throughout its entire lifespan, nor can it adapt to the complex changes in operating conditions during actual operation. Once significant performance drift occurs, the only solution is to periodically replace the injector or return it to the factory for recalibration, which increases maintenance costs and inconvenience for users.

[0005] The second approach is model-based online compensation. This method attempts to establish a mathematical-physical model that describes the fuel injection process, predicts injection behavior through the model, and adjusts control commands online. However, fuel injection is a complex process involving strong coupling of multiple physical fields such as electromagnetics, fluid dynamics, and thermodynamics, including nonlinear phenomena such as fuel cavitation, flash boiling, and high-speed dynamic response of the needle valve. Establishing a high-precision physical model that accurately describes all these phenomena is extremely difficult, and such models often involve enormous computational demands, making real-time operation in vehicle ECUs with limited computing power challenging. Simplified physical models, on the other hand, neglect some effects, resulting in limited predictive accuracy. Especially when the injector ages or malfunctions, the deviation between the model and reality increases dramatically, leading to compensation failure. In recent years, some solutions have introduced purely data-driven neural network models to fit the input-output relationship of the injector. While these models demonstrate good fitting ability on specific datasets, their black-box nature makes them lack physical interpretability. In edge cases not covered by training data, the reliability and generalization ability of their predictions are difficult to guarantee. Furthermore, most solutions remain at the level of performance prediction or fault diagnosis, failing to form a self-calibration system capable of real-time, proactive closed-loop optimization of the injector's control strategy. Therefore, developing a technology capable of accurate, real-time, and adaptive online calibration of fuel injectors throughout their entire lifecycle to maintain their long-term optimal performance is a significant technical challenge currently facing the field of advanced engine technology. Summary of the Invention

[0006] The technical problem to be solved by this invention is to provide a technical solution that can realize online adaptive calibration of fuel injectors throughout their entire life cycle, addressing the shortcomings of existing fuel injector performance calibration and compensation technologies, such as offline calibration being unable to adapt to performance aging, insufficient accuracy of physical models and computational complexity, and poor generalization ability and interpretability of pure data-driven models.

[0007] To address the aforementioned technical problems, one aspect of the present invention provides a fuel injector performance self-calibration method integrating digital twins and neural networks. This method constructs a physical-data hybrid driven digital twin model that is synchronized in real-time with the physical injector's state. This high-precision twin model is used for performance prediction and ideal trajectory planning. Furthermore, an inverse-solution neural network model capable of handling control ambiguity problems is used to calculate the optimal calibration control command required to achieve ideal performance. Ultimately, this forms an adaptive closed-loop control and self-evolving learning system encompassing prediction, decision-making, execution, and feedback.

[0008] The specific technical solution provided by this invention may include the following steps:

[0009] Step one involves constructing a physics-data hybrid-driven digital twin model of the fuel injector. The goal of this step is to create a virtual computational model capable of accurately and efficiently reproducing the comprehensive dynamic characteristics of the physical fuel injector. The overall input to this digital twin model consists of control commands sent to the physical injector, such as drive pulse width and drive voltage waveforms, as well as real-time boundary condition parameters, such as common rail pressure, fuel temperature, and ambient temperature. Its output is the prediction of key injector performance indicators, such as predicted single-cycle injection quality, instantaneous injection rate curve, and Sauter Mean Diameter (SMD). This model does not employ a single modeling paradigm but organically integrates a physical mechanism model and a data-driven model. Specifically, a simplified physical model describing the core working mechanism of the fuel injector is first established. This simplified physical model is based on recognized fundamental equations of fluid mechanics and electromagnetism. For example, it describes the relationship between the driving current and electromagnetic force using an RLC circuit model of an electromagnetic coil, describes the dynamic response of the electromagnetic force and the needle valve lift using a second-order dynamic model of the needle valve, and describes the relationship between the needle valve lift, fuel pressure, and injection flow rate using Bernoulli's equation or orifice flow rate formula. This simplified physical model ensures that the digital twin possesses basic physical logic correctness and interpretability, while its computational complexity is low, making it suitable for real-time computation. To compensate for the complex nonlinear effects that this simplified physical model cannot capture due to its simplification assumptions, such as fuel cavitation, the influence of carbon deposits on injection flow characteristics, and changes in fuel viscosity and density caused by temperature variations, this invention further constructs a residual neural network model. The input of this residual neural network model can include real-time operating parameters of the injector (such as fuel pressure and driving voltage) and internal state variables of the simplified physical model (such as the predicted maximum needle valve lift). Its output is the deviation between the prediction results of the simplified physical model and the actual performance of the physical entity, i.e., residual terms or correction terms. Finally, the output of the simplified physical model is fused with the output of the residual neural network, for example, through linear superposition or nonlinear combination through a small network, to form the final high-precision prediction output of the digital twin model. Preferably, the residual neural network can employ a deep network structure with strong nonlinear temporal feature capture capabilities, such as a Long Short-Term Memory (LSTM) network with an attention mechanism or a Gated Recurrent Unit (GRU), to effectively learn and compensate for the complex dynamic characteristics of the jetting process.

[0010] Step two involves achieving real-time synchronization between the digital twin model and the physical injector's state. The purpose of this step is to ensure that the digital twin model can dynamically track and accurately reflect changes in the physical injector's real-time health status and performance due to factors such as wear and aging. Specifically, a series of physical measurements related to injection performance are collected in real-time using a sensor array deployed on the physical injector system or its associated systems. These measurements can be divided into direct and indirect measurements. Direct measurements may include the current and voltage waveforms of the drive coil, fuel common rail pressure sensor signals, etc., which directly reflect the electromagnetic and hydraulic drive state of the injector. Indirect measurements may include combustion process feedback signals obtained through in-cylinder pressure sensors, knock sensors, or wide-range oxygen sensors, which reflect the final effect of fuel injection. The collected real-time sensor data is input into a state estimation algorithm module. This module compares the actual sensor measurements with the predicted output values ​​of the digital twin model under the same input conditions. This invention also addresses the bias caused by this comparison by employing advanced filtering algorithms suitable for nonlinear systems, such as Unscented Kalman Filter (UKF) or Particle Filter (PF). Based on the bias, this filtering algorithm performs real-time optimal estimation and correction of the internal state variables of the digital twin model and some key network parameters of the residual neural network. An important aspect is that this invention also includes some physical characteristic parameters that are difficult to measure directly but change slowly with aging, such as the effective flow area of ​​the nozzle, the stiffness of the needle valve return spring, or the dynamic response delay time of the needle valve, as part of the state vector for online estimation. In this way, the digital twin can quantify and track changes in physical characteristics caused by mechanical wear or nozzle carbon buildup in real time. Optionally, to more effectively utilize indirect measurements such as combustion feedback, the state estimation algorithm module can further integrate a simplified combustion model. This model is used to establish the correlation between injection performance parameters such as injection quantity and injection rate and combustion result indicators such as in-cylinder pressure and combustion heat release rate, thereby enabling more accurate back-calculation and correction of the twin model's state based on the final combustion effect. Through this step, the digital twin is transformed from a static simulation model into a dynamic, living mirror image that grows and evolves synchronously with the physical entity.

[0011] Step three involves performance prediction and ideal trajectory planning based on the synchronized digital twin model. In this step, the digital twin model, already highly matched to the physical entity's performance, serves as a precise virtual testbed that requires no physical experimental setup. First, based on the engine ECU's current performance requirements—such as the target cycle injection quantity, desired spray penetration distance, or desired combustion heat release rate—the system can perform multiple sets of advanced simulation calculations on the digital twin model. These simulations can quickly and accurately predict the actual injection performance (such as actual injection quantity, injection rate curve, Sauter mean diameter, etc.) corresponding to a series of different control parameter combinations (e.g., different main injection pulse widths, pre-injection pulse widths, injection start points, and intervals between multiple injections) under the current operating conditions and injector health status. Based on this massive amount of virtual simulation data, the system can deeply analyze the deviation between the current physical injector's performance and its new or ideal design state. Furthermore, the system can plan an ideal performance trajectory in this virtual space based on combustion and emission control optimization criteria. For example, design an ideal injection rate curve that can achieve optimal fuel atomization and minimum soot generation while meeting the total injection quantity requirements.

[0012] Step four involves constructing and running a neural network model for inverse control parameter solving to generate calibration instructions. This step is the core of adaptive calibration, aiming to solve for the optimal control parameters that achieve the ideal performance planned in step three. To realize this inverse mapping from target to control, this invention constructs a dedicated inverse solving neural network model. This invention effectively addresses the inherent multi-solution problem in control mapping by constructing and employing a Mixture Density Network (MDN) as the inverse control parameter solving model. In fuel injection control, various combinations of control parameters may produce similar injection effects, representing a one-to-many mapping relationship that traditional neural networks with deterministic outputs struggle to handle. The output of the Mixture Density Network (MDN) is not a deterministic control parameter value, but rather a set of parameters from a Gaussian mixture model, specifically including the mean vector, covariance matrix, and mixing weight coefficients for each Gaussian component. This set of parameters collectively describes the probability distribution of all possible control parameter solutions for a given performance target. The training data for this reverse-engineering MDN model can be entirely derived from the control parameter-performance data pairs generated in step three through large-scale virtual simulation experiments using a digital twin model, thus avoiding expensive and time-consuming physical experiments. During system operation, the ideal performance trajectory planned in step three (e.g., target fuel injection quantity and target Sauter average diameter) is used as input to the MDN, which immediately outputs the probability distribution of the optimal control parameters.

[0013] Step five involves risk assessment, closed-loop control, and continuous online optimization of the model. This step transforms the solution generated by MDN into actual control actions, forming a closed-loop and continuous learning mechanism. Specifically, the system first analyzes the probability distribution of the control parameters output by MDN and selects one or more probability peak points as candidate control commands. Preferably, before finally selecting and executing a candidate command, the system uses the digital twin model, which is highly synchronized with the entity in Step two, to perform a rapid virtual stress test or risk pre-assessment on the top N high-probability candidate control commands output by MDN. This virtual stress test includes injecting typical transient disturbances of engine operating conditions, such as instantaneous fluctuations in fuel pressure, into the digital twin model, and performing multi-step advance prediction and quantitative evaluation of the performance stability of each candidate control command under disturbances, calculating a control robustness score. Ultimately, the system selects not simply the theoretically optimal solution given by MDN, but the risk-adjusted optimal control command obtained after a weighted balance between theoretical optimality and control robustness score. This command is then sent to the injector drive unit and executed. After execution, the real-time status synchronization mechanism in step two immediately captures the actual physical effects of the injection, forming a complete prediction-decision-execution-feedback closed loop. Furthermore, the system continuously accumulates new and effective data during long-term operation. This data accurately records the relationship between the injector's control input and performance output under different operating conditions and aging levels. This data is used as incremental learning samples to periodically optimize the residual neural network and the inverse MDN model in the digital twin model online. This continuous online learning mechanism endows the entire calibration system with self-evolution capabilities, enabling it to adapt to longer-term and more severe performance degradation, and even failure modes not anticipated in the initial model design phase, thus ensuring the system's robustness throughout its entire lifecycle.

[0014] Another aspect of the present invention provides a fuel injector performance self-calibration system integrating digital twins and neural networks. This system may include: a data acquisition module configured to collect sensor data from the injector and associated systems; a digital twin module configured to run the physical-data hybrid-driven digital twin model; a state synchronization module configured to synchronize the digital twin model with the physical injector state using a state estimation algorithm; a reverse calibration module configured to run the control parameter inverse solution neural network model and generate calibration instructions; and a control execution module configured to execute the calibration instructions. These modules are configured to work collaboratively to perform the corresponding steps in the aforementioned method.

[0015] Compared with existing technologies, this invention has the following advantages: By constructing a physical-data hybrid driven digital twin model and utilizing advanced nonlinear filtering algorithms to achieve real-time synchronization of key physical parameters and physical entities, this invention solves the problems of insufficient accuracy of traditional physical models and poor generalization ability of pure data models, achieving accurate tracking and quantification of injector performance degradation throughout its entire life cycle. By employing a hybrid density network (MDN) for inverse solution of control parameters and combining it with risk pre-assessment based on digital twins, this invention effectively solves the problems of multiple solutions and robustness in complex control systems, generating calibration commands that balance optimal performance and control stability. Finally, by combining high-precision twin prediction, intelligent inverse decision-making, and continuous online learning, this invention constructs a complete, closed-loop, adaptive, and self-evolving online self-calibration system, which can significantly improve engine fuel efficiency throughout its entire life cycle, reduce harmful emissions, and enhance system reliability and environmental adaptability. Attached Figure Description

[0016] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0017] Figure 1 This is a flowchart illustrating a fuel injector performance self-calibration method that integrates digital twins and neural networks according to an embodiment of the present invention.

[0018] Figure 2 This is a schematic diagram of a digital twin model structure of a fuel injector driven by a physical-data hybrid approach according to an embodiment of the present invention.

[0019] Figure 3 This is a schematic diagram illustrating the real-time synchronization process between a digital twin model and the physical injector state according to an embodiment of the present invention.

[0020] Figure 4 This is a functional block diagram of a fuel injector performance self-calibration system that integrates digital twins and neural networks according to an embodiment of the present invention.

[0021] Figure 5 This is a schematic diagram comparing the prediction accuracy of a digital twin model and a simplified physical model according to an embodiment of the present invention.

[0022] Figure 6 This is a schematic diagram illustrating the online tracking effect of the effective flow area of ​​the nozzle according to an embodiment of the present invention.

[0023] Figure 7 This is a schematic diagram of the dynamic lift response curve of a fuel injector needle valve according to an embodiment of the present invention.

[0024] Figure 8This is a schematic diagram illustrating the convergence process of the unscented Kalman filter algorithm for state estimation of the effective flow area of ​​a nozzle according to an embodiment of the present invention.

[0025] Figure 9 This is a schematic diagram illustrating the effect of fuel temperature on injection performance according to an embodiment of the present invention.

[0026] Figure 10 This is a schematic cross-sectional view of the physical structure of a fuel injector according to an embodiment of the present invention.

[0027] Figure 11 This is a cross-sectional schematic diagram of the flow field and cavitation region inside the nozzle according to an embodiment of the present invention. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this invention clearer, the specific embodiments of this invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0029] Example 1

[0030] This embodiment provides a self-calibration method for fuel injector performance that integrates digital twins and neural networks. (Refer to...) Figure 1 This method is a closed-loop adaptive control method. It constructs a digital twin that is synchronized with the physical injector in real time, and uses the twin to make intelligent decisions, ultimately achieving online calibration of the injector control commands.

[0031] Please refer to Figure 1 The method described in this invention may specifically include the following steps:

[0032] Step S100: Construct a physics-data hybrid-driven digital twin model of the fuel injector. The purpose of this step is to create a high-fidelity virtual injector model. (Refer to...) Figure 10 A typical fuel injector includes core components such as an electromagnetic coil, needle valve, spring, and nozzle. (Refer to...) Figure 2The construction of this digital twin model 200 integrates two technical approaches. First, a simplified physical model 210 is established. This model, based on explicit physical laws, mathematically describes the core dynamic process of the fuel injector. Specifically, a resistor-inductor-capacitor (RLC) circuit model can be used to characterize the relationship between the voltage waveform applied by the drive unit and the drive current generated in the injector's electromagnetic coil. Subsequently, a second-order dynamic equation considering the needle valve mass, spring preload, hydraulic pressure, and electromagnetic force, such as m*d^2x / dt^2 + c*dx / dt + k*x = F_em(t) - F_hyd(t) - F_spring, is used to describe the dynamic lift response of the needle valve, where m is the equivalent moving mass of the needle valve, c is the damping coefficient, k is the spring stiffness, x is the needle valve lift, F_em(t) is the electromagnetic force, F_hyd(t) is the fuel hydraulic pressure, and F_spring is the spring preload. (Refer to...) Figure 7 The figure illustrates the typical dynamic process of needle valve lift changing over time under the influence of driving voltage, including the rising stage driven by electromagnetic force, the stable stage reaching maximum lift, and the closing stage returning to its original position under spring force. Finally, the instantaneous injection rate at a specific needle valve lift is calculated using Bernoulli's equation or a modified orifice flow formula, such as Q_rate(t) = C_d * A_eff * sqrt(2*ΔP(t) / ρ), where Q_rate(t) is the instantaneous volumetric flow rate, C_d is the flow coefficient, A_eff is the effective flow area of ​​the injection orifice, ΔP(t) is the pressure difference inside and outside the injection orifice, and ρ is the fuel density. This simplified physical model 210 ensures that the predictions of the digital twin model 200 have a solid physical basis and good extrapolation ability. However, to compensate for the complex nonlinear phenomena neglected by the simplification of this model, such as the compressibility of fuel under high pressure and the reference... Figure 11 The cavitation effect shown in the nozzle, reference Figure 9To illustrate the effects of temperature changes on fuel properties (such as dynamic viscosity) and the alteration of flow characteristics by carbon deposits, this invention further constructs a residual neural network model 220. The inputs to this residual neural network model 220 include control commands 201 sent to the physical injector (e.g., drive pulse width), real-time boundary condition parameters 202 (e.g., common rail pressure, fuel temperature), and internal state variables of the simplified physical model 210 (e.g., predicted maximum needle valve lift). Its output is the deviation between the simplified physical model's prediction result 211 and the actual performance of the physical entity, i.e., the residual term 221. Preferably, the residual neural network model 220 can employ a Long Short-Term Memory (LSTM) network structure to effectively capture the temporal dependence and dynamic nonlinear characteristics of the injection process. Finally, the final predicted output 203 of the digital twin model 200, such as the predicted single-cycle injection quality, is obtained by fusing 230 the output 211 of the simplified physical model 210 with the output 221 of the residual neural network model 220. This fusion can be a simple linear superposition or a nonlinear combination achieved through a small, fully connected network, thus obtaining a final prediction result that combines physical interpretability and data fitting accuracy. (See reference...) Figure 5 The figure visually demonstrates that the predicted output 203 (e.g., instantaneous fuel injection rate) of the digital twin model 200 described in this invention can more accurately match the actual value of the physical injector compared to the predicted result 211 of the simplified physical model 210. The key is that the residual neural network model 220 successfully compensates for the residual term 221 that the simplified physical model cannot describe.

[0033] Step S200 involves real-time synchronization of the digital twin model's state with that of the physical injector. This step ensures that the virtual digital twin model can dynamically track and reflect the actual performance evolution of the physical injector 100. (Refer to...) Figure 3A series of physical measurements 301 are acquired in real time through a sensor array deployed on the physical system. These measurements may include drive current waveforms and common rail pressure sensor signals that directly reflect the driving state, as well as in-cylinder pressure sensor signals or wide-area oxygen sensor signals at the exhaust manifold that indirectly reflect the final injection effect. The acquired real-time sensor data is sent to a state estimation algorithm module 302 in a state synchronization module 300. This module compares the actual measured values ​​of the sensors with the predicted output values ​​of the digital twin model 200 under the same input conditions, generating an observation residual. Further, the present invention employs the unscented Kalman filter (UKF) algorithm to perform optimal estimation and online correction 303 of the internal state vector of the digital twin model 200 based on this observation residual. An important technical feature is that the state vector not only includes rapidly changing dynamic variables such as needle valve position and speed, but also includes a series of slowly varying physical parameters that characterize the long-term health state of the injector and are difficult to measure directly, such as the effective flow area of ​​the nozzle A_eff, the needle valve return spring stiffness k, or the needle valve dynamic response delay time. (Refer to...) Figure 8 The figure illustrates the convergence process of the UKF algorithm in estimating the effective flow area A_eff of the nozzle. Even starting from an initial guess with a large deviation, the estimation algorithm can quickly converge to near the true value within several injection cycles and achieve stable tracking. By estimating these physical parameters as state variables online, the digital twin model 200 can quantify and track changes in physical properties caused by aging factors such as mechanical wear or nozzle carbon buildup in real time, thus transforming it from a static model into a dynamic mirror that can evolve synchronously with the physical entity 100.

[0034] Step S300 involves performance prediction and ideal trajectory planning based on the synchronized digital twin model. In this step, the digital twin model, which is highly matched to the physical entity's performance, is used as a precise virtual test bench that does not require a physical test rig. First, based on the engine ECU's performance requirements under current operating conditions, such as the target cycle injection quantity, desired spray penetration distance, or desired combustion heat release rate pattern, the system can perform multiple sets of advanced simulation calculations on the digital twin model. These simulations can quickly and accurately predict what actual injection performance (such as actual injection quantity, injection rate curve, Sauter mean diameter, etc.) will correspond to under current operating conditions and the injector's current health state, given a series of different combinations of control parameters (e.g., different main injection pulse widths, pre-injection pulse widths, injection start points, and intervals between multiple injections). Based on this massive amount of virtual simulation data, the system can deeply analyze the degree of deviation between the current physical injector's performance and its new or ideal design state. Furthermore, the system can plan an ideal performance curve, i.e., the trajectory, in this virtual space based on the optimization criteria of combustion science and emission control. For example, design an ideal injection rate curve that can achieve optimal fuel atomization and minimum soot generation while meeting the total injection quantity requirements.

[0035] Step S400 involves constructing and running a neural network model for inverse control parameter solving to generate calibration commands. This step addresses the inverse problem of deriving control commands from desired performance. Traditional control mapping suffers from pleiosity, meaning that various combinations of control parameters (e.g., a single main injection or a pre-injection plus a main injection) may achieve similar final injection quantities. To address this issue, this invention constructs and employs a Mixture Density Network (MDN) as the inverse solution model. The input to this MDN model is the ideal performance trajectory planned in step S300, such as the target injection quantity and the target Sauter Mean Diameter (SMD). Its output is not a fixed control parameter value, but rather a probability distribution describing all possible solutions. Specifically, the output is a set of parameters for a Gaussian mixture model, including the mean vector, covariance matrix, and mixing weight coefficients for each Gaussian component. The mean vector of each Gaussian component represents a potential, feasible control command solution (e.g., a specific combination of injection pulse width and injection initiation point), while its weight coefficients reflect the probability or likelihood of that solution. The training data for this MDN model comes entirely from large-scale offline or online virtual simulations of the digital twin model. That is, by running massive amounts of control parameter-performance simulation experiments on the digital twin model, the data pairs required for training are generated, thus avoiding dependence on expensive physical test benches.

[0036] Step S500 involves performing risk assessment, closed-loop control, and continuous online optimization of the model. This step is crucial for actual calibration and the formation of self-evolutionary capabilities. First, the system selects several solutions with the highest probabilities from the control parameter probability distribution output by the MDN in step S400 as candidate control commands. Before issuing these commands to the physical injectors, this invention introduces a pre-assessment step based on a digital twin model. Specifically, the system utilizes a digital twin model highly synchronized with the physical entity to perform a rapid performance stability simulation prediction for each candidate control command. For example, the instantaneous fluctuation of the common rail pressure is simulated in the digital twin model, and then the deviation of each candidate command's injection performance (e.g., injection quantity) under this disturbance is evaluated, and a control robustness score is calculated based on this deviation. A specific calculation method is as follows: the integral of the absolute value of the difference between the target injection quantity and the predicted injection quantity after disturbance within a complete injection event simulation period is taken as the performance deviation degree D. Then, the control robustness score S_r can be calculated as S_r = 1 / (1 + D). The smaller the deviation D of the instruction, the higher its control robustness score S_r. Finally, the system combines the probability P_mdn of the candidate instructions given by MDN with the calculated control robustness score S_r in a weighted balance. For example, by calculating a comprehensive performance index I = w1 * P_mdn + w2 * S_r (where w1 and w2 are preset weight coefficients), the instruction with the highest comprehensive performance index I is selected as the optimal control instruction after risk adjustment. This optimal instruction is sent to the injector drive unit for execution. After execution, the sensor data of the physical system is immediately fed back through the state synchronization mechanism in step S200 to verify the effect of this calibration, thus forming a complete prediction-decision-execution-feedback closed-loop control. Furthermore, during long-term operation, all verified control instruction-operating condition-actual performance output data triplets are stored as incremental learning samples. The system can be configured to periodically (e.g., every 100 working hours) or when the model prediction error is detected to consistently exceed a threshold, use these newly accumulated samples to perform online parameter fine-tuning or retraining of the residual neural network model and the inverse MDN model in the digital twin model. This continuous learning capability enables the entire self-calibration system to possess self-evolutionary characteristics, continuously adapting to deeper aging of the injector, and even some unforeseen operating modes, thereby ensuring its calibration accuracy and robustness throughout its entire lifecycle.

[0037] Example 2

[0038] This embodiment provides a fuel injector performance self-calibration system integrating digital twins and neural networks, which is used to implement the method described in Embodiment 1 above. (Refer to...) Figure 4The system 400 may include: a data acquisition module 410, a digital twin module 420, a status synchronization module 430, a reverse calibration module 440, and a control execution module 450.

[0039] The data acquisition module 410 is configured to communicate with various sensors (such as current sensors, pressure sensors, temperature sensors, wide-range oxygen sensors, etc.) on the physical injector 100 and its associated systems (such as engine cylinder and fuel common rail) to collect various physical measurements reflecting the working status and boundary conditions of the injector in real time, and transmit the collected data to the status synchronization module 430 and the digital twin module 420.

[0040] The digital twin module 420 integrates a computing unit and a memory to store and run the physical-data hybrid driven digital twin model. This module receives control commands from the control execution module 450 and real-time operating parameters from the data acquisition module 410, and performs the functions of steps S100 and S300, namely, running the model to perform advanced performance prediction and planning an ideal performance trajectory according to preset optimization criteria.

[0041] The state synchronization module 430 is configured to receive real-time sensor measurements from the data acquisition module 410 and model predictions from the digital twin module 420. Internally, this module implements a state estimation algorithm, preferably an unscented Kalman filter algorithm, to perform the function of the aforementioned step S200. Specifically, by comparing the measured values ​​with the predicted values, it estimates and corrects the model state variables and key physical parameters in the digital twin module 420 in real time to ensure state synchronization between the digital twin model and the physical injector 100.

[0042] The reverse calibration module 440 integrates a computing unit and a memory to store and run the reverse-engineering neural network model of the control parameters, preferably a hybrid density network model. This module receives the ideal performance trajectory planned by the digital twin module 420, executes the decision-making functions in steps S400 and S500, that is, reverse-engineers the probability distribution of the control parameters that can achieve the ideal performance, and further combines this with the risk pre-assessment performed by the digital twin module 420 to finally generate the risk-adjusted optimal calibration command.

[0043] The control execution module 450 may be part of the engine control unit (ECU) or a separate unit that works in conjunction with it. This module receives the optimal calibration command generated by the reverse calibration module 440 and converts it into a drive signal, sending it to the drive circuit of the physical injector 100 to control the actual injection action of the physical injector 100, thereby completing closed-loop control. Simultaneously, this module also feeds back the executed command information to the digital twin module 420 for simulation.

[0044] During the operation of this system, the above modules work together to form a continuous closed loop. The digital twin module 420 provides accurate prediction, the state synchronization module 430 ensures the real-time accuracy of the prediction, the reverse calibration module 440 makes intelligent decisions, the control execution module 450 is responsible for execution, and the data acquisition module 410 constitutes a feedback path, enabling the entire system to continuously adjust and optimize its control strategy based on the real-time state changes of the physical injector 100.

[0045] Example 3

[0046] This embodiment illustrates the implementation process of the method of the present invention in a specific application scenario, particularly for a fuel injector whose performance has deteriorated due to long-term use. The scenario is set as a fuel injector in a high-pressure common rail diesel engine system that has run for the equivalent of 100,000 kilometers. Due to incomplete combustion and fuel impurities, a certain degree of carbon deposits has formed on the injector nozzle head. These carbon deposits reduce the effective flow area of ​​the nozzle orifice and may alter the spray pattern, a typical problem that conventional control strategies based on static lookup tables cannot effectively address.

[0047] In this scenario, the specific implementation steps of the method of the present invention are as follows:

[0048] First, during engine operation, the control system, based on the factory-calibrated control chart, sends a standard drive pulse command to the aging injector, intending to inject 25 mg of fuel. However, due to reduced actual flow area caused by carbon buildup in the nozzle, the actual amount of fuel injected is less than expected, for example, only 22 mg. This deviation results in a leaner air-fuel mixture in the cylinder, and the changes in the combustion process are detected by the in-cylinder pressure sensor and the wide-range oxygen sensor at the exhaust manifold.

[0049] Subsequently, the state synchronization module (step S200) in the method of this invention begins to function. This module continuously compares the actual measured values ​​from the sensors (e.g., the calculated peak combustion heat release rate is lower than expected) with the predicted output values ​​of the digital twin model under the same input command. Initially, because the effective flow area (A_eff) parameter of the nozzle inside the digital twin model is still the ideal value at the factory, its predicted combustion process deviates significantly from the actual measurement. The state estimation algorithm (e.g., unscented Kalman filtering) uses this deviation as a correction signal to update the state vector of the digital twin model online. In this process, the algorithm identifies the effective flow area (A_eff) of the nozzle as the most likely root cause of the prediction error and iteratively corrects it, gradually reducing its estimated value from the initial 100% to, for example, 90% that matches the current physical injector state. (Refer to...) Figure 6 The figure illustrates the effect of the method of the present invention on long-term online tracking of the effective flow area A_eff of the nozzle. The online estimate of UKF in the figure closely follows the actual value range that continuously declines due to aging. Thus, the digital twin model has achieved precise synchronization with the physical state of the aging injector.

[0050] Next, in the next injection cycle, when the engine control unit again requests the target injection quantity of 25 mg, the system enters the performance prediction and ideal trajectory planning step (step S300). The digital twin model that has completed state synchronization is invoked, which can accurately predict that if the original standard drive pulse is continued to be used, only 22 mg of injection quantity will be obtained. To compensate for this deviation, the system plans a new ideal performance trajectory based on the synchronized twin model. The goal of this trajectory is to accurately achieve an injection quantity of 25 mg, and may include optimization requirements for the injection rate pattern to improve atomization quality deteriorated by carbon deposits.

[0051] Subsequently, the system runs the control parameters and solves the neural network model in reverse (step S400). Using the aforementioned ideal performance trajectory as input to a hybrid density network (MDN), the MDN model, based on its understanding of the current injector characteristics (i.e., A_eff is 90%), calculates the control commands required to achieve the target. The output of the MDN may be a probability distribution indicating that a longer drive pulse width than the original factory command (e.g., extending the pulse width from 1200 microseconds to 1310 microseconds) is the optimal solution to achieve the target.

[0052] Finally, after performing the risk assessment (step S500), the calibrated new drive pulse command (1310 microseconds) is sent to the injector for execution. Under this new command, the physical injector actually injected approximately 25 milligrams of fuel, the combustion process returned to normal, and feedback from the cylinder pressure and oxygen sensor signals verified the success of this self-calibration. The entire process forms a closed loop, and the successful calibration data (i.e., 1310 microsecond pulse width corresponding to 25 milligrams of fuel injection under the condition of A_eff of 90%) is recorded by the system for future continuous online optimization of the model, enabling the system to adapt to further performance evolution of the injector.

[0053] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for self-calibrating fuel injector performance by integrating digital twins and neural networks, characterized in that, Includes the following steps: Step 1: Build and run a physics-data hybrid driven digital twin model of the fuel injector. The digital twin model includes a simplified physical model and a residual neural network model. The output of the simplified physical model and the output of the residual neural network model are fused to generate a predicted output of the physical injector performance. Step 2: Achieve real-time synchronization between the digital twin model and the physical injector's state. This step specifically involves: collecting sensor data from the physical injector system associated with the physical injector; comparing the actual measured values ​​of the sensor data with the predicted output values ​​generated by the digital twin model; and using an unscented Kalman filter algorithm to estimate and correct online a set of slowly varying physical parameters within the digital twin model that characterize the aging state of the physical injector due to wear or carbon buildup, based on the deviation generated by the comparison. Step 3: Based on the digital twin model that has completed state synchronization, predict the injection performance corresponding to different combinations of control parameters under the current working condition and aging condition, and plan an ideal performance trajectory accordingly. Step 4: Run the control parameter inverse solution neural network model, taking the ideal performance trajectory as input, and inversely solve to generate calibration instructions that can achieve the ideal performance trajectory; the control parameter inverse solution neural network model is a hybrid density network model; Step 5: Execute the calibration command and output data based on the actual performance of the physical injector to perform online optimization of the residual neural network model and the control parameter inverse solution neural network model; Step five, before executing the calibration command, also includes a risk pre-assessment step, which specifically includes: From the probability distribution of the control parameter solutions output by the hybrid density network model, a preset number of high-probability candidate calibration commands are selected; Using the digital twin model that has completed state synchronization in step two, the performance stability of each candidate calibration command is simulated and predicted under the condition of injected operating disturbance, and its control robustness score is calculated. For each candidate calibration command, its probability output by the hybrid density network model is weighted and balanced with the control robustness score, and the risk-adjusted optimal control command is selected and executed accordingly.

2. The method according to claim 1, characterized in that, In step one, the simplified physical model includes: an electromagnetic coil RLC circuit model for describing the relationship between driving current and electromagnetic force, a second-order dynamic model of the needle valve for describing the dynamic response of the needle valve lift and the electromagnetic force, and an orifice flow formula for describing the relationship between the needle valve lift, fuel pressure, and injection orifice flow. The input of the residual neural network model includes the real-time operating parameters of the injector and the internal state variables of the simplified physical model, and the output is the residual term between the prediction result of the simplified physical model and the actual performance of the physical entity.

3. The method according to claim 1, characterized in that, The hybrid density network model is configured to receive the ideal performance trajectory as input and output a set of parameters of a Gaussian mixture model. The parameters of the Gaussian mixture model include the mean vector, covariance matrix and mixing weight coefficients of each Gaussian component. These parameters together describe the probability distribution of all possible control parameter solutions for a given ideal performance trajectory, where the mean vector of each Gaussian component represents a candidate calibration command.

4. The method according to claim 3, characterized in that, The training data for the hybrid density network model comes from the control parameter-performance data pairs generated by conducting large-scale virtual simulation experiments using the digital twin model described in step three.

5. The method according to claim 1, characterized in that, In step five, the online optimization includes: using the verified control command-operating condition-actual performance output data triplet accumulated during long-term operation as incremental learning samples, and retraining the residual neural network model and the control parameter inverse solution neural network model online under preset triggering conditions.

6. The method according to claim 1, characterized in that, In step two, the slowly varying physical parameters include at least one of the following: effective flow area of ​​the nozzle, stiffness of the needle valve return spring, and dynamic response delay time of the needle valve.

7. The method according to claim 1, characterized in that, In step three, the ideal performance trajectory is an ideal fuel injection rate curve, which is planned according to the optimization criteria of combustion and emission control.

8. The method according to claim 1, characterized in that, In step two, the sensor data includes: the drive coil current waveform and fuel common rail pressure sensor signal as direct measurements, and the in-cylinder pressure sensor signal or wide-range oxygen sensor signal as indirect measurements.

9. A fuel injector performance self-calibration system integrating digital twins and neural networks, for performing the method as described in any one of claims 1-8, wherein the system is configured to perform performance self-calibration on a physical injector, characterized in that, include: The data acquisition module is configured to collect a set of sensor data from the physical injector-related system. The digital twin module is configured to run a physical-data hybrid-driven digital twin model of the fuel injector to generate a predictive output of the physical injector's performance and to plan an ideal performance trajectory. The state synchronization module is configured to receive the sensor data and the prediction output, and use an unscented Kalman filter algorithm to correct a set of slowly varying physical parameters characterizing the aging state of the physical injector in the digital twin model online according to the deviation between the two, so as to achieve state synchronization between the model and the physical injector. The reverse calibration module is configured to run a neural network model for reverse solving of control parameters, using the ideal performance trajectory as input, and to generate calibration instructions through reverse solving. A control execution module is configured to execute the calibration instructions to control the spraying action of the physical injector.

Images