An engine electronic control method based on an AI agent

Abstract

The application discloses an engine electronic control method based on an AI agent, and relates to the field of engine electronic control.The method comprises the following steps: S1, based on bottom support, collecting engine state signals, receiving vehicle instructions, collecting deviation data, and executing control instructions; the bottom support comprises a heterogeneous multi-core SoC hardware architecture; S2, based on the heterogeneous multi-core SoC hardware architecture, at least three physically isolated computing domains are divided, and the following is executed respectively: reaction agent logic, planning agent logic and learning agent logic; S6, further comprising deploying a shared digital twin engine: a digital twin model based on an integrated physical mechanism model and a neural network model, which simulates the running state and performance parameters of the engine in real time, and provides a prediction basis for the planning agent logic and an update object for the learning agent logic.

Description

Technical Field

[0001] This application relates to the field of engine electronic control, and in particular to engine electronic control methods based on AI agents, engine electronic control units (ECUs) based on AI agents, electronic devices, storage media, and vehicle development platforms. Background Technology

[0002] Currently, the mainstream development process for engine electronic control units (ECUs) is based on a model-based approach. The core reliance is on preset parameters, PID algorithms, and MAP lookup table methods to achieve open-loop / closed-loop control. This approach mainly falls into two technical categories:

[0003] Conventional real-time control ECUs only have basic sensor signal acquisition, filtering, analysis, and rule-based execution functions. They process sensor signals such as speed, temperature, pressure, knock, and oxygen concentration through a pre-calibrated MAP diagram on the bench and control logic. Based on fixed conditions, they execute emergency actions such as engine shutdown, torque limiting, and fuel cut-off. They have no predictive, planning, or self-optimization capabilities, and their control strategies rely entirely on prior offline calibration.

[0004] Preliminary intelligent ECUs: Some are equipped with lightweight machine learning models or simplified mathematical models, which can achieve dynamic adaptive adjustment of a small number of parameters; digital twin models are mostly deployed in the cloud, back-end servers or external domain controllers, and cannot be embedded in the ECU for real-time interaction; a few ECUs with learning capabilities directly modify control output commands through learning algorithms, which does not comply with automotive functional safety specifications.

[0005] Therefore, the existing technology has the following problems:

[0006] Real-time emergency response is completely disconnected from forward-looking planning, resulting in severe control lag: It can only passively respond to real-time sensor signals, lacks the ability to predict operating conditions and provide risk warnings, and is prone to control delays under transient operating conditions, leading to problems such as knocking, oil dilution, slow power response, and instantaneous emissions exceeding standards.

[0007] The digital twin is disconnected from the ECU local control and lacks real-time optimization capabilities: The digital twin model is deployed outside the ECU, resulting in communication delays and data transmission losses, making it impossible to simulate and predict the control effect based on the real-time needs of the whole vehicle; the control strategy relies on offline calibration, making it difficult to adapt to complex road conditions and changing driving habits, and the control accuracy and adaptability to operating conditions are insufficient.

[0008] Intelligent learning conflicts with functional safety, posing potential driving safety hazards: ECUs with learning functions can directly modify control outputs, and learning deviations and algorithm anomalies can lead to control command errors that do not meet functional safety standards; AI tasks and real-time control tasks are not physically isolated, and memory overflows and infinite loops can directly interrupt actuator control, causing safety accidents.

[0009] The control strategy is rigid and cannot adapt to the stringent emission regulations: Traditional ECUs have no online model correction capability. After long-term operation, the model deviation increases, and emissions of particulate matter, hydrocarbons and other substances are prone to exceed the limits, making it difficult to meet the requirements of China VI B and Euro VII regulations. Moreover, it is impossible to balance power performance, fuel economy and emission compliance.

[0010] Calibration involves a large workload, long development cycle, and high cost: As engine technology becomes more complex, the number of MAP diagrams and control parameters that need to be calibrated increases exponentially. Bench tests and vehicle calibration are time-consuming and labor-intensive, resulting in low product iteration efficiency.

[0011] Insufficient fault degradation capability and low system reliability: The control unit has no independent safety domain and there is no perfect degradation mechanism after the core module fails, which can easily lead to engine loss of control and stalling, affecting the overall vehicle operation safety.

[0012] Therefore, an engine electronic control strategy based on an AI agent is designed to improve and optimize the electronic control unit (ECU), achieving a closed-loop process of real-time emergency response, forward-looking planning, and online correction, thus overcoming the shortcomings of passive response and lack of prediction. A digital twin model is embedded locally into the ECU to solve the problem of the disconnect between the twin model and the control unit, enabling real-time optimization of the control strategy. An isolated intelligent learning mechanism is constructed to avoid the safety risks of the learning algorithm directly modifying the control output, meeting functional safety requirements. Control accuracy and adaptability are improved, reducing the risk of exceeding emission standards, balancing the three objectives of power, fuel consumption, and emissions. A heterogeneous multi-core physically isolated architecture is adopted to solve the architectural bottleneck of resource conflicts and mutual interference between AI tasks and safety control tasks. Online self-correction of the model is achieved, reducing offline calibration workload, shortening the development cycle, and lowering calibration costs. A fault degradation mechanism is improved to enhance the reliability of the ECU under extreme operating conditions and module failures. Summary of the Invention

[0013] The purpose of this invention is to provide an engine electronic control method based on an AI agent, an engine electronic control unit (ECU) based on an AI agent, an electronic device, a storage medium, and a vehicle development platform, thereby solving at least one of a number of technical problems.

[0014] Key technical issues: Real-time emergency response and forward-looking planning are completely disconnected, resulting in severe control lag. Digital twins and local ECU control are disconnected, lacking real-time optimization capabilities. Intelligent learning conflicts with functional safety, posing potential driving safety hazards. Fixed control strategies cannot adapt to stringent emission regulations. Calibration workload is large, development cycle is long, and costs are high. Insufficient fault degradation capabilities lead to low system reliability.

[0015] This invention provides the following solution:

[0016] According to a first aspect of the present invention, an engine electronic control method based on an AI agent is provided, comprising:

[0017] Based on the underlying support, it collects engine status signals, receives vehicle commands, collects deviation data, and executes control commands.

[0018] The underlying support includes a heterogeneous multi-core SoC hardware architecture;

[0019] Based on a heterogeneous multi-core SoC hardware architecture, at least three physically isolated computing domains are divided to execute: reactive agent logic, planning agent logic, and learning agent logic, respectively.

[0020] The reactive agent logic runs on a lockstep computation core with a preset high security level. It is used to collect engine status signals, detect emergency conditions, and generate real-time emergency control commands at intervals less than or equal to a preset first sampling period threshold.

[0021] The planning intelligent agent logic runs on a high-performance application core with a preset functional safety level. It is used to receive instructions from the vehicle's high-level management, call the local digital twin engine to simulate multiple candidate control strategies, and generate the optimal control parameters through multi-objective optimization.

[0022] The system learns the logic of the intelligent agent and runs on an isolated coprocessor at a preset quality management level. It is used to collect control deviation data in the background and incrementally update the parameters of the digital twin model.

[0023] It also includes deploying a shared digital twin engine: a digital twin model based on an integrated physical mechanism model and a neural network model, which simulates the engine's operating status and performance parameters in real time, providing a predictive basis for planning the intelligent agent's logic and providing an update object for learning the intelligent agent's logic.

[0024] Furthermore, the underlying support also includes:

[0025] The system acquires engine status signals through the sensor interface unit; receives vehicle commands through the communication interface unit; acquires deviation data through the data acquisition and buffering unit; executes control commands through the safety execution unit; and provides hardware and software support through the basic support unit.

[0026] According to a second aspect of the present invention, an engine electronic control unit (ECU) based on an AI agent is provided, comprising:

[0027] The sensor interface layer, communication interface layer, data acquisition and caching module, security execution layer, and basic support platform participate in building the underlying support;

[0028] Based on the underlying support, the ECU adopts a heterogeneous multi-core SoC hardware architecture, dividing the computing domains into at least three physically isolated domains, which are deployed separately:

[0029] The reactive agent runs on a lockstep computation kernel with a preset high security level. It is used to collect engine status signals at intervals less than or equal to a preset first sampling period threshold, detect emergency conditions, and generate real-time emergency control commands. The reactive agent does not receive any input from the learning agent.

[0030] The planning intelligent agent runs on a high-performance application core with a preset functional safety level. It is used to receive instructions from the vehicle's high-level management, call the local digital twin engine to simulate multiple candidate control strategies, and generate the optimal control parameters through multi-objective optimization.

[0031] The learning agent runs on an isolated coprocessor at a preset quality management level. It is used to collect control deviation data in the background and only incrementally updates the parameters of the digital twin model without directly interfering with the engine's control output commands.

[0032] It also includes a shared digital twin engine deployed in the ECU's local memory. The engine integrates a physical mechanism model and a neural network model to simulate the engine's operating status and performance parameters in real time, providing a basis for prediction for the planning agent and an update object for the learning agent.

[0033] Furthermore, including:

[0034] Inter-core communication between planning agents and reactive agents is achieved using structured instruction data frames;

[0035] The data frame includes:

[0036] The control parameter field is used to store the engine's execution control parameters;

[0037] The confidence field is used to mark the degree of confidence of the planning instruction, and the agent will only adopt the instruction when the confidence level is greater than a preset confidence threshold.

[0038] The validity period field defines the preset validity period window for the instruction; instructions that exceed this window will be automatically discarded.

[0039] The security verification field is used to store the verification code and hardware signature to ensure data integrity.

[0040] Furthermore, including:

[0041] The neural network of the shared digital twin engine has an input layer feature vector including crankshaft angle time series features, and an output layer prediction vector including oil dilution rate prediction and transient emission prediction.

[0042] Furthermore, it also includes fault degradation control logic:

[0043] When the planning agent fails to respond beyond the preset response time threshold, the reaction agent automatically switches to the preset conservative control mapping table and smoothly reduces the target torque according to the preset torque adjustment slope to prevent power interruption shock.

[0044] Furthermore, it also includes the online learning triggering logic of the learning agent:

[0045] Based on the vehicle's power status and the ECU's real-time calculated load rate, the sampling period of the learning agent is dynamically adjusted. When the ECU's real-time calculated load exceeds the preset load threshold, the sampling frequency is reduced using a preset automatic frequency conversion stepping strategy.

[0046] When the real-time computing load of the ECU is below a preset load threshold, the sampling frequency is restored using a preset automatic frequency conversion step strategy.

[0047] Furthermore, the reactive agent is also equipped with an emergency takeover unit:

[0048] The emergency takeover unit is used to suspend the output of the planning agent by notifying it of an inter-core interruption when a preset dangerous condition is detected, and to prioritize the execution of emergency control commands.

[0049] Furthermore, it also includes the model update process for learning intelligent agents:

[0050] The simulation verification of the new parameters is completed by calling the digital twin copy in isolated memory;

[0051] Once verification is successful, the securely signed incremental update package will be written to the area of ​​the model to be activated, without directly modifying the running digital twin model.

[0052] Furthermore, the ECU's shared memory is divided into at least three independent partitions protected by a memory protection unit (MPU):

[0053] Independent partitions are used for the reactive agent to write sensor data, the planning agent to write control instructions, and the learning agent to write model update packages.

[0054] According to a third aspect of the present invention, an electronic device is provided, comprising: a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other via the communication bus;

[0055] The memory stores a computer program, which, when executed by a processor, causes the processor to perform steps such as those in an AI-based engine electronic control method.

[0056] According to a fourth aspect of the present invention, a computer-readable storage medium is provided, comprising: storing a computer program executable by an electronic device, wherein when the computer program is run on the electronic device, the electronic device causes the electronic device to perform steps such as an engine electronic control method based on an AI agent.

[0057] According to a fifth aspect of the present invention, a vehicle development platform is provided, comprising:

[0058] Electronic devices for implementing steps such as engine electronic control methods based on AI agents;

[0059] The processor runs programs, and when the programs are running, they execute steps such as engine electronic control methods based on AI agents, based on data output from electronic devices.

[0060] Storage medium for storing programs that, when running, execute steps such as an AI-based engine electronic control method based on data output from an electronic device.

[0061] The above solution achieves the following beneficial technical effects:

[0062] This application features a response agent that can handle emergency situations in milliseconds, with an independent safety core ensuring ASIL-D level protection. It executes actions such as torque limiting and engine shutdown immediately, preventing equipment damage caused by knocking or overheating, and promoting a faster and more reliable safety response.

[0063] This application plans an intelligent agent that relies on local digital twin real-time simulation to optimize the output of optimal control parameters through multi-objective optimization, improves adaptability under complex working conditions, significantly reduces emissions of pollutants such as NOx, and makes the control strategy more precise and adaptive.

[0064] This application's learning agent only fine-tunes model parameters / reward functions without directly interfering with control output, fully complying with ISO26262; heterogeneous multi-core physical isolation avoids AI tasks interfering with safety control, achieving perfect compatibility between intelligence and functional safety.

[0065] This application constructs a "prediction-execution-correction" closed loop to continuously correct model deviations without the need for frequent offline calibration. Over long-term operation, the control accuracy is continuously improved, thus forming a fully closed-loop self-optimizing system.

[0066] Online optimization and precise control meet China VI B and Euro VII emission requirements, and are suitable for operation in all scenarios such as mountainous areas, low temperatures, congestion, and high loads, thus easily adapting to stringent regulations and complex working conditions.

[0067] This application's self-learning capability reduces reliance on bench calibration, reduces MAP calibration workload by more than 50%, enables faster product iteration, and can significantly reduce calibration costs and development cycles.

[0068] This application features a robust degradation mode that automatically switches to basic MAP control when the AI ​​module fails, ensuring stable engine operation and improving system reliability and fault tolerance. Attached Figure Description

[0069] Figure 1 This is a schematic diagram of an engine electronic control unit (ECU) system architecture based on an AI agent, provided in a specific embodiment of the present invention.

[0070] Figure 2 This is a schematic diagram of the internal structure of a reactive intelligent agent provided in a specific embodiment of the present invention.

[0071] Figure 3 This is a schematic diagram of the internal structure of a planning agent provided in a specific embodiment of the present invention.

[0072] Figure 4 This is a schematic diagram of the internal structure of a learning agent provided in a specific embodiment of the present invention.

[0073] Figure 5 This is a schematic diagram of a local digital twin engine structure provided in a specific embodiment of the present invention.

[0074] Figure 6 This is a flowchart of an engine electronic control method based on an AI agent provided by one or more embodiments of the present invention.

[0075] Figure 7 This is a structural diagram of an engine electronic control unit (ECU) based on an AI agent, provided by one or more embodiments of the present invention.

[0076] Figure 8 This is a block diagram of an electronic device structure for an engine electronic control method based on an AI agent, provided in one or more embodiments of the present invention. Detailed Implementation

[0077] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0078] Figure 6 This is a flowchart of an engine electronic control method based on an AI agent provided by one or more embodiments of the present invention.

[0079] like Figure 6 The engine electronic control method based on AI agents shown includes:

[0080] Step S1: Based on the underlying support, collect engine status signals, receive vehicle commands, collect deviation data, and execute control commands.

[0081] The underlying support includes a heterogeneous multi-core SoC hardware architecture;

[0082] Step S2: Based on the heterogeneous multi-core SoC hardware architecture, at least three physically isolated computing domains are divided to execute: reactive agent logic, planning agent logic, and learning agent logic, respectively.

[0083] Step S3: The reactive agent logic runs in a lockstep computation kernel with a preset high security level. It is used to collect engine status signals, detect emergency conditions, and generate real-time emergency control commands at intervals less than or equal to the preset first sampling period threshold.

[0084] Step S4: Plan the intelligent agent logic, run on a high-performance application core with a preset functional safety level, to receive instructions from the vehicle's high-level management, call the local digital twin engine to simulate multiple candidate control strategies, and generate the optimal control parameters through multi-objective optimization.

[0085] Step S5: Learn the agent logic and run it in an isolated coprocessor at a preset quality management level. This coprocessor is used to collect control deviation data in the background and incrementally update the parameters of the digital twin model.

[0086] Step S6 also includes deploying a shared digital twin engine: a digital twin model based on an integrated physical mechanism model and a neural network model, which simulates the engine's operating status and performance parameters in real time, providing a predictive basis for planning the intelligent agent's logic and providing an update object for learning the intelligent agent's logic.

[0087] In this embodiment, the underlying support also includes:

[0088] The system acquires engine status signals through the sensor interface unit; receives vehicle commands through the communication interface unit; acquires deviation data through the data acquisition and buffering unit; executes control commands through the safety execution unit; and provides hardware and software support through the basic support unit.

[0089] Figure 7 This is a structural diagram of an engine electronic control unit (ECU) based on an AI agent, provided by one or more embodiments of the present invention.

[0090] like Figure 7 The AI-based engine electronic control unit (ECU) shown includes:

[0091] The sensor interface layer (sensor interface unit), communication interface layer (communication interface unit), data acquisition and caching module (data acquisition and caching unit), security execution layer (security execution unit), and basic support platform (basic support unit) participate in building the underlying support;

[0092] Based on the underlying support, the ECU adopts a heterogeneous multi-core SoC hardware architecture, dividing the computing domains into at least three physically isolated domains, which are deployed separately:

[0093] The reactive agent runs on a lockstep computation kernel with a preset high security level. It is used to collect engine status signals at intervals less than or equal to a preset first sampling period threshold, detect emergency conditions, and generate real-time emergency control commands. The reactive agent does not receive any input from the learning agent.

[0094] The planning intelligent agent runs on a high-performance application core with a preset functional safety level. It is used to receive instructions from the vehicle's high-level management, call the local digital twin engine to simulate multiple candidate control strategies, and generate the optimal control parameters through multi-objective optimization.

[0095] The learning agent runs on an isolated coprocessor at a preset quality management level. It is used to collect control deviation data in the background and only incrementally updates the parameters of the digital twin model without directly interfering with the engine's control output commands.

[0096] It also includes a shared digital twin engine deployed in the ECU's local memory. The engine integrates a physical mechanism model and a neural network model to simulate the engine's operating status and performance parameters in real time, providing a basis for prediction for the planning agent and an update object for the learning agent.

[0097] In this embodiment, it includes:

[0098] Inter-core communication between planning agents and reactive agents is achieved using structured instruction data frames;

[0099] The data frame includes:

[0100] The control parameter field is used to store the engine's execution control parameters;

[0101] The confidence field is used to mark the degree of confidence of the planning instruction, and the agent will only adopt the instruction when the confidence level is greater than a preset confidence threshold.

[0102] The validity period field defines the preset validity period window for the instruction; instructions that exceed this window will be automatically discarded.

[0103] The security verification field is used to store the verification code and hardware signature to ensure data integrity.

[0104] In this embodiment, it includes:

[0105] The neural network of the shared digital twin engine has an input layer feature vector including crankshaft angle time series features, and an output layer prediction vector including oil dilution rate prediction and transient emission prediction.

[0106] In this embodiment, fault degradation control logic is also included:

[0107] When the planning agent fails to respond beyond the preset response time threshold, the reaction agent automatically switches to the preset conservative control mapping table and smoothly reduces the target torque according to the preset torque adjustment slope to prevent power interruption shock.

[0108] In this embodiment, online learning triggering logic for the learning agent is also included:

[0109] Based on the vehicle's power status and the ECU's real-time calculated load rate, the sampling period of the learning agent is dynamically adjusted. When the ECU's real-time calculated load exceeds the preset load threshold, the sampling frequency is reduced using a preset automatic frequency conversion stepping strategy.

[0110] When the real-time computing load of the ECU is below a preset load threshold, the sampling frequency is restored using a preset automatic frequency conversion step strategy.

[0111] In this embodiment, the reactive agent is also equipped with an emergency takeover unit:

[0112] The emergency takeover unit is used to suspend the output of the planning agent by notifying it of an inter-core interruption when a preset dangerous condition is detected, and to prioritize the execution of emergency control commands.

[0113] In this embodiment, a model update process for the learning agent is also included:

[0114] The simulation verification of the new parameters is completed by calling the digital twin copy in isolated memory;

[0115] Once verification is successful, the securely signed incremental update package will be written to the area of ​​the model to be activated, without directly modifying the running digital twin model.

[0116] In this embodiment, the shared memory of the ECU is further divided into at least three independent partitions protected by a memory protection unit (MPU):

[0117] Independent partitions are used for the reactive agent to write sensor data, the planning agent to write control instructions, and the learning agent to write model update packages.

[0118] Specifically, the engine electronic control unit (ECU) based on AI intelligent agents in this application includes: a sensor interface layer, a communication interface layer, a data acquisition and caching module, a safety execution layer, and a basic support platform;

[0119] The ECU adopts a heterogeneous multi-core SoC hardware architecture, dividing the computing domains into three physically isolated domains, which are deployed separately:

[0120] The reactive agent, running on an ASIL-D level lockstep computing core, is used to acquire engine status signals at a period of ≤1ms, detect emergency conditions, and generate real-time emergency control commands. The reactive agent does not receive any input from the learning agent.

[0121] The planning agent runs on a high-performance application core at ASIL-B level. It is used to receive instructions from the vehicle's high-level management, call the local digital twin engine to simulate multiple candidate control strategies, and generate the optimal control parameters through multi-objective optimization.

[0122] The learning agent runs on a QM-level isolated coprocessor to collect control deviation data in the background. It only incrementally updates the parameters of the digital twin model and does not directly modify the engine control output.

[0123] It also includes a shared digital twin engine deployed in the ECU's local memory. The engine integrates a physical mechanism model and a lightweight neural network to simulate the engine's combustion, thermodynamics, emissions, and oil dilution behavior in real time, providing predictive basis for planning agents and update objects for learning agents.

[0124] This also includes the use of structured instruction data frames for inter-core communication between the planning agent and the reaction agent. The data frames include:

[0125] The control parameter field is used to store the ignition advance angle, injection pulse width, and EGR valve opening parameters.

[0126] The confidence level is used to mark the degree of confidence of the planning instruction. The agent will only adopt the instruction when the confidence level is greater than a preset threshold.

[0127] The validity period field defines the valid time window for the instruction. Instructions that have not been refreshed before the timeout will be automatically discarded.

[0128] The security verification field is used to store the CRC checksum and hardware signature to ensure data integrity.

[0129] It also includes a lightweight neural network that shares a digital twin engine, with input layer feature vectors containing crankshaft angle time series features and output layer prediction vectors containing predicted values ​​for oil dilution rate, HC emissions, and NOx emissions.

[0130] It also includes fault degradation control logic: when the planning agent fails to respond within a timeout period, the reaction agent automatically switches to the preset conservative MAP and smoothly reduces the target torque according to the preset slope to prevent power interruption impact.

[0131] It also includes the online learning triggering logic of the learning agent, which includes: dynamically adjusting the sampling period of the learning agent according to the power status of the vehicle and the real-time calculated load rate of the ECU, and automatically reducing the sampling frequency when the load is too high.

[0132] It also includes an emergency takeover port, which, when a dangerous condition is detected, notifies the planning agent via an inter-core interrupt to suspend policy output and prioritize the execution of emergency control commands.

[0133] It also includes the model update process for learning intelligent agents, which includes: calling the digital twin copy in isolated memory to complete the simulation verification of new parameters; after the verification is passed, writing the incremental update package signed by the hardware security module into the model area to be activated, without directly modifying the running digital twin model.

[0134] It also includes that the ECU's shared memory is divided into three independent partitions protected by the MPU: Zone_A is used for responding agents to write sensor data, Zone_B is used for planning agents to write control instructions, and Zone_C is used for learning agents to write model update packages.

[0135] Furthermore, it also includes:

[0136] The learning agent also supports federated learning and cloud collaboration, uploading local incremental model updates to the cloud platform through an encrypted and secure channel, and receiving global model updates aggregated from the cloud to achieve collaborative model optimization among multiple vehicles.

[0137] It also includes a diagnostic agent running in an independent computing domain, used for fault diagnosis and predictive maintenance based on full operational data. The diagnostic agent is physically isolated from other agents and does not interfere with the core control process.

[0138] The digital twin engine also integrates a simulation model of the hybrid powertrain system, adapts to the control optimization of hybrid / range-extended engines, and achieves multi-objective collaborative optimization of the powertrain system.

[0139] The digital twin simulation of the planning intelligent agent is adapted to automotive-grade neural network processors (NPUs), supporting parallel simulation of larger-scale digital twin models and adapting to iterative upgrades in hardware computing power.

[0140] The physically isolated computing domain can be replaced by a hardware architecture with two independent MCUs: a security MCU deploys reactive agents, and a high-performance MCU deploys planning agents and learning agents, implementing the same hierarchical isolation logic.

[0141] The three-layer intelligent agent architecture can be integrated into the vehicle domain controller as a sub-controller of the domain controller, physically isolated from other domain control units, to achieve cross-controller architecture reuse.

[0142] Furthermore, it may also include:

[0143] The independent partition is configured with concurrent access locks and a priority scheduling mechanism. For access requests from different agents, scheduling is performed according to the priority of the reactive agent > the planning agent > the learning agent, thus solving the resource contention problem of shared memory.

[0144] The learning agent is also equipped with a model rollback mechanism. When new parameters are updated and the prediction deviation of the digital twin model is detected to exceed the preset deviation threshold, the digital twin model is automatically rolled back to the previous stable version to avoid control deviation caused by model update errors.

[0145] The safety detection logic of the reactive agent also includes a multi-parameter combination risk detection unit, which matches a preset combination risk threshold to the combined state of multiple control parameters and identifies complex dangerous conditions that cannot be detected by a single parameter threshold.

[0146] The learning agent is also equipped with an aging deviation correction unit, which is used to identify the aging deviation of engine parts during long-term operation, and to perform offset correction on the basic parameters of the digital twin model according to the aging state, so as to adapt to the long-term aging characteristics of the engine.

[0147] The shared digital twin engine is equipped with an access priority scheduling mechanism. The simulation call priority of the planning agent is higher than the model update call priority of the learning agent, which ensures the real-time performance of the planning agent and avoids simulation task timeouts.

[0148] It is worth noting that although this system / device only discloses the above-mentioned modules / units, it does not mean that this system / device is limited to the above-mentioned basic functional modules. On the contrary, what this invention intends to express is that, based on the above-mentioned basic functional modules, those skilled in the art can add one or more functional modules in combination with the prior art to form an infinite number of embodiments or technical solutions. That is to say, this system is open rather than closed. It cannot be assumed that the scope of protection of the claims of this invention is limited to the above-disclosed basic functional modules just because this embodiment only discloses a few basic functional modules.

[0149] In one specific embodiment, an engine electronic control strategy based on AI intelligent agents is disclosed, which adopts an ECU architecture of "software and hardware collaboration + three-layer AI intelligent agents + local digital twin + heterogeneous multi-core physical isolation". The hardware is based on heterogeneous multi-core SoC to achieve safety domain isolation, and the software is divided into three layers: reaction intelligent agent, planning intelligent agent and learning intelligent agent. Together with the local digital twin engine, safety execution layer, communication layer and hardware support layer, it realizes closed-loop control of the whole process.

[0150] This embodiment includes, such as Figure 1As shown, the AI ​​intelligent agent ECU system includes: a sensor interface layer, a communication interface layer, a data acquisition and caching module, a reactive intelligent agent, a planning intelligent agent, a learning intelligent agent, a shared digital twin engine, a safe execution layer, and a basic support platform.

[0151] in:

[0152] The sensor interface layer is used to collect engine operating status signals in real time, including but not limited to crankshaft speed, knock signal, engine oil temperature, intake manifold pressure (MAP), and oxygen sensor signal; its output is connected to the reactive agent.

[0153] The communication interface layer supports CAN FD and automotive Ethernet protocols to receive high-level commands such as torque requests and driving modes from the vehicle controller (VCU) or transmission controller (TCU); its output is connected to the planning agent.

[0154] The data acquisition and caching module is used to record the deviation data between the actual output parameters of the engine (such as emissions, fuel consumption, and vibration) and the digital twin predicted values, and to label them with operating conditions; its output is connected to the learning agent.

[0155] The reactive agent is a low-latency, rule-driven agent configured in a high-security-level computing core (ASIL-D) to handle emergency events and generate real-time control commands; its output is connected to the secure execution layer.

[0156] The planning agent is a multi-objective optimization agent, configured in the non-safety core (ASIL-B), and calls the shared digital twin engine to jointly evaluate fuel economy, emission compliance, NVH performance and oil dilution risk to generate the optimal control target; its output is also connected to the safety execution layer.

[0157] The learning agent is an online learning agent that runs in an isolated security domain. Based on feedback data provided by the data acquisition and caching module, it incrementally updates the model parameters or reward function of the shared digital twin engine; it is bidirectionally connected to the shared digital twin engine.

[0158] The shared digital twin engine is deployed in the ECU's local memory, integrating a physical mechanism model and a lightweight neural network for real-time simulation of engine combustion processes, thermodynamic behavior, emission generation, and oil dilution evolution; its inputs are connected to the planning agent and the learning agent, respectively, and its outputs are fed back to the planning agent.

[0159] The safety execution layer includes actuator drive circuits and a safety monitoring unit, which are used to drive actuators such as fuel injectors, VVT phasers, EGR valves, and ignition coils, and have hardware watchdog and voltage / temperature over-limit protection functions; it receives final control commands from reactive agents or planning agents.

[0160] The basic support platform provides underlying software and hardware support (the connotation of the basic support platform in the current embodiment is limited) including but not limited to multi-core microcontrollers (such as AURIX TC3xx series), AUTOSAR compatible real-time operating systems, secure boot mechanisms, encrypted OTA firmware update modules, and non-volatile memory for storing agent policies, digital twin models, and calibration parameters.

[0161] also, Figure 1 It also includes, showing the external connectivity relationships: the ECU communicates bidirectionally with the vehicle network through the communication interface layer; the learning agent can interact with the cloud platform / calibration bench for model updates through a secure channel.

[0162] This embodiment also includes: a hardware architecture based on physical isolation design of heterogeneous multi-core: relying on automotive-grade heterogeneous multi-core microcontroller, physical isolation of intelligent agents is achieved through hardware resource partitioning.

[0163] Reactive Agent: Runs on the Lockstep Core (CPU0+1) and configured at ASIL-D level. It has exclusive access to the PFLASH code area, ECC-protected DSCRATCH tightly coupled memory, a dedicated ADC sampling channel, and a high-speed PWM output module. Hardware watchdog and program flow monitoring (PFL) are enabled to ensure deterministic code execution flow.

[0164] Deliberative Agent: Runs on the high-performance application core (CPU2) and is configured to ASIL-B level.

[0165] Connect external DDR / LPBDDR memory to load large digital twin models and enable hardware neural network accelerators (HWA) or floating-point units (FPU) for parallel computing.

[0166] Learning Agent: Runs on a low-power coprocessor or independent security domain (CPU3), configured at QM (Quality Management) level. It has exclusive access to a large-capacity Data Flash for storing historical runtime logs and incremental package of model parameters, accessing the shared data area via a DMA controller.

[0167] This embodiment also includes: software layering logic:

[0168] (1) Reactive agent

[0169] Functional positioning: The underlying safety execution layer is responsible for millisecond-level real-time control and emergency response to ensure the basic safety of the engine.

[0170] Deployment environment: Independent ASIL-D lockstep core, hardware ECC tightly coupled memory, dedicated sensor sampling channel, and hardware firewall isolation.

[0171] Core components: signal acquisition and filtering module, security rule base, fast decision-making module, and security execution driver module.

[0172] Workflow: Acquire sensor signals such as speed, knock, oil temperature, water pressure, and oxygen concentration with a sampling period of ≤1ms; filter, verify, and detect anomalies in the signals; match the rule base, and synchronize the data to the shared memory if no anomalies are found; if dangerous conditions such as knock, over-temperature, or over-speed are detected, immediately execute actions such as torque limiting, engine shutdown, and fuel cut-off, with a response time of ≤5ms.

[0173] like Figure 2 As shown, the reactive agent is deployed on a lockstep computation core that meets ISO 26262 ASIL-D standards and includes the following functional modules:

[0174] Real-time sensor input interface: Receives millisecond-level signals from the engine, including knock intensity, crankshaft speed, oil pressure, and coolant temperature;

[0175] Safety event detector: Based on preset thresholds and logical rules (e.g., "If the detonation signal > threshold T1 and the duration > 2ms, then trigger detonation protection").

[0176] Rule base: Stores hard-coded control policies that have been certified for functional safety, including flameout protection, overheat torque limiting, and low oil pressure response;

[0177] Lightweight decision tree engine: Runs a decision tree model (<10KB) generated by bench calibration for rapid parameter generation in non-emergency situations;

[0178] Instruction Arbitration and Verification Module: Performs priority arbitration on the rule base output and decision tree output, and adds timestamps, validity periods and CRC check codes;

[0179] Safe execution output interface: The final control instruction is written to the shared memory area A and can be directly driven by hardware signals to drive the actuator drive circuit.

[0180] Emergency takeover signal output port: Sends an inter-core interrupt (IPI) to the planning agent, notifying it to suspend policy output.

[0181] All modules operate within an ASIL-D memory domain protected by the MPU, and do not receive any input from the learning agent, ensuring the determinism and isolation of the safety-critical path.

[0182] (2) Planning intelligent agents

[0183] Functional positioning: Mid-level decision-making core, responsible for analyzing vehicle requirements, forward-looking planning of control strategies, and multi-objective optimization.

[0184] Deployment environment: ASIL-B high-performance application core, equipped with hardware AI accelerator, floating-point operation unit, and external DDR memory.

[0185] Core components: vehicle requirements analysis module, digital twin engine, multi-objective optimization algorithm, and instruction output module.

[0186] Workflow: Receive power demand, driving mode, and operating condition commands from VCU / TCU / BMS; read real-time sensor data from shared memory; call the local digital twin model to simulate power, fuel consumption, emissions, oil dilution, and NVH parameters for multiple control strategies; construct the cost function. Solve for the optimal control trajectory that minimizes J; (Fuel is fuel consumption) Write the instructions to the shared memory instruction buffer and trigger a hardware doorbell interrupt to notify the responding agent.

[0187] like Figure 3 As shown, the planned intelligent agent is deployed on an independent application core that meets ASIL-B standards, and includes the following functional modules:

[0188] Vehicle demand input interface: Receives torque requests, driving modes, and regeneration commands sent by VCU / TCU through the communication interface layer;

[0189] Current operating condition perception module: integrates sensor interface layer data with status information fed back by the intelligent agent (such as actual torque and limiting reasons);

[0190] Target decomposer: Breaks down high-level requirements into multi-dimensional control targets, including target torque, allowable NOx emission limit, and maximum oil dilution tolerance;

[0191] Candidate strategy generator: Generates multiple sets of candidate control parameters (EGR rate, ignition advance angle, injection timing) based on calibrated MAP or heuristic algorithms.

[0192] Digital twin call interface: Initiate synchronous calls to the shared digital twin engine to obtain the prediction results (fuel consumption, emissions, dilution risk) for each group of candidate strategies.

[0193] Multi-objective optimizer: Constructs a weighted scoring function and selects the optimal policy;

[0194] Structured instruction encapsulation module: encapsulates the optimal parameters into a message with an expiration date (e.g., 5ms), confidence level (0~1), and version number;

[0195] Cooperative communication output interface: Writes instructions to shared memory region A and listens for emergency takeover signals from reactive agents.

[0196] (3) Learning agents;

[0197] Functional positioning: Core of upper-level optimization, responsible for data collection, model deviation correction, and online iterative optimization.

[0198] Deployment environment: QM low-power coprocessor, connected to non-volatile memory, running independently in the background.

[0199] Core components: Deviation acquisition module, incremental learning module, parameter correction module, security verification module, and OTA interface.

[0200] Workflow:

[0201] Monitor shared memory and data bus to collect digital twin predictions and actual engine operating values;

[0202] Calculate deviations in key parameters such as torque, emissions, and combustion efficiency;

[0203] Online gradient descent or transfer learning algorithms are used to update only the parameters of the sub-models with large deviations in the digital twin.

[0204] The updated model takes effect after security verification and supports OTA remote update strategy.

[0205] like Figure 4 As shown, the learning agent is deployed on a QM (or ASIL-A) level isolated computing core and includes the following functional modules:

[0206] Operation log input interface: Reads the structured operation log written by the planning agent from the shared memory area B, including operating condition labels, digital twin prediction values, and actual sensor feedback;

[0207] Cloud update package receiving interface: Receives encrypted and signed model incremental update packages via vehicle Ethernet;

[0208] Data preprocessing module: performs condition clustering (such as cold start, high-speed cruise, GPF regeneration), outlier removal, and feature normalization;

[0209] Sub-model identifier: Compares prediction biases to identify digital twin sub-models that need optimization (such as the "low-temperature fuel evaporation model").

[0210] Incremental learning engine: Employs online learning algorithms (such as stochastic gradient descent) to update only the parameters of significantly biased sub-models;

[0211] Simulation verification module: Calls the digital twin copy in isolated memory to verify whether the new model reduces prediction error without exceeding the safety boundary;

[0212] Secure Release Controller: Generates differential update packages and submits them to the Hardware Security Module (HSM) for signing;

[0213] Model writing interface: Writes the HSM-signed update package to the "model to be activated area" in Flash without directly modifying the running model.

[0214] This embodiment also includes: a local digital twin engine;

[0215] Deployment method: Local embedded deployment of ECU, model size <100MB, no cloud dependency.

[0216] Model architecture: A fusion of a physical mechanism model and a lightweight neural network (PINN), including sub-models for thermodynamics, combustion, emissions, oil dilution, and friction losses. (e.g.) Figure 5 (As shown)

[0217] Core capabilities: Real-time prediction of emissions, oil dilution rate, torque output, and fuel consumption rate; supports multi-strategy parallel simulation; inference time <20ms.

[0218] This embodiment also includes: an inter-core communication mechanism;

[0219] Shared memory: Divided into three MPU-protected partitions: Zone_A: Reactive agents write sensor data, other cores read-only; Zone_B: Planning agents write control instructions, reactive agents read; Zone_C: Learning agents write model parameter update packages.

[0220] Inter-core interrupt: triggered by a hardware doorbell, instruction delivery delay <10μs.

[0221] EDMA background migration: Data acquisition by learning agents does not occupy bus resources.

[0222] Hardware CRC + timestamp: ensures the integrity and timing determinism of communication data.

[0223] This embodiment also includes: a fault degradation mechanism;

[0224] Planning agent failure: If the reaction agent detects a timeout and fails to respond, it automatically switches to "limp home" mode (preset conservative MAP map).

[0225] Learning agent failure: Stop model updates and continue using the parameters of the current stable version.

[0226] In case of agent failure: the hardware watchdog triggers a system reset to ensure the engine operates at a minimum level.

[0227] This embodiment also includes: core data interaction structure and model definition;

[0228] The inter-core communication data frame structure includes: a specific structured instruction data frame is defined to ensure secure communication between the planning agent and the reaction agent;

[0229] The data frame contains the following fields:

[0230] Control parameter field: Includes specific execution values ​​such as ignition advance angle, injection pulse width, and EGR valve opening;

[0231] Confidence Level: A value between 0 and 1 output by the planning agent, indicating that the agent will only adopt the instruction when the confidence level is greater than a preset threshold (e.g., 0.7).

[0232] Time-to-Live (UTL): Defines the valid time window for instructions (e.g., 5ms). Instructions that are not refreshed within the time limit will be automatically discarded by the reactive agent to prevent deadlock.

[0233] The security verification field, including the CRC checksum and hardware signature information, is used to prevent data tampering.

[0234] The specific components of a digital twin model are as follows:

[0235] The lightweight neural network model in the shared digital twin engine has an input layer feature vector X that contains at least the following physical quantities: X = { N, MAP, T_intake, λ, VVT_pos, EGR_rate, t_crank}, where N: engine speed (rpm), a basic parameter reflecting the current operating state; MAP: intake manifold absolute pressure (kPa), which determines the amount of air entering the cylinder; T_intake: intake air temperature (°C), which affects air density and combustion efficiency; λ: air-fuel ratio, the ratio of the actual amount of air to the amount of air required for theoretical complete combustion; VVT_pos: variable valve timing position (°CA), which controls the opening timing of intake and exhaust valves; EGR_rate: exhaust gas recirculation rate (%), which affects combustion temperature and NOx formation; and t_crank: crankshaft angle time series, which captures transient combustion fluctuation characteristics.

[0236] Its output layer prediction vector Y is used to predict transient performance indicators: Y = { Torque_pred, BSFC_pred, HC_pred, NOx_pred, Dilution_pred, Noise_pred}, where Torque_pred: predicted torque (Nm) and BSFC_pred: predicted specific fuel consumption (g / kWh), used to monitor combustion integrity and adjust control parameters in real time; HC_pred: predicted hydrocarbon emissions (ppm) and NOx_pred: predicted nitrogen oxide emissions (ppm), used to assess combustion temperature control and emission levels; Dilution_pred: predicted oil dilution rate (%), used to prevent oil performance degradation; and Noise_pred: predicted combustion noise (dB), used to optimize driving comfort.

[0237] The input vector X is transformed into the output vector Y through a lightweight neural network model, achieving: mapping from physical parameters to performance indicators; capturing transient combustion fluctuation characteristics that are difficult to describe by traditional physical models; and providing decision-making basis for planning agents to optimize engine control strategies.

[0238] In another specific embodiment, Embodiment 1 of multi-objective optimization control under normal driving conditions is disclosed:

[0239] 1. Hardware configuration;

[0240] Processor: Based on heterogeneous multi-core SoC (4-core architecture: CPU0+1 lockstep core, CPU2 high-performance application core, CPU3 low-power coprocessor);

[0241] Memory configuration:

[0242] CPU0+1: 128KB ECC-protected DSCRATCH tightly coupled memory;

[0243] CPU2: 512MB LPDDR4 external memory;

[0244] CPU3: 512KB Data Flash;

[0245] Storage: 256MB NVM Flash, partitioned storage for agent strategies, digital twin models, and calibration parameters;

[0246] Communication interface: Supports CAN FD (1Mbps) and vehicle Ethernet (100Mbps).

[0247] Actuators: fuel injector, VVT phaser, EGR valve, ignition coil drive circuit;

[0248] 2. Software configuration;

[0249] Reactive Agent (ASIL-D): Runs on the lockstep core (CPU0+1) with a 1ms sampling period;

[0250] Planning Agent (ASIL-B): Runs on CPU2, configured with a 20ms sampling period;

[0251] Learning agent (QM): running on CPU3, configured with a 100ms sampling period;

[0252] Digital Twin Engine: Integrates physical mechanism models (thermodynamics, combustion process) with lightweight neural networks (PINN), with a model size of 85MB;

[0253] 3. Work process;

[0254] (1) Initial state;

[0255] The engine is in the warm-up phase 30 seconds after a cold start, the vehicle speed is 50km / h, and the engine speed is 2000rpm.

[0256] The reactive agent is in basic safety mode, while the planning agent is in initial optimization mode.

[0257] The learning agent is in a stable model state and has not been updated.

[0258] (2) Step 1: Data acquisition and operating condition identification;

[0259] Sensor interface layer (101) collects data:

[0260] Crankshaft speed: 2000 rpm, intake manifold pressure (MAP): 50 kPa, oxygen sensor signal: 0.45V, oil temperature: 85℃, knock signal: 0.1 (no knock);

[0261] The data acquisition and caching module preprocesses the data and identifies the current operating condition as "medium load cruise";

[0262] Operating condition label: "Medium load cruise_85℃ oil temperature";

[0263] (3) Step 2: Plan the agent to optimize decision-making;

[0264] The planning agent receives the VCU command: target torque 120Nm, driving mode is "economy mode";

[0265] Reading real-time data from shared memory:

[0266] Actual torque: 115 Nm, Current emissions: HC 0.05 g / km, Oil dilution rate: 0.03%;

[0267] Call the shared digital twin engine to perform simulation of multiple strategies: as shown in Table 1, which is the simulation strategy table of the shared digital twin engine.

[0268] Table 1

[0269] Candidate strategy Ignition advance angle (°) EGR rate (%) Predicted fuel consumption (L / 100km) Predicted HC (g / km) Predicted oil dilution rate (%) Strategy A 12 10 6.2 0.05 0.03 Strategy B 15 5 6.0 0.04 0.02 Strategy C 10 15 6.3 0.06 0.04

[0270] Construct the cost function: J = 0.4×Fuel + 0.3×HC + 0.2×Dilution + 0.1×Comfort;

[0271] Calculate the cost of each strategy:

[0272] J_A = 0.4×6.2 + 0.3×0.05 + 0.2×0.03 + 0.1×0.8 = 2.51;

[0273] J_B = 0.4×6.0 + 0.3×0.04 + 0.2×0.02 + 0.1×0.9 = 2.45;

[0274] J_C = 0.4×6.3 + 0.3×0.06 + 0.2×0.04 + 0.1×0.7 = 2.59;

[0275] Select strategy B as the optimal control objective and generate control commands:

[0276] Ignition advance angle: 15°, EGR rate: 5%, expected torque: 120Nm, shelf life: 5ms, confidence level: 0.95;

[0277] (4) Step 3: The reactive agent executes the action;

[0278] The reactive agent receives instructions from the planning agent and verifies the validity of the instructions.

[0279] Check safety boundaries: Ignition advance angle of 15° is within the safe range (10°-20°), EGR rate of 5% is within the safe range (0%-15%).

[0280] Add timestamps and CRC checks via the instruction arbitration and verification module;

[0281] Write control instructions to shared memory region B;

[0282] Drive the actuator through the safety execution layer:

[0283] Adjust the ignition advance angle to 15°, adjust the EGR valve opening to 5%, and achieve the target torque of 120Nm through fuel injector control;

[0284] (5) Step 4: Data feedback and learning;

[0285] The data acquisition and caching module records actual operational data:

[0286] Actual fuel consumption: 6.05L / 100km, actual HC emissions: 0.042g / km, actual oil dilution rate: 0.025%;

[0287] Calculate the deviation:

[0288] Fuel consumption deviation: +0.05L / 100km, HC emission deviation: +0.002g / km, oil dilution deviation: +0.005%;

[0289] Deviation data, along with operating condition labels ("medium load cruise_85℃ oil temperature"), is sent to the learning agent.

[0290] (6) Step 5: Learning agent update;

[0291] Data preprocessing for learning intelligent agents:

[0292] Operating condition clustering: Confirmed as "medium load cruise" operating condition;

[0293] Feature normalization: Normalizes the bias data to the range [0,1].

[0294] Sub-model identification: The identification of the "low-temperature fuel evaporation model" and the "emission generation model" showed a large discrepancy;

[0295] Incremental learning:

[0296] Parameters updated for the "low-temperature fuel evaporation model": Δθ = -0.05; Parameters updated for the "emission generation model": Δθ = +0.03; Simulation verification: Verification was performed by calling a digital twin copy in isolated memory; New model predicts fuel consumption: 6.02L / 100km (deviation -0.03L / 100km); New model predicts HC emissions: 0.041g / km (deviation -0.001g / km); No safety boundary breaches were observed.

[0297] Secure release: Generate a differential update package and submit it to the Hardware Security Module (HSM) for signing;

[0298] Model writing: Write the update package to the "model to be activated area" in Flash;

[0299] 4. The beneficial technical effects achieved in this embodiment are as follows:

[0300] Improved fuel economy: Through multi-objective optimization, fuel consumption is reduced by 1.2% under normal driving conditions (compared to traditional ECUs).

[0301] Emissions optimization: HC emissions reduced by 5.7%, meeting the China VI b emission standard;

[0302] Engine oil dilution control: The engine oil dilution rate is controlled below 0.03% to extend engine life;

[0303] System stability: Through the collaboration of three-layer intelligent agents, the performance fluctuations of traditional ECUs under complex operating conditions are avoided;

[0304] In another specific embodiment, Embodiment 2 of emergency response and optimization under high-load driving scenarios is disclosed:

[0305] 1. Hardware configuration;

[0306] Processor: Same as in Example 1; Memory configuration: Same as in Example 1; Storage: Same as in Example 1; Communication interface: Same as in Example 1; Actuator: Same as in Example 1.

[0307] 2. Software configuration;

[0308] Responsive agent: runs on the lockstep core (CPU0+1) with a sampling period of 0.5ms; Planning agent: runs on CPU2 with a sampling period of 50ms (descent stage in emergency situations); Learning agent: runs on CPU3 with a sampling period of 200ms.

[0309] 3. Work process;

[0310] (1) Initial state;

[0311] The engine is operating at high speed, with a vehicle speed of 120 km / h and an engine speed of 4500 rpm; this is a high-load driving scenario, and the planning agent is in high-load optimization mode; the reaction agent is in monitoring mode, and the learning agent is in model stabilization mode.

[0312] (2) Step 1: Emergency operating condition detection;

[0313] Sensor interface layer (101) collects data:

[0314] Crankshaft speed: 4500 rpm, knock signal: 0.8 (exceeding threshold T1=0.5), oil pressure: 1.2 bar (below the safe threshold of 1.5 bar), coolant temperature: 105℃ (close to the critical value of 110℃);

[0315] The reactive agent security event detector detected:

[0316] Knock signal > threshold T1 (0.8 > 0.5) and duration > 2ms, oil pressure < 1.5 bar, coolant temperature > 100℃;

[0317] (3) Step 2: Emergency response of the reactive agent;

[0318] Rule base matching emergency events:

[0319] Knock protection: Implements torque limiting and ignition delay;

[0320] Low oil pressure protection: Reduces engine load;

[0321] Overheat protection: Reduce engine speed;

[0322] A lightweight decision tree engine generates fast-response strategies:

[0323] Ignition advance angle: delayed from 15° to 5°; fuel injection quantity: reduced by 20%; target speed: reduced from 4500rpm to 3500rpm.

[0324] The command arbitration and verification module prioritizes arbitration, confirming that emergency response has the highest priority.

[0325] The emergency takeover signal output port sends an inter-core interrupt (IPI) to the planning agent.

[0326] The secure execution output interface writes control instructions to shared memory region A;

[0327] The actuator is immediately driven through the safety actuator layer: ignition is delayed to 5°, fuel injection is reduced by 20%, and valve timing is adjusted through the VVT ​​phaser;

[0328] (4) Step 3: Plan the agent's response and optimize it;

[0329] The planning agent receives an emergency takeover signal from the response agent; it immediately suspends the current optimization process and switches to "limp home" mode.

[0330] Read the preset conservative MAP from the rule base:

[0331] Maximum torque: 80 Nm, maximum speed: 3500 rpm, ignition advance angle: 5°, EGR rate: 0%;

[0332] The planning agent generates temporary control instructions and writes them into shared memory region B;

[0333] Ensure the execution of control commands through a secure execution layer;

[0334] (5) Step 4: Data collection and learning;

[0335] The data acquisition and caching module records emergency operating data:

[0336] Emergency operating condition labels: "High load knocking + low oil pressure + overheating";

[0337] Actual torque: 75 Nm, actual speed: 3450 rpm, actual oil pressure: 1.3 bar, actual coolant temperature: 102℃;

[0338] Deviation calculation:

[0339] Deviations from the preset MAP: Torque deviation -5Nm, speed deviation -50rpm; oil pressure deviation +0.1bar, coolant temperature deviation -3℃;

[0340] (6) Step 5: Learning agent update;

[0341] Data preprocessing for learning intelligent agents:

[0342] Operating condition clustering: confirmed as a composite operating condition of "high load knocking + low oil pressure + overheating"; Feature normalization: normalized the deviation data; Sub-model identification: identified that the "oil dilution model" and "thermal management model" need to be updated.

[0343] Incremental learning:

[0344] Parameters updated for "oil dilution model": Δθ = +0.15; Parameters updated for "thermal management model": Δθ = -0.2; Simulation verification: Verification was performed by calling a digital twin copy in isolated memory; New model predicts oil pressure: 1.35 bar (deviation +0.05 bar); New model predicts coolant temperature: 101.5℃ (deviation -0.5℃); No safety boundary breaches were exceeded.

[0345] Secure release: Generate a differential update package and submit it to the Hardware Security Module (HSM) for signing;

[0346] Model writing: Write the update package to the "model to be activated area" in Flash;

[0347] 4. Beneficial effects achieved in this embodiment

[0348] Emergency response capability: Knock protection response time ≤5ms, 4-6 times faster than traditional ECU (20-30ms); System safety: Avoids engine damage risk under high-load emergency conditions; Driving continuity: Ensures safe driving to the repair shop without immediate stopping through the "limp home" mode; Long-term optimization: Learning from emergency condition data enables the system to prevent knocking in similar subsequent conditions, reducing the probability of knocking by 35%.

[0349] The English fields in this solution are annotated as follows: ASIL-D refers to Automotive Safety Integrity Level D as defined by the ISO 26262 Functional Safety Standard, which is the highest safety level under this standard and is used for safety-critical control modules; ASIL-B refers to Automotive Safety Integrity Level B, which is a medium safety level and is used for conventional functional control modules; QM refers to Quality Management Level, which is the lowest level under the ISO 26262 standard, requiring no functional safety certification and used for non-safety-critical background tasks; SoC refers to System-on-Chip, an automotive-grade processor that integrates multiple CPU cores, peripherals, and interfaces onto a single chip; MPU refers to Memory Protection Unit, a hardware-level memory access control module used to isolate different memory partitions and prevent unauthorized access; MCU refers to Microcontroller Unit, which is an automotive-grade control microcontroller; NPU refers to Neural Processing Unit, a hardware unit specifically designed to accelerate neural network inference and meet AI computing power requirements; ECC refers to Error Correction and Detection, a hardware error correction mechanism for memory used to ensure the reliability of safety-critical data; DMA refers to Direct Memory Access, which does not require CPU intervention. This allows for the transfer of memory data, reducing CPU usage during background data acquisition. EDMA (Enhanced Direct Memory Access) is a high-performance DMA controller that enables background data transfer without consuming bus resources. HSM (Hardware Security Module) is a hardware unit used for encryption and signing operations, ensuring data and firmware security. AUTOSAR (Automotive Open System Architecture) is a standardized automotive-grade software architecture used to improve software compatibility and reusability. OTA (Over-The-Air) updates the firmware and model of the ECU without disassembling the vehicle.

[0350] ECU refers to Electronic Control Unit, which is the engine control computer, commonly known as the vehicle's computer; VCU refers to Vehicle Controller Unit, which is the master controller of the vehicle's power domain, responsible for issuing high-level commands to various sub-controllers; TCU refers to Transmission Control Unit, responsible for transmission shift control; BMS refers to Battery Management System, the battery management module in hybrid / electric vehicles; CAN FD refers to Flexible Data Rate Controller Area Network, a high-speed automotive-grade communication bus used to transmit vehicle control data; PID refers to Proportional-Integral-Derivative, a closed-loop control algorithm commonly used in traditional ECUs; MAP can refer to either intake manifold absolute pressure, the core operating parameter of the engine, or the calibration mapping table in a traditional ECU; VVT refers to Variable Valve Timing, the engine's valve train mechanism used to adjust the opening timing of intake and exhaust valves to improve power and fuel efficiency; EGR refers to Exhaust Gas Recirculation, a technology that reintroduces a portion of exhaust gas into the intake air to reduce combustion temperature and NOx. Emissions; BSFC refers to brake-to-fuel ratio, a core indicator for measuring engine fuel economy, measured in grams per kilowatt-hour; HC refers to hydrocarbons, one of the pollutants emitted by engines, namely incompletely burned fuel; NOx refers to nitrogen oxides, another pollutant emitted by engines, produced by high-temperature combustion, and is a core target of emission regulations; NVH refers to noise, vibration, and harshness, an indicator for measuring engine operating comfort; Ramp Rate refers to torque adjustment slope, used in degraded control to smooth the rate of torque change and prevent power interruption shocks. AI refers to Artificial Intelligence, which is the technical foundation of the hierarchical intelligent control in this solution; PINN refers to Physical Information Neural Network, which is a model architecture that combines physical mechanisms with neural networks, and is the core model architecture of the digital twin in this solution; IPI refers to Inter-core Interrupt, which is a hardware interrupt signal between different cores in a multi-core chip, used to notify other cores to handle emergency events; CRC refers to Cyclic Redundancy Check, which is a data verification algorithm used to ensure the integrity of data transmission and prevent tampering or errors; Confidence Level refers to the confidence level, which is the credibility of the instructions output by the planning agent, used to provide decision-making reference for the response agent; Time-to-Live refers to the validity period, which is the valid time window of the instruction. Expired instructions are automatically discarded to prevent expired instructions from affecting control.Fuel refers to fuel consumption, a fuel economy indicator in multi-objective optimization; Dilution refers to oil dilution, an indicator of oil performance degradation caused by fuel mixing with oil, and one of the core optimization objectives of this solution; Comfort refers to NVH-related comfort indicators, and one of the objectives of multi-objective optimization; J refers to the cost function, a weighted scoring function used in multi-objective optimization to evaluate the effectiveness of a strategy, used to solve for the optimal control strategy; w refers to the weight coefficient, the weight of each optimization objective in multi-objective optimization, used to adjust the priority of different objectives. AURIX TC3xx refers to Infineon's automotive-grade multi-core microcontroller series, the hardware platform exemplified in this solution; GPF refers to a gasoline particulate filter, an after-treatment device used to filter particulate matter from engine emissions to meet emission regulations; PFLASH refers to program flash memory, non-volatile memory used to store program code; DSCRATCH refers to tightly coupled Scratch memory, high-speed on-chip memory used to store safety-critical real-time data; HW Watchdog refers to a hardware watchdog timer, a hardware monitoring module used to detect program crashes and trigger a system reset; PFL refers to program flow monitoring, a hardware program execution flow monitoring used to detect abnormal program execution; FPU refers to a floating-point unit, a hardware unit used to accelerate floating-point operations and improve the simulation speed of digital twins; HWA refers to a hardware neural network accelerator, a hardware unit used to accelerate neural network inference; LPDDR refers to low-power double data rate memory, low-power external memory used to load large digital twin models; NVM refers to non-volatile memory, storage that retains data even when power is lost, used to store model, parameter, and other data.

[0351] Figure 8 This is a block diagram of an electronic device structure for an engine electronic control method based on an AI agent, provided in one or more embodiments of the present invention.

[0352] like Figure 8 As shown, this application provides an electronic device, including: a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus;

[0353] The memory stores a computer program that, when executed by a processor, causes the processor to perform steps of an AI-based engine electronic control method.

[0354] This application also provides a computer-readable storage medium storing a computer program executable by an electronic device, which, when run on the electronic device, causes the electronic device to perform the steps of an AI-based engine electronic control method.

[0355] This application also provides a vehicle development platform, including:

[0356] Electronic equipment for implementing an AI-based engine electronic control method;

[0357] The processor runs a program that, when running, executes steps of an AI-based engine electronic control method based on data output from electronic devices.

[0358] Storage medium for storing programs that, when running, execute steps of an AI-based engine electronic control method based on data output from electronic devices.

[0359] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An engine electronic control method based on an AI intelligent agent, characterized in that, The AI-based engine electronic control method includes: Based on the underlying support, it collects engine status signals, receives vehicle commands, collects deviation data, and executes control commands. The underlying support includes a heterogeneous multi-core SoC hardware architecture; Based on a heterogeneous multi-core SoC hardware architecture, at least three physically isolated computing domains are divided to execute: reactive agent logic, planning agent logic, and learning agent logic, respectively. The reactive agent logic runs on a lockstep computation core with a preset high security level. It is used to collect engine status signals, detect emergency conditions, and generate real-time emergency control commands at intervals less than or equal to a preset first sampling period threshold. The planning intelligent agent logic runs on a high-performance application core with a preset functional safety level. It is used to receive instructions from the vehicle's high-level management, call the local digital twin engine to simulate multiple candidate control strategies, and generate the optimal control parameters through multi-objective optimization. The system learns the logic of the intelligent agent and runs on an isolated coprocessor at a preset quality management level. It is used to collect control deviation data in the background and incrementally update the parameters of the digital twin model. It also includes deploying a shared digital twin engine: a digital twin model based on an integrated physical mechanism model and a neural network model, which simulates the engine's operating status and performance parameters in real time, providing a predictive basis for planning the intelligent agent's logic and providing an update object for learning the intelligent agent's logic.

2. The engine electronic control method based on AI intelligent agent according to claim 1, characterized in that, The underlying support also includes: The system acquires engine status signals through the sensor interface unit; receives vehicle commands through the communication interface unit; acquires deviation data through the data acquisition and buffering unit; executes control commands through the safety execution unit; and provides hardware and software support through the basic support unit.

3. An engine electronic control unit (ECU) based on an AI agent, characterized in that, The AI-based engine electronic control unit (ECU) includes: The sensor interface layer, communication interface layer, data acquisition and caching module, security execution layer, and basic support platform participate in building the underlying support; Based on the underlying support, the ECU adopts a heterogeneous multi-core SoC hardware architecture, dividing the computing domains into at least three physically isolated domains, which are deployed separately: The reactive agent runs on a lockstep computation kernel with a preset high security level. It is used to collect engine status signals, detect emergency conditions and generate real-time emergency control commands at intervals less than or equal to a preset first sampling period threshold. The reactive agent does not receive any input from the learning agent. The planning intelligent agent runs on a high-performance application core with a preset functional safety level. It is used to receive instructions from the vehicle's high-level management, call the local digital twin engine to simulate multiple candidate control strategies, and generate the optimal control parameters through multi-objective optimization. The learning agent runs on an isolated coprocessor with a preset quality management level. It is used to collect control deviation data in the background and only incrementally updates the parameters of the digital twin model. It does not directly interfere with the engine's control output commands. It also includes a shared digital twin engine deployed in the ECU's local memory. This engine integrates a physical mechanism model and a neural network model to simulate the engine's operating status and performance parameters in real time, providing a predictive basis for the planning agent and an update object for the learning agent.

4. The engine electronic control unit (ECU) based on an AI agent according to claim 3, characterized in that, include: The inter-core communication between the planning agent and the reaction agent uses structured instruction data frames; The data frame includes: The control parameter field is used to store the engine's execution control parameters; The confidence field is used to mark the degree of confidence of the planning instruction, and the agent will only adopt the instruction when the confidence level is greater than a preset confidence threshold. The validity period field defines the preset validity period window for the instruction; instructions that exceed this window will be automatically discarded. The security verification field is used to store the verification code and hardware signature to ensure data integrity.

5. The engine electronic control unit (ECU) based on an AI agent according to claim 3, characterized in that, include: The neural network of the shared digital twin engine has an input layer feature vector including crankshaft angle time series features, and an output layer prediction vector including oil dilution rate prediction and emission transient prediction.

6. The engine electronic control unit (ECU) based on an AI agent according to claim 3, characterized in that, It also includes fault degradation control logic: When the planning agent fails to respond beyond the preset response time threshold, the reaction agent automatically switches to the preset conservative control mapping table and smoothly reduces the target torque according to the preset torque adjustment slope to prevent power interruption shock.

7. The engine electronic control unit (ECU) based on an AI agent according to claim 3, characterized in that, It also includes the online learning triggering logic of the learning agent: Based on the vehicle's power status and the ECU's real-time calculated load rate, the sampling period of the learning agent is dynamically adjusted. When the ECU's real-time calculated load exceeds the preset load threshold, the sampling frequency is reduced using a preset automatic frequency conversion stepping strategy. When the real-time computing load of the ECU is below a preset load threshold, the sampling frequency is restored using a preset automatic frequency conversion step strategy.

8. The engine electronic control unit (ECU) based on an AI agent according to claim 3, characterized in that, The reactive agent is also equipped with an emergency takeover unit: The emergency takeover unit is used to, when a preset dangerous condition is detected, notify the planning agent to suspend the output of the strategy through an inter-core interruption and prioritize the execution of emergency control commands.

9. The engine electronic control unit (ECU) based on an AI agent according to claim 3, characterized in that, It also includes the model update process for the learning agent: The simulation verification of the new parameters is completed by calling the digital twin copy in isolated memory; Once verification is successful, the securely signed incremental update package will be written to the activated model area, without directly modifying the running digital twin model.

10. The engine electronic control unit (ECU) based on an AI agent according to claim 3, characterized in that, This also includes the fact that the shared memory of the ECU is divided into at least three independent partitions protected by a memory protection unit (MPU): The independent partitions are respectively used for the reaction agent to write sensor data, the planning agent to write control instructions, and the learning agent to write model update packages.

Images