A diesel transient calibration optimization method and system based on machine learning

Abstract

This invention discloses a machine learning-based transient calibration optimization method and system for diesel engines, relating to the field of diesel engine electronic control calibration technology. Its key technical points are: by collecting data in segments across the entire transient operating condition time sequence, dynamic feature data from pre-loading, loading transition to unloading and fallback is obtained, filling the feature gap of existing technologies that only collect single-point result data; furthermore, through differentiated operating condition clustering, the constant-speed loading basic operating condition and speed-torque coupled variable operating condition are separated, and a multi-stage architecture combining a basic constant-speed loading response model and an incremental generalization model is specifically constructed. The basic model accurately adapts to the core transient calibration operating condition, while the incremental generalization model extends to cover complex variable operating conditions on actual roads, solving the technical problem of excellent bench calibration results but excessive emissions on actual roads in existing technologies, achieving full-scenario adaptation from bench calibration to actual road applications.

Description

Technical Field

[0001] This invention relates to the field of diesel engine electronic control calibration technology, and in particular to a diesel engine transient calibration optimization method and system based on machine learning. Background Technology

[0002] Diesel engines, as core power units in transportation, construction machinery, agricultural machinery, and marine power, occupy an irreplaceable position in national economic development due to their advantages of high thermal efficiency, strong power, and good reliability. With the continued global focus on climate change and ecological environmental protection, countries have successively introduced increasingly stringent fuel consumption and emission regulations. my country's implementation of the National VI b emission standard and the fourth stage fuel consumption limit for heavy commercial vehicles have placed higher demands on the control of pollutants and fuel economy of diesel engines under all operating conditions, especially strengthening the emission control of transient conditions during actual driving. This poses an unprecedented challenge to diesel engine electronic control calibration technology.

[0003] The widespread adoption of high-pressure common rail electronic control systems has provided a technical foundation for the refined control of core parameters of diesel engines, such as injection pressure, injection timing, multiple injection strategies, intake throttling, and exhaust gas recirculation. It has also introduced dozens of controllable calibration parameters that are mutually coupled and have strong nonlinear relationships. The calibration and optimization of these parameters is the core link that determines the power, economy, and emission compliance of diesel engines.

[0004] Traditional diesel engine calibration technology relies heavily on engineers' experience. First, calibration engineers, based on past experience, break down the diesel engine's operating conditions into hundreds of steady-state operating points, ignoring transient conditions that account for over 70% of actual operating time. Then, on an engine test bench, single-factor variable tests are conducted for each steady-state operating point, with calibration parameters manually adjusted and performance and emissions data tested, relying on experience to achieve local optimization. After optimization for all operating points, a calibration MAP is generated through interpolation fitting. Finally, only a few typical transient conditions are simply verified, and the steady-state calibration parameters are directly written into the engine's electronic control unit. This process is not only time-consuming and costly, but also, by completely ignoring the dynamic characteristics of transient conditions, it directly leads to problems such as a sudden increase in carbon emissions during acceleration and loading, and NOx control failure, failing to meet current regulations for transient condition control.

[0005] In recent years, artificial intelligence technologies, represented by machine learning, have provided a new direction for solving the pain points of traditional calibration technologies. Industry-specific technical solutions applying machine learning to diesel engine calibration have emerged. First, test operating points are planned using Design of Experiments (DOE) methods, and bench tests are conducted to collect basic data. Then, machine learning algorithms are used to train a surrogate model for engine performance prediction, replacing some physical bench tests. Finally, combined with global optimization algorithms, parameters are optimized to obtain calibration parameters, targeting power and emissions performance. However, the following problems exist:

[0006] First, it only collects single-point results such as T90 response time and smoke peak during constant speed loading, without collecting data in segments throughout the entire transient working condition time sequence. It completely ignores core dynamic features such as torque change rate and air-oil response lag during dynamic processes such as loading transition and unloading fallback. The surrogate model trained can only be adapted to the fixed constant speed loading condition in the laboratory environment. For the complex transient working conditions of speed-torque coupling changes in actual road driving, the prediction error generally exceeds 10%, which cannot support the calibration and optimization of actual working conditions.

[0007] Secondly, setting only fixed engine safety boundary constraints fails to consider the gradient changes in calibration parameters between adjacent operating conditions, leading to abrupt changes in the optimized parameters during operating condition switching. This can cause torque fluctuations, emission jumps, and even trigger engine thermal protection during actual diesel engine operation, making the calibration results unsuitable for direct engineering application. Furthermore, it fails to consider the cumulative effect of engine thermal load under continuous transient operating conditions. Either the constraints are too strict, preventing the full release of the optimization potential of the calibration parameters, or the constraints are too loose, affecting engine reliability. Simultaneously, the existing model training is a one-time closed-loop process. When it is necessary to expand the operating condition range or adapt to different engine models, it is necessary to redesign the experiment, collect all data, and complete a full model retraining, making incremental model updates impossible. The calibration iteration cycle remains long and cannot meet the engineering requirements for rapid engine iteration. Therefore, we propose a machine learning-based transient calibration optimization method and system for diesel engines. Summary of the Invention

[0008] (a) Technical problems to be solved

[0009] To address the shortcomings of existing technologies, this invention provides a machine learning-based method and system for optimizing transient calibration of diesel engines, solving the technical problems of poor generalization, insufficient parameter smoothness, and long calibration cycle in existing transient calibration techniques.

[0010] (II) Technical Solution

[0011] To achieve the above objectives, the present invention provides the following technical solution:

[0012] A machine learning-based transient calibration optimization method for diesel engines, applied to high-pressure common rail diesel engines, based on an automatic engine calibration system, is characterized by comprising:

[0013] The diesel engine performance testing system acquires multi-dimensional transient operating condition source data of the target diesel engine across the entire operating condition range. At the same time, it performs full-stage segmented acquisition of the transient operating condition time sequence process to obtain full-time-series characteristic data of the transient operating condition.

[0014] Based on the multi-dimensional transient operating condition source data and full-time-series feature data, obtain the differential operating condition clustering results and the feature weight allocation information of the core calibration parameters of the diesel engine;

[0015] Based on the clustering results of the differential operating conditions and the feature weight allocation information, a multi-level generalized machine learning proxy model for the transient response of diesel engines is constructed based on the active experimental design method.

[0016] With optimal transient response performance and optimal emission performance of diesel engine as optimization objectives and engine safe operation boundary as constraint condition, a multi-objective optimization function is constructed that embeds operating condition ride comfort constraint and dynamic safety constraint. Multiple sets of optimized solution sets of transient operating condition calibration parameters are obtained by solving the aforementioned machine learning proxy model.

[0017] Based on the smoothness constraint of the operating conditions embedded in the multi-objective optimization function, the parameters of adjacent operating points are adaptively corrected for the optimization solution set of the multiple sets of transient operating condition calibration parameters, and a continuous and smooth calibration correction MAP map is generated.

[0018] The optimized solution set of the corrected calibration parameters is integrated and verified through multi-objective verification by steady-state operating condition verification and transient cycle of statutory emission test to obtain the evaluation value of calibration optimization effect; at the same time, the verification measured data is input into the machine learning agent model to complete the online incremental update of the model.

[0019] Preferably, the diesel engine performance testing system includes an electric dynamometer, emission analysis equipment, particulate matter collection equipment, fuel consumption testing equipment, intake parameter testing equipment, engine electronic control unit, and calibration system;

[0020] The multi-dimensional transient operating condition source data includes diesel engine operating status data, fuel injection system data, intake and exhaust system data, after-treatment system data, and emission performance data;

[0021] The transient operating condition timing process is a constant speed loading transient timing process, which includes four stages: MAP filling, full throttle loading, data measurement, and unloading stabilization. Among them, the full throttle loading stage covers the complete statistical process of acceleration response time from no-load torque to 90% of the external characteristic torque.

[0022] The full-time characteristic data includes pre-stabilized operating condition baseline parameters, dynamic change parameters during loading, steady-state verification parameters, and unloading and fallback process parameters.

[0023] Preferably, based on multi-dimensional transient operating condition source data and full-time feature data, the diesel engine speed and torque time-domain variation features, operating condition switching dynamic features, and air-fuel circuit response matching features are extracted;

[0024] Based on the extracted features, a full-time-series correlation index of working conditions and timestamps and transient working condition labels are established; based on the working condition switching frequency, coverage and dynamic response degree of transient working condition labels, multi-dimensional basic classification dimensions for working condition clustering are generated.

[0025] Based on the operating condition response sensitivity of the fuel injection system and the intake and exhaust system, the first weighting coefficient of the core calibration parameters is obtained; based on the degree of influence of emission performance and transient response performance parameters, the second weighting coefficient of the core calibration parameters is obtained; the two sets of weighting coefficients are weighted and integrated to generate feature weight allocation information.

[0026] Based on the basic classification dimension and feature weight allocation information, unsupervised density clustering is performed on the source data and feature data across the entire operating condition range to generate differential operating condition clustering results.

[0027] Preferably, based on the clustering results of different working conditions and the feature weight allocation information, the basic model training task and the incremental generalization model training task are split into two parts.

[0028] For the basic model training task, an active experimental design method is used to automatically generate a list of constant-speed loading test conditions and training configuration parameters. Within the set boundary constraints, the core calibration parameter combination is iteratively adjusted according to the optimization goal to generate continuous test condition points.

[0029] For the incremental generalization model training task, test points of speed-torque coupling variable working condition are added on the basis of constant speed loading condition to generate an extended test condition list.

[0030] Based on the training configuration parameters and the test condition list, the corresponding training dataset and validation dataset are obtained; after preprocessing the dataset, the standard test condition dataset is obtained.

[0031] The basic model was trained and iteratively optimized using a standard working condition dataset under constant speed loading conditions. When the error between the model prediction and the measured result was within 5%, the basic model convergence verification was completed.

[0032] Based on transfer learning, the incremental generalization model is trained using a dataset of varying working conditions. The incremental generalization model uses the base model as a pre-framework and is trained incrementally on the feature space of varying working conditions.

[0033] An incremental iterative update mechanism for the model is established, and new measured data is added to perform incremental training on the corresponding feature space.

[0034] Preferably, the optimization objective is to minimize the acceleration response time and the smoke emission, and the auxiliary optimization objective is to ensure that the fluctuation of nitrogen oxide emissions does not exceed a preset limit.

[0035] The safe operating boundary of the engine is that the diesel engine's explosion pressure and exhaust temperature do not exceed the safety limits of the corresponding emission regulations.

[0036] The smoothness constraint of the operating condition is to preset the maximum allowable gradient threshold for the core calibration parameters, and introduce a penalty term for gradients exceeding the threshold during the optimization process.

[0037] The dynamic safety constraint is based on the cumulative change of in-cylinder temperature and exhaust temperature under continuous operating conditions to classify the heat load level and dynamically adjust the safety constraint boundary accordingly.

[0038] The multi-objective optimization function also incorporates post-processing system adaptation constraints, dynamically adjusting the nitrogen oxide emission constraint boundary based on the post-processing system temperature.

[0039] Preferably, the constant speed loading verification condition sequence, statutory emission test transient cycle condition and multi-dimensional verification indicators are determined based on the optimized solution set of calibration parameters, covering the entire speed range.

[0040] Based on the verification conditions, actual test data of diesel engine bench were obtained; the test data were classified and analyzed to obtain three types of verification results: power performance, fuel economy performance, and emission compliance performance; based on the three types of verification results, three types of evaluation values ​​were obtained: full-condition adaptability continuity, dynamic condition response consistency, and thermal load safety margin.

[0041] The weight coefficients of each performance dimension are determined by the analytic hierarchy process (AHP), and the performance compliance evaluation value is obtained by combining the verification results. The three evaluation values ​​are then weighted and integrated with the performance compliance evaluation value to obtain the calibration optimization effect evaluation value.

[0042] Preferably, the method further includes:

[0043] Based on the calibration optimization effect evaluation value, obtain the calibration optimization qualification judgment value, and determine whether the calibration optimization qualification judgment value is greater than the preset qualification threshold. If yes, write the calibration correction MAP map corresponding to the set of calibration parameters into the engine electronic control unit. If no, iteratively correct the set of calibration parameters until the calibration optimization qualification judgment value is greater than the preset qualification threshold.

[0044] A machine learning-based transient calibration and optimization system for diesel engines, built upon an automatic engine calibration system and adapted for high-pressure common rail diesel engines, includes a data acquisition unit, a model building unit, a calibration and optimization unit, and an experimental verification unit. The system also includes:

[0045] The first acquisition module, which interfaces with the data acquisition unit, is used to acquire multi-dimensional transient operating condition source data of the target diesel engine within the full operating condition range through the diesel engine performance testing system, and simultaneously complete the full time-series characteristic data acquisition of the transient operating condition time-series process;

[0046] The second acquisition module, which interfaces with the first acquisition module, is used to acquire the clustering results of differential operating conditions and the feature weight allocation information of the core calibration parameters of the diesel engine based on multi-dimensional transient operating condition source data and full-time-series feature data.

[0047] The model building module, which interfaces with the second acquisition module, is used to build a multi-level generalized machine learning proxy model for the transient response of diesel engines based on the active experimental design method.

[0048] The optimization solution module interfaces with the model building module to construct multi-objective optimization functions and obtain multiple sets of optimized solutions for transient operating condition calibration parameters through machine learning proxy models.

[0049] The smoothness correction module, which interfaces with the optimization solution module, is used to adaptively correct the parameters of adjacent working points on the calibration parameter optimization solution set, and generate a continuous and smooth calibration correction MAP map.

[0050] The experiment verification module, which interfaces with the smoothness correction module and the model building module, is used to perform multi-objective integrated verification of the calibration parameter optimization solution set, obtain the calibration optimization effect evaluation value, and simultaneously complete the online incremental update of the machine learning agent model.

[0051] Preferably, the first acquisition module has a built-in data acquisition and access unit, an experiment execution unit, a time series feature acquisition unit, and a data classification and judgment unit, which are respectively used to connect to the test system acquisition node, execute constant speed loading transient time series experiment, acquire full time series feature data, and classify and process source data;

[0052] The model building module includes an experimental design unit, a data preprocessing unit, a model training unit, and an incremental update unit, which are used for generating an experimental condition list, preprocessing the dataset, achieving model training convergence, and incremental iterative updating of the model, respectively.

[0053] Preferably, the system further includes a calibration result determination and closed-loop execution unit, which includes a third acquisition module and a judgment execution module;

[0054] The third acquisition module is connected to the experimental verification module and is used to obtain the calibration optimization qualification judgment value based on the calibration optimization effect evaluation value.

[0055] The judgment execution module is connected to the third acquisition module to determine whether the calibration optimization qualified judgment value is greater than the preset qualified threshold. If so, the calibration correction MAP map is written into the engine electronic control unit. Otherwise, the calibration parameters are iteratively corrected until the threshold requirement is met.

[0056] (III) Beneficial Effects

[0057] 1. By collecting data in segments throughout the entire transient operating condition time sequence, dynamic feature data of the entire process from preloading, loading transition to unloading and fallback was obtained, filling the feature gap of existing technologies that only collect single-point result data. Then, through differential operating condition clustering, the constant speed loading basic operating condition and speed-torque coupled variable operating condition were separated, and a multi-level architecture combining the basic constant speed loading response model and the incremental generalization model was constructed. The basic model is accurately adapted to the core transient calibration operating condition, and the incremental generalization model is extended to cover the complex variable operating conditions of actual roads. This solves the technical problem that the existing technology has excellent bench calibration results but exceeds emission standards in actual roads, and realizes full-scenario adaptation from bench calibration to actual road application.

[0058] 2. During the parameter optimization phase, the smoothness constraints, transient thermal load dynamic safety constraints, and aftertreatment system adaptation constraints are directly embedded into the multi-objective optimization function. This allows for simultaneous performance target optimization, parameter smoothness verification, and dynamic adaptation of safety boundaries. Furthermore, based on the embedded smoothness constraints, adaptive correction of parameters at adjacent operating points is achieved, ultimately generating a continuous and smooth calibration correction MAP. This optimizes transient response and emission performance while reducing parameter fluctuations at adjacent operating points, avoiding emission jumps and torque limitations caused by sudden parameter changes. Simultaneously, the dynamic safety constraints can adjust boundaries in real time based on accumulated thermal load, maximizing the optimization potential of calibration parameters without exceeding engine safety operating limits. This balances the core contradiction between performance optimization and operational reliability. The optimized calibration parameters can be directly written into the ECU for practical application without secondary manual correction.

[0059] 3. Through proactive experimental design, high-value test conditions can be automatically generated iteratively based on optimization objectives, significantly reducing the number of invalid tests. Simultaneously, a high-precision machine learning proxy model replaces a large number of physical bench tests, compressing the tens of thousands of bench tests required for traditional calibration to less than a thousand. Coupled with an incremental model update mechanism, when expanding the operating range or adjusting the model to suit different aircraft types, only incremental training of the feature space corresponding to the new data is needed, eliminating the need for retraining with all data. Furthermore, a complete transient calibration closed-loop process is constructed, ensuring that calibration results meet emission regulations through bench verification of transient cycles in statutory emission tests. Simultaneously, the measured data collected during verification is used for online incremental updates of the model, continuously improving model accuracy and forming a positive iteration of calibration results. Attached Figure Description

[0060] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, the preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0061] Figure 1 This is a flowchart illustrating the core process of the diesel engine transient calibration and optimization method in this embodiment of the invention.

[0062] Figure 2 This is a simplified flowchart of the construction of the multi-level generalized proxy model in this embodiment of the invention;

[0063] Figure 3 This is a simplified flowchart of multi-constraint, multi-objective parameter optimization in an embodiment of the present invention;

[0064] Figure 4 This is a diagram of a bench testing system in an embodiment of the present invention. Detailed Implementation

[0065] This application provides a machine learning-based diesel engine transient calibration optimization method and system, which effectively solves the technical problems of poor generalization, insufficient parameter smoothness, and long calibration cycle of existing transient calibration technologies.

[0066] During the development of diesel engine transient calibration technology, we conducted a systematic study on common technical problems in the industry and found that the existing technology has two unavoidable core defects. First, the model generalization ability is seriously insufficient, and it can only be adapted to fixed speed loading conditions. It cannot cover the complex transient conditions of speed-torque coupling changes in actual road driving, which easily leads to the problem of excellent bench calibration results but excessive emissions in actual road driving. Second, the parameter optimization logic has inherent defects. The method of first optimizing and then correcting cannot simultaneously take into account power performance, smoothness of operation and safety of the whole machine during the optimization process, and the calibration results cannot be directly applied.

[0067] Based on this, we established a solution that includes full-dimensional data collection, refined chemical condition classification, multi-level generalization model construction, multi-constraint collaborative optimization, and full-process closed-loop verification. The main purpose is to comprehensively improve the transient response performance of diesel engines while significantly shortening the calibration cycle under the premise of strictly meeting the requirements of China VI b emission regulations, thereby improving the practicality and feasibility of the calibration solution.

[0068] Using a six-cylinder, four-stroke, water-cooled, high-pressure common rail diesel engine that meets the China VI b emission standard as the verification object, the core technical parameters of this engine are: displacement of 7.7L, compression ratio of 19.5, intake method of turbocharging and intercooling, firing order of 1-5-3-6-2-4, rated power of 257kW / 2000r / min, and calibration conditions covering the full speed range of 800r / min to 2200r / min with 100r / min intervals. This is based on the paper "Research on Transient Calibration of Common Rail Diesel Engine Based on Machine Learning Model".

[0069] The test conditions during implementation were uniformly set as follows: commercially available China VI 0# diesel fuel was used; the engine coolant outlet temperature was controlled at 88±3℃; and the temperature after intercooling was controlled at 49±2℃. The test system was built based on the AVLCAMEO 5.3 engine automatic calibration system. The test equipment included an electric dynamometer, air flow meter, fuel consumption analyzer, gas emission analyzer, micro-carbon smoke meter, particulate matter collection system, engine electronic control unit, and INCA calibration system. The calibration optimization parameters were the three core calibration parameters of the fuel injection system: smoke limit value, main injection advance angle, and rail pressure correction value. The optimization objectives were to minimize the acceleration response time T90 under constant speed loading conditions, minimize smoke emissions, and ensure that NOx emission fluctuations did not exceed the preset limits. The basic constraints were that the engine exhaust temperature and combustion pressure did not exceed the safety limits corresponding to the China VI b emission regulations. All test conditions and parameter settings were consistent with the actual situation of diesel engine calibration.

[0070] Example 1

[0071] After the diesel engine undergoes basic data acquisition via the bench testing system, the calibration optimization process proceeds through seven stages: transient data acquisition, operating condition clustering and weight allocation, multi-level generalization model construction, multi-objective parameter optimization, parameter smoothness correction, bench verification and model update, and closed-loop execution of calibration results. The details are as follows:

[0072] First, multi-dimensional transient operating condition source data of the target diesel engine is acquired through a diesel engine performance testing system across the entire operating range. Simultaneously, full-stage segmented acquisition is performed on the four stages of constant speed loading transient time series: MAP filling, full throttle loading, data measurement, and unloading stabilization to obtain full-time characteristic data of transient operating conditions. Among them, the multi-dimensional transient operating condition source data includes five categories of data: diesel engine operating status, fuel injection system, intake and exhaust system, aftertreatment system, and emission performance. This provides a complete and accurate data source for the entire calibration process, solving the problem of incomplete data acquisition in existing technologies.

[0073] Next, dynamic features of working conditions are extracted based on the collected source data and full-time series feature data, and a full-time series association index is established. Differential working condition clustering results are obtained through unsupervised density clustering. At the same time, feature weight allocation information of core calibration parameters is calculated based on the working condition response sensitivity and performance impact degree of calibration parameters. Complex working conditions are classified, and the influence weight of different calibration parameters is determined, providing targeted guidance for subsequent model training and parameter optimization.

[0074] Then, based on the clustering results and weight allocation information, the training tasks of the basic model and the incremental generalization model are split. The training of the basic constant speed loading response model and the speed-torque coupled variable working condition incremental generalization model are completed respectively using the ActiveDOE active experimental design method based on the CAMEO system. A multi-level generalized machine learning proxy model is constructed and an incremental iterative update mechanism is established to solve the technical problem of insufficient generalization ability of the existing model, while taking into account the basic accuracy and working condition expansion capability of the model.

[0075] Subsequently, with the optimal transient response performance and emission performance as the core optimization objectives and the engine safe operation boundary as the basic constraint, a multi-objective optimization function was constructed, which incorporates the gradient smoothness constraint of adjacent operating condition parameters, the dynamic constraint of transient thermal load accumulation, and the adaptation constraint of the aftertreatment system. The global optimization solution was obtained by using a trained machine learning proxy model to obtain multiple sets of optimized solutions for transient operating condition calibration parameters. This approach abandons the inherent logic of optimizing first and then correcting in existing technologies, and simultaneously considers the balance of multiple objectives during the optimization process.

[0076] Subsequently, based on the embedded operating condition smoothness constraints, the parameters of adjacent operating point points of the optimized solution set are adaptively corrected. For parameters exceeding the gradient threshold, linear interpolation is performed to smooth them and generate a continuous and smooth calibration correction MAP, further ensuring the operating condition continuity of the calibration parameters and avoiding parameter abrupt changes during operating condition switching. Then, multi-objective bench integration verification is carried out through constant speed loading steady-state operating conditions and WHTC statutory emission test transient cycles. The measured data is analyzed in multiple dimensions to obtain the calibration optimization effect evaluation value. At the same time, the measured data is input into the model to complete online incremental updates, verifying the practical application effect of the calibration results, and continuously optimizing the model accuracy through measured data. Finally, a pass judgment value is obtained based on the calibration optimization effect evaluation value. If it is greater than the preset pass threshold, the calibration correction MAP is written into the engine electronic control unit. If it is less than the preset pass threshold, the calibration parameters are iteratively corrected until the requirements are met, thus forming a complete calibration process loop, ensuring that the final output calibration result always meets the relevant standards.

[0077] Example 2

[0078] Based on the overall process of Example 1, regarding the data acquisition part, we found during the research and development process that the model accuracy of the existing calibration scheme is insufficient. The root cause is that there is a serious shortcoming in the data acquisition link. It can only collect single-point result data of transient working conditions and cannot capture the dynamic response characteristics of the entire transient process. At the same time, it relies on manual experience to divide fixed working conditions and allocate parameter weights, which cannot match the actual response characteristics of the working conditions, resulting in a lack of high-quality data support for subsequent model training.

[0079] Based on this discovery, we developed a full-time-series segmented data acquisition scheme and a feature-weight-based unsupervised work condition clustering scheme to classify work conditions and rationally allocate parameter weights, providing high-quality, highly relevant data sources for subsequent model training. Details are as follows:

[0080] First, the diesel engine bench test system was connected and debugged. Through the data acquisition access node, the signal connection between various test equipment and the engine electronic control unit and calibration system was completed to ensure the synchronous sampling frequency of 100Hz for core dynamic parameters, thus ensuring the synchronization and accuracy of data acquisition. Basic data acquisition was carried out according to the constant speed loading transient timing test specification. First, MAP filling was completed to make the engine reach a stable operating state at the target speed. Then, full throttle loading was started and data was collected synchronously with timing. When the torque was loaded to 90% of the external characteristics, T90 timing was stopped. After the set loading time was reached, emission data acquisition was stopped. After waiting for the time limit, unloading was performed. After the torque returned to no-load stability, a single cycle was completed. The test conditions covered the full speed range of 800r / min to 2200r / min, and at least 3 repeated tests were completed at a single speed. After removing abnormal data, the average value was taken as the basic data source. Five multi-dimensional transient condition source data of diesel engine operating status, fuel injection system, intake and exhaust system, after-treatment system, and emission performance were collected.

[0081] While collecting basic data, the system performs full-stage segmented data collection for the four stages of constant-speed loading. During the MAP filling stage, it collects pre-stabilized operating condition benchmark parameters to provide a benchmark for dynamic analysis of subsequent operating condition switching. During the full-throttle loading stage, it collects high-frequency dynamic response characteristic data to accurately capture the dynamic response mismatch characteristics of the air-oil circuit during loading. During the data measurement stage, it collects steady-state and dynamic data of core verification indicators. During the unloading stabilization stage, it collects engine dynamic fall-off parameters, covering the entire process of transient operating conditions from pre-stabilization to loading and then to unloading fall-off.

[0082] Subsequently, based on the collected source data and full-time-series feature data, three types of features were extracted: time-domain changes in speed and torque, dynamic switching of operating conditions, and matching of air-oil circuit response. A full-time-series correlation index of operating conditions and timestamps was established, and three types of labels were generated for four stages of time-series: single-speed constant loading, speed-torque coupled variable operating conditions, and constant-speed loading, to classify transient operating conditions. Based on the operating condition switching frequency, coverage, and dynamic response degree of the labels, a basic classification dimension for clustering was generated. According to the operating condition response sensitivity and performance impact degree of the calibration parameters, two types of weight coefficients were calculated and weighted and integrated to obtain feature weight allocation information, which replaced the traditional manual experience allocation method and made the weight allocation more in line with the actual response characteristics of the operating conditions.

[0083] Finally, unsupervised density clustering was used to generate two types of clustering results: constant speed and constant loading, and speed-torque coupled variable working condition. The constant speed and constant loading working condition cluster completely matched the test working condition points, providing a data foundation with clear boundaries and well-defined features for subsequent multi-level model training.

[0084] Example 3

[0085] Based on the data sources and operating condition classification results of Examples 1 and 2, after completing data collection and operating condition classification, we found through multiple sets of comparative experiments that the existing single-structure calibration model cannot simultaneously take into account the prediction accuracy of basic operating conditions and the generalization ability of complex variable operating conditions. If more operating conditions are to be covered, the accuracy of basic operating conditions will inevitably decrease. At the same time, the existing parameter optimization adopts the method of first optimizing and then correcting, and the corrected parameters will deviate from the optimal solution. Furthermore, it cannot simultaneously consider smoothness and safety constraints during the optimization process, resulting in the calibration results not being directly applicable.

[0086] To address this, we devised a multi-stage architecture combining a basic constant-rate loading response model with an incremental generalization model. This significantly improves the model's generalization ability to complex and variable working conditions without sacrificing the accuracy of the basic working condition. Simultaneously, we designed a multi-objective optimization function with directly embedded multi-constraints to achieve synergistic balance during the optimization process. Details are as follows:

[0087] Based on the clustering results and weight allocation information obtained in Example 2, the training tasks of the basic model and the incremental generalization model are split into two independent tasks. The boundaries of the two tasks are clarified to avoid data interference during the training process. For the basic model training task, the ActiveDOE method is used to generate a list of constant speed loading test conditions and training configuration parameters. Within the preset boundary constraints, the calibration parameter combination is iteratively adjusted to generate the DOE list. After obtaining the training and validation datasets, the standard test condition dataset is obtained through preprocessing such as outlier removal, standardization, time alignment, and dimension matching. The standard dataset of constant speed and constant loading conditions is integrated to complete the supervised training and iterative optimization of the basic model. When the error between the model prediction result and the bench test result is within 5%, the convergence verification is completed to ensure the prediction accuracy of the basic test conditions.

[0088] To address the incremental generalization model training task, an extended working condition list is generated by supplementing the constant speed loading condition with speed-torque coupling variable working condition test points. After obtaining the variable working condition dataset and preprocessing it through the same process, the core weight parameters of the basic model are frozen and the core weight parameters of the basic model are frozen. Incremental training is only performed on the variable working condition feature space, keeping the prediction accuracy and output logic of the basic model unchanged. This achieves lossless extension of the basic model and solves the problem that the accuracy and generalization of a single-structure model cannot be balanced.

[0089] Finally, an incremental iterative update mechanism for the model was established, and the triggering conditions for incremental updates were set. When the difference between the feature distribution of the newly added measured data and the existing training dataset exceeds a preset threshold, the update is triggered. During the update process, incremental training is only performed on the feature space corresponding to the newly added data, without the need to retrain the model completely, which greatly reduces the time cost of model iteration.

[0090] After the model was built, we set the shortest acceleration response time (T90) and the lowest peak smoke opacity and integral value as the optimization objectives, and set NOx emission fluctuations not exceeding the preset limit as the auxiliary optimization objective. The basic constraints were that the burst pressure and exhaust temperature did not exceed the safety limits of the China VI b regulations, which fully met the core requirements of the calibration. The gradient smoothness constraints of adjacent operating condition parameters, the dynamic constraints of transient heat load accumulation, and the aftertreatment system adaptation constraints were directly embedded into the multi-objective optimization function. Among them, the smoothness constraint was achieved by setting the maximum allowable gradient threshold and introducing a penalty term when the threshold was exceeded, ensuring that the optimization results were consistent. While meeting the smoothness requirements, the dynamic safety constraints dynamically adjust the constraint boundaries according to the cumulative heat load level under continuous operating conditions to avoid exceeding the engine heat load limit under continuous transient operating conditions. The after-treatment adaptation constraints dynamically adjust the NOx emission constraints according to the inlet temperature of the SCR after-treatment system to ensure emission compliance under all operating conditions. Based on the optimization function that has completed the constraint embedding, a genetic algorithm is used for global optimization. The population size is set to 100, the number of iterations to 200, the crossover probability to 0.8, and the mutation probability to 0.05. Finally, multiple sets of Pareto optimal calibration parameter optimization solutions are obtained, realizing the coordinated balance of multiple objectives.

[0091] Example 4

[0092] Based on the calibration parameter optimization results of Examples 1-3, this embodiment constructs a multi-dimensional calibration effect verification and evaluation system, and establishes a closed-loop execution mechanism for calibration results. This comprehensively and objectively evaluates the calibration effect, and through closed-loop iteration, ensures that the calibration results continuously meet design requirements and regulatory standards. Specifically:

[0093] The verification conditions and indicators are determined by optimizing the solution set of calibration parameters obtained from the solution. The constant speed loading verification conditions cover the entire speed range and three repeated tests are set for each speed. The statutory emission test adopts WHTC transient cycle to complete the complete test of cold start and hot start. The multi-dimensional verification indicators are divided into three categories: power performance, fuel economy performance, and emission compliance performance. Key indicators include acceleration response time T90, peak smoke opacity and integral value, NOx emission integral value, PM and PN emission data, which comprehensively cover the evaluation dimensions of calibration effect.

[0094] Based on defined verification conditions, diesel engine bench tests were conducted to obtain multiple sets of measured verification data. These data were then categorized and analyzed to obtain verification results for power performance, fuel economy, and emission compliance. Based on these three types of verification results, three evaluation values ​​were calculated: full-condition adaptability continuity, dynamic operating condition response consistency, and thermal load safety margin. The weighting coefficients of power performance, fuel economy, emission compliance, and thermal load safety margin in the calibration optimization effect evaluation were determined using the analytic hierarchy process (AHP). Combined with the corresponding verification results, a performance compliance evaluation value was obtained. Finally, the three evaluation values—full-condition adaptability continuity, dynamic operating condition response consistency, and thermal load safety margin—were weighted and integrated with the performance compliance evaluation value to obtain the calibration optimization effect evaluation value. This multi-dimensional quantitative evaluation of the calibration effect replaces the traditional single-indicator verification method.

[0095] The measured data collected during the bench verification process are synchronously input into the trained machine learning agent model. Following the incremental iterative update mechanism established in Example 3, the model is updated online incrementally to continuously optimize the model's prediction accuracy for actual working conditions and form a positive iteration of calibration effect.

[0096] We preset a calibration optimization qualification threshold based on the limits of China VI b emission regulations, engine design safety limits, and overall engine performance targets. Based on the calibration optimization effect evaluation value obtained from the above steps, we calculate the calibration optimization qualification judgment value. This value is then compared with the preset qualification threshold. If the value is greater than the threshold, the calibration parameters are deemed to meet the calibration requirements, and the corresponding calibration correction MAP is written into the engine electronic control unit, completing the calibration process. If the value is less than the threshold, the parameters are deemed not to meet the requirements. Based on the verification test data, we fine-tune the optimization objectives and constraints, iteratively correct the parameters, and repeat the entire process of model optimization, parameter solving, and bench verification until the value is greater than the threshold, forming a closed-loop execution of the entire calibration process. This fundamentally ensures the stability and compliance of the calibration results.

[0097] Example 5

[0098] Based on the calibration methods in Examples 1-4, a machine learning-based transient calibration optimization system for diesel engines is proposed. This system is built upon the AVLCAMEO calibration system and is adapted to the transient calibration requirements of high-pressure common rail diesel engines. It includes four basic units: data acquisition, model building, calibration optimization, and experimental verification; and seven functional modules: first acquisition, second acquisition, model building, optimization solution, ride comfort correction, experimental verification, calibration result determination, and closed-loop execution. Details are as follows:

[0099] The first acquisition module interfaces with the data acquisition unit and has built-in data acquisition access, test execution, time series feature acquisition, and data classification and judgment units, which respectively realize the functions of multi-source data synchronous acquisition, test process automated control, transient dynamic feature accurate capture, and data standardization classification, corresponding to the data acquisition link in the method.

[0100] The second acquisition module connects with the first acquisition module. Based on the source data and full-time-series feature data, it extracts the core features of the working conditions, establishes the association index, and generates clustering results and weight allocation information to provide data support for model construction. This corresponds to the working condition clustering and weight allocation steps in the method.

[0101] The model building module is connected to the second acquisition module and has built-in experimental design, data preprocessing, model training and incremental update units. These units respectively realize the functions of automatic generation of experimental condition list, dataset standardization, basic and incremental model training convergence and model incremental iterative update, corresponding to the multi-order generalization model building link in the method.

[0102] The optimization solution module is integrated with the model building module. It has built-in optimization target setting, constraint embedding, and global optimization units to complete the functions of optimization target configuration, multi-constraint embedding integration, and calibration parameter optimization solution set solving, corresponding to the multi-objective parameter optimization link in the method.

[0103] The ride comfort correction module is integrated with the optimization solution module to perform adaptive parameter correction and smoothing on the optimization solution set, and generate a calibration correction MAP map compatible with the engine electronic control unit calibration system, corresponding to the parameter ride comfort correction step in the method.

[0104] The experimental verification module is connected with the smoothness correction module and the model building module to carry out automated bench verification, analyze measured data, calculate calibration optimization effect evaluation value, and simultaneously complete online incremental updates of the model with measured data, corresponding to the bench verification and model update links in the method.

[0105] The calibration result determination and closed-loop execution unit includes a third acquisition and judgment execution module. The third acquisition module calculates the qualified judgment value based on the calibration optimization effect evaluation value. The judgment execution module compares the qualified judgment value with the preset threshold. If the requirements are met, the calibration correction MAP map is written into the engine electronic control unit to complete the calibration. If the requirements are not met, iterative correction is performed until the standard is met. This corresponds to the calibration result closed-loop execution link in the method.

[0106] Finally, it should be noted that the above embodiments are merely examples for clearly illustrating the present invention and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A machine learning-based transient calibration optimization method for diesel engines, applied to high-pressure common rail diesel engines, based on an automatic engine calibration system, characterized in that... include: The diesel engine performance testing system acquires multi-dimensional transient operating condition source data of the target diesel engine across the entire operating condition range. At the same time, it performs full-stage segmented acquisition of the transient operating condition time sequence process to obtain full-time-series characteristic data of the transient operating condition. Based on the multi-dimensional transient operating condition source data and full-time-series feature data, obtain the differential operating condition clustering results and the feature weight allocation information of the core calibration parameters of the diesel engine; Based on the clustering results of the differential operating conditions and the feature weight allocation information, a multi-level generalized machine learning proxy model for the transient response of diesel engines is constructed based on the active experimental design method. With optimal transient response performance and optimal emission performance of diesel engine as optimization objectives and engine safe operation boundary as constraint condition, a multi-objective optimization function is constructed that embeds operating condition ride comfort constraint and dynamic safety constraint. Multiple sets of optimized solution sets of transient operating condition calibration parameters are obtained by solving the aforementioned machine learning proxy model. Based on the smoothness constraint of the operating conditions embedded in the multi-objective optimization function, the parameters of adjacent operating points are adaptively corrected for the optimization solution set of the multiple sets of transient operating condition calibration parameters, and a continuous and smooth calibration correction MAP map is generated. The optimized solution set of the corrected calibration parameters is integrated and verified through multi-objective verification by steady-state operating condition verification and transient cycle of statutory emission test to obtain the evaluation value of calibration optimization effect; at the same time, the verification measured data is input into the machine learning agent model to complete the online incremental update of the model.

2. The machine learning-based transient calibration optimization method for diesel engines according to claim 1, characterized in that, The diesel engine performance testing system includes an electric dynamometer, emission analysis equipment, particulate matter collection equipment, fuel consumption testing equipment, intake parameter testing equipment, engine electronic control unit, and calibration system. The multi-dimensional transient operating condition source data includes diesel engine operating status data, fuel injection system data, intake and exhaust system data, after-treatment system data, and emission performance data; The transient operating condition timing process is a constant speed loading transient timing process, which includes four stages: MAP filling, full throttle loading, data measurement, and unloading stabilization. Among them, the full throttle loading stage covers the complete statistical process of acceleration response time from no-load torque to 90% of the external characteristic torque. The full-time characteristic data includes pre-stabilized operating condition baseline parameters, dynamic change parameters during loading, steady-state verification parameters, and unloading and fallback process parameters.

3. The machine learning-based transient calibration optimization method for diesel engines according to claim 1, characterized in that, Based on multi-dimensional transient operating condition source data and full-time series feature data, the time-domain variation features of diesel engine speed and torque, dynamic features of operating condition switching, and air-oil circuit response matching features are extracted. Based on the extracted features, a full-time-series correlation index of working conditions and timestamps and transient working condition labels are established; Based on the operating condition switching frequency, coverage, and dynamic response degree of transient operating condition labels, multi-dimensional basic classification dimensions for operating condition clustering are generated; Based on the operating condition response sensitivity of the fuel injection system and the intake and exhaust system, the first weighting coefficient of the core calibration parameters is obtained; based on the degree of influence of emission performance and transient response performance parameters, the second weighting coefficient of the core calibration parameters is obtained; the two sets of weighting coefficients are weighted and integrated to generate feature weight allocation information. Based on the basic classification dimension and feature weight allocation information, unsupervised density clustering is performed on the source data and feature data across the entire operating condition range to generate differential operating condition clustering results.

4. The machine learning-based transient calibration optimization method for diesel engines according to claim 1, characterized in that, Based on the clustering results of the differential chemical conditions and the feature weight allocation information, the basic model training task and the incremental generalization model training task are separated. For the basic model training task, an active experimental design method is used to automatically generate a list of constant-speed loading test conditions and training configuration parameters. Within the set boundary constraints, the core calibration parameter combination is iteratively adjusted according to the optimization goal to generate continuous test condition points. For the incremental generalization model training task, test points of speed-torque coupling variable working condition are added on the basis of constant speed loading condition to generate an extended test condition list. Based on the training configuration parameters and the list of experimental conditions, obtain the corresponding training dataset and validation dataset; After preprocessing the dataset, a standard working condition dataset is obtained; The basic model was trained and iteratively optimized using a standard working condition dataset under constant speed loading conditions. When the error between the model prediction and the measured result was within 5%, the basic model convergence verification was completed. Based on transfer learning, the incremental generalization model is trained using a dataset of varying working conditions. The incremental generalization model uses the base model as a pre-framework and is trained incrementally on the feature space of varying working conditions. An incremental iterative update mechanism for the model is established, and new measured data is added to perform incremental training on the corresponding feature space.

5. The machine learning-based transient calibration optimization method for diesel engines according to claim 1, characterized in that, The specific optimization objectives are to accelerate the shortest response time and the lowest smoke emission, while the auxiliary optimization objective is to ensure that the fluctuation of nitrogen oxide emissions does not exceed the preset limit. The safe operating boundary of the engine is that the diesel engine's explosion pressure and exhaust temperature do not exceed the safety limits of the corresponding emission regulations. The smoothness constraint of the operating condition is to preset the maximum allowable gradient threshold for the core calibration parameters, and introduce a penalty term for gradients exceeding the threshold during the optimization process. The dynamic safety constraint is based on the cumulative change of in-cylinder temperature and exhaust temperature under continuous operating conditions to classify the heat load level and dynamically adjust the safety constraint boundary accordingly. The multi-objective optimization function also incorporates post-processing system adaptation constraints, dynamically adjusting the nitrogen oxide emission constraint boundary based on the post-processing system temperature.

6. The machine learning-based transient calibration optimization method for diesel engines according to claim 1, characterized in that, Based on the optimized solution set of calibration parameters, a constant speed loading verification condition sequence covering the entire speed range, a legal emission test transient cycle condition, and multi-dimensional verification indicators were determined. Based on the verification conditions, actual test data of the diesel engine bench were obtained. The measured data are classified and analyzed to obtain three types of verification results: power performance, fuel economy performance, and emission compliance performance. Based on the three types of verification results, three types of evaluation values ​​are obtained: full-condition adaptability continuity, dynamic condition response consistency, and thermal load safety margin. The weight coefficients of each performance dimension are determined by the analytic hierarchy process (AHP), and the performance compliance evaluation value is obtained by combining the verification results. The three evaluation values ​​are then weighted and integrated with the performance compliance evaluation value to obtain the calibration optimization effect evaluation value.

7. The machine learning-based transient calibration optimization method for diesel engines according to claim 1, characterized in that, The method also includes: Based on the calibration optimization effect evaluation value, obtain the calibration optimization qualification judgment value, and determine whether the calibration optimization qualification judgment value is greater than the preset qualification threshold. If yes, write the calibration correction MAP map corresponding to the set of calibration parameters into the engine electronic control unit. If no, iteratively correct the set of calibration parameters until the calibration optimization qualification judgment value is greater than the preset qualification threshold.

8. A machine learning-based transient calibration optimization system for diesel engines, built upon an automatic engine calibration system and adapted for high-pressure common rail diesel engines, comprising: The first acquisition module is used to acquire multi-dimensional transient operating condition source data of the target diesel engine within the full operating condition range through the diesel engine performance testing system, and simultaneously complete the full time-series feature data acquisition of the transient operating condition time-series process; The second acquisition module, which interfaces with the first acquisition module, is used to acquire the clustering results of differential operating conditions and the feature weight allocation information of the core calibration parameters of the diesel engine based on multi-dimensional transient operating condition source data and full-time-series feature data. The model building module, which interfaces with the second acquisition module, is used to build a multi-level generalized machine learning proxy model for the transient response of diesel engines based on the active experimental design method. The optimization solution module interfaces with the model building module to construct multi-objective optimization functions and obtain multiple sets of optimized solutions for transient operating condition calibration parameters through machine learning proxy models. The smoothness correction module, which interfaces with the optimization solution module, is used to adaptively correct the parameters of adjacent working points on the calibration parameter optimization solution set, and generate a continuous and smooth calibration correction MAP map. The experiment verification module, which interfaces with the smoothness correction module and the model building module, is used to perform multi-objective integrated verification of the calibration parameter optimization solution set, obtain the calibration optimization effect evaluation value, and simultaneously complete the online incremental update of the machine learning agent model.

9. The machine learning-based transient calibration and optimization system for diesel engines according to claim 8, characterized in that, The first acquisition module has a built-in data acquisition and access unit, an experiment execution unit, a time series feature acquisition unit, and a data classification and judgment unit, which are respectively used to connect to the test system acquisition node, execute constant speed loading transient time series experiment, acquire full time series feature data, and classify and process source data; The model building module includes an experimental design unit, a data preprocessing unit, a model training unit, and an incremental update unit, which are used for generating an experimental condition list, preprocessing the dataset, achieving model training convergence, and incremental iterative updating of the model, respectively.

10. The machine learning-based transient calibration and optimization system for diesel engines according to claim 8, characterized in that, The system also includes a calibration result determination and closed-loop execution unit, which includes a third acquisition module and a judgment execution module; The third acquisition module is connected to the experimental verification module and is used to obtain the calibration optimization qualification judgment value based on the calibration optimization effect evaluation value. The judgment execution module is connected to the third acquisition module to determine whether the calibration optimization qualified judgment value is greater than the preset qualified threshold. If so, the calibration correction MAP map is written into the engine electronic control unit. Otherwise, the calibration parameters are iteratively corrected until the threshold requirement is met.

Images