Diesel engine comprehensive liquid consumption optimization method and system based on machine learning

Abstract

This invention discloses a method and system for optimizing comprehensive fluid consumption of diesel engines based on machine learning, relating to the field of diesel engine control technology. Its key technical points are: a steady-state-transient dual-domain performance response proxy model is constructed based on an adaptive block-weighted neural network. Through adaptive adjustment of the working condition block granularity, dynamic allocation of sub-model weights, and the design of transient time-series correlation sub-models, the data fitting accuracy of steady-state conventional working conditions and dynamically switching working conditions is simultaneously ensured. This effectively improves the problem of excessive prediction deviation under actual road spectrum working conditions, accurately representing the correlation between energy consumption and emissions under all working conditions. Simultaneously, using the original emission nitrogen oxide ratio as the core constraint, multi-objective collaborative optimization is carried out by combining the hard boundary of overall engine operation safety, the dynamic constraint of SCR temperature field coupled with urea crystallization, and the pre-weight constraint of transient road spectrum. This not only compensates for the shortcomings of traditional optimization dimensions being one-sided and detached from actual working conditions, but also achieves quantitative prediction of urea crystallization risk, avoiding post-treatment blockage faults from the source of operation, and taking into account emission compliance, overall engine safety, and adaptability to actual working conditions.

Description

Technical Field

[0001] This invention relates to the field of diesel engine control technology, and in particular to a method and system for optimizing the comprehensive fluid consumption of a diesel engine based on machine learning. Background Technology

[0002] As the core power unit for commercial vehicles, construction machinery, and stationary generator sets, diesel engines, in actual service, together with fuel consumption and SCR system urea consumption, constitute the overall liquid usage cost of the entire machine. At the same time, the exhaust emissions of the entire machine must meet the China VI b standard and the current phased emission control standards of the industry. Currently, the industry has formed a mature routine implementation process for diesel engine energy consumption calibration and emission synergy optimization, but there are still systemic shortcomings in actual application.

[0003] On the one hand, existing technologies, when in use, only define the basic steady-state operating condition range from idle speed to rated speed, collect simple operating parameters such as speed and basic fuel injection quantity, and directly use basic neural networks to uniformly fit and process the data of the entire operating condition; however, they do not distinguish the differences between the steady-state smooth operation of diesel engines and the frequent switching of transient operating conditions, nor do they combine the spatiotemporal correlation characteristics between operating conditions for partitioning processing, nor do they supplement dedicated correction logic for transient dynamic changes; resulting in poor fitting effect for the operating condition switching stage. Even if the steady-state verification accuracy is qualified, the prediction deviation will be significantly amplified when applied to the dynamic road spectrum of real roads, making it difficult to accurately reflect the linkage change law of energy consumption and emissions under the entire operating condition;

[0004] On the other hand, traditional methods only use fixed extreme values ​​of in-cylinder explosion pressure and turbine exhaust temperature as basic safety limits, and mostly focus on optimizing fuel consumption indicators. They do not combine the coupling reaction mechanism of the original nitrogen oxide generation characteristics and the temperature field of the SCR box, and cannot dynamically quantify the real risk of urea crystallization. They also do not set pre-weights based on the proportion and switching frequency of high-speed road spectrum, CHTC road spectrum, and WHTC cycle, and rely on static standards to complete the optimization design. This not only leads to a single optimization dimension, which cannot simultaneously take into account the smoke emission control effect, but also makes it difficult to avoid the urea crystallization problem of the SCR system in advance, which can easily lead to pipeline blockage and engine failure alarms. At the same time, the optimization results are out of touch with the actual driving conditions, the steady-state calibration parameters are not suitable for vehicle installation, and the field implementation effect is not ideal. Therefore, we propose a diesel engine comprehensive fuel consumption optimization method and system based on machine learning. Summary of the Invention

[0005] (a) Technical problems to be solved

[0006] To address the shortcomings of existing technologies, this invention provides a machine learning-based method and system for optimizing the comprehensive fluid consumption of diesel engines, solving the technical problems of insufficient accuracy in existing diesel engine fluid consumption calibration modeling, poor transient adaptability, uncontrollable risk of urea crystallization, lack of compliance restrictions on model iteration, and lack of comprehensive evaluation standards.

[0007] (II) Technical Solution

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] A machine learning-based method for optimizing the overall fluid consumption of a diesel engine, comprising the following steps:

[0010] Multiple sets of operating source data are acquired within the full operating range of the diesel engine. The operating source data includes power operating parameters, fuel injection system parameters, emission characteristic parameters, after-treatment system operating parameters, operating condition boundary constraint parameters, and operating condition switching timing characteristic parameters.

[0011] Based on the aforementioned source data, an adaptive block weighted neural network algorithm is used to construct a diesel engine steady-state-transient dual-domain performance response proxy model. After model training and full-condition accuracy verification, a comprehensive liquid consumption prediction machine learning model is obtained.

[0012] With the original nitrogen oxide emission ratio as the core constraint, the dual optimization objectives are the lowest comprehensive liquid consumption cost and the best smoke emission. The main injection advance angle and injection rail pressure are the coupled optimization variables. Combining the preset operating safety boundary, the dynamic constraint of urea crystallization risk coupled with the original nitrogen oxide emission and SCR temperature field, and the pre-constraint of transient road spectrum operating conditions, a steady-state-transient dual-domain coupled multi-objective optimization function is constructed. The optimal scheme of injection control parameters under multiple sets of different original nitrogen oxide emission ratio targets is obtained by solving the multi-objective optimization algorithm.

[0013] The optimal solutions for multiple sets of fuel injection control parameters were written into the diesel engine electronic control unit. Verification tests were conducted on the engine bench under steady-state characteristic conditions and typical transient road spectrum conditions. Data on fuel consumption, urea consumption, comprehensive liquid consumption cost, smoke emissions, and after-treatment system operation status were collected for each solution.

[0014] By analyzing and verifying test data and combining the road spectrum scenario adaptive multi-dimensional comprehensive evaluation system, the optimal original emission nitrogen oxide ratio control range and the corresponding target fuel injection control parameter optimal scheme were determined.

[0015] The optimal solution for the target fuel injection control parameters is sent to the execution units of the diesel engine fuel injection system and after-treatment system; the actual full-condition operation data of the diesel engine after execution is collected, and the comprehensive fuel consumption prediction machine learning model is iteratively updated through the incremental learning method of full life cycle compliance boundary constraints.

[0016] Preferably, acquiring multiple sets of operating source data across the entire operating range of the diesel engine includes:

[0017] The operating range of the diesel engine under all operating conditions is defined, with the maximum torque curve corresponding to each speed from idle speed to rated speed as the boundary of steady-state operating conditions, and the operating condition switching sequence of high-speed road spectrum, CHTC road spectrum, and WHTC cycle as the boundary of transient operating conditions.

[0018] It connects to multiple data acquisition access points across the entire operating range of the diesel engine, synchronously collects raw operating data under all operating conditions, and records the characteristic data of steady-state operating conditions and the time-series dynamic data of the transient operating condition switching process;

[0019] The raw operating data is subjected to feature identification and classification to obtain power operating parameters, fuel injection system parameters, emission characteristic parameters, after-treatment system operating parameters, operating condition boundary constraint parameters, and operating condition switching timing characteristic parameters.

[0020] The power operation parameters include engine speed, cyclic fuel injection quantity, output torque, in-cylinder detonation pressure, and turbine exhaust temperature; the fuel injection system parameters include main injection advance angle and injection rail pressure; the emission characteristic parameters include the proportion of nitrogen oxides in the original emissions, the concentration of nitrogen oxides in the exhaust gas, and the smoke concentration; the aftertreatment system operation parameters include urea injection quantity, urea consumption rate, SCR box temperature, and ammonia slip concentration; and the operating condition switching timing characteristic parameters are the change rate of speed, torque, and fuel injection quantity in adjacent operating conditions.

[0021] Preferably, the construction of a diesel engine steady-state-transient dual-domain performance response proxy model includes:

[0022] Preprocessing of the source data includes outlier removal, missing value completion, data time-series alignment, and feature association to form the model training dataset;

[0023] Using power operation parameters, fuel injection system parameters, and operating condition switching timing characteristic parameters as the model input layer, and fuel consumption rate, original nitrogen oxide emission ratio, urea consumption rate, and smoke concentration as the model output layer, a basic neural network topology is built.

[0024] The entire working condition range is divided into grid blocks, incorporating the correlation features of adjacent working conditions, calculating the nonlinearity and working condition correlation of each block of data, and adaptively adjusting the block granularity and the membership function weights of the corresponding sub-models.

[0025] A time-series correlation sub-model is constructed for the transient operating condition switching interval, with the dynamic change characteristics of adjacent operating conditions as the core input;

[0026] The combination of fuel injection control parameters is generated by the basic experimental design method. After collecting basic steady-state operating condition data, the system enters the machine learning iteration stage of active experimental design to supplement the sample data of transient operating condition switching points.

[0027] The model training dataset is input into the neural network topology for iterative training. An iteration termination condition is set for surrogate model training. After bench verification confirms that the model accuracy meets the standard, the comprehensive liquid consumption prediction machine learning model is obtained.

[0028] Preferably, the optimal solution for constructing a steady-state-transient dual-domain coupled multi-objective optimization function and solving for the optimal injection control parameters includes:

[0029] The preset operating safety boundaries for the diesel engine are set as follows: the cylinder explosion pressure should not exceed 195 bar and the turbine exhaust temperature should not exceed 720°C.

[0030] A dynamic constraint model for urea crystallization risk coupled with the original emission nitrogen oxides and the SCR temperature field is constructed. The original emission nitrogen oxide ratio, SCR box temperature and urea injection rate are used as inputs to quantify the bidirectional influence of the original emission nitrogen oxide ratio on the SCR temperature field and output the real-time urea crystallization risk coefficient.

[0031] The transient road spectrum conditions of the target application scenario are decomposed into intervals, and the time proportion and switching frequency of each condition interval are statistically analyzed to generate a transient road spectrum condition weight distribution table as a pre-constraint.

[0032] Multiple sets of gradient distributions are set as the target values ​​for the proportion of nitrogen oxide emissions. Each set of target values ​​is a fixed constraint. Combined with hard constraints on operational safety, soft constraints on urea crystallization risk, and road spectrum weight preconditions, the steady-state-transient dual-domain coupled multi-objective optimization function is constructed.

[0033] The optimization function is solved by a multi-objective optimization algorithm. The generated parameter schemes are then subjected to real-time pre-verification of urea crystallization risk and transient adaptability assessment. After filtering out invalid schemes, the optimal fuel injection control parameters corresponding to the original nitrogen oxide emission ratio target value for each group are obtained and summarized to form an optimal scheme set.

[0034] Preferably, the verification test includes:

[0035] The steady-state characteristic conditions for the verification test were determined to be rated conditions, minimum fuel consumption conditions, and low-speed, low-load conditions, while the typical transient road spectrum conditions were high-speed road spectrum, CHTC road spectrum, and WHTC cycle conditions.

[0036] The main injection advance angle and injection rail pressure pulse spectrum corresponding to the optimal scheme of each group of injection control parameters were written into the diesel engine electronic control unit, and steady-state characteristic conditions and typical transient road spectrum conditions of each scheme were tested on the engine bench in sequence.

[0037] Data on fuel consumption rate, urea consumption rate, comprehensive liquid consumption cost, smoke emission, and after-treatment system operation status were collected during the test. Data on SCR box temperature, ammonia slip, and urea crystallization risk during transient operating condition switching were recorded simultaneously. During the test, the urea injection quantity was adjusted in a closed loop to maintain the concentration of nitrogen oxides in diesel engine exhaust gas in compliance with current emission standards. The single variable was controlled as the original nitrogen oxide emission ratio, and the diesel engine operating safety boundary parameters were recorded simultaneously.

[0038] Preferably, determining the optimal initial nitrogen oxide emission ratio control range and iteratively updating the comprehensive liquid consumption prediction machine learning model includes:

[0039] Analyze and verify the test data to clarify the quantitative impact of the original nitrogen oxide emission ratio on fuel consumption, urea consumption, comprehensive liquid consumption cost, and smoke emission of diesel engines under steady-state and transient operating conditions;

[0040] Construct an adaptive multi-dimensional comprehensive evaluation system for road spectrum scenarios, and adjust the weight coefficients of each evaluation indicator in the scheme evaluation according to the transient road spectrum working condition weight distribution table of the target application scenario.

[0041] Based on the criteria of a slowing decline in overall liquid consumption cost, no sharp increase in urea consumption, controllable risk of urea crystallization, significant reduction in smoke emissions, and emission compliance margin meeting the requirements of the entire life cycle, and combined with the quantitative scoring of the comprehensive evaluation system, the optimal original emission nitrogen oxide ratio control range and the corresponding optimal solution for target fuel injection control parameters are determined.

[0042] The optimal solution of the target fuel injection control parameters is decomposed into execution control commands, which are then sent to the corresponding execution units of the diesel engine fuel injection system and after-treatment system to achieve real-time closed-loop control of fuel injection and after-treatment parameters.

[0043] The actual full-condition operating data of the diesel engine is collected according to the preset sampling period. The actual operating data, model prediction results and measured results are combined to form an incremental training sample set. The current emission regulations boundary, engine operating safety boundary and urea crystallization risk threshold are used as hard constraints. Combined with the engine aging and decay coefficient, the comprehensive liquid consumption prediction machine learning model is iteratively updated through the incremental learning method of full life cycle compliance boundary constraints.

[0044] A machine learning-based diesel engine integrated fluid consumption optimization system, the system comprising:

[0045] The full-condition data acquisition module is used to acquire multiple sets of operating source data within the full operating range of the diesel engine. The operating source data includes power operating parameters, fuel injection system parameters, emission characteristic parameters, after-treatment system operating parameters, operating condition boundary constraint parameters, and operating condition switching timing characteristic parameters.

[0046] The dual-domain model construction module is used to construct a diesel engine steady-state-transient dual-domain performance response proxy model based on the running source data and through an adaptive block weighted neural network algorithm. After model training and full-condition accuracy verification, it outputs a comprehensive liquid consumption prediction machine learning model.

[0047] The coupled optimization module is used to take the original nitrogen oxide emission ratio as the core constraint, the lowest comprehensive liquid consumption cost and the best smoke emission as the dual optimization objectives, and the main injection advance angle and injection rail pressure as coupled optimization variables. It combines the preset operating safety boundary, the dynamic constraint of urea crystallization risk coupled with the original nitrogen oxide emission ratio and the SCR temperature field, and the transient road spectrum operating condition weight pre-constraint to construct a steady-state-transient dual-domain coupled multi-objective optimization function. The optimal solution of injection control parameters under multiple sets of different original nitrogen oxide emission ratio targets is obtained by solving the multi-objective optimization algorithm.

[0048] The bench verification module is used to write the optimal schemes of multiple sets of fuel injection control parameters into the diesel engine electronic control unit, and to conduct verification tests on the engine bench under steady-state characteristic conditions and typical transient road spectrum conditions. It collects relevant data on fuel consumption, urea consumption, comprehensive liquid consumption cost, smoke emission and after-treatment system operation status corresponding to each scheme.

[0049] The optimal solution decision module is used to analyze and verify test data, and combine the road spectrum scenario adaptive multi-dimensional comprehensive evaluation system to determine the optimal original emission nitrogen oxide ratio control range and the corresponding target fuel injection control parameters optimal solution.

[0050] The model iteration module is used to send the optimal solution of the target injection control parameters to the execution unit of the diesel engine fuel injection system and after-treatment system; and to collect the actual full-condition operation data of the diesel engine after execution, and to iteratively update the comprehensive fuel consumption prediction machine learning model through the incremental learning method of full life cycle compliance boundary constraints.

[0051] Preferably, the full-condition data acquisition module and the dual-domain model construction module include:

[0052] The operating condition boundary delineation unit is used to delineate the steady-state and transient operating condition boundaries of the diesel engine under all operating conditions.

[0053] The data acquisition and preprocessing unit is used to connect to multiple data acquisition access points, synchronously collect raw running data and preprocess it, and output the model training dataset.

[0054] The neural network training unit is used to build the basic topology of the neural network, perform working condition grid segmentation, adaptive weight allocation, construction of transient time series correlation sub-models and iterative training of the model, and output the comprehensive liquid consumption prediction machine learning model after full working condition accuracy verification.

[0055] Preferably, the coupling optimization module and the bench verification module include:

[0056] The constraint weight configuration unit is used to store the preset operating safety boundary of the diesel engine, the original emission nitrogen oxide ratio constraint target, run the urea crystallization risk dynamic constraint model and output the real-time risk coefficient, and generate the transient road spectrum operating condition weight distribution table.

[0057] The optimization unit is used to construct a steady-state-transient dual-domain coupled multi-objective optimization function, solves and filters invalid schemes through a multi-objective optimization algorithm, and generates a set of optimal schemes for fuel injection control parameters.

[0058] The bench test execution unit is used to set the operating conditions and procedures for the verification test, write parameters to the diesel engine electronic control unit, and simultaneously collect various data during the test process, as well as parameters related to diesel engine operation safety and after-treatment system operation.

[0059] Preferably, the optimal solution decision module and the model iteration module include:

[0060] The experimental data analysis unit is used to quantitatively analyze the impact of the original nitrogen oxide emission ratio on the comprehensive liquid consumption cost and emission indicators of diesel engines under all operating conditions.

[0061] The evaluation and decision-making unit is used to construct an adaptive multi-dimensional comprehensive evaluation system for road spectrum scenarios, adjust the weights of evaluation indicators and perform quantitative scoring, and determine the optimal original emission nitrogen oxide ratio control range and the optimal solution for target fuel injection control parameters.

[0062] The execution control unit is used to decompose the optimal solution of the target fuel injection control parameters into execution control commands and send them to the corresponding execution units to realize real-time closed-loop control of the diesel engine operating parameters;

[0063] The incremental iteration unit is used to collect actual full-condition operating data of the diesel engine to construct an incremental training sample set. With compliance and safety boundaries as hard constraints, and combined with the engine aging and attenuation coefficient, the comprehensive fluid consumption prediction machine learning model is iteratively updated.

[0064] (III) Beneficial Effects

[0065] 1. A steady-state and transient dual-domain performance response proxy model is constructed based on an adaptive block-weighted neural network. By adaptively adjusting the granularity of the operating conditions and dynamically allocating the weights of the sub-models, combined with the design of transient time-series correlation sub-models, the data fitting accuracy of steady-state conventional operating conditions and dynamically switching operating conditions is ensured simultaneously. This effectively improves the problem of excessive prediction deviation of the model under actual road spectrum operating conditions and accurately represents the correlation between energy consumption and emissions under all operating conditions. At the same time, taking the original emission nitrogen oxide ratio as the core constraint, multi-objective collaborative optimization is carried out by combining the hard boundary of the whole machine operation safety, the dynamic constraint of SCR temperature field coupled with urea crystallization, and the transient road spectrum pre-weight constraint. This not only makes up for the shortcomings of traditional optimization dimensions being one-sided and detached from actual operating conditions, but also realizes the quantitative prediction of urea crystallization risk, avoids post-treatment blockage failures from the source of operation, and takes into account emission compliance, whole machine safety and adaptability to actual operating conditions.

[0066] 2. Construct an adaptive multi-dimensional comprehensive evaluation system for road spectrum scenarios. Relying on quantitative evaluation standards to replace the traditional subjective judgment method, it objectively balances the comprehensive fluid consumption cost of diesel engines, emission compliance margin, service life of downstream components, and the overall operational stability of the entire engine throughout its life cycle, eliminating the reliance on extensive practical experience in calibration work. On this basis, the optimal control strategy is transformed into standardized commands for execution, completing the coordinated closed-loop control of the fuel injection system and after-treatment system, accurately correcting the overall engine operating parameters in real time, and further improving the operational stability of the diesel engine across all operating conditions. Secondly, the iteration process strictly adheres to legal and equipment hard boundary constraints such as emission regulations, mechanical safety operating thresholds, and urea crystallization critical standards. At the same time, it combines the performance degradation law of the engine in long-term service for compensation and correction. This is different from the traditional extensive update method that only corrects prediction errors, avoiding the risk of inaccurate control parameters and excessive operating conditions after model iteration, and ensuring that the diesel engine operates stably and compliantly from factory calibration to the entire service life stage. Attached Figure Description

[0067] 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.

[0068] Figure 1 This is an overall flowchart of an embodiment of the present invention;

[0069] Figure 2 This is a flowchart illustrating the construction process of the steady-state-transient dual-domain prediction model in an embodiment of the present invention.

[0070] Figure 3 This is a multi-constraint coupled multi-objective optimization graph in an embodiment of the present invention;

[0071] Figure 4 This is a schematic diagram of the bench testing system in an embodiment of the present invention. Detailed Implementation

[0072] This application provides a machine learning-based method and system for optimizing the comprehensive fluid consumption of diesel engines, addressing the technical problems of insufficient accuracy in existing diesel engine fluid consumption calibration modeling, poor transient adaptability, uncontrollable risk of urea crystallization, lack of compliance restrictions on model iteration, and absence of comprehensive evaluation standards.

[0073] Based on the mainstream academic research results in the fields of internal combustion engine calibration, diesel engine energy consumption-emission synergistic optimization, and machine learning in internal combustion engine performance modeling, this study addresses the problems commonly mentioned in existing journal papers, such as high accuracy of laboratory models but poor adaptability to engineering implementation, complex mechanism models that cannot be embedded in real-time optimization processes, and the disconnect between steady-state calibration and transient real vehicle conditions. It combines the research with the needs of engineering implementation.

[0074] Example 1

[0075] refer to Figures 1 to 3 As shown, this embodiment uses a mainstream commercial diesel engine as the test platform, equipped with a China VI b standard SCR aftertreatment system. All tests were completed on an engine performance calibration bench, following the common single-factor control and parallel repeat test specifications in the internal combustion engine field to minimize test errors. Details are as follows:

[0076] During the basic data collection, the operating boundaries of the diesel engine under all operating conditions are first defined. The steady-state operating condition boundary is defined by the complete speed range from the engine's idle speed of 600 r / min to the rated speed of 2200 r / min, and the maximum torque curve corresponding to each speed. For example, the maximum torque is 2500 N·m at 1000 r / min and the rated torque is 1900 N·m at 2200 r / min, which fully covers the stable operation scenario of the engine. The transient operating condition boundary is based on the operating condition switching sequence of the WHTC global unified transient cycle specified in GB / T27840-2011, the CHTC commercial vehicle comprehensive road spectrum specified in JT / T1244-2019, and the standard road spectrum of highway trunk transportation. This conforms to more than 95% of the actual vehicle dynamic operation scenarios such as trunk transportation and urban distribution, avoiding the defects of only collecting steady-state data and deviating from the actual operating conditions.

[0077] Subsequently, it connects to multiple data acquisition access points across the entire operating range of the diesel engine, including: the original diesel engine electronic control unit, Xiangyi electric dynamometer, AVL fuel consumption analyzer, Horiba gas emission analyzer, micro carbon smoke meter, intake flow meter, bench integrated calibration system, SCR box inlet / bed / outlet three-stage temperature sensor, and Bosch urea injection controller. All devices complete synchronous trigger acquisition with a unified 10ms timestamp, synchronously retaining static characteristic data of steady-state operating conditions and time-series dynamic data of transient operating condition switching process, solving the technical problem of misaligned acquisition time of multiple devices and inability to correlate and analyze data.

[0078] After completing the raw data collection, we performed feature identification and standardized classification on the data, and fully implemented the technical connotation and design logic of each type of parameter. Among them, the power operation parameters include engine speed, cyclic fuel injection quantity, output torque, in-cylinder knock pressure, and turbine exhaust temperature. These are not only direct representations of the diesel engine's power output state, but also the core basis for setting subsequent operational safety boundaries. Fuel injection system parameters, i.e., core coupled optimization variables, such as main injection advance angle and injection rail pressure, directly affect fuel atomization and combustion efficiency, the amount of raw emissions generated in the cylinder, and the basic fuel consumption level. They are the core control levers of the entire optimization method. Emission characteristic parameters are based on the proportion of raw nitrogen oxides, and also include the concentration of nitrogen oxides in the exhaust terminal emissions, smoke... Unlike traditional technologies that only control exhaust emissions, this system balances fuel consumption and urea reducing agent dosage at the source. Smoke density is used to verify combustion integrity and avoid excessive optimization that could worsen combustion conditions. Aftertreatment system operating parameters include urea injection quantity, urea consumption rate, SCR box multi-stage temperature, and ammonia slip concentration, providing a continuous input basis for building a dynamic constraint model for urea crystallization risk. Operating condition boundary constraint parameters include engine coolant temperature, intercooler intake temperature, ambient temperature and humidity, and atmospheric pressure, used to compensate for data drift errors caused by different environmental conditions and overall engine thermal decay. Operating condition switching timing characteristic parameters include the rate of change of speed, torque, and fuel injection quantity between adjacent operating conditions, laying a solid foundation for subsequent modeling.

[0079] After completing the dataset, we can proceed to the construction of the steady-state-transient dual-domain performance response proxy model. This step addresses the shortcomings of a single global neural network, which has accurate steady-state fitting but excessive transient prediction bias. We first perform standardized preprocessing on the source data, including outlier removal, missing value completion, data temporal alignment, and spatiotemporal feature association. Specifically, we use the 3σ criterion from internal combustion engine data processing to remove abnormal jump data that exceed the normal distribution range, use linear interpolation to complete short-term missing fields, and rely on unified time-series nodes to complete multi-source data alignment and spatiotemporal feature association of adjacent operating conditions, forming a standardized and usable model training dataset to avoid model training distortion caused by abnormal data and temporal misalignment.

[0080] Subsequently, a basic neural network topology was built. Based on the input and output definitions, the input layer was connected to power operation parameters, fuel injection system parameters, and operating condition switching timing characteristic parameters. The output layer corresponds to four core evaluation indicators: fuel consumption rate, original nitrogen oxide emission ratio, urea consumption rate, and smoke concentration, which maps the inherent correlation between operating condition parameters and fuel consumption and emission indicators.

[0081] To address the inherent limitations of a single global network, an adaptive block-weighted neural network architecture is adopted. First, the entire speed-load range of the aircraft is divided into initial uniform grid blocks of 100 r / min × 20% load. The spatiotemporal correlation characteristics of adjacent operating conditions are incorporated into the block division. The nonlinearity and operating condition correlation of the data within each block are calculated, i.e., the fluctuation amplitude of the data fitting residual and the parameter change linkage coefficient between that block and adjacent blocks. The block granularity is adaptively adjusted based on the nonlinearity value. For blocks with nonlinearity > 0.7 and operating condition correlation > 0.6 in the rapid acceleration and deceleration operating conditions, we refine them into a fine grid of 25 r / min × 5% load. For blocks with a nonlinearity of <0.3 concentrated in steady-state conditions such as idling and cruising, we relaxed the grid to a coarse grid of 200 r / min × 40% load. At the same time, based on the effective number of samples, data dispersion coefficient, and operating condition coverage ratio of each block, we adaptively allocated the membership function weights of the corresponding low-order linear fitting sub-models for each block. For example, the 1100-1500 r / min economic speed range block with more than 5000 samples covering the core operating conditions of trunk transportation was set to a weight of 0.4, while the extreme operating condition block with a smaller sample size was set to a weight of 0.05. This ensured that the fitting accuracy of the core operating conditions was prioritized, while taking into account the overall fitting accuracy of the model and computational efficiency.

[0082] To further address the shortcomings in transient fitting, we constructed a separate transient time-series correlation sub-model according to the established parameters. Using the time-series characteristic parameters of operating condition switching as the core input, we captured the dynamic response patterns during the operating condition switching process. In the sample expansion phase, we first generated a gradient combination array of the main injection advance angle and injection rail pressure through orthogonal experimental design. After completing the basic steady-state operating condition data collection, we entered the machine learning iteration phase of active experimental design to supplement the scarce sample data of the critical points of operating condition switching for WHTC and CHTC road spectrums, increasing the operating condition coverage of the dataset by 30%. During model training, we executed an iteration termination condition and adopted a common model evaluation metric: after five consecutive iterations, the steady-state operating condition model's coefficient of determination R0... 2 Stable R value above 0.85, transient operating condition model 2 If the stability is higher than 0.8 and the difference between the maximum and minimum values ​​of the iteration results is less than 0.1, the training is terminated. After training, multiple sets of parameters from typical steady-state working conditions and transient switching points are randomly selected for bench benchmarking verification. Only when the steady-state mean relative error (MAPE) of the model is controlled within 5% and the transient MAPE is controlled within 6% is the model accuracy deemed to be up to standard, and the comprehensive liquid consumption prediction machine learning model is finally obtained.

[0083] During use, it was found that existing optimization constraints were too simplistic, lacked post-processing coupling mechanisms, and were detached from real-world road conditions. Therefore, we also investigated multi-objective optimization solutions. Details are as follows:

[0084] First, based on the mechanical structure durability limit of this engine model, we implemented clearly defined preset operational safety boundaries, namely, setting the in-cylinder explosion pressure to no more than 195 bar and the turbine inlet exhaust temperature to no more than 720°C. The 195 bar cylinder pressure limit is based on the engine block design durability limit of 200 bar with a 5 bar safety margin, and the 720°C turbine inlet exhaust temperature limit is based on the turbocharger's high temperature resistance limit of 750°C with a 30°C safety margin. These serve as insurmountable hard constraints throughout the optimization process, preventing mechanical damage during optimization. Second, addressing the pain point that existing studies on urea crystallization risk often rely on complex CFD simulations and cannot be embedded in real-time optimization processes, we constructed a model for monitoring the original emissions of nitrogen oxides and SCR temperature. A dynamic constraint model for urea crystallization risk coupled with temperature field is proposed. This model is based on the fundamental mechanism of the exothermic reaction of SCR catalytic reduction. The inputs are the original emission nitrogen oxide ratio, the inlet temperature of the SCR chamber, the bed temperature of the SCR chamber, and the urea injection rate. The model quantifies the bidirectional influence of the original emission nitrogen oxide ratio on the SCR temperature field. That is, an increase in the original nitrogen oxide concentration will aggravate the exothermic reaction of SCR and increase the temperature inside the chamber, while an excessively low concentration will lead to temperature field instability and cooling. Both of these will increase the risk of urea crystallization. The model finally outputs the real-time urea crystallization risk coefficient in the 0-1 range, realizing the quantitative prediction of crystallization risks. It retains the accuracy of the mechanism in academic research and meets the real-time requirements of engineering optimization.

[0085] To address the pre-constraints on the weights of transient road spectrum operating conditions, we decomposed the high-speed road spectrum, CHTC road spectrum, and WHTC cyclic operating conditions into intervals, statistically analyzed the time proportion and switching frequency of each operating condition interval, and generated a transient road spectrum operating condition weight distribution table. For the trunk line transportation application scenario of this aircraft model, the weight of the high-speed steady-state operating condition was set to 70%, and the weight of the transient acceleration / deceleration operating condition was set to 30%, pre-embedded in the optimization logic. Subsequently, we set five sets of gradient-distributed original NOx emission ratio target values ​​(200ppm, 300ppm, 400ppm, 500ppm, 600ppm) as the core constraints for optimization. Each set of original NOx emission ratio target values ​​was a fixed constraint, with the dual optimization objectives of lowest overall liquid consumption cost and optimal smoke emission, and the main injection advance angle and injection rail pressure as coupling optimization factors. By transforming variables and combining hard constraints on operational safety, soft constraints on urea crystallization risk coefficient ≤0.3, and the precondition of transient path spectrum weight, a steady-state-transient dual-domain coupled multi-objective optimization function is constructed. The mainstream NSGA-Ⅲ multi-objective optimization algorithm in the field of internal combustion engine optimization is used for full-domain iterative solution. The solution follows filtering rules, and the urea crystallization risk is pre-verified in real time for each group of generated parameter schemes. At the same time, based on the transient adaptability of the schemes evaluated by transient path spectrum weight, invalid schemes with excessive crystallization risk and inadequate transient adaptability are filtered out. Finally, the combination of main injection advance angle and injection rail pressure with the lowest comprehensive liquid consumption cost and the best smoke emission under the original nitrogen oxide emission target value is calculated for each group, forming the optimal scheme of injection control parameters for the corresponding target value. The complete set of optimal schemes for 5 groups of schemes is obtained.

[0086] After receiving the test scheme set, we conducted bench benchmarking verification tests. The test design needed to follow the comparative test specifications in the field of internal combustion engines. That is, firstly, we determined the verification test conditions, covering steady-state characteristic conditions and typical transient road spectrum conditions. Among them, the steady-state characteristic conditions included rated conditions, minimum fuel consumption conditions, and low-speed low-load conditions, while the typical transient road spectrum conditions included high-speed road spectrum, CHTC road spectrum, and WHTC cycle conditions. We covered the core steady-state points of the engine model's daily operation and the mainstream dynamic scenarios of actual operation to ensure that the verification was comprehensive. We then identified the optimal solution for each set of injection control parameters and the corresponding main... The injection advance angle and injection rail pressure calibration pulse spectrum were written into the diesel engine electronic control unit. Steady-state characteristic condition test and transient road spectrum condition test of each scheme were carried out sequentially on the engine test bench. Each test was set up with 3 parallel repetitions. Fuel consumption data, urea consumption data, comprehensive liquid consumption cost data, smoke emission data and after-treatment system operation status data were collected simultaneously for each scheme. For transient conditions, additional data on SCR box temperature change, ammonia slip concentration and urea crystallization risk correlation during the condition switching process were collected to ensure that the data covers all evaluation dimensions.

[0087] Throughout the experiment, the single-variable control principle was followed. The opening of the urea injection metering valve was adjusted in real time using a PID closed-loop algorithm to consistently keep the terminal nitrogen oxide emission concentration within the limit of 350 mg / kW·h specified in the China VI b standard. The only variable was the original nitrogen oxide emission ratio. At the same time, boundary conditions such as engine coolant temperature, intake pressure, and ambient temperature and humidity were kept constant throughout the experiment. Safety boundary parameters such as cylinder knock pressure and turbine exhaust temperature were monitored simultaneously to ensure that all differences in the experimental data were caused only by changes in the original nitrogen oxide emission ratio, avoiding interference from other variables and ensuring the authenticity and comparability of the experimental data. Finally, the experimental data passed the significance test commonly used in academic research, verifying the effectiveness of the optimization scheme.

[0088] Based on the complete data collected from the experiment, the scheme decision-making, closed-loop control, and model iteration update are carried out. In the scheme decision-making stage, the verification test data is first analyzed in depth to clarify the quantitative impact of the original nitrogen oxide emission ratio on the fuel consumption, urea consumption, comprehensive liquid consumption cost, and smoke emission of the diesel engine under steady-state and transient operating conditions. An adaptive multi-dimensional comprehensive evaluation system for road spectrum scenarios is constructed. According to the weight distribution table of transient road spectrum operating conditions of the target application scenario, the weight coefficients of comprehensive liquid consumption cost, smoke emission, urea crystallization risk, emission compliance margin, and operational safety margin are adaptively adjusted. For the trunk transportation scenario, we set the weight of comprehensive liquid consumption cost to 0.4, the weight of emission compliance margin to 0.2, the weight of urea crystallization risk to 0.2, the weight of smoke emission to 0.1, and the weight of operational safety margin to 0.1.

[0089] The optimal range was determined based on the established standards: the overall liquid consumption cost reduction slowed down, urea consumption did not increase sharply, the risk of urea crystallization was controllable, the smoke emission reduction was significant, and the emission compliance margin met the requirements of the entire life cycle. Combined with the quantitative scoring of the comprehensive evaluation system, the optimal original emission nitrogen oxide ratio control range for this model was finally determined to be 300-400ppm. The optimal solution of the fuel injection control parameters corresponding to this range was determined as the final target solution.

[0090] During closed-loop control, the optimal solution of the target injection control parameters is decomposed into execution control commands that can be recognized by the diesel engine fuel injection system and after-treatment system, and sent to the corresponding execution units such as the injector and urea injection metering valve. This enables real-time coordinated closed-loop control of diesel engine injection parameters and after-treatment parameters, corrects parameter deviations caused by operating condition fluctuations in real time, and keeps the closed-loop control response speed within 50ms, ensuring the stability and consistency of diesel engine operating status.

[0091] During model iteration, an incremental learning method with full lifecycle compliance boundary constraints is implemented. Addressing the shortcomings of existing research where model iteration focuses solely on accuracy while neglecting compliance, actual full-condition operating data of the diesel engine under parameter control is collected monthly at a preset sampling period. This data, along with model predictions and measured results, forms an incremental training sample set. During iteration, the China VI b emission regulations boundary, engine operating safety boundary, and urea crystallization risk threshold are set as insurmountable hard constraints. Any model weight updates exceeding these constraints are directly locked. Simultaneously, the engine aging degradation coefficient obtained through a 1500-hour whole-engine aging test (e.g., a 1.2% reduction in output torque every 1000 hours) is used to compensate for model predictions in real time. The incremental learning method completes the iterative update of the comprehensive fluid consumption prediction machine learning model, preventing the model from exceeding the compliance and safe operating range after iteration. Verified by a 1000-hour aging test, the model's coefficient of determination R after long-term service is [value missing]. 2 It can still maintain a value above 0.8, ensuring the model's adaptability and reliability throughout the entire life cycle of the diesel engine.

[0092] Example 2

[0093] The modular integrated system built in this embodiment is compatible with that in Embodiment 1. The system hardware uses the NIPXIe measurement and control platform as the core controller, and the software is jointly developed based on LabVIEW and Python. It completes communication adaptation with the engine bench measurement and control system, diesel engine electronic control unit, and various acquisition sensors to form a complete automated operation link.

[0094] The overall system architecture consists of seven functional modules. Among them, the full-condition data acquisition module is responsible for collecting the front-end full-condition operation source data, providing basic data support for the subsequent whole process.

[0095] The dual-domain model building module uses the collected operational source data as the training basis and completes the construction of the steady-state-transient dual-domain performance response proxy model through the adaptive block weighted neural network algorithm. After iterative training and full-condition accuracy verification, it outputs a finalized comprehensive liquid consumption prediction machine learning model.

[0096] The coupled optimization module takes the original nitrogen oxide emission ratio as the core constraint, the lowest comprehensive liquid consumption cost and the best smoke emission as the dual optimization objectives, and the core fuel injection parameters as coupled optimization variables. It completes the construction of a dual-domain coupled multi-objective optimization function by combining multi-layer constraint conditions, and outputs the optimal fuel injection control scheme under different nitrogen oxide targets through a multi-objective optimization algorithm.

[0097] The bench verification module is responsible for writing multiple optimal solutions into the diesel engine electronic control unit, completing the verification test under steady-state and transient conditions on the bench, and simultaneously collecting full-dimensional measured data corresponding to each solution.

[0098] The optimal solution decision module performs in-depth analysis of the verification test data and, combined with the road spectrum scenario adaptive multi-dimensional comprehensive evaluation system, determines the optimal nitrogen oxide ratio control range and corresponding target solution that takes into account economy, compliance, reliability and life cycle stability.

[0099] The closed-loop control module sends the target scheme to the execution unit of the fuel injection and aftertreatment system to realize real-time closed-loop control of the diesel engine operating parameters;

[0100] The model iteration module collects actual operating data after parameter control and completes the iterative update of the model through an incremental learning method with full lifecycle compliance boundary constraints.

[0101] To enable the module to be implemented, the full-condition data acquisition module and the dual-domain model construction module are equipped with three types of dedicated functional units. The working condition boundary delineation unit has a built-in library of various standard road spectrum working conditions specified by national standards. It can automatically match the corresponding steady-state and transient working condition boundaries according to the application scenario of the target model. For example, for mining engineering machinery models, it can automatically supplement the boundary definitions of special working conditions such as idling standby and high torque excavation, and accurately standardize the working condition range of data acquisition and model training.

[0102] The data acquisition and preprocessing unit uses a unified 10ms timestamp to synchronously trigger all acquisition devices. It incorporates a 3σ outlier removal algorithm, a linear interpolation completion algorithm, and a temporal alignment algorithm, enabling fully automated preprocessing of raw data without manual intervention. This significantly improves data processing efficiency compared to manual processing, ultimately outputting a compliant and usable model training dataset. The neural network construction and training unit incorporates an adaptive block weighting algorithm module, a transient temporal correlation sub-model module, an iterative convergence judgment module, and a bench benchmarking verification module. This enables fully automated operation of the model from training to finalization, ultimately outputting a comprehensive liquid consumption prediction machine learning model with satisfactory accuracy.

[0103] The coupling optimization module and bench verification module are equipped with four dedicated functional units: The constraint configuration unit integrates a parameter storage module and a dynamic model calculation module. It can permanently store the diesel engine's preset operating safety boundaries and multiple sets of gradient distributions of original nitrogen oxide emission ratio constraints. In the background, it runs a dynamic constraint model for urea crystallization risk in real time and outputs a quantified risk coefficient, while automatically generating a transient road spectrum operating condition weight distribution table. The road spectrum weight unit integrates an automatic road spectrum decomposition algorithm, which can decompose the transient road spectrum operating conditions of the target application scenario into intervals, statistically analyze the time proportion and switching frequency of each operating condition interval, and provide a preliminary weight basis for optimization solutions. For example, in urban delivery scenarios, it can automatically decompose low-speed, low-load... The weight of operating conditions has been increased from 30% to 60%. The optimization solution unit has a built-in multi-objective optimization function construction module, NSGA-Ⅲ optimization algorithm module, and invalid scheme filtering module. It can automatically complete the construction of dual-domain coupled multi-objective optimization functions, global iterative solution, and high-risk invalid scheme filtering, and finally generate a standardized set of optimal solutions for fuel injection control parameters. The test execution unit has a built-in standardized test process library, electronic control parameter one-click writing module, and multi-channel synchronous data acquisition module. It can automatically set the operating condition type and execution process of the verification test, complete the calibration pulse spectrum of the diesel engine electronic control unit with one click, and simultaneously collect full-dimensional measured data such as fuel consumption, emissions, aftertreatment, and safety boundaries during the test.

[0104] The optimal solution decision-making module, closed-loop control module, and model iteration module are comprised of four dedicated functional units: The data analysis unit incorporates a multi-dimensional data correlation and parsing algorithm, capable of quantifying the impact of the original nitrogen oxide emission ratio on the comprehensive fuel consumption cost and emission indicators of the diesel engine under all operating conditions, providing precise data support for solution decision-making; the evaluation and decision-making unit incorporates a road spectrum scenario adaptive multi-dimensional comprehensive evaluation system, which can automatically adjust the weight ratio of each evaluation indicator according to the target application scenario, complete the quantitative scoring and ranking of solutions, and ultimately determine the optimal original nitrogen oxide emission ratio control range and the corresponding optimal solution for target fuel injection control parameters, eliminating reliance on the manual experience of calibration engineers; the execution control unit incorporates a control command decomposition module and a real-time communication module. The block-type, closed-loop PID control module can decompose the top-level target optimization scheme into execution control commands that can be recognized by the underlying hardware, and send them to the corresponding execution units of the fuel injection system and after-treatment system to achieve real-time coordinated closed-loop control of diesel engine operating parameters, with a control response speed of less than 50ms. The incremental iteration unit has built-in incremental sample construction module, compliance boundary locking module, and incremental learning training module, which can automatically collect actual full-condition operating data of diesel engine to build an incremental training sample set. During the iteration process, emission regulations, mechanical safety, and crystallization threshold are used as insurmountable hard constraints. At the same time, compensation and correction are completed by combining the engine aging and decay coefficient, realizing the compliant incremental iterative update of the comprehensive liquid consumption prediction machine learning model, ensuring the stability and adaptability of the model in long-term service.

[0105] 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. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A method for optimizing the overall fluid consumption of a diesel engine based on machine learning, characterized in that, The steps of this method are as follows: Multiple sets of operating source data are acquired within the full operating range of the diesel engine. The operating source data includes power operating parameters, fuel injection system parameters, emission characteristic parameters, after-treatment system operating parameters, operating condition boundary constraint parameters, and operating condition switching timing characteristic parameters. Based on the aforementioned source data, an adaptive block weighted neural network algorithm is used to construct a diesel engine steady-state-transient dual-domain performance response proxy model. After model training and full-condition accuracy verification, a comprehensive liquid consumption prediction machine learning model is obtained. With the original nitrogen oxide emission ratio as the core constraint, the dual optimization objectives are the lowest comprehensive liquid consumption cost and the best smoke emission. The main injection advance angle and injection rail pressure are the coupled optimization variables. Combining the preset operating safety boundary, the dynamic constraint of urea crystallization risk coupled with the original nitrogen oxide emission and SCR temperature field, and the pre-constraint of transient road spectrum operating conditions, a steady-state-transient dual-domain coupled multi-objective optimization function is constructed. The optimal scheme of injection control parameters under multiple sets of different original nitrogen oxide emission ratio targets is obtained by solving the multi-objective optimization algorithm. The optimal solutions for multiple sets of fuel injection control parameters were written into the diesel engine electronic control unit. Verification tests were conducted on the engine bench under steady-state characteristic conditions and typical transient road spectrum conditions. Data on fuel consumption, urea consumption, comprehensive liquid consumption cost, smoke emissions, and after-treatment system operation status were collected for each solution. By analyzing and verifying test data and combining the road spectrum scenario adaptive multi-dimensional comprehensive evaluation system, the optimal original emission nitrogen oxide ratio control range and the corresponding target fuel injection control parameter optimal scheme were determined. The optimal solution for the target fuel injection control parameters is sent to the execution units of the diesel engine fuel injection system and after-treatment system; the actual full-condition operation data of the diesel engine after execution is collected, and the comprehensive fuel consumption prediction machine learning model is iteratively updated through the incremental learning method of full life cycle compliance boundary constraints.

2. The method for optimizing the comprehensive fluid consumption of a diesel engine based on machine learning according to claim 1, characterized in that, Acquiring multiple sets of operating source data across the entire operating range of the diesel engine includes: The operating range of the diesel engine under all operating conditions is defined, with the maximum torque curve corresponding to each speed from idle speed to rated speed as the boundary of steady-state operating conditions, and the operating condition switching sequence of high-speed road spectrum, CHTC road spectrum, and WHTC cycle as the boundary of transient operating conditions. It connects to multiple data acquisition access points across the entire operating range of the diesel engine, synchronously collects raw operating data under all operating conditions, and records the characteristic data of steady-state operating conditions and the time-series dynamic data of the transient operating condition switching process; The raw operating data is subjected to feature identification and classification to obtain power operating parameters, fuel injection system parameters, emission characteristic parameters, after-treatment system operating parameters, operating condition boundary constraint parameters, and operating condition switching timing characteristic parameters. The power operation parameters include engine speed, cyclic fuel injection quantity, output torque, in-cylinder detonation pressure, and turbine exhaust temperature; the fuel injection system parameters include main injection advance angle and injection rail pressure; the emission characteristic parameters include the proportion of nitrogen oxides in the original emissions, the concentration of nitrogen oxides in the exhaust gas, and the smoke concentration; the aftertreatment system operation parameters include urea injection quantity, urea consumption rate, SCR box temperature, and ammonia slip concentration; and the operating condition switching timing characteristic parameters are the change rate of speed, torque, and fuel injection quantity in adjacent operating conditions.

3. The method for optimizing the comprehensive fluid consumption of a diesel engine based on machine learning according to claim 1, characterized in that, The construction of a diesel engine steady-state-transient dual-domain performance response proxy model includes: Preprocessing of the source data includes outlier removal, missing value completion, data time-series alignment, and feature association to form the model training dataset; Using power operation parameters, fuel injection system parameters, and operating condition switching timing characteristic parameters as the model input layer, and fuel consumption rate, original nitrogen oxide emission ratio, urea consumption rate, and smoke concentration as the model output layer, a basic neural network topology is built. The entire working condition range is divided into grid blocks, incorporating the correlation features of adjacent working conditions, calculating the nonlinearity and working condition correlation of each block of data, and adaptively adjusting the block granularity and the membership function weights of the corresponding sub-models. A time-series correlation sub-model is constructed for the transient operating condition switching interval, with the dynamic change characteristics of adjacent operating conditions as the core input; The combination of fuel injection control parameters is generated by the basic experimental design method. After collecting basic steady-state operating condition data, the system enters the machine learning iteration stage of active experimental design to supplement the sample data of transient operating condition switching points. The model training dataset is input into the neural network topology for iterative training. An iteration termination condition is set for surrogate model training. After bench verification confirms that the model accuracy meets the standard, the comprehensive liquid consumption prediction machine learning model is obtained.

4. The method for optimizing the comprehensive fluid consumption of a diesel engine based on machine learning according to claim 1, characterized in that, Constructing a steady-state-transient dual-domain coupled multi-objective optimization function and solving for the optimal injection control parameters includes: The preset operating safety boundaries for the diesel engine are set as follows: the cylinder explosion pressure should not exceed 195 bar and the turbine exhaust temperature should not exceed 720°C. A dynamic constraint model for urea crystallization risk coupled with the original emission nitrogen oxides and the SCR temperature field is constructed. The original emission nitrogen oxide ratio, SCR box temperature and urea injection rate are used as inputs to quantify the bidirectional influence of the original emission nitrogen oxide ratio on the SCR temperature field and output the real-time urea crystallization risk coefficient. The transient road spectrum conditions of the target application scenario are decomposed into intervals, and the time proportion and switching frequency of each condition interval are statistically analyzed to generate a transient road spectrum condition weight distribution table as a pre-constraint. Multiple sets of gradient distributions are set as the target values ​​for the proportion of nitrogen oxide emissions. Each set of target values ​​is a fixed constraint. Combined with hard constraints on operational safety, soft constraints on urea crystallization risk, and road spectrum weight preconditions, the steady-state-transient dual-domain coupled multi-objective optimization function is constructed. The optimization function is solved by a multi-objective optimization algorithm. The generated parameter schemes are then subjected to real-time pre-verification of urea crystallization risk and transient adaptability assessment. After filtering out invalid schemes, the optimal fuel injection control parameters corresponding to the original nitrogen oxide emission ratio target value for each group are obtained and summarized to form an optimal scheme set.

5. The method for optimizing the comprehensive fluid consumption of a diesel engine based on machine learning according to claim 1, characterized in that, The verification test includes: The steady-state characteristic conditions for the verification test were determined to be rated conditions, minimum fuel consumption conditions, and low-speed, low-load conditions, while the typical transient road spectrum conditions were high-speed road spectrum, CHTC road spectrum, and WHTC cycle conditions. The main injection advance angle and injection rail pressure pulse spectrum corresponding to the optimal scheme of each group of injection control parameters were written into the diesel engine electronic control unit, and steady-state characteristic conditions and typical transient road spectrum conditions of each scheme were tested on the engine bench in sequence. Data on fuel consumption rate, urea consumption rate, comprehensive liquid consumption cost, smoke emission, and after-treatment system operation status were collected during the test. Data on SCR box temperature, ammonia slip, and urea crystallization risk during transient operating condition switching were recorded simultaneously. During the test, the urea injection quantity was adjusted in a closed loop to maintain the concentration of nitrogen oxides in diesel engine exhaust gas in compliance with current emission standards. The single variable was controlled as the original nitrogen oxide emission ratio, and the diesel engine operating safety boundary parameters were recorded simultaneously.

6. The method for optimizing the comprehensive fluid consumption of a diesel engine based on machine learning according to claim 1, characterized in that, Determining the optimal initial nitrogen oxide emission ratio control range and iteratively updating the comprehensive liquid consumption prediction machine learning model includes: Analyze and verify the test data to clarify the quantitative impact of the original nitrogen oxide emission ratio on fuel consumption, urea consumption, comprehensive liquid consumption cost, and smoke emission of diesel engines under steady-state and transient operating conditions; Construct an adaptive multi-dimensional comprehensive evaluation system for road spectrum scenarios, and adjust the weight coefficients of each evaluation indicator in the scheme evaluation according to the transient road spectrum working condition weight distribution table of the target application scenario; Based on the criteria of a slowing decline in overall liquid consumption cost, no sharp increase in urea consumption, controllable risk of urea crystallization, significant reduction in smoke emissions, and emission compliance margin meeting the requirements of the entire life cycle, and combined with the quantitative scoring of the comprehensive evaluation system, the optimal original emission nitrogen oxide ratio control range and the corresponding optimal solution for target fuel injection control parameters are determined. The optimal solution of the target fuel injection control parameters is decomposed into execution control commands, which are then sent to the corresponding execution units of the diesel engine fuel injection system and after-treatment system to achieve real-time closed-loop control of fuel injection and after-treatment parameters. The actual full-condition operating data of the diesel engine is collected according to the preset sampling period. The actual operating data, model prediction results and measured results are combined to form an incremental training sample set. The current emission regulations boundary, engine operating safety boundary and urea crystallization risk threshold are used as hard constraints. Combined with the engine aging and decay coefficient, the comprehensive liquid consumption prediction machine learning model is iteratively updated through the incremental learning method of full life cycle compliance boundary constraints.

7. A diesel engine integrated fluid consumption optimization system based on machine learning, characterized in that, The system includes: The full-condition data acquisition module is used to acquire multiple sets of operating source data within the full operating range of the diesel engine. The operating source data includes power operating parameters, fuel injection system parameters, emission characteristic parameters, after-treatment system operating parameters, operating condition boundary constraint parameters, and operating condition switching timing characteristic parameters. The dual-domain model construction module is used to construct a diesel engine steady-state-transient dual-domain performance response proxy model based on the running source data and through an adaptive block weighted neural network algorithm. After model training and full-condition accuracy verification, it outputs a comprehensive liquid consumption prediction machine learning model. The coupled optimization module is used to take the original nitrogen oxide emission ratio as the core constraint, the lowest comprehensive liquid consumption cost and the best smoke emission as the dual optimization objectives, and the main injection advance angle and injection rail pressure as coupled optimization variables. It combines the preset operating safety boundary, the dynamic constraint of urea crystallization risk coupled with the original nitrogen oxide emission ratio and the SCR temperature field, and the transient road spectrum operating condition weight pre-constraint to construct a steady-state-transient dual-domain coupled multi-objective optimization function. The optimal solution of injection control parameters under multiple sets of different original nitrogen oxide emission ratio targets is obtained by solving the multi-objective optimization algorithm. The bench verification module is used to write the optimal schemes of multiple sets of fuel injection control parameters into the diesel engine electronic control unit, and to conduct verification tests on the engine bench under steady-state characteristic conditions and typical transient road spectrum conditions. It collects relevant data on fuel consumption, urea consumption, comprehensive liquid consumption cost, smoke emission and after-treatment system operation status corresponding to each scheme. The optimal solution decision module is used to analyze and verify test data, and combine the road spectrum scenario adaptive multi-dimensional comprehensive evaluation system to determine the optimal original emission nitrogen oxide ratio control range and the corresponding target fuel injection control parameters optimal solution. The model iteration module is used to send the optimal solution of the target injection control parameters to the execution unit of the diesel engine fuel injection system and after-treatment system; and to collect the actual full-condition operation data of the diesel engine after execution, and to iteratively update the comprehensive fuel consumption prediction machine learning model through the incremental learning method of full life cycle compliance boundary constraints.

8. The diesel engine integrated fluid consumption optimization system based on machine learning according to claim 7, characterized in that, The full-condition data acquisition module and the dual-domain model construction module include: The operating condition boundary delineation unit is used to delineate the steady-state and transient operating condition boundaries of the diesel engine under all operating conditions. The data acquisition and preprocessing unit is used to connect to multiple data acquisition access points, synchronously collect raw running data and preprocess it, and output the model training dataset. The neural network training unit is used to build the basic topology of the neural network, perform working condition grid segmentation, adaptive weight allocation, construction of transient time series correlation sub-models and iterative training of the model, and output the comprehensive liquid consumption prediction machine learning model after full working condition accuracy verification.

9. A diesel engine integrated fluid consumption optimization system based on machine learning according to claim 7, characterized in that, The coupling optimization module and bench verification module include: The constraint weight configuration unit is used to store the preset operating safety boundary of the diesel engine, the original emission nitrogen oxide ratio constraint target, run the urea crystallization risk dynamic constraint model and output the real-time risk coefficient, and generate the transient road spectrum operating condition weight distribution table. The optimization solution unit is used to construct a steady-state-transient dual-domain coupled multi-objective optimization function, solve and filter invalid schemes through a multi-objective optimization algorithm, and generate the optimal scheme set of fuel injection control parameters. The bench test execution unit is used to set the operating conditions and procedures for the verification test, write parameters to the diesel engine electronic control unit, and simultaneously collect various data during the test process, as well as parameters related to diesel engine operation safety and after-treatment system operation.

10. A diesel engine integrated fluid consumption optimization system based on machine learning according to claim 7, characterized in that, The optimal solution decision module and the model iteration module include: The experimental data analysis unit is used to quantitatively analyze the impact of the original nitrogen oxide emission ratio on the comprehensive liquid consumption cost and emission indicators of diesel engines under all operating conditions. The evaluation and decision-making unit is used to construct an adaptive multi-dimensional comprehensive evaluation system for road spectrum scenarios, adjust the weights of evaluation indicators and perform quantitative scoring, and determine the optimal original emission nitrogen oxide ratio control range and the optimal solution for target fuel injection control parameters. The execution control unit is used to decompose the optimal solution of the target fuel injection control parameters into execution control commands and send them to the corresponding execution units to realize real-time closed-loop control of the diesel engine operating parameters; The incremental iteration unit is used to collect actual full-condition operating data of the diesel engine to construct an incremental training sample set. With compliance and safety boundaries as hard constraints, and combined with the engine aging and attenuation coefficient, the comprehensive fluid consumption prediction machine learning model is iteratively updated.

Images