A diesel engine transient condition XGBoost-LSTM time series prediction control method

Abstract

The present application belongs to the technical field of industrial process control and time series prediction, and in particular to a diesel engine transient condition XGBoost-LSTM time series prediction control method. The method comprises the following steps: S1: obtaining a multi-dimensional sensor operation data set of a diesel engine under typical transient conditions; S2: constructing a classification training data set for identifying a model; S3: performing spontaneous identification and endogenous constraint training of an XGBoost model on a risk boundary of the diesel engine under transient conditions; S4: performing generalization performance quantization verification of the XGBoost model; S5: performing LSTM end-to-end prediction training; S6: performing generalization performance quantization verification of the LSTM model; and S7: performing architecture development and performance evaluation of the XGBoost-LSTM prediction model. The present application can effectively solve the problems of signal phase difference and response delay of the diesel engine under transient conditions, and greatly improves the real-time performance and accuracy of the prediction instruction of the diesel engine under transient conditions.

Description

Technical Field

[0001] This invention relates to the field of industrial process control and timing prediction technology, specifically to a diesel engine transient condition XGBoost-LSTM timing prediction control method. Background Technology

[0002] In the field of complex thermodynamic machinery, exemplified by diesel engines, the combined effects of the inertia of fluid transmission in the complex intake and exhaust gas paths, multiphase flow coupling, and the physical inertia of electromechanical actuators result in significant signal phase differences and time delays in diesel engine systems under transient conditions such as sudden load increases, decreases, or accelerations. Due to significant electronic control delays, mechanical frictional resistance, and nonlinear dynamic response lags in key air valves and fluid control actuators (such as exhaust gas turbochargers and thermostats), they cannot promptly follow sudden changes in command. Furthermore, combined with the inherent volumetric effect and transport lag in the intake and exhaust pipes, this multi-layered physical delay, primarily driven by actuator response lag, causes traditional control strategies to exhibit severe lag and distortion when handling dynamic transient changes. This makes it extremely easy for key dynamic and thermodynamic performance parameters of diesel engines to experience control overshoot or severe oscillations during transient transitions.

[0003] Currently, ensemble learning and recurrent neural networks are widely used in nonlinear time series prediction. However, these purely data-driven models often treat the diesel engine as a pure "black box," completely ignoring the inherent physical and dynamic boundaries of air valves and fluid control actuators. Due to the lack of underlying constraints in physical logic, the predicted instruction sequences often fail at the physical execution level because they cannot adapt to the real hysteresis characteristics of mechanical components, resulting in jumps that do not conform to the laws of physical evolution. Furthermore, conventional models exhibit significant "logical inertia" in their perception of sudden changes in diesel engine operating conditions, making it difficult to spontaneously achieve effective feedforward cancellation of fluid transmission and mechanical response delays on the time axis.

[0004] Existing single methods cannot achieve a good balance between prediction accuracy, computational speed, and control cost. Therefore, the field of diesel engine transient control urgently needs a technical solution that can overcome the above-mentioned shortcomings. This solution can be achieved by deeply embedding the unique fluid transport response mechanism of the diesel engine and the first-order hysteresis dynamics of the actuator as intrinsic constraints into heterogeneous neural network architectures such as XGBoost-LSTM. By constructing end-to-end feedforward target reconstruction and phase lead anchoring at the algorithm's bottom layer, physical feedforward compensation for the phase difference of the system's multidimensional response can be achieved. This ensures high-precision prediction of sudden change risks under severe transient conditions and ensures the smooth dynamic execution of the final physical execution command, thereby maintaining the optimal performance level of the diesel engine during transient operation. Summary of the Invention

[0005] The purpose of this section is to outline some aspects of the embodiments of the present invention and to briefly describe some preferred embodiments. Simplifications or omissions may be made in this section, as well as in the abstract and title of this application, to avoid obscuring the purpose of these documents; however, such simplifications or omissions should not be construed as limiting the scope of the invention.

[0006] To address the aforementioned technical problems, according to one aspect of the present invention, the present invention provides the following technical solution:

[0007] A transient XGBoost-LSTM time-series predictive control method for diesel engines includes the following steps:

[0008] S1: Build a transient operation monitoring platform for diesel engines and obtain multi-dimensional sensor operation datasets of diesel engines under typical transient operating conditions;

[0009] S2: Based on the physical dynamics limits and transient fluctuation characteristics of diesel engines, the running dataset is preprocessed and adaptively labeled to construct a classification training dataset for the identification model;

[0010] S3: Introduce a threshold endogenous self-calibration mechanism based on the characteristics of sample distribution proportion to spontaneously identify and train the XGBoost model on the risk boundary of transient operating conditions of diesel engines.

[0011] S4: Conduct quantitative verification of the generalization performance of the XGBoost model based on multi-criteria evaluation metrics;

[0012] S5: LSTM end-to-end prediction training is performed using an LSTM architecture that integrates deep hysteresis physics formulas and penalty incentive mechanisms.

[0013] S6: Perform quantitative verification of the generalization performance of the LSTM model based on multiple evaluation criteria;

[0014] S7: Develop the architecture and evaluate the performance of the XGBoost-LSTM prediction model.

[0015] As a preferred embodiment of the XGBoost-LSTM timing predictive control method for transient diesel engine conditions described in this invention, the multi-dimensional sensor in S1 is a multi-dimensional industrial sensor configured at key physical nodes of the diesel engine to synchronously collect environmental state signals and physical feedback signals of the actuators at a preset sampling period; the state signals cover physical characteristics such as fluid dynamic parameters, thermodynamic temperature, or mechanical displacement during system operation; the physical feedback signals are the actual response action parameters of the actuators under control commands;

[0016] The method for obtaining the running dataset is to run a preset dynamic cycle test of the diesel engine and collect multi-dimensional operating parameters of the diesel engine in real time with a set sampling step.

[0017] The multidimensional operating parameters include at least: core performance parameters, actuator feedback parameters, and system constraint indicators, forming the original multidimensional transient operating tensor dataset.

[0018] As a preferred embodiment of the XGBoost-LSTM time-series predictive control method for transient diesel engine conditions described in this invention, in step S2, the dataset preprocessing includes heterogeneous data cleaning and redundant data removal. The core of heterogeneous data cleaning lies in relying on the unique fluid inertia and thermodynamic limits of the diesel engine to identify and remove singular value samples that are significantly deviated from the actual operating evolution law of the diesel engine due to high-frequency severe overshoot or high-frequency disturbances under transient sudden operating conditions. This allows for the reconstruction of a high-fidelity feature tensor with physical and dynamic consistency. The redundant data removal involves dynamically resampling the cleaned transient dataset by expanding the discrete-time sampling step size. This identifies and removes physical redundant feature samples of the diesel engine under static equilibrium, quasi-steady-state power consumption, or idling conditions, achieving lightweight reconstruction of the dataset.

[0019] The rules for constructing the classification training dataset are as follows:

[0020] Remove physical redundancy data of the diesel engine under static equilibrium and normal power consumption conditions;

[0021] Set physical constraints to characterize the transient state: that is, the absolute value of the load change rate is greater than the preset change rate threshold, and the current load or power is greater than the preset load boundary threshold.

[0022] In the samples that meet the constraints, if the measured performance parameters deviate from the steady-state benchmark by a greater degree than the threshold obtained by the system, and the system constraint index deviates from the preset threshold by a greater degree than the deviation threshold, then it is determined that there is a risk of transient distortion, and the sample label is marked as 1; otherwise, it is marked as 0.

[0023] As a preferred embodiment of the XGBoost-LSTM time-series predictive control method for transient diesel engine conditions described in this invention, the specific method of S3 is as follows: quantifying the dispersion of transient and steady-state datasets in a high-dimensional feature space through cross-domain differential statistical analysis to generate a deviation matrix characterizing the disturbance intensity; subsequently, automatically solving for the optimal dynamic difference threshold based on the probability density distribution of the matrix; finally, performing logical disjunction operations on the threshold and system performance evaluation indicators to achieve simultaneous judgment of sample feature deviation and transient performance fluctuations, and automatically labeling binary classification logical tags accordingly for XGBoost model training.

[0024] As a preferred embodiment of the XGBoost-LSTM time-series predictive control method for transient diesel engine conditions described in this invention, the specific method of S4 is as follows: inputting test sample sequences independent of the training dataset into the model architecture; establishing a mapping consistency function between the discretized recognition results output by the model and the objective physical evolution trajectory of the system; deeply analyzing the model's classification robustness in high-dimensional feature space, its sensitivity to capturing transient disturbances, and its ability to suppress steady-state noise; and verifying the reliability of the XGBoost feedback results.

[0025] As a preferred embodiment of the XGBoost-LSTM time-series predictive control method for transient diesel engine conditions described in this invention, the specific method of S5 is as follows: Constructing an LSTM prediction architecture integrating physical and dynamic constraints, selecting the required performance parameters, analyzing the preprocessed multidimensional feature distribution through the model's built-in statistical sensing unit, spontaneously deriving the safety threshold based on the probability distribution density of the performance parameters, and constructing the actuator penalty multiplier by combining the principle of equivalent flow area in fluid dynamics.

[0026]

[0027] in Indicates the penalty multiplier. Indicates the safety threshold. Indicates the actual value;

[0028] The feedback signal is multiplied by a penalty multiplier to generate an unsmoothed theoretical ideal target value. Then, a first-order hysteresis loop is used to reconstruct this theoretical ideal target value to simulate physical hysteresis characteristics. The formula for calculating the exponentially weighted moving average in the discrete domain is as follows:

[0029]

[0030] in, The theoretical ideal target value at the current moment. This represents the actual physical position of the actuator at the previous moment. As a discrete smoothing factor, this formula accurately simulates the physical hysteresis characteristics in real-world situations by establishing a weight allocation between the "new instruction" and the "old state".

[0031] The frequency domain matching calculation of the discrete smoothing factor is as follows:

[0032]

[0033] in, The collection frequency of the dataset. The inherent physical time constant:

[0034]

[0035] in, Indicates the target delay time. Indicates the adjustment time.

[0036] As a preferred embodiment of the XGBoost-LSTM timing predictive control method for transient diesel engine conditions described in this invention, the specific method of S6 is to input an independent validation set into the architecture, focus on evaluating the mapping accuracy between the predicted trajectory and the actual physical evolution in the full operating condition space, and quantitatively analyze the cancellation effect of the phase compensation logic on the hysteresis characteristics of the diesel engine.

[0037] As a preferred embodiment of the XGBoost-LSTM time-series predictive control method for transient diesel engine conditions described in this invention, the specific method of S7 is as follows: the trained XGBoost and LSTM modules are logically coupled, and the dynamic discrimination threshold is automatically locked according to the classification probability distribution using a fusion architecture to achieve deep integration of the recognition logic and the prediction loop; subsequently, the response consistency and time-series mapping accuracy of the model under combined conditions are measured to quantitatively evaluate the cancellation gain and prediction gain of the collaborative architecture on the transient hysteresis characteristics of the diesel engine.

[0038] Compared with the prior art, the beneficial effects of the present invention are: the present invention can effectively solve the problems of signal phase difference and response hysteresis in diesel engines during transient processes, and greatly improve the real-time performance and accuracy of predicted commands in diesel engines during transient processes. Attached Figure Description

[0039] To more clearly illustrate the technical solutions of the embodiments of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and detailed embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0040] Figure 1 This is a flowchart of a diesel engine transient operating condition XGBoost-LSTM timing predictive control method according to the present invention;

[0041] Figure 2 This is an XGBoost confusion matrix diagram in Embodiment 1 of the XGBoost-LSTM timing predictive control method for transient diesel engine conditions according to the present invention;

[0042] Figure 3 The image shows the XGBoost ROC curve in Embodiment 1 of the XGBoost-LSTM timing predictive control method for transient diesel engine conditions according to the present invention.

[0043] Figure 4 This is a comparison of EGR rates before and after LSTM prediction in Embodiment 1 of the XGBoost-LSTM timing predictive control method for transient diesel engine conditions according to the present invention.

[0044] Figure 5 This is a comparison diagram of the EGR valve opening before and after LSTM prediction in Embodiment 1 of the XGBoost-LSTM timing predictive control method for transient diesel engine conditions according to the present invention.

[0045] Figure 6 This is an XGBoost prediction curve diagram in Embodiment 2 of the XGBoost-LSTM timing predictive control method for transient diesel engine conditions according to the present invention.

[0046] Figure 7 The graph shows the variation curve of N (extinction smoke, particulate matter concentration) in Example 2 of the XGBoost-LSTM timing predictive control method for transient diesel engine conditions according to the present invention.

[0047] Figure 8 This is a graph showing the EGR rate suppression effect in Embodiment 2 of the XGBoost-LSTM timing predictive control method for transient diesel engine conditions according to the present invention.

[0048] Figure 9 This is a comparison diagram of the EGR valve response before and after the prediction model in Embodiment 2 of the XGBoost-LSTM timing predictive control method for transient diesel engine conditions according to the present invention. Detailed Implementation

[0049] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0050] Secondly, the present invention is described in detail with reference to the schematic diagrams. When detailing the embodiments of the present invention, for ease of explanation, the cross-sectional views illustrating the device structure may be partially enlarged, not according to the usual scale. Furthermore, the schematic diagrams are merely examples and should not limit the scope of protection of the present invention. In addition, actual fabrication should include three-dimensional spatial dimensions of length, width, and depth.

[0051] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0052] like Figure 1 As shown, an embodiment of the present invention provides a diesel engine transient EGR control method based on an XGBoost-LSTM prediction model, comprising the following steps:

[0053] Step 1: Set up a diesel engine transient measurement and control test platform to acquire multi-dimensional sensor operation data sets of the diesel engine under transient cyclic conditions; specifically:

[0054] Using a high dynamic response diesel engine transient test and control bench, with a heavy-duty diesel engine as the test prototype, and running the WHTC (Worldwide Harmonized Transient Test Cycle) standard, the engine's fully mapped transient operating conditions under real road conditions were simulated. The characteristics of the intake, exhaust, and emission ends were captured in real time at a synchronous sampling frequency of 10Hz. The collected multi-dimensional operating parameters included: engine speed, torque, actual EGR rate, EGR opening, and nitrogen (extinction smoke, particulate matter concentration). Emissions, power, etc. The collected data are combined to form the original multidimensional transient operating condition dataset.

[0055] Step 2: Based on physical boundaries and transient operating condition characteristics, clean the runtime dataset and remove redundant data; specifically:

[0056] The abnormal data of negative torque under idling conditions caused by sensor acquisition problems on the test platform were cleaned, and the step size was increased from 0.1s to 0.2s to remove redundant data in the middle time and form a reconstructed transient dataset.

[0057] Step 3: Input the reconstructed transient dataset into XGBoost for XGBoost training; specifically:

[0058] The cleaned transient dataset is fed into the XGBoost feature processing module, which uses built-in logical discriminant operators to classify and identify real-time operating conditions. By monitoring the instantaneous fluctuations in torque change rate and output power, the system automatically identifies sample points with a torque change rate greater than 20 Nm / s and an absolute torque value exceeding 974.3 Nm or a real-time power exceeding 80 kW as variable operating condition feature regions.

[0059] Within the variable operating condition feature region, the module further introduces multi-dimensional performance constraint discrimination. By comparing the dispersion of the real-time EGR rate with the steady-state benchmark, and combining the response intensity of N (extinction smoke opacity, particulate matter concentration), feature samples with an EGR rate deviation greater than 1% and smoke opacity fluctuation exceeding 10% are automatically extracted and labeled as '1', while the remaining samples are labeled as '0', forming binary classification samples to drive the XGBoost classification engine to complete the deep feature extraction of the transient risk boundary;

[0060] Step 4: Input the reconstructed transient dataset into the LSTM for LSTM training; specifically:

[0061] Select the required performance parameters for LSTM: speed, torque, EGR rate, EGR valve opening, N (extinction smoke level, particulate matter concentration), input them into the LSTM processing module, and the system sets the EGR rate based on the dataset. The penalty percentage is updated with 8% as the safety threshold. :

[0062]

[0063] The original EGR valve opening command set by the ECU is multiplied by the trimmed penalty multiplier to generate the unsmoothed theoretical ideal EGR valve target opening.

[0064] The adjustment time is set to 3 seconds, and the system updates the physical inherent time constant by taking 0.9 seconds as the delay time based on the data.

[0065]

[0066] This leads to the smoothing factor:

[0067]

[0068] The ideal opening sequence is reconstructed using the updated exponentially weighted moving average algorithm:

[0069]

[0070] A mapping database from real-time sensing features to advance compensation targets has been established, providing structured data support for the LSTM architecture to capture the evolution law of nonlinear hysteresis systems.

[0071] Step 5: Build the XGBoost-LSTM prediction model; specifically:

[0072] A heterogeneous collaborative architecture is constructed, using XGBoost as the prior discrimination unit and LSTM as the dynamic regression unit. The system acquires the transient risk probability distribution output by XGBoost in real time and adaptively locks the decision threshold to 0.5 using the built-in logical optimization operator. When the predicted probability exceeds this threshold, the internal logic of the model automatically triggers the high-dimensional state flag "1", realizing seamless closed-loop transmission of the working condition identification signal and the subsequent prediction loop.

[0073] After capturing the decision flag, the LSTM prediction module immediately activates the prediction logic, which has the capability to cancel out physical delays. Based on the input real-time feature vector and the preset physical time constant, the module performs advance compensation calculations for the nonlinear response characteristics of the actuator, and finally outputs an EGR valve opening control command that conforms to the actual physical evolution of the controlled object.

[0074] Two specific embodiments are provided below to verify the effectiveness of this method.

[0075] Example 1:

[0076] This embodiment focuses on the independent performance analysis of the two models. It uses multi-dimensional performance indicators to quantify the independent verification process of the XGBoost transient condition identification model and the LSTM predictive control model, demonstrating the accuracy of the endogenous self-optimization mechanism in identifying transient risk boundaries and the physical consistency and smoothness of predictive control of complex systems.

[0077] Step 1: According to Table 1, XGBoost's training metrics are all above 90%, demonstrating excellent statistical inference capabilities. Figure 2 The classification confusion matrix shows that for data with a true value of 0, the ratio of correct to incorrect is 2175:248, and for data with a true value of 1, the ratio of correct to incorrect is 267:22. This effectively verifies the endogenous self-optimization mechanism's ability to keenly capture risk boundaries under unbalanced feature distributions. Figure 3 The ROC curve shows that the area under the curve (AUC) is close to 1.0, which further confirms from the perspective of measurement theory that the architecture maintains high global classification accuracy while having extremely high detection sensitivity and classification robustness for a very small number of key transient features.

[0078] Table 1: XGBoost Training Result Metrics

[0079] Accuracy (%) Recall rate (%) AUC (%) 90.4 92.39 96.32

[0080] Based on the regression performance evaluation parameters shown in Table 2, the LSTM prediction architecture exhibits high goodness of fit and generalization robustness across multiple dimensions. Figure 4 Analysis shows that during periods of drastic fluctuations in system operating conditions, the LSTM model achieved significant phase lead intervention thanks to its built-in physical hysteresis compensation operator. This mechanism successfully offset the mechanical response delay of the actuator, precisely constraining the measured performance parameters within a preset safety threshold of 8%, thus verifying the model's real-time interception capability for transient peaks. Figure 5 The control trajectory shown further confirms that the actuator opening command output by the model effectively suppresses overshoot while exhibiting excellent dynamic smoothness. The output sequence is logically rigorous, completely avoiding high-frequency oscillations and physical saturation phenomena, and strongly supports the physical consistency and smoothness of the LSTM architecture for predictive control of complex systems from an engineering measurement perspective.

[0081] Table 2: LSTM Training Result Metrics

[0082] Mean square error 6.5100 0.9734

[0083] Example 2:

[0084] This embodiment focuses on the analysis of the combined effect of the two models. Through quantitative chart analysis of the XGBoost-LSTM collaborative architecture, it systematically verifies that the method significantly improves the performance of XGBoost-LSTM predictive timing control in real physical systems.

[0085] Combination Figure 6 , Figure 7 When a diesel engine enters a severe transient operating condition that causes a surge in EGR overshoot risk (with a significant peak in extinction smoke N under the baseline strategy), the XGBoost classification model can accurately capture the physical precursors of emission deterioration. Based on this, the system triggers the logic discrimination operator in real time, seamlessly switching the drive gas path regulation authority from steady-state MAP logic to LSTM predictive control loop execution, demonstrating excellent operating condition sensitivity and timely state capture.

[0086] Based on this, combine Figure 8 , Figure 9 LSTM not only effectively suppresses the EGR rate to below 8%, but its output EGR valve opening command also exhibits high trajectory smoothness and logical continuity. There is no valve opening jump that does not conform to the actual situation due to the model predicting the timing control according to the ideal. This shows that the method significantly improves the execution performance of XGBoost-LSTM predictive timing control in real physical systems.

[0087] Although the present invention has been described above with reference to embodiments, various modifications can be made and components can be replaced with equivalents without departing from the scope of the invention. In particular, as long as there is no structural conflict, the features in the disclosed embodiments can be combined with each other in any manner. The lack of an exhaustive description of these combinations in this specification is merely for the sake of brevity and resource conservation. Therefore, the present invention is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A diesel engine transient operating condition XGBoost-LSTM timing predictive control method, characterized in that, Includes the following steps: S1: Build a transient operation monitoring platform for diesel engines and obtain multi-dimensional sensor operation datasets of diesel engines under typical transient operating conditions; S2: Based on the physical dynamics limits and transient fluctuation characteristics of diesel engines, the running dataset is preprocessed and adaptively labeled to construct a classification training dataset for the identification model; S3: Introduce a threshold endogenous self-calibration mechanism based on the characteristics of sample distribution proportion to spontaneously identify and train the XGBoost model on the risk boundary of transient operating conditions of diesel engines. S4: Conduct quantitative verification of the generalization performance of the XGBoost model based on multi-criteria evaluation metrics; S5: LSTM end-to-end prediction training is performed using an LSTM architecture that integrates deep hysteresis physics formulas and penalty incentive mechanisms. S6: Perform quantitative verification of the generalization performance of the LSTM model based on multiple evaluation criteria; S7: Develop the architecture and evaluate the performance of the XGBoost-LSTM prediction model.

2. The XGBoost-LSTM timing predictive control method for transient operating conditions of a diesel engine according to claim 1, characterized in that, The multi-dimensional sensor in S1 is a multi-dimensional industrial sensor configured at key physical nodes of the diesel engine to synchronously collect environmental state signals and physical feedback signals of the actuators at a preset sampling period; the state signals cover physical characteristics such as fluid dynamic parameters, thermodynamic temperature or mechanical displacement during system operation. The physical feedback signal is the actual response action parameter of the actuator under the control command; The method for obtaining the running dataset is to run a preset dynamic cycle test of the diesel engine and collect multi-dimensional operating parameters of the diesel engine in real time with a set sampling step. The multidimensional operating parameters include at least: core performance parameters, actuator feedback parameters, and system constraint indicators, forming the original multidimensional transient operating tensor dataset.

3. The XGBoost-LSTM timing predictive control method for transient operating conditions of a diesel engine according to claim 1, characterized in that, In S2, dataset preprocessing includes heterogeneous data cleaning and redundant data removal. The core of heterogeneous data cleaning lies in relying on the unique fluid inertia and thermodynamic limits of diesel engines to identify and remove singular value samples that are distorted by sensor measurements due to high-frequency severe overshoot or high-frequency disturbances under transient sudden operating conditions of diesel engines, and that significantly deviate from the actual operating evolution law of diesel engines. This allows for the reconstruction of a high-fidelity feature tensor with physical and dynamic consistency. Redundant data removal involves dynamically resampling the cleaned transient dataset by expanding the discrete-time sampling step size. This identifies and removes physical redundant feature samples of diesel engines under static equilibrium, quasi-steady-state power consumption, or idling conditions, achieving lightweight reconstruction of the dataset. The rules for constructing the classification training dataset are as follows: Remove physical redundancy data of the diesel engine under static equilibrium and normal power consumption conditions; Set physical constraints to characterize the transient state: that is, the absolute value of the load change rate is greater than the preset change rate threshold, and the current load or power is greater than the preset load boundary threshold. In the samples that meet the constraints, if the measured performance parameters deviate from the steady-state benchmark by a greater degree than the threshold obtained by the system, and the system constraint index deviates from the preset threshold by a greater degree than the deviation threshold, then it is determined that there is a risk of transient distortion, and the sample label is marked as 1; otherwise, it is marked as 0.

4. The XGBoost-LSTM timing predictive control method for transient operating conditions of a diesel engine according to claim 1, characterized in that, The specific method of S3 is as follows: quantify the dispersion of transient and steady-state datasets in a high-dimensional feature space through cross-domain differential statistical analysis to generate a deviation matrix characterizing the intensity of the disturbance; then, automatically solve the optimal dynamic difference threshold based on the probability density distribution of the matrix; finally, perform logical disjunction operation on the threshold and the system performance evaluation index to achieve synchronous judgment of sample feature deviation and transient performance fluctuation, and automatically label the binary classification logical labels accordingly for XGBoost model training.

5. The XGBoost-LSTM timing predictive control method for transient operating conditions of a diesel engine according to claim 1, characterized in that, The specific method of S4 is to input test sample sequences independent of the training dataset into the model architecture, establish a mapping consistency function between the discretized recognition results output by the model and the objective physical evolution trajectory of the system, deeply analyze the model's classification robustness in high-dimensional feature space, its sensitivity to capturing transient disturbances and its ability to suppress steady-state noise, and verify the reliability of the XGBoost feedback results.

6. The XGBoost-LSTM timing predictive control method for transient operating conditions of a diesel engine according to claim 1, characterized in that, The specific method of S5 is as follows: An LSTM prediction architecture integrating physical and dynamic constraints is constructed; the required performance parameters are selected; the preprocessed multidimensional feature distribution is analyzed through the model's built-in statistical sensing unit; a safety threshold is spontaneously derived based on the probability distribution density of the performance parameters; and an actuator penalty multiplier is constructed by combining the principle of equivalent flow area in fluid mechanics. ； in Indicates the penalty multiplier. Indicates the safety threshold. Indicates the actual value; The feedback signal is multiplied by a penalty multiplier to generate an unsmoothed theoretical ideal target value. Then, a first-order hysteresis loop is used to reconstruct this theoretical ideal target value to simulate physical hysteresis characteristics. The formula for calculating the exponentially weighted moving average in the discrete domain is as follows: ； in, The theoretical ideal target value at the current moment. This represents the actual physical position of the actuator at the previous moment. As a discrete smoothing factor, this formula accurately simulates the physical hysteresis characteristics in real-world situations by establishing a weight allocation between the "new instruction" and the "old state". The frequency domain matching calculation of the discrete smoothing factor is as follows: ； in, The collection frequency of the dataset. The inherent physical time constant: ； in, Indicates the target delay time. Indicates the adjustment time.

7. The XGBoost-LSTM timing predictive control method for transient operating conditions of a diesel engine according to claim 1, characterized in that, The specific method of S6 is to input an independent verification set into the architecture, focus on evaluating the mapping accuracy between the predicted trajectory and the actual physical evolution in the full operating space, and quantitatively analyze the cancellation effect of the phase compensation logic on the hysteresis characteristics of the diesel engine.

8. The XGBoost-LSTM timing predictive control method for transient operating conditions of a diesel engine according to claim 1, characterized in that, The specific method of S7 is to logically couple the trained XGBoost with the LSTM module, and use the fusion architecture to automatically lock the dynamic discrimination threshold according to the classification probability distribution to achieve deep integration of the recognition logic and the prediction loop. Then, by measuring the response consistency and time-series mapping accuracy of the model under combined working conditions, the cancellation gain and prediction gain of the collaborative architecture on the transient hysteresis characteristics of the diesel engine are quantitatively evaluated.

Images