Method for predicting and optimizing performance of diesel hydraulic brake system based on machine learning

Abstract

This invention belongs to the technical field of digital design, performance prediction, and parameter optimization of diesel engine hydraulic braking systems, and particularly to a machine learning-based method for performance prediction and optimization of diesel engine hydraulic braking systems. First, a one-dimensional parametric model of the diesel engine hydraulic braking system is established. A subset of candidate features is constructed based on feature importance. Then, a target regression model matching each target performance index is selected from multiple candidate regression models. Subsequently, the target regression model replaces the one-dimensional parametric model to quickly predict candidate points. The probability of predicting the normal braking state is used as a hard constraint condition of the feasible region to perform multi-objective optimization. Finally, a closed-loop optimization process is constructed by simulating representative solutions, with backtesting samples flowing unidirectionally into the training set, and a fixed test set for isolation to avoid contamination. This invention can improve the effectiveness of candidate parameter selection and the engineering verifiability of optimization results in the parameter design and performance optimization of diesel engine hydraulic braking systems.

Description

Technical Field

[0001] This invention relates to the field of digital design, performance prediction and parameter optimization technology for diesel engine hydraulic braking systems, specifically a method for performance prediction and optimization of diesel engine hydraulic braking systems based on machine learning. Background Technology

[0002] Diesel engine hydraulic braking systems are primarily used in heavy-duty vehicles, and their braking performance affects the lifespan of the vehicle's fundamental braking components during long downhill braking, thus impacting occupant safety. Specifically, the time it takes for the actuator valve to reach its lower limit, the volume of oil leakage in the actuator valve chamber within a single stable load cycle, and the displacement fluctuation of the actuator valve during the loading phase all affect braking performance. The time it takes for the actuator valve to reach its lower limit, i.e., the response time, determines whether the energy compressed within the cylinder is released precisely near the optimal moment to achieve the best braking effect. The volume of oil leakage in the actuator valve chamber within a single stable load cycle affects the pressure maintenance capacity within the actuator valve body, thus affecting the valve's ability to maintain its braking position. The displacement fluctuation of the actuator valve during the loading phase reflects the magnitude of the valve displacement fluctuation in its working position under the combined action of the cam, valve system, oil pressure, spring, and cyclic load, and also indirectly reflects the fluctuation in the exhaust valve opening degree, affecting braking stability. Therefore, it is necessary to achieve a synergistic balance among multiple objectives through the design of structural parameters, oil viscosity parameters, and operating condition parameters.

[0003] Conventional approaches often involve machine learning to directly perform regression fitting on whether braking is normal within the sample space without differentiation, combined with multi-objective optimization. This approach results in relatively few invalid samples when dealing with low-dimensional parameter combinations. However, the three optimization objectives mentioned earlier are not determined by a single parameter. In diesel engine hydraulic braking systems, with the coupling of multiple physical parameters such as fuel supply pressure, multi-stage spring stiffness, oil viscosity, load force magnitude, and load application timing, changes in a few parameters can easily lead to abnormal operating conditions that do not meet engineering requirements, such as the actuator valve failing to reach its lower limit or drastic fluctuations in actuator valve displacement during load application. During the optimization process, the Pareto front may contain mathematically optimal parameter combinations that fail to achieve normal braking. Furthermore, the three objectives have different physical characteristics, resulting in varying levels of difficulty in fitting regression models, often making it impossible to directly apply the same method to construct a predictive model for all objectives. To design a better parameter combination that satisfies normal braking conditions in the early stages, it is necessary to consider the nonlinear and multi-parameter coupling characteristics of the diesel engine hydraulic system during multi-objective optimization, imposing constraints and limiting invalid parameter combinations during the optimization process. Summary of the Invention

[0004] 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.

[0005] To address the aforementioned technical problems, according to one aspect of the present invention, the present invention provides the following technical solution: a method for performance prediction and optimization of a diesel engine hydraulic braking system based on machine learning, comprising the following steps:

[0006] Step S100: Establish a one-dimensional parametric model of the diesel engine hydraulic braking system;

[0007] Step S200: Generate parameter samples in the parameter space of the design variables, and input the parameter samples into the one-dimensional parameterized model to perform batch time series simulation to obtain the original time series data corresponding to each parameter sample;

[0008] Step S300: Calculate and obtain the target performance index based on the original time series data, and generate sample state labels and failure causes by combining physical boundary conditions and running results to form a training table;

[0009] Step S400: Build a sample state classification model using all samples in the training table, output the predicted probability of the sample's braking normal state, construct a candidate subset based on the importance of the input features output by the random forest regression model, and build a target regression surrogate model for the effective samples.

[0010] Step S500: In the multi-objective optimization process, the regression model replaces the one-dimensional parameterized model to predict the candidate parameter solution set. The sample state classification model outputs the braking normal state prediction probability of the candidate parameter solution, and uses this as a hard constraint condition for entering the feasible region optimization iteration. When the braking normal state prediction probability is lower than the preset threshold, the candidate solution is prevented from entering the feasible region.

[0011] Step S600: Select representative parameter solutions from the non-dominated candidate parameter solution set, input them into the one-dimensional parameterized model for simulation backtesting, perform post-processing calculations on the original time series data, and obtain the simulation values ​​of the target performance index corresponding to the representative parameter solutions;

[0012] Step S700: Calculate the relative error between the simulated and predicted performance index values ​​of the representative solution, and compare them with the preset threshold.

[0013] As a preferred embodiment of the machine learning-based performance prediction and optimization method for diesel engine hydraulic braking systems described in this invention, the specific method for establishing the one-dimensional parametric model of the diesel engine hydraulic braking system in S100 is as follows: Based on the structure and working principle of the diesel engine hydraulic braking system, a one-dimensional parametric model including an oil supply system, control valve, ball valve, actuator valve, and load control module is established; relevant component parameters are set according to the braking system component drawings, and the load cycle and application method of the periodic load are set; among the settable parameters, at least the actuator valve spring stiffness, control valve spring stiffness, ball valve spring stiffness, oil viscosity, load application start time, oil supply pressure, and load force parameters are selected as design variables for subsequent regression model training; and the index corresponding to the actuator valve reaching the lower limit time is selected. The corresponding index of oil leakage volume in the valve cavity during a single stable load cycle. Corresponding indexes for valve displacement fluctuation during the load application phase As a performance metric for subsequent multi-objective optimization, the design variable vector is represented as:

[0014]

[0015] Where x represents the candidate parameter solution or the combination of design variables; , … These represent the specific design variables in the following categories: actuator spring stiffness, ball valve spring stiffness, control valve spring stiffness, oil viscosity, load initiation time, oil supply pressure, and load force. 'd' represents the dimension of the design variables. (The last part, "x", appears to be a typo and can be left as is.) v (t) is defined as the valve displacement at time t, q(t) is defined as the leakage flow rate at time t, and x L Defined as the lower threshold of the actuator valve, T end Defined as the simulation end time, T c Defined as a preset load cycle.

[0016] As a preferred embodiment of the machine learning-based diesel engine hydraulic braking system performance prediction and optimization method described in this invention, wherein the... The calculation formula is as follows:

[0017]

[0018] Among them, t reach Indicates the moment when the valve displacement first reaches the preset lower threshold, where inf represents the earliest moment the condition is met, and t represents the continuous time variable, [0, T]. end [] indicates the simulation time interval. (t) represents the valve displacement at time t. This indicates the preset lower threshold of the actuator valve; if no threshold is met... ≥ At that moment, Undefined, the reason for sample failure is that the lower limit of the actuator valve was not reached.

[0019] As a preferred embodiment of the machine learning-based diesel engine hydraulic braking system performance prediction and optimization method described in this invention, wherein the... The sampling time at which the lower limit threshold is first reached is used directly without introducing interpolation. Furthermore, no additional interpolation function is introduced, and the time t is the departure time from the lower limit. leave The definition is as follows:

[0020]

[0021] In the above formula, t leave τ represents the time after the lower limit is reached; τ represents any time after t; this definition means that the valve has reached the lower limit, and we are looking for the earliest time after which the valve displacement is always below the lower limit threshold.

[0022] As a preferred embodiment of the machine learning-based diesel engine hydraulic braking system performance prediction and optimization method described in this invention, the stable load cycle determination rule is as follows: Let the stable band B... s and the kth complete load cycle W k They are respectively:

[0023]

[0024] In the above formula, B s Indicates the stable band, δ s W represents the half-width of the stable band. k This represents the k-th complete load cycle, where t0 represents the start point of the load cycle, k represents the cycle number, and T... c Indicates the preset load period, if W k If the displacement of the internal actuator valve meets the preset stable zone constraint, the system is determined to have entered the stable load stage.

[0025] As a preferred embodiment of the machine learning-based diesel engine hydraulic braking system performance prediction and optimization method described in this invention, the determination rule for the stable load cycle is based on a stable band percentage threshold and a cycle mean deviation threshold, and the determination formula is as follows:

[0026]

[0027] Where, η k The valve displacement falls into the stable zone B during the kth cycle. s The percentage of sampling points; μ k It is the average displacement of the actuator valve within the k-th cycle; η th It is the preset threshold for the proportion of stable bands; δ μIt is a preset mean deviation threshold, when η k ≥η th And |μ k -x L |≤δ μ When, determine W k The stability condition is met;

[0028] Calculate using the following formula:

[0029]

[0030] In the formula, t stable T represents the starting point of a stable period. c Let q(t) represent a complete load cycle, and q(t) represent the leakage flow at time t. If the starting point of a stable cycle cannot be identified, then... Undefined;

[0031] Use I Y3 express Statistical interval:

[0032]

[0033] The displacement fluctuation Y3 of the actuator valve during the period from reaching the lower limit to leaving the lower limit is calculated as follows:

[0034]

[0035] In the formula, I Y3 for The statistical interval; max is the maximum displacement of the statistical interval, and min is the minimum displacement of the statistical interval.

[0036] As a preferred embodiment of the machine learning-based performance prediction and optimization method for diesel engine hydraulic braking systems described in this invention, in step S400, for valid samples, a subset containing the top K important features is constructed based on feature importance. Combining cross-validation with a prediction evaluation index, the model with the better performance is selected as the regression model for that performance index. Feature importance is output using a random forest regression model. The candidate feature subset includes the full feature subset and a subset containing the top K important features constructed based on feature importance ranking. All input feature sets F and the top K high-importance feature subsets F are included. K They are respectively:

[0037]

[0038] In the formula, F is the set of all sample input features, f j For the j-th input feature, the top K subsets of dimensionality-reduced features, sorted by feature importance from highest to lowest, are denoted as F. KThe top K features, sorted by importance from highest to lowest, are f. （1） to f （k） As a preferred embodiment of the machine learning-based diesel engine hydraulic braking system performance prediction and optimization method described in this invention, the normal state prediction probability and hard constraint conditions are defined as follows, and the hard constraint conditions are calculated using the constructed hard constraint function:

[0039]

[0040]

[0041] In the above formula, p ok (x) represents the probability that the state s of the candidate parameter solution x is the normal braking state, output by the sample state classification model, and the preset normal state probability threshold p is calculated using the feasible region hard constraint function c(x). th The probability p of the normal braking state of the sample ok The difference between (x) and p, when c(x)≤0, i.e. ok (x)≥p th When a candidate parameter solution x is determined to satisfy the hard constraints of the feasible region, in the process of multi-objective optimization, whenever a new candidate parameter solution is generated, it must satisfy the above hard constraints of the feasible region before subsequent sorting, selection and the next iteration update process can be carried out.

[0042] As a preferred embodiment of the machine learning-based diesel engine hydraulic braking system performance prediction and optimization method described in this invention, the representative optimal solution in the candidate parameter solution set in S600 includes at least... Minimum solution Minimum solution Minimum solution and comprehensive compromise optimal solution; among which, The minimum solution refers to the time it takes for the actuator valve to reach its lower limit. Candidate parameter solutions that yield the minimum value; The minimum solution is the solution that minimizes the leakage volume of the valve chamber oil during a single stable load cycle. Candidate parameter solutions that yield the minimum value; The minimum solution refers to the amount of displacement fluctuation of the actuator valve during the loading phase. Candidate parameter solutions that yield the minimum value;

[0043] Let P denote the set of candidate parameter solutions that satisfy the hard constraints, and x be a candidate parameter solution. , , These are the target regression surrogate model pairs , and Let argmin represent the predicted value, and let argmin represent the candidate parameter solution that minimizes the objective function. , , The minimum solution is expressed as:

[0044]

[0045] The comprehensive compromise optimal solution is defined as:

[0046]

[0047]

[0048] In the formula, x bal This represents the optimal solution through a comprehensive compromise. , , They represent respectively to , , The predicted value after normalization; , , Let represent the weight coefficients corresponding to the three objectives, and satisfy . + + =1.

[0049] As a preferred embodiment of the machine learning-based diesel engine hydraulic braking system performance prediction and optimization method described in this invention, in step S700, if the output conditions are met, a recommended parameter combination and corresponding performance indicators are output; if the output conditions are not met, the representative parameter solutions verified by simulation are added to the training sample set and the regression model is trained again and the next round of closed-loop optimization is performed; the fixed test set remains unchanged after the initial partitioning and is only used for the final evaluation of the model prediction accuracy. It does not participate in the training and validation in the regression model construction, nor is it mixed with the representative parameter solutions verified by simulation; the representative parameter solutions verified by simulation are used as validation samples and are only added to the training set according to the one-way inflow principle, without flowing into the validation set and the fixed test set.

[0050] Compared with the prior art, the beneficial effects of the present invention are: 1. When post-processing the original time series data, the target index definition and classification model are combined to divide the samples into normal braking conditions and abnormal braking conditions, so that the feature subset space of subsequent regression modeling is the normal braking space, which reduces the large amount of calculation caused by the large number of samples when fitting the full sample space to a certain extent.

[0051] 2. By sorting the features by importance, a candidate subset containing the top K important features of each target is generated. Redundant features are removed from target performance indicators suitable for dimensionality reduction, while all input features are retained for target performance indicators coupled with complex physical features, thereby improving the suitability of regression models for different indicators.

[0052] 3. By adding the probability of the candidate parameter solution set being in a normal braking state as a hard constraint in the target optimization process, the parameter solution set optimization is carried out within the normal braking space, which improves the effectiveness of the non-dominated feasible solution set, thereby improving the efficiency of finding better braking conditions to a certain extent and reducing the simulation backtesting of invalid samples.

[0053] 4. The Random Forest model, as a regression model, outputs a ranking of feature importance, indicating the contribution of each valve body spring stiffness, oil supply pressure, load force, oil viscosity, and load initiation time of the diesel engine hydraulic braking system to the prediction of target performance indicators. It also serves as a preliminary basis for constructing a candidate subset containing the top K important features. When used as a classification model, it also enables the entire method to identify whether a sample can brake normally. Attached Figure Description

[0054] 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:

[0055] Figure 1 This is a flowchart of the machine learning-based performance prediction and optimization method for diesel engine hydraulic braking systems according to the present invention.

[0056] Figure 2 This is a flowchart illustrating the construction of feature subsets and the building of a better regression model for prediction in an embodiment of the present invention;

[0057] Figure 3 This is a schematic diagram illustrating the unidirectional backflow relationship between the training set, validation set, fixed test set, and validation samples in an embodiment of the present invention.

[0058] Figure 4 This is a time-series diagram of unidentified abnormal samples in the stable phase in an embodiment of the present invention;

[0059] Figure 5 This is a time-series diagram of abnormal samples that did not reach the lower limit in an embodiment of the present invention;

[0060] Figure 6 As described in the embodiments of the present invention A schematic diagram of the importance of random forest features corresponding to the time when the actuator reaches the lower limit;

[0061] Figure 7 As described in the embodiments of the present invention A schematic diagram of the importance of random forest features corresponding to the volume of oil leakage in the valve chamber during a single stable load cycle;

[0062] Figure 8 As described in the embodiments of the present invention A schematic diagram of the importance of random forest features corresponding to the displacement fluctuation of the actuator valve during the load-addition phase. Detailed Implementation

[0063] 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.

[0064] 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.

[0065] A machine learning-based method for performance prediction and optimization of diesel engine hydraulic braking systems includes the following steps:

[0066] Step S100: Based on the structure and working principle of the diesel engine hydraulic braking system, establish a one-dimensional parametric model including the oil supply system, control valve, ball valve, actuator valve, and load control module. Set the relevant component parameters according to the braking system part drawings, and define the load cycle and application method for the periodic load. Select at least the following parameters as design variables for subsequent machine learning: actuator valve spring stiffness, control valve spring stiffness, ball valve spring stiffness, oil viscosity, load application start time, oil supply pressure, and load force, and define the target performance index. For the valve to reach its lower limit time, The volume of oil leakage in the valve chamber during a single stable load cycle. The valve displacement fluctuation during the loading phase is defined as follows; to facilitate a unified representation of multidimensional design variables, target surrogate models, and hard gating constraints, the design variable vector is set as follows:

[0067]

[0068] Where x represents the candidate parameter solution or the combination of design variables, , … These represent the specific design variables among the actuator valve parameters, ball valve parameters, control valve parameters, oil viscosity, load initiation time, oil supply pressure, and load force, respectively. d represents the dimension of the design variables. v (t) is defined as the valve displacement at time t, q(t) is defined as the leakage flow rate at time t, and x L Defined as the lower threshold of the actuator valve, T endDefined as the simulation end time, T c Defined as the preset load period. The above symbols are only used to unify mathematical expression and do not change the original physical meaning of each design variable.

[0069] valve reaches lower limit time The calculation formula is:

[0070]

[0071] in, Indicates the time t for the actuator to reach the lower limit. reach Indicates the moment when the valve displacement first reaches the preset lower threshold, where inf represents the earliest moment the condition is met, and t represents the continuous time variable, [0, T]. end ] indicates the simulation time interval (t) represents the valve displacement at time t. This indicates the preset lower threshold value for the actuator valve. If no threshold value is met... ≥ At that moment, Undefined, the reason for the failure of this sample is that the actuator valve did not reach the lower limit.

[0072] In this embodiment, the response time is 10. -2 The simulation step size is on the order of s, and the simulation step size is 1×10. -5 The sampling time step (s) meets the engineering accuracy requirements for determining when the actuator reaches its lower limit. Therefore, no additional interpolation function is introduced. This sampling time step is an optimal implementation parameter and can be evaluated and adjusted based on the accuracy requirements of the simulation model and the simulation calculation time.

[0073] The valve leaves the lower limit time t leave The calculation formula is:

[0074]

[0075] In the formula t leave Let t represent the time of departure from the lower limit, and τ represent any time after t. Due to periodic loads, the valve displacement is prone to fluctuations, continuously and briefly departing from the lower threshold. To avoid incorrect departure times in post-processing, this time is defined as the earliest time after which the valve will not return to the lower threshold. The calculation of Y2 involves determining the stable load period. Let the load application start time be t0, and the stable band B... s and the kth complete load cycle W k They are respectively:

[0076]

[0077] B s It is a stable band, δ sAs the stable band half-width, W k This represents the k-th complete load cycle, where t0 is the start point of the load cycle. k is the cycle number, and T... c This is a preset load cycle. In this embodiment, x L It is 3.5mm, δ s If the value is 0.1mm, the stable band is [3.4mm, 3.6mm]. If W k If the displacement of the actuator valve meets the preset stable zone constraint within the time period, the system is determined to have entered the stable load stage.

[0078] In a preferred embodiment, due to simulation error, the sampling points of the valve displacement curve may exhibit slight numerical fluctuations that deviate slightly from the stable band. Load cycles where the valve displacement is within the stable band are considered unstable. To avoid this situation, a stable band percentage threshold and a cycle mean deviation threshold are used for determination, with the following formula:

[0079]

[0080] Since the valve displacement curve is not constant after entering the stable zone, the valve displacement mean deviation threshold can be adjusted according to the specific simulation accuracy and engineering stability requirements.

[0081] t stable T represents the start time of entering a stable load cycle. c Given a complete load cycle, after determining the starting point of the stable load cycle, the leakage flow rate q(t) at time t is calculated. The volume of oil leakage in the valve chamber during a single stable load cycle:

[0082]

[0083] This represents the displacement fluctuation of the actuator valve during the load application phase. After determining the moment when the actuator valve leaves the lower limit, let... The statistical interval is I Y3 In a preferred embodiment, the following may be adopted:

[0084]

[0085] The displacement fluctuation of the actuator valve after applying a load during this statistical interval is:

[0086]

[0087] This is to reflect the stability of the diesel engine hydraulic braking system's ability to maintain the braking position.

[0088] Step S200: Generate parameter samples in the parameter space of the design variables, and input the parameter samples into the one-dimensional parameterized model for batch simulation to obtain the original time series data corresponding to each parameter sample.

[0089] In a preferred embodiment, the displacement curve of the actuator valve and the oil leakage flow rate curve of the actuator valve cavity are obtained, and the actuator valve displacement value and actuator valve cavity leakage flow rate value corresponding to each sampling time point of the sample are exported by writing a program. Step S300: Extract the target performance index based on the original time series data, and generate sample state labels and failure causes by combining physical boundary conditions and calculation results to form a training table; For abnormal samples that do not meet the normal braking conditions, their missing performance indicators are not interpolated or filled.

[0090] In a preferred embodiment, when simulation fails and valid original time-series data cannot be obtained, the sample status is labeled "simulation failed state," marked as "simulation_failed" in the program, and the cause of failure is marked as "simulation failed." When the calculated sample valve displacement consistently fails to reach the lower displacement threshold, the sample status is labeled "lower limit not reached state," marked as "failed_reach_lowerlimit" in the program, and the cause of failure is marked as "lower limit not reached." When the sample valve displacement fails to meet the stability criteria within the time between reaching and leaving the lower limit, the sample status is labeled "stable phase not identified state," marked as "failed_stable_not_found" in the program, and the cause of failure is marked as "stable state not identified." Only when the target indices Y1, Y2, and Y3 are all calculated to obtain valid values ​​as defined above, the sample status is labeled "normal braking state," recorded as "ok" in the program, and marked as "normal state." The above English labels are merely for program implementation purposes and are not intended to limit the scope of protection.

[0091] In this invention, the reason for not interpolating or imputing the missing target index values ​​in abnormal operating condition samples is that such samples do not physically meet the aforementioned calculation definitions. It is impossible to calculate all specific values ​​such as the valve reaching its lower limit time, the oil leakage volume in the valve chamber within a single stable load cycle, and the displacement fluctuation of the valve under load. These are all conditions where normal braking is impossible. If data processing is performed through interpolation or imputation, abnormal braking conditions, because they have corresponding definite values, mathematically become conditions with braking capability, but do not conform to physical boundaries and cannot achieve normal braking. This will cause the model to learn incorrect mapping relationships between design parameters such as multi-stage spring stiffness and oil viscosity parameters and the three target performance indicators, interfering with the accuracy of the regression model's predictions. Furthermore, in multi-objective optimization, because the surrogate model learns incorrect braking condition boundaries, the recommended representative solution may result in situations where normal braking is impossible.

[0092] Step S400: A sample state classification model is established using all samples in the training table, outputting the predicted probability of the sample's normal braking state. A random forest regression model is used to calculate the importance of the input features. Then, a candidate subset consisting of the top K important features is constructed from the effective samples based on feature importance. Cross-validation is then used to select and establish a target regression surrogate model that matches each target performance index from among multiple candidate regression models.

[0093] In a preferred embodiment, the classification model employs a random forest classification model, and the feature importance is calculated and output by a random forest regression model. The candidate feature subset includes at least the full feature subset and a subset containing the top K most important features, constructed based on feature importance ranking. Let F be the total input feature set and FK be the top K most important feature subsets. K They are respectively:

[0094]

[0095] F is a set containing a certain number of input features, f j Let F be the j-th input feature. The top K subsets of dimensionality-reduced features, sorted by feature importance from highest to lowest, are F. K The top K features, ranked from highest to lowest importance according to the output of the random forest regression model, are f. （1） to f （k）。In a preferred embodiment, different input features contribute differently to the prediction of different target performance indicators based on feature importance. A subset is initially selected based on feature importance, consisting of the top K most important features for each performance indicator, and this subset, along with the full feature subset, serves as a candidate subset. This approach aims to address the fact that different performance indicators have different physical meanings and therefore different dominant input features. Since the diesel engine hydraulic braking system is a physical system coupled with multiple input features, it is not feasible to delete some input features solely based on feature importance. Low-importance features may only contribute less to the target prediction, but they could be indispensable input variables for learning the normal and abnormal braking boundaries. Therefore, all input features are retained as a candidate feature subset, thereby removing redundant features for target performance indicators suitable for dimensionality reduction and retaining all input features for target performance indicators coupled with complex physical features, thus improving prediction accuracy.

[0096] In a preferred embodiment, regression models are independently established for each candidate model scheme. The reasons are as follows: As mentioned earlier, different input features have varying degrees of importance in predicting different target performance indicators. For example, Y1 reflects the braking response speed; the oil supply pressure is used in the early stages of the actuator valve's movement to overcome the elasticity of each stage of springs, the resistance generated by oil viscosity, and the load force. Y2 reflects the leakage level, directly affected by the pressure difference across the channel and oil viscosity; the oil supply pressure indirectly affects the leakage level. Y3 reflects the degree of fluctuation in the actuator valve's working position; the magnitude of the fluctuation is the result of a balance between the load force, oil viscosity resistance, and the support force at the actuator valve's limit end. The oil supply pressure is not the dominant factor for displacement fluctuation after the actuator valve reaches its working position. Because the three targets have different physical definitions, their physical failure conditions also differ. When the actuator valve does not reach its lower limit, the specific values ​​for all three targets cannot be calculated. When the system cannot enter a stable state under a certain parameter condition, the volume of leaked oil in the actuator valve cavity within a single stable load cycle cannot be calculated. Therefore, each sample may not be able to calculate all outputs simultaneously. For the reasons mentioned above, if a unified model is used to construct a regression prediction model with multidimensional inputs and multi-objective outputs, the model will predict three different physical indicators under the same mapping relationship, which will affect the prediction accuracy of the regression model.

[0097] Step S500: In the multi-objective optimization process, the objective regression model replaces the one-dimensional parameterized model to quickly predict the objective performance indicators of the parameter solution set. Simultaneously, the candidate solution set is fed into the classification model, and the predicted probability of normal braking state is output. A preset threshold is set as a hard constraint condition for entering the feasible region. When the predicted probability of normal braking state is greater than the preset threshold, the point is allowed to enter the feasible candidate parameter solution set for braking and subsequent iterative optimization is performed.

[0098] The normal state prediction probability and hard constraints are defined as follows, with the hard constraints calculated using the constructed hard constraint function:

[0099]

[0100]

[0101] In the above formula, p ok (x) represents the probability that the state s of the candidate parameter solution x is the normal braking state. The preset normal state probability threshold p is calculated using the feasibility constraint function c(x). th The probability p of the normal braking state of the sample ok The difference between (x) and p, when c(x)≤0, i.e. ok (x)≥p th When a candidate parameter solution x is determined to satisfy the hard constraints of the feasible region, subsequent iterations of optimization continue. In the process of multi-objective optimization, whenever a new candidate parameter solution is generated, it must satisfy the above-mentioned hard constraints of the feasible region before subsequent sorting, selection, and next iteration updates can be performed.

[0102] In a preferred embodiment, p th The value is set to 0.60. To ensure the representative solution remains in a normal braking state during the optimization process, and to avoid overlooking potential parameter solutions with better optimization results, p... th The selection should be based on the training performance of the sample state classification model. If p th If the value of p is low, the abnormal braking parameter solution may still enter the non-dominated candidate parameter solution. th The parameter solution with a higher value and better target performance within the normal braking range may be excluded. In this embodiment, p th The value of 0.60 is determined based on the feature dimension of the parameters, the number of samples, and the prediction accuracy after the classification model is trained. It can be adjusted according to specific circumstances.

[0103] Step S600: After the multi-objective optimization iteration is completed, a representative parameter solution is selected from the non-dominated candidate parameter solution set of the optimization results, and the corresponding simulation backtest is performed on the one-dimensional parameterized model. The original time series data is post-processed and calculated to obtain the simulation value of the target performance index of the backtest point.

[0104] In a preferred embodiment, to enable the diesel engine hydraulic braking system to respond faster, reduce oil leakage in the actuator chamber, and stabilize actuator displacement fluctuations, the representative solutions in the candidate parameter solution set should include at least the following: Minimum solution Minimum solution Minimum solution and comprehensive compromise optimal solution. The minimum solution is the one that minimizes the time it takes for the valve to reach its lower limit. The minimum solution refers to the solution that minimizes the volume of oil leakage in the valve chamber during a single stable load cycle. The minimum solution refers to the solution that minimizes the displacement fluctuation of the actuator valve during the loading phase. In the parameter solution set P where the probability of the normal braking state is greater than a preset threshold, the objective regression surrogate model respectively... , and Prediction, denoted as , , argmin is the candidate parameter solution that minimizes the objective function. The minimum values ​​of the three objective performance indicators are as follows:

[0105]

[0106] Meanwhile, in order to balance all objectives and achieve better overall braking performance of the diesel engine hydraulic braking system, a compromise optimal solution x is selected. bal， , , The predicted values ​​after normalization are respectively , , Each is assigned a weight. , , , + + =1. The weighting coefficient can be adjusted according to engineering requirements. In a preferred embodiment, it is considered that... , , Equally important, therefore = = =1 / 3. The optimal compromise solution is...

[0107]

[0108] .

[0109] Step S700: Calculate the relative error between the simulated and predicted values ​​of the representative parameter solution target performance index. If the relative error is less than a preset threshold, output the recommended parameter combination and corresponding performance index; if the relative error is greater than the preset threshold, supplement the backtest samples into the training samples for a new round of regression model training and multi-objective optimization and validation. The fixed test set remains unchanged after the initial partitioning and is only used for the final test, not participating in the intermediate training and validation of the regression model. Therefore, the backtest samples can correct the bias of the Pareto front prediction and the braking state boundary in the previous round of regression model prediction. The fixed test set samples are not contaminated by subsequent sample supplementation and are used to finally evaluate the model prediction accuracy after multiple rounds of training.

[0110] In one optimized implementation, sample data is labeled to ensure its intended use is fixed and it doesn't become mixed up, thus preventing contamination. For regular samples, they are pre-classified as either allowed into the training set, allowed into the validation set, or designated as a fixed test set. For example, `is_frozen_test=1` indicates the sample is labeled as a fixed test sample, `eligible_for_training=0` indicates the sample is not allowed for model training, and `eligible_for_validation=0` indicates the sample is not allowed for model validation. Fixed test samples, after being labeled using this procedure, are only used for the final evaluation of model prediction accuracy and are prohibited from being included in the training or validation sets.

[0111] For feedback samples, `is_feedback_point=1` indicates that the sample is marked as a feedback sample. The additional flags `eligible_for_training=1`, `eligible_for_validation=0`, and `is_frozen_test=0` ensure that the sample is only added to the training set and does not enter the validation set or the fixed test set. The above field names are merely one implementation method and do not constitute a limitation on the scope of protection of this invention.

[0112] The technical solution of the present invention will be described below through a specific embodiment and related drawings:

[0113] Figure 1 To illustrate the sequential relationship between the various steps, this invention forms a closed-loop process following the order of modeling, sampling, simulation, post-processing to extract indicators, state label generation, classification and regression modeling, multi-objective optimization under hard constraints, representative solution backtesting, and sample backflow to update the training set. The following steps will explain the key technical aspects of each step:

[0114] Step S100: In this embodiment, a one-dimensional parametric model of the diesel engine hydraulic braking system is constructed using Simcenter Amesim. Based on the structural composition and working principle of the diesel engine hydraulic braking system, the system is divided into an oil supply system, control valves, ball valves, actuator valves, and a load control module. According to this division, the oil supply system is established, the valve body is built using a hydraulic component design library, mechanical devices such as springs and mass blocks are added, and control signals are set. Based on the part drawings, relevant parameters, the load period of the periodic load, and the application method are set. In this embodiment, constructing a one-dimensional parametric model reduces the reliance on numerous bench tests in the early stages of parameter design. It also describes the interaction relationships between the hydraulic source, valve body, springs, and load, thereby enabling batch simulation to obtain a corresponding number of sample data for machine learning. Simulation backtesting is also performed on the representative solutions obtained from multi-objective optimization. The design variables used for machine learning include the actuator valve spring stiffness, ball valve spring stiffness, control valve spring stiffness, oil supply pressure, oil viscosity parameters, load force, and load start time. The valve reaches its lower limit time. The volume of oil leakage in the valve chamber during a single stable load cycle and The displacement fluctuation of the actuator valve during the load application phase is selected as the target performance indicator.

[0115] In step S200, this embodiment employs Latin hypercube sampling to generate samples with good coverage within the space defined by various parameter ranges. This ensures that the samples include both normal and abnormal braking conditions, providing sample data for the subsequent sample state classification model that can learn the boundaries of the physically feasible and infeasible regions. The samples are then input into a one-dimensional parameterized model for batch simulation to obtain the corresponding original time-series data.

[0116] Step S300: Calculate the original time-series data of the sample to obtain the target performance value, and mark the sample state and failure cause according to the target definition. For diesel engine hydraulic braking systems, the missing data of abnormal samples is not mainly due to data discrete errors, but rather because the system cannot achieve normal braking function under the combination of parameters such as multi-stage valve body structural parameters and applied forces. For example, when the oil supply pressure is insufficient, the liquid pressure cannot enable the valve core to overcome the spring force and static friction. The pressure response between each stage of the valve body cannot be transmitted step by step according to the normal braking process, and the actuator valve cannot reach the lower limit. In the original time-series diagram of the actuator valve displacement, this is manifested as the actuator valve displacement never reaching the lower displacement threshold. Therefore, the time Y1 for the actuator valve to reach the lower limit cannot be obtained according to the definition. When the oil supply pressure is sufficient to help the actuator valve overcome resistance and reach the lower limit, but due to the high viscosity of the engine oil, a large damping is generated, and the actuator valve therefore responds slowly. At this time, under frequent load action, the actuator valve cannot achieve force balance in time, resulting in strong displacement fluctuations. In the timing diagram of the valve displacement, the displacement curve can never enter the stable zone, so the displacement fluctuation Y3 of the valve under load cannot be calculated according to the definition.

[0117] like Figure 4 The sample valve can reach the lower limit, but when it is at the lower limit, the valve displacement fluctuates drastically, exceeding the stable band half-width threshold of 0.1 mm. This means the hydraulic braking system cannot reach a stable operating state, and the program marks it as "failed_stable_not_found". Figure 5 The sample could not reach the lower limit threshold of 3.5mm for the actuator valve displacement, meaning the hydraulic braking system could not complete the braking normally, and the program marked it as "failed_reach_lowerlimit".

[0118] Step S400: In this embodiment, a random forest is used as the classification model to process all samples, outputting the predicted probability that a sample is in a normal braking state. Based on the random forest regression model for three performance indicators, the corresponding feature importance ranking is obtained. Using the feature importance of the random forest regression model as a basis, the strength of the correlation between each input parameter and each performance indicator is initially determined, and the top K features with higher importance for each indicator are selected accordingly. The candidate feature subset consists of the top K important feature subsets for each indicator and the entire feature subset. Subsequently, regression models are independently established for Y1, Y2, and Y3. Among candidate models such as random forest, gradient boosting tree, support vector regression, Gaussian process regression, and BP neural network, the prediction evaluation indicators obtained through cross-validation are compared, and the model with better prediction performance under the corresponding objective is selected as the subsequent multi-objective optimization regression surrogate model.

[0119] The dominant input parameters for predicting different target performance indicators are not the same, such as... Figure 6 As shown, for The time it takes for the actuator to reach its lower limit, oil viscosity, ball valve spring stiffness, and oil supply pressure are of relatively high importance, indicating... It is mainly affected by the combined influence of oil damping, ball valve spring force, and oil supply pressure; such as Figure 7 As shown, for The volume of oil leakage in the valve chamber during a single stable load cycle is most significantly affected by oil viscosity, followed by ball valve spring stiffness. This indicates that oil flow resistance and the ball valve's opening and closing status influence the volume of oil leakage in the valve chamber. Figure 8 As shown, for During the loading phase, the displacement fluctuation of the actuator valve is significantly contributed by the control valve spring stiffness, oil viscosity, actuator valve spring stiffness, and load force parameters, indicating that... It is affected by the coupling of valve system stiffness, oil damping and load force.

[0120] , , The reason for using independent model selection is that the three objectives are different in terms of physical meaning, dominant parameters, and input-output mapping relationship. It is mainly controlled by the force balance between oil supply pressure, spring force and valve core resistance, and the changes are usually relatively continuous with relatively clear monotonicity. It reflects the level of oil leakage under stable load conditions of the actuator valve; Reflecting the dynamic stability of the actuator valve near its lower limit, it is easily affected by the coupling effects of cyclic loads, local damping, and valve body resistance, exhibiting nonlinear fluctuations. Since the physical definitions and dominant characteristics during prediction differ among the three types of targets, using a single model and feature set for modeling will easily lead to mutual interference in the accuracy of the regression model's predictions.

[0121] Step S500: Using the NSGA-II multi-objective genetic algorithm, an initial population is generated within the space defined by the parameter range. The target regression model, which predicted better performance in the previous step, replaces the one-dimensional parameterized model to predict the target performance indicators of the candidate points in the initial population. At the same time, the candidate points are input into the braking state classification model, which outputs the prediction probability p of the normal braking state. ok (x). In this embodiment, a preset threshold p is set. th =0.60 is used as a hard constraint condition for entering the braking feasible region. According to the hard constraint function c(x)=p th -p ok (x), when c(x)≤0, p ok (x)≥p thWhen the predicted probability of normal braking state is greater than a preset threshold, the point is allowed to enter the feasible region and proceed to the next NSGA-II iterative optimization. Whenever a new candidate parameter solution is generated through selection, crossover, or mutation, it is still necessary to predict the probability of normal braking state and determine hard constraints. When the probability of normal state is greater than a preset threshold, subsequent sorting, selection, and the next iteration update are performed until the iterative optimization ends.

[0122] This invention employs hard constraints instead of penalizing abnormal braking candidate points in the population. This is because candidate points that do not meet normal braking performance requirements are not necessarily due to poor optimization during iterations, but rather because, under this parameter combination, the actuator valve cannot reach the lower limit (i.e., braking cannot respond normally), or the actuator valve, even when at the lower limit, still fluctuates beyond the stability zone, affecting the stability of the exhaust valve opening process and resulting in unstable braking performance. These candidate points all impact the braking performance of real-world diesel engine braking systems. If a penalty term were applied, abnormal samples would participate in multi-objective ranking as poorly performing samples. The regression surrogate model might provide seemingly reasonable target performance predictions for such samples, but these predictions would conflict with the physical definitions used in actual post-processing calculations.

[0123] Step S600: After the iterative optimization is completed, representative parameter solutions are selected from the Pareto front, including... Minimum solution Minimum solution Minimum solution and comprehensive compromise optimal solution , , The predicted values ​​after normalization are respectively , , In this embodiment, the three target performances are given equal weights, as shown in the following expression:

[0124]

[0125] The representative solution is fed into a one-dimensional parametric model for simulation verification to obtain simulated values ​​of the target performance index. This step is to prevent the output of the surrogate model from being directly used as the final result. The surrogate model is not a one-dimensional parametric model and can only make approximate predictions; therefore, it needs to be verified through simulation backtesting.

[0126] Step S700: Calculate the relative error between the simulated and predicted values ​​of the representative parameter solution's performance index. If it is lower than a preset threshold, it indicates that the prediction accuracy for this region does not meet the requirements. Then... Figure 3As shown, the backtested parameter points are used as feedback to unidirectionally supplement the training set, increasing the sample density in that region. This allows the regression model to correct the prediction bias in that region during the next training round. However, the data is not allowed to flow to the fixed test and validation sets, preventing samples from being mixed into other sample sets and creating the illusion of good generalization. If the data exceeds a preset threshold, it is recommended for output.

[0127] 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 method for performance prediction and optimization of diesel engine hydraulic braking systems based on machine learning, characterized in that, Includes the following steps: Step S100: Establish a one-dimensional parametric model of the diesel engine hydraulic braking system; Step S200: Generate parameter samples in the parameter space of the design variables, and input the parameter samples into the one-dimensional parameterized model to perform batch time series simulation to obtain the original time series data corresponding to each parameter sample; Step S300: Calculate and obtain the target performance index based on the original time series data, and generate sample state labels and failure causes by combining physical boundary conditions and running results to form a training table; Step S400: Build a sample state classification model using all samples in the training table, output the predicted probability of the sample's braking normal state, construct a candidate subset based on the importance of the input features output by the random forest regression model, and build a target regression surrogate model for the effective samples. Step S500: In the multi-objective optimization process, the regression model replaces the one-dimensional parameterized model to predict the candidate parameter solution set. The sample state classification model outputs the braking normal state prediction probability of the candidate parameter solution, and uses this as a hard constraint condition for entering the feasible region optimization iteration. When the braking normal state prediction probability is lower than the preset threshold, the candidate solution is prevented from entering the feasible region. Step S600: Select representative parameter solutions from the non-dominated candidate parameter solution set, input them into the one-dimensional parameterized model for simulation backtesting, perform post-processing calculations on the original time series data, and obtain the simulation values ​​of the target performance index corresponding to the representative parameter solutions; Step S700: Calculate the relative error between the simulated and predicted performance index values ​​of the representative solution, and compare them with the preset threshold.

2. The method for performance prediction and optimization of diesel engine hydraulic braking system based on machine learning according to claim 1, characterized in that, The specific method for establishing the one-dimensional parametric model of the diesel engine hydraulic braking system in S100 is as follows: Based on the structure and working principle of the diesel engine hydraulic braking system, a one-dimensional parametric model including the oil supply system, control valve, ball valve, actuator valve, and load control module is established; relevant component parameters are set according to the brake system part drawings, and the load cycle and application method of the periodic load are set; among the settable parameters, at least the actuator valve spring stiffness, control valve spring stiffness, ball valve spring stiffness, oil viscosity, load application start time, oil supply pressure, and load force parameters are selected as design variables for subsequent regression model training, and the index corresponding to the actuator valve reaching the lower limit time is selected. The corresponding index of oil leakage volume in the valve cavity during a single stable load cycle. Corresponding indexes for valve displacement fluctuation during the load application phase As a performance metric for subsequent multi-objective optimization, the design variable vector is represented as: Where x represents the candidate parameter solution or the combination of design variables; , … These represent the specific design variables in the following categories: actuator spring stiffness, ball valve spring stiffness, control valve spring stiffness, oil viscosity, load initiation time, oil supply pressure, and load force. 'd' represents the dimension of the design variables. (The last part, "x", appears to be a typo and can be left as is.) v (t) is defined as the valve displacement at time t, q(t) is defined as the leakage flow rate at time t, and x L Defined as the lower threshold of the actuator valve, T end Defined as the simulation end time, T c Defined as a preset load cycle.

3. The method for performance prediction and optimization of diesel engine hydraulic braking system based on machine learning according to claim 2, characterized in that, The The calculation formula is as follows: Among them, t reach Indicates the moment when the valve displacement first reaches the preset lower threshold, where inf represents the earliest moment the condition is met, and t represents the continuous time variable, [0, T]. end [] indicates the simulation time interval. (t) represents the valve displacement at time t. This indicates the preset lower threshold of the actuator valve; if no threshold is met... ≥ At that moment, Undefined, the reason for sample failure is that the lower limit of the actuator valve was not reached.

4. The method for performance prediction and optimization of diesel engine hydraulic braking system based on machine learning according to claim 2, characterized in that, The The sampling time at which the lower limit threshold is first reached is used directly without introducing interpolation. Furthermore, no additional interpolation function is introduced, and the time t is the departure time from the lower limit. leave The definition is as follows: In the above formula, t leave τ represents the time after the lower limit is reached; τ represents any time after t; this definition means that the valve has reached the lower limit, and we are looking for the earliest time after which the valve displacement is always below the lower limit threshold.

5. The method for performance prediction and optimization of diesel engine hydraulic braking system based on machine learning according to claim 2, characterized in that, The stable load cycle determination rule is as follows: Let the stable band be B. s and the kth complete load cycle W k They are respectively: In the above formula, B s Indicates the stable band, δ s W represents the half-width of the stable band. k This represents the k-th complete load cycle, where t0 represents the start point of the load cycle, k represents the cycle number, and T... c Indicates the preset load period, if W k If the displacement of the internal actuator valve meets the preset stable zone constraint, the system is determined to have entered the stable load stage.

6. The method for performance prediction and optimization of a diesel engine hydraulic braking system based on machine learning according to claim 5, characterized in that, The determination rule for the stable load cycle is based on the stable band percentage threshold and the cycle mean deviation threshold, and the determination formula is as follows: Where, η k The valve displacement falls into the stable zone B during the kth cycle. s The percentage of sampling points; μ k It is the average displacement of the actuator valve within the k-th cycle; η th It is the preset threshold for the proportion of stable bands; δ μ It is a preset mean deviation threshold, when η k ≥η th And |μ k -x L |≤δ μ When, determine W k The stability condition is met; Calculate using the following formula: In the formula, t stable T represents the starting point of a stable period. c Let q(t) represent a complete load cycle, and q(t) represent the leakage flow at time t. If the starting point of a stable cycle cannot be identified, then... Undefined; Use I Y3 express Statistical interval: The displacement fluctuation Y3 of the actuator valve during the period from reaching the lower limit to leaving the lower limit is calculated as follows: In the formula, I Y3 for The statistical interval; max is the maximum displacement of the statistical interval, and min is the minimum displacement of the statistical interval.

7. The method for performance prediction and optimization of diesel engine hydraulic braking system based on machine learning according to claim 1, characterized in that, In step S400, for valid samples, a subset containing the top K important features is constructed based on feature importance. Combining this with a cross-validation prediction evaluation metric, the model with the better performance is selected as the regression model for that performance metric. Feature importance is output using a random forest regression model. The candidate feature subset includes the full feature subset and a subset containing the top K important features constructed based on feature importance ranking. All input feature sets F and the top K high-importance feature subsets Fi are also included. K They are respectively: In the formula, F is the set of all sample input features, f j For the j-th input feature, the top K subsets of dimensionality-reduced features, sorted by feature importance from highest to lowest, are denoted as F. K The top K features, sorted by importance from highest to lowest, are f. （1） to f （k） .

8. The method for performance prediction and optimization of a diesel engine hydraulic braking system based on machine learning according to claim 1, characterized in that, The normal state prediction probability and hard constraints are defined as follows, with the hard constraints calculated using the constructed hard constraint function: In the above formula, p ok (x) represents the probability that the state s of the candidate parameter solution x is the normal braking state, output by the sample state classification model, and the preset normal state probability threshold p is calculated using the feasibility hard constraint function c(x). th The probability p of the normal braking state of the sample ok The difference between (x) and p, when c(x)≤0, i.e. ok (x)≥p th When a candidate parameter solution x is determined to satisfy the hard constraints of the feasible region, in the process of multi-objective optimization, whenever a new candidate parameter solution is generated, it must satisfy the above hard constraints of the feasible region before subsequent sorting, selection and the next iteration update process can be carried out.

9. The method for performance prediction and optimization of a diesel engine hydraulic braking system based on machine learning according to claim 2, characterized in that, The representative optimal solutions in the candidate parameter solution set of S600 include at least the following: Minimum solution Minimum solution Minimum solution and comprehensive compromise optimal solution; among which, The minimum solution refers to the time it takes for the actuator valve to reach its lower limit. Candidate parameter solutions that yield the minimum value; The minimum solution is the solution that minimizes the leakage volume of the valve chamber oil during a single stable load cycle. Candidate parameter solutions that yield the minimum value; The minimum solution is the amount of displacement fluctuation of the actuator valve during the loading phase. Candidate parameter solutions that yield the minimum value; Let P denote the set of candidate parameter solutions that satisfy the hard constraints, and x be a candidate parameter solution. , , These are the target regression surrogate model pairs , and Let argmin represent the predicted value, and let argmin represent the candidate parameter solution that minimizes the objective function. , , The minimum solution is expressed as: The comprehensive compromise optimal solution is defined as: In the formula, x bal This represents the optimal solution through a comprehensive compromise. , , They represent respectively to , , The predicted value after normalization; , , Let represent the weight coefficients corresponding to the three objectives, and satisfy . + + =1.

10. The method for performance prediction and optimization of a diesel engine hydraulic braking system based on machine learning according to claim 1, characterized in that, In S700, if the relative error is less than a preset threshold, the recommended parameter combination and corresponding performance index are output; if the relative error is greater than the preset threshold, the representative parameter solution verified by simulation is added to the training sample set and the regression model is trained again and the next round of closed-loop optimization is performed. The fixed test set remains unchanged after the initial partitioning and is used only for the final evaluation of the model's prediction accuracy. It does not participate in the training and validation of the regression model construction, nor does it include the representative parameter solutions that have been backtested through simulation. The representative parameter solutions that have been backtested through simulation are used as backtest samples and are added to the training set only according to the one-way inflow principle, without flowing into the validation set or the fixed test set.

Images