A computer-aided engine torque model simulation construction method

Abstract

The present application relates to the technical field of computer-aided simulation model construction, and discloses a computer-aided engine torque model simulation construction method, comprising: constructing and training a model cluster and a dissenter model with different learning principles; a vector representing the divergence mode of the model cluster is constructed and input into a health state representation model to output a preliminary health state; when the model cluster is highly consistent internally but deviates from the dissenter model prediction as a whole, the preliminary health state is corrected, the present application establishes a correlation between the divergence mode in the model and the physical health state of the engine, and perceives the physical entity health state by using a pure software method, the built-in heterogeneous model verification mechanism can also identify and avoid overall cognitive blind spots caused by training data source pollution, and the reliability of health state perception is improved.

Description

TECHNICAL FIELD

[0001] The application relates to a computer-aided engine torque model simulation construction method and belongs to the technical field of computer-aided simulation model construction. BACKGROUND

[0002] At present, an engine torque model is constructed by a numerical method using a sparse data set collected from an engine bench test, which has become a common technical means to support simulation test and calibration of an engine electronic control unit software, and the value of the method in improving research and development efficiency and reducing test cost has become a common technical cognition. However, in actual engineering application, an inherent limitation of this data processing method is that no matter what algorithm is used, the design goal of the existing model is to give a unique deterministic torque point estimate value for any input working condition. When facing the training data which is inevitably sparse due to the limitation of test cost, this single numerical value output mode introduces a technical risk: when a new working condition point that has never been run on the bench is input into the model, the model will still return an accurate numerical value, but the numerical value is essentially a speculation result based on mathematical interpolation or extrapolation. The model itself cannot evaluate the degree of deviation of the speculation result from the true physical value, so objectively, the prediction behavior of the model in the data unknown area is to output a deterministic speculation value whose basis strength cannot be declared.

[0003] Moreover, even if advanced algorithms such as machine learning are introduced at the data processing level, the existing technology still cannot escape the traditional framework of seeking a unique optimal solution at the bottom level, so it cannot avoid the risk brought by the cognitive blind spot. For example, a Chinese invention patent with the authorized publication number CN117189381A discloses a method and device for improving an engine charge control model. The scheme uses machine learning to optimize a single existing charge control model to correct the calibration values of its pulse spectrum table and interpolation lookup table. This method is essentially a fine tuning of the parameters of a deterministic model, and its goal is to improve the fitting accuracy of the model in the training data coverage area, but it does not change the structure of outputting a single deterministic result. When the training data itself has an unnoticed systematic deviation or is sparse in some working condition area, the model trained by the method will still confidently give an incorrect result that is seriously deviated from the physical truth, and the model itself does not have the internal mechanism to identify such cognitive blind spots and express uncertainty about the prediction result.

[0004] To improve the model performance, increasing the bench test density to expand the training data is a direct solution, but this will cause the research and development cost to rise sharply, and the feasibility is limited in engineering practice, using a more complex nonlinear fitting model with stronger data processing ability also does not change the basic way of point estimation of model output, a more complex model may only produce a more complex mathematical speculation in the sparse data area, which cannot answer the question of how high the credibility of the current prediction is, the reason is that the objective function of the existing model construction usually only focuses on reducing the error in the data supported area, and does not contain the corresponding mechanism for expressing the uncertainty in the data missing area, the existing technology mainly has the following deficiencies in this problem: 1, the data structure of the model output is single, only contains the estimated value of the physical quantity, lacks the quantitative evaluation dimension of the credibility of the estimated value itself, and the provided information is incomplete; 2, in the data sparse or boundary extrapolation area, the deterministic prediction value of the model output will introduce potential risks to the downstream simulation test or control strategy, a prediction value with a false high-precision appearance may be more misleading than an estimated value with an explicit error range; 3, the existing technology improvement path mainly focuses on improving the prediction accuracy of the model in the data coverage area, and there is no effective technical solution with feasible calculation for how to make the model have the ability to quantitatively output the credibility index of the prediction. Therefore, how to design a data processing method, so that the torque model constructed by the method can quantitatively output the credibility index of the estimated value itself while outputting the estimated value of the torque, thereby avoiding the generation of misleading deterministic speculation in the data sparse area, becomes a technical problem to be solved by the present application. SUMMARY

[0005] The present application provides a computer-aided engine torque model simulation construction method, which mainly aims to solve the problem that the prediction result of the existing torque model in the data sparse area is only a single deterministic value, which cannot quantitatively output its own credibility, and thus has the problems of incomplete information and potential misleading risk.

[0006] To achieve the above purpose, the present application provides a computer-aided engine torque model simulation construction method, which comprises the following computer-executed steps:

[0007] Step a, based on the training data set stored in the computer readable medium, a model cluster and a dissenter model are constructed and offline trained, the model cluster comprises a plurality of basic model units, and the learning principle of the dissenter model is different from that of the basic model unit;

[0008] Step b, receiving a real-time input working condition, and obtaining the independent torque prediction value output by each basic model unit in the model cluster for the input working condition and the dissenter torque prediction value output by the dissenter model;

[0009] Step c, determining a torque estimation value based on statistical central tendency of the plurality of independent torque predictions, and constructing a divergence mode vector representing divergence mode within the model cluster;

[0010] Step d, inputting the divergence mode vector into a health state representation model to output a preliminary representation result corresponding to the physical health state of the engine;

[0011] Step e, when the statistical dispersion degree of the plurality of independent torque predictions is lower than a preset first threshold value, and the difference between the torque estimation value and the dissenting torque prediction value is higher than a preset second threshold value, then the preliminary representation result is corrected to a preset cognitive conflict state.

[0012] Preferably, in step a, a plurality of sub-training sets are generated by self-aggregated sampling of the training data set, and the plurality of sub-training sets are used to independently train a plurality of base model units respectively.

[0013] Preferably, in step c, the divergence mode vector is constructed according to the rule that the difference between each independent torque prediction value and the torque estimation value is taken as an independent component of the divergence mode vector; and in step d, the health state representation model stores a nonlinear mapping relationship from the divergence mode vector space to the physical health state space of the engine.

[0014] Preferably, the method further comprises: identifying whether the engine is in a transient operating condition; when it is identified that the engine is in a transient operating condition, capturing a plurality of divergence mode vectors arranged in time sequence within a time window of the transient operating condition to form a divergence response trajectory; updating the preliminary representation result based on dynamic characteristic analysis of the divergence response trajectory; the dynamic characteristic analysis includes extracting the maximum value of the modulus of the divergence mode vectors in the divergence response trajectory, and the time difference from the start of the transient operating condition to the appearance of the maximum value.

[0015] Preferably, the method further comprises: storing a plurality of divergence mode vectors arranged in time sequence within a time window to form a divergence historical trajectory; performing dynamic evolution characteristic analysis on the divergence historical trajectory to extract a trajectory feature vector representing the drift rate and fluctuation amplitude thereof; determining a prediction result related to the remaining useful life of the engine based on the trajectory feature vector.

[0016] Preferably, the method further comprises: obtaining a context vector representing the running context of the engine; before inputting the divergence mode vector into the health state representation model, performing a normalization process on the divergence mode vector based on the context vector, the normalization process being configured to correct divergence mode changes not related to health state caused by changes in running altitude or ambient temperature.

[0017] Preferably, the method further includes: determining in real time the current macroscopic operating domain of the engine; arbitrating the preliminary characterization result output in step d or the cognitive conflict state output in step e based on a physical rule that defines the feasibility of different physical health states under different macroscopic operating domains; and only using the preliminary characterization result or the cognitive conflict state as the final output when the arbitration is passed.

[0018] Preferably, in step a, the basic model unit is a shallow neural network, and the dissenter model is a gradient boosting decision tree model or a support vector regression model; in step c, the statistical central tendency is the mean of multiple independent torque prediction values, and the statistical dispersion is the standard deviation of multiple independent torque prediction values.

[0019] Preferably, in the trajectory feature vector, the drift rate The calculation rules follow the following relationship: ,in, Let this be the divergence pattern vector at the current moment. for The divergence pattern vector from a time unit ago, For a preset time window length, Calculate the vector magnitude.

[0020] Preferably, the preset first threshold and the preset second threshold are obtained by offline calibration based on the benchmark dataset collected under engine health conditions.

[0021] Compared with the prior art, the beneficial effects of the present invention are:

[0022] 1. By constructing a model cluster containing multiple basic model units and using a self-administered aggregation sampling method to differentiate the learning processes of each unit, when responding to the same input condition, it can not only determine the torque estimate based on the statistical central trend of multiple independent torque predictions, but also simultaneously use the statistical dispersion between these predictions as an indicator of the reliability of the torque estimate output. This process changes the structure of the torque model in traditional data processing methods, which only outputs a single deterministic value, thus expanding the data dimension of the output results. Downstream electronic control units or simulation systems receive the torque estimate and the uncertainty quantification information of the estimate under the current operating condition, thereby providing a decision-making basis for subsequent control strategies or testing processes that cannot be obtained in existing technologies.

[0023] 2、The application further constructs the deviation between each independent prediction value in the model cluster and the final torque estimation value into a divergence mode vector representing the divergence form of the internal opinion in the model cluster; since the change of the physical health state of the engine will cause systematic change of its torque response, which will be stably reflected on the specific form of the divergence mode vector, therefore, by inputting this vector into a pre-trained health state representation model, a direct correlation path from the internal calculation state of the torque model to the external physical health of the engine is established, so that the data processing method not only completes its core torque simulation task, but also outputs a representation result about the health state of the engine in parallel using the internal data generated during the operation, providing a new information source for predictive maintenance.

[0024] 3、While constructing the model cluster composed of multiple basic model units with the same learning principle, an independent dissenter model with different learning principle is trained, and the deviation between the torque estimation value of the model cluster and the prediction value of the dissenter model is monitored during torque prediction; when the statistical dispersion degree within the model cluster shows high confidence, but the deviation presented by it is higher than the preset state, the system will correct the confidence index to low confidence, which establishes an internal checking loop based on different data interpretation methods to identify and alert the cognitive blind spot of the model cluster as a whole which may be caused by the common deviation of the training data source, avoiding the risk of the system outputting an error answer deviating from the physical reality while providing high confidence guarantee. BRIEF DESCRIPTION OF DRAWINGS

[0025] Fig. 1 The flow chart of the torque simulation and health evaluation of the application based on model cluster divergence and heterogeneous verification;

[0026] Fig. 2 The system core application scenario use case diagram of the calibration engineer in the simulation environment of the application;

[0027] Fig. 3 The system deployment architecture diagram covering data collection offline training and hardware-in-the-loop simulation of the application. DETAILED DESCRIPTION

[0028] In order to make the purpose, technical scheme and advantages of the application more clear, the technical scheme of the application will be described in detail below, but those skilled in the art can understand that the technical scheme of the application can be modified or replaced equivalently without departing from the spirit and scope of the application, and these modifications and replacements are within the protection scope of the application.

[0029] The application relates to a computer-aided engine torque model simulation construction method, which is applied to the software function development and simulation test link of an engine electronic control unit, and provides a data processing framework which can not only predict engine torque, but also synchronously evaluate the uncertainty of the prediction result and further represent the physical health state of the engine; the method is realized by a software system deployed and executed in a computer, mainly includes four core stages of cooperative work, namely, an offline construction stage of a model group based on sparse training data, a parallelized prediction and divergence representation stage for real-time input working conditions, a preliminary inference stage of the physical health state based on the divergence mode, and a cognitive conflict identification and result correction stage based on heterogeneous verification, thereby forming a process for data processing and information enhancement; in an engineering data processing task, a technical problem objectively existing is that the reliability of the output result of a numerical model constructed by using a sparse data set with limited high-cost collection cannot be self-evaluated when facing unknown input regions which are not covered by the training data; in order to solve the technical problem, the method is configured to first perform step a, that is, based on an initial training data set stored in a computer readable medium, a model cluster and a dissenter model are constructed and offline trained; the data structure of the training data set is a series of data points, each data point contains a group of working condition parameters for representing the engine running state as input, and a corresponding torque value obtained by high-precision measurement as a label, a data point can be expressed as an input vector [2000, 75, 90] corresponding to the speed 2000 r / min, the load 75% and the coolant temperature 90 The construction of the model cluster is specifically that a plurality of sub-training sets with the same structure but different contents are generated by self-aggregation sampling of the training data set, if the original training data set contains M data points, the generation process of a sub-training set is that M times are randomly drawn from the original data set with replacement, a new training set containing M data points is formed, and N sub-training sets are obtained by repeating the process N times; then, the N different sub-training sets are used to independently train N basic model units, and the N basic model units jointly constitute the model cluster.

[0030] In one specific implementation, the base model unit is a shallow neural network with a structure consisting of an input layer, a hidden layer with 10 neurons, and an output layer. Simultaneously, a dissident model is independently trained using the unsampled, complete original training dataset. It's important to note that the learning principle of the dissident model differs from that of the base model unit. If the base model unit uses a shallow neural network, the dissident model can be configured as a gradient boosting decision tree model or a support vector regression model. Through this step, a model cluster consisting of multiple base model units with different configurations and a dissident model with different learning principles are established in the computer's memory, providing a data processing entity for quantifying uncertainty by analyzing the differences in opinions between them. To quantitatively define the differences in learning principles between the base model unit and the dissident model, a difference verification procedure is executed during the model selection phase. This procedure first prepares a dataset containing various known functions, including linear, polynomial, and step functions. A baseline numerical dataset of relationships is used, and then candidate basic model unit algorithms and dissenting model algorithms are independently trained using this dataset. The prediction error distribution of the two algorithms when facing different functional relationships is recorded. When the error distribution of the two algorithms shows a systematic difference that can be identified by statistical tests on different function types, for example, one algorithm has a lower residual variance in the linear region than the other, but a higher overshoot in the step region, then it is determined that the learning principles of the two algorithms have the differences required for subsequent heterogeneous verification. To determine the number N of basic model units contained in the model cluster, an offline calibration procedure is performed to achieve a certain engineering balance between uncertainty quantification capability and computational cost. This procedure first sets a sequence of candidate N values, such as 3, 5, 10, 15, 20, and uses a dense validation set containing interpolation and extrapolation test points. For each N value in the sequence, a corresponding model cluster is independently trained and its statistical dispersion of all test points in the interpolation region is calculated when processing all test points in the validation set. mean Statistical dispersion of all operating points in the extrapolation region mean Then the resolution ratio of the two was calculated. And select the ratio that makes the resolution so high. The minimum N value at which the growth reaches saturation for the first time, meaning that further increasing N no longer causes [a specific effect]. The corresponding value at the time of the upgrade is used as the quantity in the final deployment.

[0031] After the model population is completed, the next data processing task is how to convert multiple parallel prediction results into a structured output with uncertainty information; in view of this, the system receives a real-time input working condition, and executes steps b and c, that is, obtains the independent torque prediction value of each basic model unit in the model cluster and the dissenting torque prediction value of the dissenter model, and performs statistical processing based on the prediction values; when an input working condition vector [2500, 80, 92] is received, N basic model units (taking N = 5 as an example) in the model cluster are called in parallel, and five independent torque prediction values are output, denoted as , and the specific values may be {180.5, 181.2, 179.8, 180.9, 179.1} N·m; at the same time, the dissenter model also processes the input working condition, and outputs a dissenting torque prediction value , and the value may be 180.6 N·m; the system then determines a final torque estimate value based on the statistical central tendency of multiple independent torque prediction values, and the statistical central tendency is configured to calculate the mean value, that is, N·m; in addition, the method of the application does not discard the difference information between the independent prediction values, but further constructs a divergence mode vector characterizing the divergence mode in the model cluster, and the construction rule is to take the difference between each independent torque prediction value and the torque estimate value as an independent component of the vector, that is, , according to the aforementioned values, it can be obtained that ; the execution of this step enables the system to not only obtain a more robust central estimate value than any single model prediction, but also convert the divergence between models into a structured data for characterizing the current prediction uncertainty state of the system ; in order to associate the abstract data state inside the model with the health state of the external physical entity, the system is configured to execute step d, that is, input the divergence mode vector generated in step c into a health state characterization model to output a preliminary characterization result corresponding to the physical health state of the engine; the health state characterization model is a pre-off-line trained pattern recognizer, which internally stores a nonlinear mapping relationship from the divergence mode vector space to the engine physical health state space; the training process of the model is as follows: first, an extended training data set is prepared, which is obtained by simulating on a test bench or collecting working condition and torque data of the engine in different known health states, such as benchmark health, slight carbon deposition, injector aging, etc., then, the model cluster in the method of the application processes each data point in the data set to generate its corresponding divergence mode vector , and the divergence mode vector is input into the health state characterization model to output a preliminary characterization result corresponding to the physical health state of the engine The health state representation model is trained as input with its corresponding known health state as label; in real-time operation, when a new divergence mode vector is generated, the model compares it with the stored mode and outputs a preliminary health state representation result, such as a baseline health; through the data processing mode, a representation result associated with the internal state of the physical entity can be obtained.

[0032] If there is a common, unperceived systematic deviation in the original training data in a certain local area, all basic model units may learn this common error, resulting in a serious deviation from the physical reality when they give a highly consistent but incorrect answer; at this time, only analyzing the internal divergence of the model cluster cannot identify such collective cognitive blind spots; to avoid this risk, the method of the present application sets step e, an internal arbitration mechanism based on heterogeneous verification; the triggering condition of the mechanism is designed to satisfy two logical judgments at the same time: the first judgment is that the statistical dispersion degree of multiple independent torque prediction values is lower than a preset first threshold, and the statistical dispersion degree here is configured to calculate the standard deviation of the prediction values ; the second judgment is that the consensus result presented by the model cluster to the outside, that is, the absolute value of the difference between the torque estimation value and the dissenting torque prediction value is higher than a preset second threshold; only when both conditions are met at the same time, the system will determine that this is a high-risk pseudo-consensus event, and will forcibly modify the preliminary representation result output in step d, regardless of its content, to a preset cognitive conflict state for alerting the user; the specific values of the first threshold and the second threshold are determined by an offline calibration procedure, which is to collect a baseline data set in the full operating range of an engine confirmed to be in a healthy state, input the data set into the method of the present application, record the values of and at all operating points, and calculate the statistical distribution of the two groups of values respectively, and finally set the result of adding three times the standard deviation to the mean value of each to the first threshold and the second threshold respectively; the mechanism provides a mechanism for identifying and avoiding overall misjudgment caused by data source deviation by introducing an independent decision reference system with different learning principles.

[0033] In the case of no conflict, to further improve the dimension of engine physical state perception, the method of the present application can also include a series of extended data processing steps; one extension is that after outputting the preliminary health state in step d, an adaptive normalization processing link based on the running context can be introduced, which first uses the existing sensors of the engine to obtain a context vector representing the current running altitude or environmental temperature, and then uses the context vector to normalize the divergence mode vector​ Before inputting into the health state characterization model, a normalization network conditioned on the context vector is used to A mathematical transformation is performed to correct the divergence pattern change caused by external environment changes, which is not related to the health state, so that the health state judgment has higher consistency under different operating environments; another extension is to introduce analysis of the divergence pattern vector time series. The system can continuously store multiple divergence pattern vectors arranged in time sequence in a first-in-first-out sliding time window to form a divergence history trajectory, and analyze the dynamic evolution characteristics of the trajectory, such as calculating the drift rate according to the relationship is the drift rate, is the divergence pattern vector at the current time, is the divergence pattern vector a time unit ago, is a preset time window length, is a vector length calculation, and finally based on the extracted trajectory feature vector composed of drift rate and fluctuation amplitude, a prediction result related to the remaining effective life of the engine is further determined; in addition, to improve the detection sensitivity of early weak faults, the method of the present application can also be configured to identify in real time whether the engine is in transient operating conditions such as sudden acceleration or sudden deceleration. When the transient operating condition is identified, a series of divergence pattern vectors are captured synchronously within the transient time window to form a divergence response trajectory, and the preliminary characterization result is updated based on the dynamic characteristic analysis of the trajectory; finally, to prevent the data-driven model from producing judgments that violate physical common sense, the method of the present application can also add an arbitration step based on physical rules before the final output. The step first determines the current operating domain of the engine in real time, such as the idle domain or the high load domain, and then performs a reasonableness check on the characterization result output by the previous step based on a rule base that defines the feasibility of different physical health states under different operating domains. Only when the result is determined to be physically possible in the operating domain will it be taken as the final output.

[0034] ​Embodiment 1: This embodiment is a specific operational example of the technical solution described in a specific industrial application scenario; in a hardware-in-the-loop simulation test environment of an engine electronic control unit software, a calibration engineer is verifying a new control algorithm, the running of which relies on a torque model trained through a sparse data set collected from engine bench tests, in the test process, the engineer inputs a parameter combination of a working condition that simulates a transient working condition, which is theoretically correct but not included in the original training data set; when the input working condition is sent to the simulation system, the model cluster in the system receives the input working condition, and the N basic model units contained therein simultaneously perform independent torque prediction; since the working condition point is located in the sparse region of the training data set, each basic model unit outputs N independent torque prediction values that are different from each other based on the slightly different function surfaces formed in the training process; the system then processes the N prediction values, on the one hand, calculates the mean value to obtain a torque estimate , on the other hand, calculates the standard deviation of the N prediction values and constructs a corresponding divergence mode vector At this time, due to the unknown input working condition, the highly inconsistent internal prediction of the model cluster, the calculated standard deviation value exceeds the baseline value generated when processing known working condition points of the training data.

[0035] Further, the system does not output the single torque estimate as the final result directly to the downstream control algorithm, but outputs and the standard deviation as a data pair as a reverse confidence index; the simulation test environment receiving the data pair is configured with internal logic to trigger a low confidence warning when value exceeds the preset threshold, so the calibration engineer observes not a single torque value but a torque range explicitly marked as a low confidence speculation on the software interface; at the same time, the divergence mode vector The health state characterization model is sent into, since its morphology is not different from the divergence mode of the benchmark health state, the preliminary characterization result output by the system is still the benchmark health, and after comparing with the dissent torque prediction value output by the independently running dissenter model, the system does not trigger the correction of the cognitive conflict state; in this data processing flow, the quantification mechanism of the opinion divergence within the model cluster operates in parallel with the health state characterization mechanism, the former is responsible for evaluating the uncertainty of the current predicted behavior, and the latter independently evaluates the physical state of the simulated object, both of which together provide the user with a decision basis containing multiple information dimensions; finally, according to the low confidence warning output by the system, the calibration engineer does not adopt the torque estimation value obtained in this simulation as the calibration basis of the control algorithm, but marks this input working condition point as a key area that needs additional physical bench test to supplement data, thereby avoiding the risk of introducing a speculative value based on mathematical extrapolation and high uncertainty into the subsequent research and development process as a design benchmark; the application of this method makes the divergence between models explicit by measuring and outputting this information, adding a self-awareness dimension to the data processing model, allowing data users to distinguish between reliable calculations based on data and cognitive boundaries based on inference, and the entire simulation test work is therefore based on a technical foundation that has a quantitative awareness of model uncertainty.

[0036] Example 2: To objectively verify the effectiveness of the method claimed in the present application in quantifying the prediction uncertainty when dealing with sparse data sets, a numerical comparative experiment was designed and performed; the data basis of the experiment came from a physical engine test bench equipped with a torque sensor with an accuracy of ±0.5 N·m and a data acquisition system with a sampling frequency of 100 Hz, by running a series of steady-state operating points, the original data set covering the main operating range of the engine was collected; based on the original data set, two subsets were constructed, one was a sparse training set composed of 100 data points, which was sparsely distributed in the operating space and had data blank areas, the other was a dense validation set composed of 1000 data points, which uniformly covered the entire operating space, and the torque values measured with high precision in the dense validation set were used as the benchmark true values to evaluate the prediction accuracy of the model; based on the aforementioned data sets, two test groups for comparison were constructed, namely a control group and a test group; the control group used a commonly used data processing method in the prior art, that is, a single deep neural network model with three hidden layers was trained using the sparse training set, which only outputs a deterministic torque point estimate for any input operating condition; the test group used the method claimed in the present application, using the same sparse training set, according to the above procedure, a model cluster containing 10 basic model units and a gradient boosting decision tree dissident model were constructed and trained; the test process is to select several representative test operating points from the dense validation set, these operating points are divided into two categories, one is the interpolation operating point inside the envelope of the sparse training set data points, the other is the extrapolation operating point outside the envelope, then the input parameters of these test operating points are sent into the control group model and the test group method at the same time, and the output results of each are recorded.

[0037] During the experiment, there were quantitative differences in the data output of the two test groups when dealing with different types of operating points, see Table 1, which lists some test operating points and their corresponding test data; for test points 1 to 4, these interpolation operating points, because there is support around the training data, the deep neural network model of the control group and the model cluster of the test group both give prediction results with smaller deviations from the true torque value, at the same time, the statistical dispersion degree of the prediction values calculated by the test group method within the model cluster is at a lower level, at test point 2, the ​The value is 0.21, and the system determines that it is a high-confidence prediction; for test points 5 to 8, these extrapolation working points, due to the lack of training data support, the torque prediction value output by the control group model has a large deviation from the true value, and at test point 7, the prediction deviation reaches 15.8 N·m, but its output form is still a deterministic value without any uncertainty warning; the test group method in processing these extrapolation working points, each basic model unit in the model cluster lacks consistency constraint and produces quantitative prediction divergence, which is directly reflected in the value of the statistical dispersion degree The value of the test point 7 increases to 4.52, and the system determines that it is a low-confidence prediction, thereby providing a synchronous quantitative warning for the reliability of the torque estimate value .

[0038] Table 1: Comparison of test data at different test working points.

[0039]

[0040] The test data shows that by using the method claimed in the application, through the parallelization of prediction and divergence quantization, a quantitative index that can effectively represent the uncertainty of the estimated value can be output simultaneously ; the index remains low when the model prediction is in the data support area, and increases when it enters the extrapolation area with sparse data; the test results confirm that the method of the application can add a quantitative uncertainty index to a single torque estimate value, thereby avoiding the risk of outputting misleading results in the data sparse area.

[0041] Embodiment 3: This embodiment combines Figs. 1 to 3 , a computer-aided engine torque model simulation construction method is described, such as Fig. 1 ​As shown, the process begins with a training dataset containing engine operating parameters and corresponding torque values. This dataset is used to construct the model cluster and the dissenting model. The model cluster consists of multiple basic model units with the same learning principle, while the dissenting model has a different learning principle. Both are trained through bootstrap aggregation sampling and different learning principles. Upon receiving real-time input operating conditions such as engine speed, load, and temperature, the system performs parallel torque prediction. The model cluster and the dissenting model make predictions simultaneously. The output is, on the one hand, a torque estimate is generated based on the central trend predicted by the model cluster, and on the other hand, a torque estimate is generated independently by the dissenting model. Seeing the torque prediction value, the system also constructs a divergence pattern vector based on the divergence patterns predicted within the model cluster, and inputs this vector into a health status representation model to establish the correlation between the divergence patterns and the engine's physical health status, thereby outputting a preliminary engine health status representation result. Subsequently, the system enters the heterogeneous verification and arbitration stage to determine whether the model cluster is highly consistent and whether its overall prediction deviates from the dissenting model. If so, the result is corrected to a cognitive conflict state that warns of potential overall cognitive blind spots. If not, the preliminary representation result output from the previous stage is directly adopted, and the system outputs the merged torque estimate and health status.

[0042] like Fig. 2 As shown, this illustrates how a calibration engineer, acting as a user, interacts with a system that embeds this method within a simulation test environment. The engineer performs three core use cases: torque simulation prediction, inferring engine health status, and predicting remaining effective life. During torque simulation prediction, the system implicitly performs sub-functions to quantify prediction uncertainty and identify cognitive conflicts. Similarly, during engine health inference, the system implicitly arbitrates physical rules. The final results of these interactions all affect the simulation test environment. Fig. 3 As shown, the architecture comprises three core components: an engine test bench for acquiring raw data, an engineer workstation for offline processing, and a hardware-in-the-loop (HIL) simulation environment for real-time operation. The engine test bench acquires raw data through its data acquisition system and high-precision sensors and provides it to the engineer workstation, which stores sparse or benchmark datasets and deploys an offline model building and training system to generate a trained model population, namely model clusters and dissenting models. The trained models are then deployed to the hardware-in-the-loop (HIL) simulation environment, where a torque simulation and health assessment application relies on an engine electronic control unit (ECU) to run in real time.

[0043] Embodiment 4: This embodiment provides a reproducible engineering procedure for the offline calibration of the first and second thresholds used to identify the cognitive conflict state; the values of these two thresholds directly affect the sensitivity and specificity of the cognitive conflict state determination mechanism, so a systematic calibration procedure is needed to determine their specific values; the initial conditions for the execution of this calibration procedure are a data processing environment that has been deployed in the computer and contains a model cluster and a dissenter model constructed and trained according to the specific embodiment method, and a benchmark data set for calibration; the benchmark data set is obtained by collecting data on a physical engine test bench that is confirmed to be in a benchmark healthy state, and the data set contains 1000 steady-state operating points covering the full operating range of the engine, and the data structure of each operating point contains a set of operating input parameters and a high-precision torque measurement true value corresponding to it; the calibration procedure first sends the operating input parameters in the 1000 benchmark data sets one by one to the deployed data processing environment and executes steps b and c; for each operating point, the system calculates and records two core quantitative indicators, namely the statistical dispersion degree representing the degree of disagreement among the opinions in the model cluster , and the absolute value of the difference representing the degree of deviation of the model cluster's overall opinion from the dissenter model's opinion ; after processing all 1000 data points, two data sequences each containing 1000 floating-point values are obtained, one sequence is the value set under all operating points, and the other sequence is the value set under all operating points.

[0044] Next, the procedure performs independent statistical analysis on the two data sequences, i.e. calculates the mean and standard deviation of each data sequence; if the statistical analysis of the value set gives a mean of 0.35 and a standard deviation of 0.12, and the statistical analysis of the value set gives a mean of 0.40 and a standard deviation of 0.15, then according to a rule for defining the range of normal statistical fluctuations, the mean plus three times the standard deviation is taken as the boundary, to determine the final values of the two thresholds; specifically, the first threshold is determined to be , and the second threshold is determined to be ; by executing the above procedure, the values of the first and second thresholds are determined to be 0.71 and 0.85 respectively, and these two values are then fixed as the configuration parameters of the system when performing real-time cognitive conflict state determination.

[0045] Example 5: In order to make the method of the present application consistent in the health state characterization results based on the divergence mode vector under different operating altitudes or ambient temperatures, a normalization processing calibration for the context vector is performed once before system deployment; the calibration procedure is carried out on an engine test bench, multiple sets of altitude and ambient temperature combinations are set through an environmental simulation cabin, under each combination condition, the engine is controlled to operate in a reference health state and a set of test data covering all operating conditions is collected, then the method of the present application is used to process these test data, for each operating condition under each environmental combination condition, a corresponding divergence mode vector is generated , and the mean value of the divergence mode vectors of all health operating points collected under standard atmospheric pressure and 25 is taken as a reference center vector Finally, by training a normalization network with the context vector corresponding to the environmental combination condition as input and a transformation matrix that can transform the divergence mode vector under the environment to the reference center vector as output, the calibration is completed.

[0046] The construction process of the physical rule library used for arbitration in the method of the present application is as follows: first, based on the physical working principle of the engine, the continuous operating condition space of the engine is divided into several operating domains defined by clear speed and load thresholds, wherein the region with speed lower than 1000 r / min and load lower than 10% is defined as the idling domain; then, for each known physical health state of the engine, the possibility of its occurrence in each operating domain is determined according to its physical mechanism of occurrence, and the determination result is encoded into a logical lookup table, in which the rows represent different health states, the columns represent different operating domains, and the elements in the table are logical values of possible or impossible, wherein in the column corresponding to the idling domain, the element intersected with the health state of turbocharger overspeed failure is set to impossible; the lookup table is fixed in the computer memory as the basis for calling by the system when performing real-time arbitration.

[0047] Example 6: This embodiment aims to provide an offline training and construction procedure for the health state characterization model and the remaining useful life prediction model used to characterize the physical health state of the engine and predict its remaining useful life; the initial conditions of the procedure are an already constructed model cluster according to the specific embodiments, and an extended data set with time stamp and physical health state label collected through engine bench test, which covers the full life cycle data of the engine from the reference health state to various typical failure modes; the construction process of the health state characterization model is as follows: first, the aforementioned extended data set is preprocessed, each operating point in the data set is input into the already constructed model cluster, and the corresponding N-dimensional divergence mode vector , so as to transform the original working condition data set into a feature data set composed of divergence mode vectors and corresponding to the physical health state label; subsequently, a multi-classification support vector machine is selected as an algorithm architecture of the health state representation model, the transformed feature data set is taken as a training input, the support vector machine is trained through a supervised learning process, a nonlinear mapping relationship from an N-dimensional divergence mode vector space to a discrete physical health state label space is established in the support vector machine, and the model parameters after training are solidified in the computer memory.

[0048] The construction of the remaining effective life prediction model is mainly to quantitatively extract and model the dynamic evolution characteristics of the divergence historical trajectory; the process first sets a time window with a fixed length k, and in the process of engine operation, the system continuously records the divergence mode vectors in the last k time units in a first-in-first-out cache in chronological order to form a divergence historical trajectory ; then, the system calculates the trajectory and extracts a trajectory feature vector composed of two components, wherein the first component is the drift rate defined in the specific embodiment , and the second component is the fluctuation amplitude representing state stability , and the calculation rule is , wherein represents the fluctuation amplitude, represents the length of the divergence mode vector at time i; finally, a linear regression model is trained by taking the trajectory feature vector extracted from the whole life cycle data as the input and taking the real remaining effective life of the engine at the corresponding time as the label, so as to establish a prediction relationship from the dynamic evolution characteristics to the remaining effective life.

[0049] In order to further verify the technical necessity and beneficial effects of introducing the dissenter model and the heterogeneous verification arbitration mechanism in the method claimed in the present application, the following comparative examples are set to compare and illustrate.

[0050] Comparative Example 1: This comparative example aims to illustrate the inherent technical defects of the conventional technical path of relying only on the statistical dispersion degree within the model cluster to evaluate the uncertainty when facing source data pollution, in the absence of the dissenter model and the heterogeneous verification arbitration mechanism specific to the present application; the test conditions, the basic model unit (shallow neural network), and the construction method of the model cluster (containing 10 basic model units, independently trained by self-aggregated sampling of the training data set) used in this comparative example are completely consistent with those in Example 2, and the essential difference lies in that the dissenter model and the heterogeneous verification arbitration mechanism defined in step e associated therewith are removed from the data processing method of this comparative example, therefore, the only basis for the system to determine the confidence of the prediction result is the statistical dispersion degree of the multiple independent torque prediction values within the model cluster ; the training data set used in the test, on the basis of the sparse training set used, a targeted modification was made to simulate the local data source pollution caused by the instantaneous calibration drift of a certain sensor during the bench data acquisition process, which was difficult to be detected in advance. Specifically, the torque label values of all working points in the original sparse training set with a speed higher than 4000 r / min and a load greater than 80% were artificially and systematically reduced by 12%. Except for the local area, the other parts of the data set were completely the same as those in Example 2. The test procedure was to input an input parameter of a new extrapolation working point (test point 9: speed 4200 r / min, load 85%) located in the above-mentioned data pollution area but not included in the original training set into the system constructed by the comparative example without the heterogeneous verification mechanism, and record the output result. The physical true torque value of the working point after independent calibration by a high-precision sensor was 265.7 N·m. The test results are shown in Table 2.

[0051] Table 2: Test data table under specific data pollution working conditions.

[0052]

[0053] The test results show that since all 10 basic model units in the model cluster learn the training data containing systematic bias, they produce a torque estimation value (234.1 N·m) with a deviation of up to 31.6 N·m from the physical true value when facing the test point 9 in the data pollution area. More importantly, since all the basic model units learn the same wrong cognition, their predictions are highly consistent, resulting in a statistical dispersion degree of the calculated value (0.51) at a very low level. According to the judgment logic of the method, the system mistakenly labels the prediction result seriously deviating from the true value as high confidence. The test data confirm that when the key heterogeneous verification mechanism of the application is absent, even if the model cluster is used to quantify uncertainty, the system cannot identify the potential prediction risk when facing the cognitive blind area of the training data source with common bias, but will output a misleading high-confidence guarantee error deterministic prediction value, thereby introducing serious technical risks for downstream simulation tests or control strategies.

[0054] It is obvious to those skilled in the art that the present application is not limited to the details of the above exemplary embodiments, and can be implemented in other specific forms without departing from the spirit or essential characteristics of the present application.

[0055] Finally, it should be noted that the above examples are merely intended to illustrate the technical solutions of the present application and not to limit the present application. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present application.

Claims

1. A computer-aided method for constructing an engine torque model simulation, characterized in that, The method includes the following computer-executed steps: Step a: Based on the training dataset stored in a computer-readable medium, construct and train a model cluster and a dissident model offline. The model cluster contains multiple basic model units. The learning principle of the dissident model is different from that of the basic model units. Step b: Receive a real-time input condition and obtain the independent torque prediction value output by each basic model unit in the model cluster for the input condition, and the dissenting torque prediction value output by the dissenting model. Step c: Based on the statistical central trend of multiple independent torque predictions, determine a torque estimate and construct a divergence pattern vector that characterizes the divergence patterns within the model cluster. Step d: Input the divergence mode vector into a health state representation model to output a preliminary representation result corresponding to the physical health state of the engine. Step e: When the statistical dispersion of multiple independent torque prediction values ​​is lower than a preset first threshold, and the difference between the torque estimate and the discrepancy torque prediction value is higher than a preset second threshold, the preliminary characterization result is corrected to a preset cognitive conflict state. In step c, the construction rule for the divergence mode vector is: the difference between each independent torque prediction value and the torque estimate value is used as an independent component to constitute the divergence mode vector; in step d, the health state characterization model stores the nonlinear mapping relationship from the divergence mode vector space to the engine physical health state space. In step a, the basic model unit is a shallow neural network, and the dissenter model is a gradient boosting decision tree model or a support vector regression model; in step c, the statistical central tendency is the mean of multiple independent torque predictions, and the statistical dispersion is the standard deviation of multiple independent torque predictions.

2. The method for constructing a computer-aided engine torque model simulation according to claim 1, characterized in that, In step a, multiple sub-training sets are generated by performing self-organized aggregation sampling on the training dataset, and multiple base model units are trained independently using the multiple sub-training sets.

3. The method for constructing a computer-aided engine torque model simulation according to claim 1, characterized in that, The method also includes: identifying whether the engine is in a transient operating condition; when the engine is identified to be in a transient operating condition, capturing multiple divergent mode vectors arranged in time sequence within the time window of the transient operating condition to form a divergent response trajectory; updating the preliminary characterization results based on the dynamic feature analysis of the divergent response trajectory; the dynamic feature analysis includes extracting the maximum value of the modulus of the divergent mode vectors in the divergent response trajectory, as well as the time difference from the start of the transient operating condition to the occurrence of the maximum value.

4. The method for constructing a computer-aided engine torque model simulation according to claim 1, characterized in that, The method also includes: storing multiple divergence pattern vectors arranged chronologically within a time window to form a divergence history trajectory; performing dynamic evolution feature analysis on the divergence history trajectory to extract a trajectory feature vector characterizing its drift rate and fluctuation amplitude; and determining a prediction result related to the engine's remaining effective life based on the trajectory feature vector.

5. The method for constructing a computer-aided engine torque model simulation according to claim 1, characterized in that, The method also includes: obtaining a context vector representing the engine operating context; and before inputting the divergence mode vector into the health status representation model, performing a normalization process on the divergence mode vector based on the context vector, the normalization process being configured to correct for divergence mode changes related to unhealthy conditions caused by changes in operating altitude or ambient temperature.

6. The method for constructing a computer-aided engine torque model simulation according to claim 1, characterized in that, The method also includes: determining in real time the current macroscopic operating domain of the engine; arbitrating the preliminary characterization result output in step d or the cognitive conflict state output in step e based on a physical rule that defines the feasibility of different physical health states under different macroscopic operating domains; and only using the preliminary characterization result or cognitive conflict state as the final output when the arbitration is passed.

7. The computer-aided engine torque model simulation construction method according to claim 4, characterized in that, In the trajectory feature vector, the drift rate The calculation rules follow the following relationship: ,in, Let this be the divergence pattern vector at the current moment. for The divergence pattern vector from a time unit ago, For a preset time window length, Calculate the vector magnitude.

8. The method for constructing a computer-aided engine torque model simulation according to claim 1, characterized in that, The preset first threshold and the preset second threshold are obtained through offline calibration based on the benchmark dataset collected under engine health conditions.

Images