A Machine Learning-Based Method for Engine Torque Anomaly Diagnosis and Key Parameter Localization

By using machine learning methods, especially the XGBoost model, combined with data acquisition and preprocessing, a standard torque evaluation benchmark was established, which solved the problem of lack of quantitative standards in engine torque diagnosis, and realized automated and accurate torque anomaly diagnosis and fault location, thus improving diagnostic accuracy and efficiency.

CN123116047APending Publication Date: 2026-11-17HARBIN DONGAN AUTO ENGINE CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611147217.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-07-30
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Current technologies for engine torque diagnosis lack a unified quantitative standard and rely on manual comparison of external characteristic curves, resulting in problems such as strong subjectivity in judgment, low level of automation, and difficulty in tracing the cause of failure.

Method used

By employing machine learning methods, particularly the XGBoost model, and combining data acquisition, preprocessing, model pre-building, and iterative training, a method for torque anomaly diagnosis and key parameter localization is established. Through the normalization and interpolation of multi-dimensional operating parameters, a standard torque evaluation benchmark is generated. The model is then used to train and generate torque state labels and analyze the contribution of parameters, thereby achieving automated diagnosis.

Benefits of technology

It enables quantitative evaluation of torque diagnosis, reduces human error, improves diagnostic accuracy and efficiency, can quickly locate fault causes, is suitable for offline performance analysis and online monitoring of the whole machine, and reduces costs.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123116047A_ABST
    Figure CN123116047A_ABST
Patent Text Reader

Abstract

This invention relates to a machine learning-based method for diagnosing engine torque anomalies and locating key parameters, belonging to the field of engine torque diagnosis technology. It improves the unified quantitative evaluation benchmark for torque diagnosis, addressing the problems of cumbersome fault cause tracing processes and poor automation and interpretability. The invention collects multi-dimensional operating parameters from the engine test bench, reads the corresponding measured maximum torque, and obtains the global standard torque based on discrete points of the engine's external characteristics through one-dimensional linear interpolation and spline interpolation fitting, establishing a quantitative torque evaluation benchmark. It calculates the torque deviation and generates torque compliance status labels, trains a classification model using the XGBoost algorithm to complete automated diagnosis, and combines global feature importance analysis, weighted normalization, and sorting to locate key parameters of torque anomalies. This invention integrates quantitative benchmarks, machine learning classification, and interpretable feature analysis, reducing subjective bias in manual judgment, improving the automation level of torque diagnosis and fault diagnosis efficiency, and demonstrating strong engineering applicability.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of engine torque diagnosis technology, specifically a machine learning-based method for engine torque anomaly diagnosis and key parameter localization. Background Technology

[0002] Engine torque is a key indicator characterizing the overall power performance of an engine. Current technologies mainly rely on manual comparison of external characteristic curves and experience to judge whether the operating conditions meet the standards, lacking a unified quantitative judgment standard and calculation basis. Traditional methods are difficult to accurately quantify the magnitude of torque deviation, nor can they quantify the influence weight of various operating parameters on torque anomalies. They suffer from drawbacks such as strong subjectivity in judgment, low level of automation, and difficulty in tracing the cause of failure. There is an urgent need for a quantifiable and interpretable intelligent diagnostic solution. Summary of the Invention

[0003] To address the problems existing in the background technology, the present invention provides a machine learning-based method for diagnosing engine torque anomalies and locating key parameters.

[0004] To achieve the above objectives, the present invention adopts the following technical solution: a machine learning-based method for diagnosing engine torque anomalies and locating key parameters, the method comprising the following steps:

[0005] S1: Data acquisition and preprocessing, obtaining multi-dimensional operating parameters of the engine test bench, reading the measured maximum torque data under the corresponding working conditions, identifying and eliminating invalid abnormal samples caused by drastic fluctuations in working conditions, and constructing sample training and test sets;

[0006] S2: Model pre-building and iterative training. Pre-build the XGBoost model framework and configure hyperparameters. Read the discrete data of external characteristics of engine performance acceptance and complete the filling of discrete points. Perform key parameter normalization processing on the multi-dimensional operating parameters obtained in S1. Perform supervised training on the model based on the processed data and determine whether the training is complete. If not, adjust the model hyperparameters and repeat the training. If complete, proceed to the next step.

[0007] S3: Model evaluation and diagnosis. First, the test set is used to complete the model torque compliance test and the contribution ranking of each parameter. Then, external working condition samples to be diagnosed are read, input into the model to generate torque status labels, and key problem parameters affecting torque anomalies are extracted to complete the diagnosis.

[0008] In S1, the multi-dimensional operating parameters include engine speed, intake manifold pressure, outlet water temperature, average fuel consumption, exhaust temperature, crankcase pressure, and turbine inlet temperature, all collected from the engine test bench.

[0009] In step S2, discrete point filling is performed using one-dimensional linear interpolation and spline interpolation to interpolate the discrete points of the engine's external characteristic speed-torque, densifying the mesh and supplementing the torque values ​​between adjacent points to smooth the data curve and obtain the standard torque value corresponding to any speed n within the operating speed range. Establish a complete torque performance evaluation benchmark; within the rated operating speed range, the torque performance judgment error based on the interpolation benchmark is ≤5%.

[0010] In S2, the key parameter normalization process involves numerically normalizing and mapping the multidimensional operating parameters to eliminate the dimensional differences between different parameters and obtain the model input feature vector.

[0011] In S2, the supervised training of the model takes multi-dimensional operating parameters as input and torque state labels generated based on torque deviation as output, iteratively updating the model accuracy until convergence; where torque deviation... The calculation formula is: in, This represents the standard torque value corresponding to rotational speed n. This represents the measured maximum torque value corresponding to rotational speed n.

[0012] The torque performance status label uses As a benchmark for compliance, the corresponding torque deviation is no greater than 10%, and binary labels are generated according to the following formula. : ;when When =1, the torque is considered to meet the standard. When the value is 0, it is judged as torque abnormality; or expanded to a multi-category label of mild abnormality / moderate abnormality / severe abnormality.

[0013] S3 further includes the following steps:

[0014] S31: Verify the accuracy of the model in judging the torque compliance status based on the test set, analyze the characteristic contribution of each operating parameter and sort them from high to low according to their weight;

[0015] S32: Read external working condition samples to be diagnosed, generate torque status labels for the samples to be diagnosed, extract key problem parameters with the highest contribution ranking, and complete torque anomaly diagnosis.

[0016] In S31, the torque compliance test is based on the test set to verify the model's classification accuracy, and the model's classification diagnostic accuracy on the test sample set is ≥90%; the parameter contribution ranking is based on the feature contribution analysis to calculate the contribution weight of each operating parameter and sort them from high to low according to the weight.

[0017] The formula for calculating the contribution weight of the parameter is as follows: in, Let be the feature contribution value corresponding to the i-th operating parameter, and k be the total number of operating parameters; select the parameters with higher contribution weights and identify them as the key parameters affecting torque anomalies.

[0018] Compared with the prior art, the beneficial effects of the present invention are:

[0019] 1. Establish a quantitative torque evaluation benchmark to reduce human judgment errors. Use one-dimensional linear interpolation and spline interpolation to obtain the standard torque across the entire range. The judgment error within the rated speed range is ≤5%, replacing manual curve comparison and reducing subjective bias.

[0020] 2. Automated quantitative diagnosis of torque anomalies with stable identification accuracy. Based on the XGBoost algorithm, it mines the nonlinear relationship between parameters and torque state, achieving a test set diagnostic accuracy of ≥90%, reducing manual operation, and adapting to scenarios such as bench testing and certification inspection.

[0021] 3. Quantify the impact of parameters to quickly pinpoint the cause of failure. By combining the importance of global features and the contribution of single-sample features, normalizing weights and ranking them, key parameters of abnormal torque can be intuitively identified, improving the efficiency of fault diagnosis.

[0022] 4. Complete workflow and good engineering adaptability. It supports both offline model training and online real-time diagnostics, is compatible with existing benchtop data acquisition systems, and can be used for offline performance analysis and online monitoring of the entire system, with low deployment costs.

[0023] In summary, this invention combines interpolation quantization benchmarks, machine learning classification, and feature contribution analysis to improve the problems of inconsistent traditional torque diagnostic standards, strong subjectivity, difficulty in tracing faults, and low automation. It balances diagnostic accuracy and efficiency and has industrial application value. Attached Figure Description

[0024] Figure 1 This is an overall flowchart of the method of the present invention;

[0025] Figure 2 This is a torque distribution diagram of the present invention;

[0026] Figure 3 This is a distribution diagram showing the influence of various parameters in this invention. Detailed Implementation

[0027] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the invention, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0028] This embodiment describes a machine learning-based method for diagnosing engine torque anomalies and locating key parameters. The method includes the following steps:

[0029] S1: Data acquisition and preprocessing, obtaining multi-dimensional operating parameters of the engine test bench, reading the measured maximum torque data under the corresponding working conditions, identifying and eliminating invalid abnormal samples caused by drastic fluctuations in working conditions, and constructing sample training and test sets;

[0030] S2: Model pre-building and iterative training. Pre-build the XGBoost model framework and configure hyperparameters. Read the discrete data of external characteristics of engine performance acceptance and complete the filling of discrete points. Perform key parameter normalization processing on the multi-dimensional operating parameters obtained in S1. Perform supervised training on the model based on the processed data and determine whether the training is complete. If not, adjust the model hyperparameters and repeat the training. If complete, proceed to the next step.

[0031] S3: Model evaluation and diagnosis. First, the test set is used to complete the model torque compliance test and the contribution ranking of each parameter. Then, external working condition samples to be diagnosed are read, input into the model to generate torque status labels, and key problem parameters affecting torque anomalies are extracted to complete the diagnosis.

[0032] In S1, the multi-dimensional operating parameters include engine speed, intake manifold pressure, outlet water temperature, average fuel consumption, exhaust temperature, crankcase pressure, and turbine inlet temperature, all collected from the engine test bench.

[0033] In step S2, discrete point filling is performed using one-dimensional linear interpolation and spline interpolation to interpolate the discrete points of the engine's external characteristic speed-torque, densifying the mesh and supplementing the torque values ​​between adjacent points to smooth the data curve and obtain the standard torque value corresponding to any speed n within the operating speed range. Establish a complete torque performance evaluation benchmark; within the rated operating speed range, the torque performance judgment error based on the interpolation benchmark is ≤5%.

[0034] In S2, the key parameter normalization process involves numerically normalizing and mapping the multidimensional operating parameters to eliminate the dimensional differences between different parameters and obtain the model input feature vector.

[0035] In S2, the supervised training of the model takes multi-dimensional operating parameters as input and torque state labels generated based on torque deviation as output, iteratively updating the model accuracy until convergence; where torque deviation... The calculation formula is: in, This represents the standard torque value corresponding to rotational speed n. This represents the measured maximum torque value corresponding to rotational speed n.

[0036] The torque performance status label uses As a benchmark for compliance, the corresponding torque deviation is no greater than 10%, and binary labels are generated according to the following formula. : ;when When =1, the torque is considered to meet the standard. When the value is 0, it is judged as torque abnormality; or expanded to a multi-category label of mild abnormality / moderate abnormality / severe abnormality.

[0037] S3 further includes the following steps:

[0038] S31: Verify the accuracy of the model in judging the torque compliance status based on the test set, analyze the characteristic contribution of each operating parameter and sort them from high to low according to their weight;

[0039] S32: Read external working condition samples to be diagnosed, generate torque status labels for the samples to be diagnosed, extract key problem parameters with the highest contribution ranking, and complete torque anomaly diagnosis.

[0040] In S31, the torque compliance test is based on the test set to verify the model's classification accuracy, and the model's classification diagnostic accuracy on the test sample set is ≥90%; the parameter contribution ranking is based on the feature contribution analysis to calculate the contribution weight of each operating parameter and sort them from high to low according to the weight.

[0041] The formula for calculating the contribution weight of the parameter is as follows: in, Let be the feature contribution value corresponding to the i-th operating parameter, and k be the total number of operating parameters; select the parameters with higher contribution weights and identify them as the key parameters affecting torque anomalies.

[0042] Taking the bench test data of a certain type of China VI gasoline engine as an example, 800 sets of operating condition data within the rated speed range of 1000~5500 r / min were collected and divided into training set and test set in a 7:3 ratio. The standard torque value at the corresponding speed was calculated by linear interpolation or spline interpolation. After verification by factory calibration and performance acceptance benchmark value, the maximum error of torque performance judgment in the entire speed range was 4.2%, which meets the accuracy requirement of ≤5%. The number of decision trees of the XGBoost model was set to 120 and the learning rate was 0.08. Formal iterative training will be carried out after the subsequent generation of labels and standardized features. After training, the accuracy rate of torque anomaly diagnosis on the test sample set was 91.3%, which can stably realize the rapid location of key parameters.

[0043] This invention addresses the shortcomings of traditional engine torque diagnosis, which relies on manual experience, lacks a unified quantitative evaluation benchmark, and makes it difficult to trace fault parameters. It constructs a complete intelligent diagnostic system based on the fusion of engine bench test data and the XGBoost machine learning algorithm. Its working principle is as follows: First, the engine bench data acquisition system acquires multi-dimensional operating parameters under all operating conditions, including engine speed, intake manifold pressure, coolant temperature, average fuel consumption, exhaust temperature, crankcase pressure, and turbine inlet temperature, as well as the measured maximum torque data under the corresponding operating conditions. After preprocessing such as outlier removal, missing value completion, and operating point alignment, an original sample dataset is constructed and divided into training and testing sets for subsequent model training and diagnosis. The analysis provides a standardized basic data source; then, based on the discrete speed-torque external characteristic calibration points of the engine at the factory, linear interpolation and spline interpolation methods are used to fit and calculate the discrete calibration points and refine the speed grid to obtain the standard torque value corresponding to any speed n within the rated operating speed range. This establishes a unified and quantifiable engine torque evaluation benchmark, where the torque deviation is characterized by the ratio of the difference between the measured and standard torque to the standard torque. The torque judgment error caused by the interpolation benchmark within the rated operating speed range can be controlled within 5%. At the same time, an engine torque operating condition distribution map can be generated, which intuitively presents the relative position of the actual operating condition point, the standard torque curve, and the 90% torque compliance benchmark line.

[0044] First, an XGBoost model framework is pre-built based on the preprocessed dataset. Model hyperparameters such as the number of decision trees, learning rate, and sampling ratio are configured to complete model initialization. In this stage, only the basic tree structure is fixed, without ground truth labels, classification loss function, gradient descent and weight update operations are not performed. Then, two processing tasks are completed in parallel. On the one hand, the deviation value is calculated by substituting the standard torque at the same speed and the measured maximum torque into the torque deviation formula. The threshold for compliance is set at 90% of the standard torque of the measured maximum torque. Binary sample labels of torque compliance L(n)=1 and torque abnormality L(n)=0 are generated by segmented formula. Alternatively, mild, moderate and severe torque abnormality can be divided according to the deviation range to generate multi-class labels.

[0045] On the other hand, the multidimensional operating parameters are normalized to eliminate dimensional differences, resulting in standardized feature vectors of uniform specifications. These standardized features, along with state labels, are input into a pre-built model to construct a classification loss function. Multiple rounds of gradient descent are performed to iteratively update the model weights. The convergence criterion is that the test set loss stably decreases to a threshold. If the convergence condition is not met, the model hyperparameters are adjusted and retrained until a mature, convergent model suitable for torque diagnosis is obtained. XGBoost is an extreme gradient boosting ensemble learning algorithm that relies on the serial construction of multiple classification and regression trees and gradient descent to optimize the loss function and fuse weak classification outputs. It has strong adaptability to industrial structural chemical condition data and boasts advantages such as fast training speed, resistance to overfitting, and the ability to output parameter contribution scores. It can accurately uncover the nonlinear correlation between multidimensional operating parameters and torque performance. After model iteration convergence, an independent test sample set is used to verify the classification effect, ensuring that the model's classification diagnostic accuracy on the test sample set is not less than 90%. Simultaneously, global feature importance analysis and... Single-sample feature contribution analysis quantifies the individual impact of each parameter on torque anomalies under a single operating condition by relying on the global feature importance statistics based on the built-in feature splitting gain of XGBoost. Both analysis results can generate parameter contribution distribution maps, addressing the shortcoming of traditional machine learning black-box models that cannot trace the root cause of faults. In the online diagnosis stage, external engine operating condition samples to be diagnosed are read, preprocessed using normalization rules identical to those used in the training set, and then fed into the trained model for inference. The output samples' torque status labels and the original feature contribution values ​​of each parameter are then calculated using the contribution weight conversion formula. All parameters are sorted from highest to lowest weight, and the parameters with the highest weights are selected as the key parameters for this torque anomaly. The entire process requires no manual comparison or experience-based judgment, automatically completing torque performance diagnosis and fault root cause location, significantly reducing fault investigation time. It can be adapted to various industrial scenarios such as engine bench test analysis, factory quality control, and whole-machine online monitoring.

[0046] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of the equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A method for diagnosing engine torque anomaly and locating key parameters based on machine learning, characterized in that, The method includes the following steps: S1: Data acquisition and preprocessing, obtaining multi-dimensional operating parameters of the engine test bench, reading the measured maximum torque data under the corresponding working conditions, identifying and eliminating invalid abnormal samples caused by drastic fluctuations in working conditions, and constructing sample training and test sets; S2: Model pre-building and iterative training. The XGBoost model framework is pre-built and hyperparameters are configured. Discrete data of external characteristics of engine performance acceptance are read and discrete points are filled in. Key parameter normalization is performed on the multi-dimensional operating parameters obtained in S1. The model is supervised training based on the processed data to determine whether training is complete. If not, the model hyperparameters are adjusted and training is repeated. If complete, proceed to the next step; S3: Model evaluation and diagnosis. First, the test set is used to complete the model torque compliance test and the contribution ranking of each parameter. Then, external working condition samples to be diagnosed are read, input into the model to generate torque status labels, and key problem parameters affecting torque anomalies are extracted to complete the diagnosis.

2. The machine learning based engine torque anomaly diagnosis and key parameter localization method according to claim 1, characterized in that: In S1, the multi-dimensional operating parameters include engine speed, intake manifold pressure, outlet water temperature, average fuel consumption, exhaust temperature, crankcase pressure, and turbine inlet temperature, all collected from the engine test bench.

3. The method for engine torque anomaly diagnosis and key parameter localization based on machine learning according to claim 1, characterized in that: In the S2, the discrete points are filled by using one-dimensional linear interpolation and spline interpolation method to interpolate the discrete points of the engine external characteristic speed-torque, to encrypt the grid and supplement the torque value between adjacent points, to smooth the data curve, and to obtain the standard torque value corresponding to any speed n in the working speed range , to establish a complete torque performance evaluation benchmark; and in the rated working speed range, the torque performance error based on the interpolation benchmark is ≤5%.

4. The machine learning based engine torque anomaly diagnosis and key parameter localization method of claim 1, wherein: In S2, the key parameter normalization process involves numerically normalizing and mapping the multidimensional operating parameters to eliminate the dimensional differences between different parameters and obtain the model input feature vector.

5. The method for engine torque anomaly diagnosis and key parameter localization based on machine learning according to claim 1, characterized in that: In S2, the model supervised training takes multi-dimensional operating parameters as input and torque state labels generated based on torque deviation as output, and iteratively updates the model accuracy until convergence. Wherein the torque deviation degree The calculation formula is: Wherein, The standard torque value corresponding to the rotation speed n, The measured maximum torque value corresponding to the rotation speed n.

6. The method for engine torque anomaly diagnosis and key parameter localization based on machine learning according to claim 5, characterized in that: The torque performance status label uses As a benchmark for compliance, the corresponding torque deviation is no greater than 10%, and binary labels are generated according to the following formula. : ;when When =1, the torque is considered to meet the standard. When the value is 0, it is judged as torque abnormality; or expanded to a multi-category label of mild abnormality / moderate abnormality / severe abnormality.

7. The method for engine torque anomaly diagnosis and key parameter localization based on machine learning according to claim 1, characterized in that: S3 further includes the following steps: S31: Verify the accuracy of the model in judging the torque compliance status based on the test set, analyze the characteristic contribution of each operating parameter and sort them from high to low according to their weight; S32: Read external working condition samples to be diagnosed, generate torque status labels for the samples to be diagnosed, extract key problem parameters with the highest contribution ranking, and complete torque anomaly diagnosis.

8. The method for engine torque anomaly diagnosis and key parameter localization based on machine learning according to claim 7, characterized in that: In S31, the torque compliance test is based on the test set to verify the model's classification accuracy, and the model's classification diagnostic accuracy on the test sample set is ≥90%; the parameter contribution ranking is based on the feature contribution analysis to calculate the contribution weight of each operating parameter and sort them from high to low according to the weight.

9. The method for engine torque anomaly diagnosis and key parameter localization based on machine learning according to claim 8, characterized in that: The formula for calculating the contribution weight of the parameter is as follows: in, Let be the feature contribution value corresponding to the i-th operating parameter, and k be the total number of operating parameters; select the parameters with higher contribution weights and identify them as the key parameters affecting torque anomalies.