A method of load testing a gasoline engine

Abstract

The application relates to the technical field of engine testing, in particular to a gasoline engine load testing method, which comprises the following steps: obtaining an expected data set; fusing a global discriminant index and a local consistency index of each index dimension to construct a hybrid driving adaptive weight vector; based on the adaptive weight vector, a hybrid driving adaptive distance is calculated to measure the difference between any two data points in the expected data set; a density peak value clustering algorithm is executed by using the hybrid driving adaptive distance, and a clustering center with the highest comprehensive expected score is determined according to a clustering result, and a control parameter combination corresponding to the clustering center is combined as an ideal performance emission balance point of the engine under a specific working condition. According to the technical scheme, the global optimal working point of the engine can be identified, and the intelligent level of the testing method is significantly improved.

Description

Technical Field

[0001] This application relates to the field of engine testing technology, and in particular to a method for testing the load of a gasoline engine. Background Technology

[0002] As the core power source for various vehicles and mechanical equipment, the performance calibration and optimization of gasoline engines are key steps in the research and development and production process. Maximizing the power and economy of engines while meeting increasingly stringent environmental emission requirements is a typical multi-objective optimization problem.

[0003] Currently, one technical approach is to employ a data-driven intelligent testing method. This method typically involves automatically scanning multiple key control parameter combinations of the engine under specific engine operating conditions using a test bench, thereby collecting massive amounts of multi-dimensional performance and emission data points. Subsequently, unsupervised learning algorithms such as Density Peak Clustering (DPC) are used to analyze this high-dimensional data. By calculating the local density and relative distance of data points, density peaks in the data space are automatically identified. These identified density peak regions are considered to be the operating range where the engine's overall performance is optimal, thus providing engineers with a basis for calibration decisions.

[0004] However, the standard DPC algorithm uses undifferentiated, purely mathematical Euclidean distance to measure the relationship between data points. In the complex physical space of engine performance and emissions, the physical meaning, numerical scale, and engineering importance of various dimensional indicators vary greatly. Traditional Euclidean distance, through equal-weighted sum-of-squares calculations, completely erases the physical connotation of these indicators. This means that the density peaks found by the algorithm may simply be geometrically crowded but mediocre in engineering performance. Conversely, performance peaks that truly meet specific engineering goals such as high power, low fuel consumption, and low emissions are ignored because they are not dense enough under undifferentiated distance measurement. Ultimately, this leads to optimization results that deviate significantly from the true optimal balance point, losing their engineering guidance value. Summary of the Invention

[0005] To address the technical problem of low optimization accuracy in existing density peak clustering algorithms, this application provides a gasoline engine load testing method.

[0006] This application provides a method for load testing of a gasoline engine, comprising: acquiring an expected dataset containing multiple performance and emission indicators generated by a combination of multiple sets of control parameters under specific operating conditions; for each data point in the expected dataset, fusing the global discriminative power index and the local consistency index of each indicator dimension to construct a hybrid-driven adaptive weight vector; calculating the hybrid-driven adaptive distance based on the adaptive weight vector to measure the difference between any two data points in the expected dataset; performing a density peak clustering algorithm using the hybrid-driven adaptive distance, and determining the cluster center with the highest comprehensive expected score based on the clustering results, and using the control parameter combination corresponding to the cluster center as the ideal performance and emission balance point of the engine under specific operating conditions.

[0007] This application can dynamically assign weights to each dimension of each data point in the space, enabling the distance metric to automatically focus on the most important performance dimension in the current region, thereby accurately discovering the globally optimal performance-emission balance point and solving the problem of inaccurate optimization by traditional methods.

[0008] In one embodiment, the process of obtaining the expected dataset is as follows: collect the original data vector of the engine under multiple sets of control parameters, and perform max-min normalization on each indicator dimension; perform expected transformation on the normalized value according to the optimization direction of the indicator, take the normalized value of the positive indicator as its expected value, and define the expected value of the negative indicator as the difference between 1 and the normalized value, thereby constructing an expected dataset in which all components are uniformly categorized as having a higher performance as the value is closer to 1.

[0009] By constructing an expectation space, all indicators with different physical meanings and optimization directions are unified under the framework of the higher the expectation value, the better. This eliminates the interference caused by the inconsistency between the dimensions and the optimization direction, and lays a solid foundation for fair comparison and weight calculation of all dimensions in the future.

[0010] In one embodiment, the process of obtaining the global discriminative power index is as follows: a Gaussian mixture model is used to fit the data distribution of each indicator dimension in the desired dataset, and the optimal number of components for that dimension is determined according to the Bayesian information criterion; the global discriminative power index is calculated based on the optimal number of components.

[0011] In one embodiment, the global discriminative power index satisfies the following relationship: ;in, The global discriminative power index is the one with dimension o. and are the optimal Gaussian component numbers of the data distribution in the 0th and mth dimensions, respectively, where M is the total number of dimensions.

[0012] In one embodiment, the process of obtaining the local consistency index is as follows: for each data point in the desired dataset, determine the local neighborhood formed by its multiple nearest neighbors, and calculate the local standard deviation of each indicator dimension within the neighborhood; calculate the local consistency index based on the local standard deviation.

[0013] By calculating the local consistency of each data point in each dimension, it is possible to identify indicators that are stable and reliable within a specific performance range, allowing the weights to adapt to the local characteristics of the data and further improving the accuracy of optimization.

[0014] In one embodiment, the local consistency index satisfies the following relationship: ;in, Let be the local consistency index of the i-th data point in dimension o. Let be the local standard deviation of dimension o within the neighborhood of the i-th data point. Let be the global standard deviation of the 0th dimension over the entire expected dataset. To prevent extremely small positive numbers with a denominator of zero.

[0015] In one embodiment, the weights of each dimension in the adaptive weight vector satisfy the following relationship: ;in, Let be the adaptive weight of the i-th data point in dimension o. and These are the global discriminative power indices for dimensions o and m, respectively. Let M be the local consistency index of the i-th data point on dimension o, and M be the total number of dimensions.

[0016] In one embodiment, the hybrid-driven adaptive distance satisfies the following relationship: ;in, For the hybrid driving adaptive distance of data point a and data point b, and These are the adaptive weights of points a and b in the o-th dimension, respectively. and These are the expected values ​​of points a and b in the o-th dimension, respectively.

[0017] By using hybrid-driven adaptive distance, differences in key dimensions can be automatically amplified while differences in non-key dimensions are suppressed, ensuring that the clustering algorithm can identify data points that are truly similar and clustered in engineering terms, rather than geometrically close points.

[0018] In one embodiment, the combination of control parameters includes air-fuel ratio and ignition advance angle.

[0019] The technical solution of this application has the following beneficial technical effects:

[0020] This application combines the global distinguishing potential of dimensions with the local stability of data points, which can both macroscopically determine which performance indicators are key to distinguishing different working states and microscopically determine the reliability of these indicators within a specific performance region, thereby generating a suitable dimension weight vector for each data point and effectively improving the intelligence and automation level of the optimization process.

[0021] Furthermore, by using hybrid drive adaptive distance, the importance of each dimension can be dynamically adjusted according to the different performance space locations of the data points, ensuring that the found balance point not only has excellent overall performance but also strong stability, providing effective support for the precise calibration of the engine and improving the accuracy and robustness of the optimization results. Attached Figure Description

[0022] Figure 1 This is a flowchart of a gasoline engine load testing method according to an embodiment of this application. Detailed Implementation

[0023] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0024] Figure 1 This is a flowchart of a gasoline engine load testing method according to an embodiment of this application. Figure 1 As shown, the gasoline engine load test method includes steps S101 to S104, which are described in detail below.

[0025] S101, obtain the expected dataset containing multiple performance and emission indicators generated by the combination of multiple sets of control parameters of the engine under specific operating conditions.

[0026] In one embodiment, at a specified engine test condition, such as 3000 rpm and 50% load, the control system of the engine test bench can automatically scan multiple sets of control parameter combinations within a preset range. The control parameters mainly refer to key variables that affect the engine combustion process, such as air-fuel ratio and ignition advance angle. For example, the air-fuel ratio can be set to vary from 12.5 to 14.7 with a step size of 0.2, and the ignition advance angle can be set to vary from 15 degrees to 30 degrees with a step size of 1 degree.

[0027] In this optional embodiment, for each set of scanned control parameters, the engine is allowed to run stably for a period of time, and then a raw data vector containing multiple performance and emission indicators is collected by sensors. For example, four performance and emission indicators can be selected: torque, fuel consumption rate, nitrogen oxide emissions, and carbon monoxide emissions.

[0028] After collecting the original dataset of all data points, to eliminate the influence of the dimensions and numerical ranges of each indicator, max-min normalization is performed on all data of each dimension o to obtain normalized values, which are then mapped to the [0,1] interval. Subsequently, to unify the optimization direction of all indicators, indicator expectation transformation is performed. For example, torque is a positive indicator where the larger the better, while NOx is a negative indicator where the smaller the better. For positive indicators, the expected value is the corresponding max-min normalized value, and for negative indicators, the expected value is the difference between 1 and the normalized value. By subtracting the normalized value of the negative indicator from 1, the expected dataset of all components is obtained with the unified standard that the closer the value is to 1, the more ideal the performance.

[0029] In this way, through data collection and expectation space construction, a dimensionless data foundation with consistent optimization direction is provided for subsequent adaptive distance measurement and cluster analysis, ensuring the fairness and effectiveness of comparisons between different indicators.

[0030] S102, for each data point in the desired dataset, integrates the global discriminative power index and the local consistency index of each indicator dimension to construct a hybrid-driven adaptive weight vector.

[0031] In one embodiment, a global discriminative power index can be calculated for each dimension, which is used to evaluate the natural clustering potential or distinguishing ability of each performance dimension from the perspective of the overall data distribution.

[0032] Specifically, for all N data points in each dimension o of the expectation space, a Gaussian mixture model is used for fitting. The Gaussian mixture model describes the overall distribution of the data by using a weighted sum of multiple Gaussian distributions. In this embodiment, different numbers of Gaussian components, such as from 1 to 5, are used to fit the data in that dimension, and each fitting result is evaluated using the Bayesian information criterion. The number of components corresponding to the model with the smallest Bayesian information criterion value is considered the optimal number of Gaussian components for that dimension. The larger the value of the optimal number of Gaussian components, the more likely the data in that dimension is to naturally form multiple independent groups, and the greater its potential to distinguish different performance characteristics.

[0033] The global discriminative power index satisfies the following relationship:

[0034]

[0035] in, The global discriminative power index is the one with dimension o. and are the optimal Gaussian component numbers of the data distribution in the 0th and mth dimensions, respectively, where M is the total number of dimensions.

[0036] Furthermore, a local consistency index can be calculated for each data point in each dimension to evaluate the performance stability of each dimension within the neighborhood of each data point. For each data point in the desired dataset, a local neighborhood consisting of its multiple nearest neighbors can be determined using Euclidean distance. Then, for each dimension o, the local standard deviation of all data points within that neighborhood in that index dimension is calculated, and the local standard deviation is converted into a local consistency index using a negative exponential function. The local consistency index satisfies the following relationship:

[0037]

[0038] in, Let be the local consistency index of the i-th data point in dimension o. Let be the local standard deviation of dimension o within the neighborhood of the i-th data point. Let be the global standard deviation of the 0th dimension over the entire expected dataset. To prevent extremely small positive numbers with a denominator of zero, for example, the value is 0.00001.

[0039] In this optional embodiment, the obtained macroscopic global discriminative power index can be combined with the microscopic local consistency index to generate a final adaptive weight for each dimension of each data point, and then normalized to form an adaptive weight vector, wherein the weights of each dimension in the adaptive weight vector satisfy the following relationship:

[0040]

[0041] in, Let be the adaptive weight of the i-th data point in dimension o. and These are the global discriminative power indices for dimensions o and m, respectively. Let M be the local consistency index of the i-th data point on dimension o, and M be the total number of dimensions.

[0042] In this way, by integrating global discriminative power and local consistency, a weight vector reflecting the characteristics of the data environment in which each data point is located can be dynamically generated, so that subsequent distance calculations can adaptively focus on the most informative and reliable indicator dimensions in the current context.

[0043] S103, based on the adaptive weight vector, calculates the hybrid-driven adaptive distance to measure the difference between any two data points in the desired dataset.

[0044] In an optional embodiment, a hybrid-driven adaptive distance for measuring the difference between any two data points in the desired dataset can be calculated using the obtained adaptive weight vector. This distance adaptively senses spatial characteristics, and the hybrid-driven adaptive distance satisfies the following relationship:

[0045]

[0046] in, For the hybrid driving adaptive distance of data point a and data point b, and These are the adaptive weights of points a and b in the o-th dimension, respectively. and These are the expected values ​​of points a and b in the o-th dimension, respectively.

[0047] Specifically, when calculating the distance between two data points, the weight of dimension o is not a fixed global value, but is determined by the average of the adaptive weights of the two data points in that dimension.

[0048] Furthermore, the calculated hybrid-driven adaptive distance is anisotropic because it varies across different locations and directions in the performance space. If two points differ in a high-weight dimension, this difference will be amplified, making them appear farther apart; conversely, differences in low-weight dimensions will be suppressed.

[0049] Thus, intelligent metrics that reflect the intrinsic structure and engineering importance of data provide crucial support for subsequent clustering algorithms to accurately identify meaningful data clusters.

[0050] S104 employs a hybrid drive adaptive distance execution density peak clustering algorithm and determines the cluster center with the highest comprehensive expected score based on the clustering results. The control parameter combination corresponding to the cluster center is then used as the ideal performance-emission balance point of the engine under specific operating conditions.

[0051] In one embodiment, a hybrid-driven adaptive distance can be used to execute the density peak clustering algorithm. The density peak clustering algorithm identifies cluster centers by calculating the local density of each data point and the minimum distance to higher density points. In this solution, all distance calculations in the density peak clustering algorithm are replaced by hybrid-driven adaptive distance.

[0052] Specifically, since the hybrid-driven adaptive distance already contains global discrimination and local consistency information, the density peak clustering algorithm can accurately identify those dense regions with small local standard deviations in key performance dimensions, resulting in high consistency and high expected values, thus exhibiting excellent overall performance. The center point of these regions is the candidate optimal balance point.

[0053] After identifying all cluster centers, the globally optimal one needs to be selected from these candidates. Therefore, the comprehensive expected score of each center point can be calculated. This represents the sum of the overall superiority or expected value of cluster centers across all performance and emission dimensions:

[0054]

[0055] in, Let be the expected value of the cluster centers in the 0th dimension. Since It is the sum of the expected values ​​of all dimensions of the corresponding center point in the expected space. Since all dimensions have been unified to be as large as possible, the higher the sum, the better the overall performance of the point.

[0056] Furthermore, the cluster center with the highest comprehensive expected score can be selected, and the original engine control parameter combination corresponding to that data point can be backtracked. This set of parameters is determined as the globally optimal ideal performance-emission balance point under the current operating conditions. Test engineers can use this result to accurately calibrate the engine.

[0057] Thus, by combining adaptive distance with density peak clustering algorithm and setting a clear optimization objective function, the optimal operating point of the engine can be automatically, accurately and robustly located from massive test data, achieving a high degree of intelligence in the testing process.

[0058] It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the scope of protection of this application. Therefore, the scope of protection of this patent application shall be determined by the appended claims.

Claims

1. A method for testing the load of a gasoline engine, characterized in that, include: Obtain the desired dataset containing multiple performance and emission indicators generated by the combination of multiple sets of control parameters of the engine under specific operating conditions; For each data point in the desired dataset, the global discriminative power index and the local consistency index of each indicator dimension are fused to construct a hybrid-driven adaptive weight vector. ;in, Let be the adaptive weight of the i-th data point in dimension o. and These are the global discriminative power indices for dimensions o and m, respectively. Let be the local consistency index of the i-th data point on dimension o, and M be the total number of dimensions; The process of obtaining the global discriminative power index is as follows: a Gaussian mixture model is used to fit the data distribution of each indicator dimension in the desired dataset, and the optimal number of components for that dimension is determined according to the Bayesian information criterion; the global discriminative power index is calculated based on the optimal number of components. Based on the adaptive weight vector, calculate the hybrid driving adaptive distance. ;in, For the hybrid driving adaptive distance of data point a and data point b, and These are the adaptive weights of points a and b in the o-th dimension, respectively. and These are the expected values ​​of points a and b in the o-th dimension, respectively, used to measure the difference between any two data points in the expected dataset; The hybrid drive adaptive distance execution density peak clustering algorithm is adopted, and the cluster center with the highest comprehensive expected score is determined according to the clustering results. The control parameter combination corresponding to the cluster center is used as the ideal performance emission balance point of the engine under specific operating conditions.

2. The gasoline engine load testing method according to claim 1, characterized in that, The process of obtaining the desired dataset is as follows: The original data vector of the engine under multiple sets of control parameters is collected, and the maximum-minimum normalization process is performed on each index dimension. Based on the optimization direction of the indicators, the normalized values ​​are transformed into expected values. The normalized values ​​of positive indicators are used as their expected values, and the expected values ​​of negative indicators are defined as the difference between 1 and the normalized values. This constructs an expected dataset where all components are considered to have better performance as the value is closer to 1.

3. The gasoline engine load testing method according to claim 1, characterized in that, The global discriminant power index satisfies the following relationship: in, The global discriminative power index is the one with dimension o. and are the optimal Gaussian component numbers of the data distribution in the 0th and mth dimensions, respectively, where M is the total number of dimensions.

4. The gasoline engine load testing method according to claim 1, characterized in that, The process of obtaining the local consistency index is as follows: For each data point in the desired dataset, determine the local neighborhood formed by its multiple nearest neighbors, and calculate the local standard deviation of each indicator dimension within that neighborhood. The local consistency index is calculated based on the local standard deviation.

5. The gasoline engine load testing method according to claim 4, characterized in that, The local consistency index satisfies the following relationship: in, Let be the local consistency index of the i-th data point in dimension o. Let be the local standard deviation of dimension o within the neighborhood of the i-th data point. Let be the global standard deviation of the 0th dimension over the entire expected dataset. To prevent extremely small positive numbers with a denominator of zero.

6. The gasoline engine load testing method according to claim 1, characterized in that, The control parameter combination includes air-fuel ratio and ignition advance angle.

Images