Multi-dimensional Weibull parameter prediction method and system based on neural network

Abstract

The invention provides a multi-dimensional Weibull parameter prediction method and system based on a neural network, and relates to the field of engine performance evaluation, and the specific scheme comprises the steps: obtaining multi-dimensional physical property working condition parameters of a to-be-evaluated engine; the multi-dimensional physical property working condition parameters are input into the trained multi-dimensional data prediction model, and multi-dimensional Weber parameters are obtained and used for evaluating the combustion performance of the engine; wherein the multi-dimensional data prediction model takes a Bayesian neural network as a bottom layer network, quantifies uncertainty by using a coverage rate and an average interval width, introduces the quantified uncertainty into a loss function, and restrains a model training process; according to the method, the structure and the loss function of the Bayesian neural network are improved, and the problems of precision imbalance distribution, prediction result uncertainty and real-time performance are solved.

Description

Technical Field

[0001] The present invention belongs to the field of engine performance evaluation, is applied to HIL (Hardware-in-the-Loop) virtual calibration scenarios, and particularly relates to a multi-dimensional Weber parameter prediction method and system based on a neural network. Background Art

[0002] As pollutant emission regulations become increasingly stringent, research on advanced power systems has received widespread attention; the combustion process of fuel in the engine cylinder directly affects the engine's power output, fuel economy and emission levels. The combustion process, a key link, is of great significance for studying engine characteristics; advanced power system engines are difficult to simulate and calculate using traditional modeling methods due to the complexity and variability of their combustion processes; combustion models are divided into zero-dimensional, quasi-dimensional and three-dimensional models according to spatial dimensions, among which the zero-dimensional Weber combustion model is the most widely used combustion heat release rate model. The experiment uses a three-Weber zero-dimensional combustion model to evaluate the heat release rate.

[0003] In the application of advanced power system engineering, it is often hoped to obtain more correspondences between engine operating parameters and Weber parameters. The three-Weber zero-dimensional combustion model is used to evaluate the heat release rate, taking the engine operating physical property parameters as input and the Weber parameters as output. Constructing a neural network prediction model is one of the effective ways to establish this correspondence.

[0004] The existing neural network prediction models have the following problems:

[0005] (1) For small sample data sets, it is difficult for neural networks to obtain good prediction performance. In addition, since the relationship between engine operating parameters and Weber parameters is multi-dimensional, multi-dimensional input and output data will also cause problems such as unbalanced distribution of accuracy when predicting neural networks.

[0006] (2) As the dimension of training data increases, the uncertainty of prediction will also increase. However, current prediction models rarely conduct uncertainty analysis on many-to-many neural networks. The lack of quantitative analysis of uncertainty has brought certain limitations to a more reasonable evaluation of model performance.

[0007] (3) The real-time simulation platform provides computing power support for model calculations, but its computing power is always limited. In order to meet the real-time requirements of the prediction model in the real-time simulation platform, it is necessary to shorten the actual running time of the online learning prediction model as much as possible; in the process of building a neural network, as the complexity of the network structure increases and the number of calculation formulas increases, the network training speed and computing efficiency also decrease; long-term training and prediction processes bring limitations to practical applications.

[0008] Therefore, when dealing with multi-input and multi-output problems, existing models have problems such as unbalanced accuracy distribution, uncertainty in prediction results, and lack of real-time performance. Summary of the invention

[0009] In order to overcome the deficiencies of the above-mentioned prior art, the present invention provides a multi-dimensional Weber parameter prediction method and system based on neural network, which improves the accuracy imbalance distribution, prediction result uncertainty and real-time problems by improving the structure and loss function of the Bayesian neural network.

[0010] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions:

[0011] The first aspect of the present invention provides a multi-dimensional Weber parameter prediction method based on a neural network.

[0012] A multi-dimensional Weibull parameter prediction method based on a neural network, comprising:

[0013] Obtaining multi-dimensional physical property parameters of the engine to be evaluated;

[0014] Input the multidimensional physical property parameters into the trained multidimensional data prediction model to obtain multidimensional Weber parameters for evaluating the combustion performance of the engine;

[0015] Among them, the multidimensional data prediction model uses the Bayesian neural network as the underlying network, uses coverage and average interval width to quantify uncertainty, introduces it into the loss function, and constrains the model training process.

[0016] Furthermore, the multi-dimensional physical property operating condition parameters include engine speed, excess air coefficient, intake pressure, intake temperature, fuel injection amount, natural gas injection amount, fuel injection time, and natural gas injection time;

[0017] The multidimensional Weber parameters include gas phase fuel heat release rate factor Fp, solid phase fuel heat release rate factor Fm, gas phase fuel mass fraction mp, solid phase fuel mass fraction mm, liquid phase fuel mass fraction mt, gas phase fuel diffusion coefficient Dp, solid phase fuel diffusion coefficient Dm, liquid phase fuel diffusion coefficient Dt, and combustion starting point SOC.

[0018] Furthermore, the multidimensional data prediction model is based on the Bayesian neural network and adds a BatchNorm layer and a DropOut layer.

[0019] Furthermore, the loss function includes mean square error, KL divergence value, coverage rate and average interval width.

[0020] Furthermore, the coverage rate and average interval width are specifically expressed as follows:

[0021]

[0022] Among them, MedianIntervalWidth_val represents coverage, Coverage_val represents the average interval width, and y i represents the true value, Represents a set of predicted values, N represents The number of groups, express The average value of represents the standard deviation, and l represents the target value ratio.

[0023] Furthermore, it also includes: introducing an adaptive weight mechanism, based on the prediction accuracy of each Weibull parameter, adaptively allocating weights between the multidimensional Weibull parameters output by the multidimensional data prediction model, and adjusting the imbalanced distribution of accuracy.

[0024] Furthermore, the training data set used for training the multidimensional data prediction model is finally constructed on the basis of small sample experimental data through data expansion and data cleaning.

[0025] The second aspect of the present invention provides a multi-dimensional Weber parameter prediction system based on a neural network.

[0026] A multi-dimensional Weber parameter prediction system based on neural network, including an acquisition module and a prediction module:

[0027] An acquisition module is configured to: acquire multi-dimensional physical property operating condition parameters of the engine to be evaluated;

[0028] The prediction module is configured to: input the multi-dimensional physical property operating condition parameters into the trained multi-dimensional data prediction model to obtain multi-dimensional Weber parameters for evaluating the combustion performance of the engine;

[0029] Among them, the multidimensional data prediction model uses the Bayesian neural network as the underlying network, uses coverage and average interval width to quantify uncertainty, introduces it into the loss function, and constrains the model training process.

[0030] The third aspect of the present invention provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps in a multi-dimensional Weber parameter prediction method based on a neural network as described in the first aspect of the present invention.

[0031] The fourth aspect of the present invention provides an electronic device, including a memory, a processor, and a program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps in a multi-dimensional Weber parameter prediction method based on a neural network as described in the first aspect of the present invention are implemented.

[0032] One or more of the above technical solutions have the following beneficial effects:

[0033] (1) The present invention is based on a Bayesian neural network. By constructing an uncertainty quantification index and integrating the uncertainty quantification index into the loss function, the reasonable output and prediction of uncertainty are achieved, thereby improving the prediction accuracy of the multi-input and multi-output model and solving the problem of uncertainty in the prediction results.

[0034] (2) By introducing an adaptive weight mechanism, the present invention significantly improves the accuracy performance imbalance of multi-dimensional output data.

[0035] (3) Integrate vectorized programming ideas into the entire code writing process and combine them with a parallel architecture for training to achieve accelerated prediction model training and efficient execution, with efficiency improved by more than 30%; maintain good online learning and real-time prediction effects in real-time machines, thereby meeting the needs of actual engineering applications.

[0036] Advantages of additional aspects of the present invention will be given in part in the following description, and in part will become obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0038] Figure 1 This is a diagram of the architecture of the first embodiment.

[0039] Figure 2 This is a model structure diagram of the first embodiment.

[0040] FIG. 3 is a result analysis diagram of the first embodiment. DETAILED DESCRIPTION

[0041] It should be noted that the following detailed descriptions are illustrative and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used in the present invention have the same meanings as those commonly understood by those skilled in the art to which the present application belongs.

[0042] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0043] Embodiment 1

[0044] In one embodiment of the present disclosure, a multi-dimensional Weber parameter prediction method based on a neural network is provided, comprising the following steps:

[0045] Step S1: obtaining multi-dimensional physical property parameters of the engine to be evaluated;

[0046] Step S2: inputting the multidimensional physical property operating condition parameters into the trained multidimensional data prediction model to obtain multidimensional Weber parameters for evaluating the combustion performance of the engine;

[0047] Among them, the multidimensional data prediction model uses the Bayesian neural network as the underlying network, uses coverage and average interval width to quantify uncertainty, introduces it into the loss function, and constrains the model training process.

[0048] Furthermore, the multi-dimensional physical property operating condition parameters include engine speed, excess air coefficient, intake pressure, intake temperature, fuel injection amount, natural gas injection amount, fuel injection time, and natural gas injection time;

[0049] The multidimensional Weber parameters include gas phase fuel heat release rate factor Fp, solid phase fuel heat release rate factor Fm, gas phase fuel mass fraction mp, solid phase fuel mass fraction mm, liquid phase fuel mass fraction mt, gas phase fuel diffusion coefficient Dp, solid phase fuel diffusion coefficient Dm, liquid phase fuel diffusion coefficient Dt, and combustion starting point SOC.

[0050] Furthermore, the multidimensional data prediction model is based on the Bayesian neural network and adds a BatchNorm layer and a DropOut layer.

[0051] Furthermore, the loss function includes mean square error, KL divergence value, coverage rate and average interval width.

[0052] Furthermore, the coverage rate and average interval width are specifically expressed as follows:

[0053]

[0054] Among them, MedianIntervalWidth_val represents coverage, Coverage_val represents the average interval width, and y i represents the true value, Represents a set of predicted values, N represents The number of groups, express The average value of represents the standard deviation, and l represents the target value ratio.

[0055] Furthermore, it also includes: introducing an adaptive weight mechanism, based on the prediction accuracy of each Weibull parameter, adaptively allocating weights between the multidimensional Weibull parameters output by the multidimensional data prediction model, and adjusting the imbalanced distribution of accuracy.

[0056] Furthermore, the training data set used for training the multidimensional data prediction model is finally constructed on the basis of small sample experimental data through data expansion and data cleaning.

[0057] The following is a detailed description of the implementation process of a multi-dimensional Weber parameter prediction method based on a neural network in this embodiment.

[0058] In order to solve the problems of unbalanced distribution of accuracy, uncertainty of prediction results and real-time performance of existing models when dealing with multi-input and multi-output problems, an accelerated Bayesian neural network prediction model architecture based on small samples was designed to predict the multi-dimensional Weber parameters of the engine with high accuracy and real-time performance. The accuracy performance of small sample data regression prediction was significantly improved, the performance balance effect was significant, the uncertainty deep fusion training process and the rationality analysis of uncertainty quantification indicators were analyzed, the computing efficiency was improved, and the running time was shortened, thereby meeting the requirements of online learning in real-time machines.

[0059] The structure of the accelerated Bayesian neural network prediction model architecture is as follows Figure 1 As shown, it includes three parts: data construction and preprocessing, construction and training of multidimensional data prediction model, model prediction and result analysis, which are explained separately below.

[0060] 1. Data Construction and Preprocessing

[0061] As described above, the input and output of the multidimensional data prediction model are multidimensional physical property parameters and multidimensional Weber parameters, respectively. Therefore, the construction of the data set for model training is crucial to the training effect of the model. In this embodiment, the training data set is finally constructed on the basis of small sample experimental data through data expansion and data cleaning.

[0062] Specifically, firstly, small sample experimental data are obtained through experiments, that is, the multidimensional physical property operating parameters of the engine that have an important impact on the prediction of combustion performance are used as input data, and the multidimensional Weber parameters are used as output data to form a small sample experimental data set.

[0063] Then, the GAN data enhancement method is used to enrich the small sample experimental data set; data cleaning is achieved through the autoencoder to improve the data quality of the expanded data set and provide a guarantee for high-quality prediction performance.

[0064] Finally, the training set and test set are divided into 8:2 ratio.

[0065] In this embodiment, based on the 14 discretized data of the small sample obtained in the experiment, the data volume reaches 1000 after data expansion, and the most valuable 950 data are obtained after data cleaning, which provides sufficient data reserve for the subsequent training of the prediction model; the Bayesian neural network with multi-dimensional input and multi-dimensional output trained using the expanded data has an accuracy performance improvement of more than 20%.

[0066] 2. Building and training multidimensional data prediction models

[0067] The underlying layer of the multidimensional data prediction model is the Bayesian neural network. Compared with the back propagation network, whose weight is a fixed value after optimization, the Bayesian neural network regards the weight as obeying a Gaussian distribution with a mean of μ and a variance of δ. Each weight obeys a different Gaussian distribution. The back propagation network optimizes the weight, and the Bayesian neural network optimizes the mean and variance of the weight. Therefore, the parameters that the Bayesian neural network needs to optimize are twice those of the back propagation network.

[0068] When predicting, the Bayesian neural network will sample from each Gaussian distribution to obtain weight values. At this time, the Bayesian neural network is equivalent to a back propagation network; multiple sampling can also be performed to obtain multiple prediction results, and the multiple prediction results can be averaged to obtain the final prediction result.

[0069] The multidimensional data prediction model in this embodiment is based on the Bayesian neural network, with the addition of a BatchNorm layer and a DropOut layer. The specific model structure is as follows: Figure 2 As shown, it consists of a Bayesian neural network layer, a ReLU activation function, a BatchNorm layer, and a DropOut layer.

[0070] Among them, the BatchNorm layer reduces the problem of internal covariate shift by standardizing the data of each small batch; adjusts the data distribution of each layer of the network input to make the network training process more efficient; reduces the difference between different batches of data, making the model training more stable and less prone to gradient explosion or gradient disappearance. The introduction of BatchNorm has improved the accuracy and computing speed of the neural network prediction model; the DropOut layer randomly deletes some neurons in each training so that the output of these neurons will not be updated in this training, thereby enhancing the robustness of the prediction model, improving the generalization ability, and avoiding overfitting of the model.

[0071] In order to solve the problem of unbalanced prediction accuracy of different dimensions of multi-dimensional data under the same number of training iterations, an adaptive weight mechanism was developed, specifically:

[0072] First, initialize the sample dimension weights, usually with uniformly distributed initial weights, that is, each sample weight is initialized to Where N is the total number of samples.

[0073] Then, with t as the current iteration number, repeat the following iterative calculation:

[0074] (1) According to the model trained with the current weight, we get the sample x i The predicted value of

[0075] (2) Calculate the weighted error rate ∈ of the current model after adding weight parameters t , the formula is as follows:

[0076]

[0077] Among them, y i (x i ) is the sample x i The true value of .

[0078] (3) Calculate the sample dimension error rate impact metric

[0079] (4) Update the sample dimension weights. Dynamically adjust the sample dimension weights based on the model’s performance in predicting data of different dimensions, so that in the next round of iterations, the weights of samples with larger errors are strengthened, while the weights of samples with smaller errors are weakened. The specific weight update formula is:

[0080]

[0081] Among them, Z t is a normalization factor that ensures the weights sum to 1.

[0082] Finally, the training process stops until the preset number of iterations T is reached or a certain accuracy is achieved.

[0083] This mechanism is based on the Adaboost idea, which introduces the idea of ​​weight distribution to different output dimensions of the same model, and corrects the weights of the Bayesian neural network in different output dimensions, so that dimensions with relatively low prediction accuracy can be reasonably allocated larger weights according to corresponding rules, thereby achieving high quality and balanced distribution of accuracy performance.

[0084] In order to improve the uncertainty prediction ability of the model, two uncertainty quantification indicators, coverage and average interval width, are used to evaluate the uncertainty in the prediction results. By introducing these two uncertainty indicators into the training loss function, uncertainty is deeply integrated into the training process, so that the model can obtain better uncertainty prediction performance. The training loss function is shown in the following formula:

[0085] loss=MSE_val+KL_val+Median Interval Width_val+Coverage_val

[0086]

[0087] Among them, MSE_val represents mean square error, KL_val represents KL divergence value, Median Interval Width_val represents coverage, Coverage_val represents average interval width, y i represents the true value, Represents a set of predicted values, N represents The number of groups, express The average value of represents the standard deviation, and l represents the target value ratio.

[0088] Based on the training data set, the ensemble learning training method is applied to tune the hyperparameters of the neural network prediction model to obtain a set of hyperparameters that can achieve the best performance.

[0089] Specifically, vectorized computing is introduced into neural networks to improve computing efficiency. Through model parallelism, the online training process is accelerated to the greatest extent, thereby effectively addressing the problems of long computing time and low efficiency in the process of approximate solution of the posterior distribution of Bayesian neural networks.

[0090] The model parallel acceleration process mainly includes the following steps:

[0091] (1) Model parallel partitioning: The neural network is partitioned according to the structure and computing requirements so that different network layers of the model are allocated to different GPUs to maximize parallel efficiency.

[0092] (2) Forward propagation settings: The neural network layers of each part of the prediction model are located on different GPU devices. The order of data transmission between GPU devices is clearly set to ensure that the data is forward propagated sequentially on each GPU device.

[0093] (3) Back propagation and parameter update: This part is similar to the ordinary regression prediction model, ensuring that the gradients are correctly calculated and aggregated.

[0094] By using model parallelism and utilizing the memory and computing power of multiple GPUs, the computing load is effectively shared, accelerating the training process. At the same time, the forward propagation and back propagation process calculations of the model remain essentially the same, and the accuracy performance of the model parallel training results is highly consistent with that of the non-parallel model training.

[0095] 3. Model prediction and result analysis

[0096] After the multidimensional data prediction model is trained, it is deployed to the real-time machine for real-time prediction of multidimensional Weber parameters. The predicted multidimensional Weber parameters are input into the three-Weber zero-dimensional combustion model to evaluate the heat release rate that characterizes the engine performance.

[0097] The effect of the method of this embodiment is analyzed below:

[0098] The frequency histogram of the normalized output data (Y1 to Y9) of different dimensions of the small sample experimental data set used for model training is shown in Figure 3(a) below. After using GAN for data expansion, the data volume has been expanded, and the data distribution is also well consistent with the original data, as shown in Figure 3(b) below. In Figure 3(a) and Figure 3(b), Y1 to Y9 represent the output data in nine dimensions, the horizontal axis represents the value after normalization of the data, and the vertical axis represents the frequency of occurrence of the data under the corresponding horizontal axis. By introducing the uncertainty index into the training loss function, the training accuracy is improved. As shown in Figure 3(c), introducing two uncertainty indicators into the loss function of the model has the best effect on improving the accuracy. However, it can still be observed that the imbalance of prediction accuracy in different output dimensions is more obvious. In Figure 3(c), the horizontal axis Y_No represents different output dimensions, namely Y1 to Y9, and the vertical axis represents the accuracy performance index determination coefficient R 2 .

[0099] By developing an adaptive weight mechanism, the weights of different output dimensions of the model are adaptively allocated, so that the output dimensions with lower accuracy are given larger weights, achieving a balanced improvement in accuracy performance; the process of weight adaptive changes with iterations is shown in Figure 3(d), where the horizontal axis represents the current number of training iterations and the vertical axis represents the value of the weight. It can be found that this mechanism captures the obvious disadvantages of output dimensions 1 and 5 in accuracy very well, and thus continuously adds higher weight values ​​to these two dimensions.

[0100] The adaptive weight mechanism has a balanced prediction accuracy effect, as shown in Figure 3(e) below. The horizontal axis Y_No represents different output dimensions, i.e., Y1 to Y9, and the vertical axis represents the accuracy performance index determination coefficient R 2 It can be found that the accuracy improvement is most obvious in output dimension 1 and output dimension 5, which is consistent with the discussion of adaptively allocating weights to improve the lower accuracy performance dimensions and achieve improved accuracy balance distribution.

[0101] Embodiment 2

[0102] In one embodiment of the present disclosure, a multi-dimensional Weber parameter prediction system based on a neural network is provided, including an acquisition module and a prediction module:

[0103] An acquisition module is configured to: acquire multi-dimensional physical property operating condition parameters of the engine to be evaluated;

[0104] The prediction module is configured to: input the multi-dimensional physical property operating condition parameters into the trained multi-dimensional data prediction model to obtain multi-dimensional Weber parameters for evaluating the combustion performance of the engine;

[0105] Among them, the multidimensional data prediction model uses the Bayesian neural network as the underlying network, uses coverage and average interval width to quantify uncertainty, introduces it into the loss function, and constrains the model training process.

[0106] Embodiment 3

[0107] The purpose of this embodiment is to provide a computer-readable storage medium.

[0108] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in a multi-dimensional Weber parameter prediction method based on a neural network as described in the first embodiment of the present disclosure.

[0109] Embodiment 4

[0110] The purpose of this embodiment is to provide an electronic device.

[0111] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, the steps in a multi-dimensional Weber parameter prediction method based on a neural network as described in Embodiment 1 of the present disclosure are implemented.

[0112] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A multidimensional Weber parameter prediction method based on neural network, characterized in that: include: Obtaining multi-dimensional physical property parameters of the engine to be evaluated; Input the multidimensional physical property parameters into the trained multidimensional data prediction model to obtain multidimensional Weber parameters for evaluating the combustion performance of the engine; Among them, the multidimensional data prediction model uses the Bayesian neural network as the underlying network, uses coverage and average interval width to quantify uncertainty, introduces it into the loss function, and constrains the model training process.

2. A multidimensional Weber parameter prediction method based on a neural network as claimed in claim 1, characterized in that: The multi-dimensional physical property operating condition parameters include engine speed, excess air coefficient, intake pressure, intake temperature, fuel injection amount, natural gas injection amount, fuel injection time, and natural gas injection time; The multidimensional Weber parameters include gas phase fuel heat release rate factor Fp, solid phase fuel heat release rate factor Fm, gas phase fuel mass fraction mp, solid phase fuel mass fraction mm, liquid phase fuel mass fraction mt, gas phase fuel diffusion coefficient Dp, solid phase fuel diffusion coefficient Dm, liquid phase fuel diffusion coefficient Dt, and combustion starting point SOC.

3. A multi-dimensional Weber parameter prediction method based on a neural network as claimed in claim 1, characterized in that: The multidimensional data prediction model is based on the Bayesian neural network and adds a BatchNorm layer and a DropOut layer.

4. A multi-dimensional Weber parameter prediction method based on a neural network as claimed in claim 1, characterized in that: The loss function includes mean square error, KL divergence value, coverage rate and average interval width.

5. A multi-dimensional Weber parameter prediction method based on a neural network as claimed in claim 4, characterized in that: The specific formula for the coverage rate and average interval width is: Among them, Median Interval Width_val represents coverage, Coverage_val represents average interval width, and y i Represents real data, Represents a set of predicted values, N represents The number of groups, express The average value of represents the standard deviation, and l represents the target value ratio.

6. A multi-dimensional Weber parameter prediction method based on a neural network as claimed in claim 1, characterized in that: It also includes: introducing an adaptive weight mechanism, based on the prediction accuracy of each Weibull parameter, adaptively allocating weights between the multidimensional Weibull parameters output by the multidimensional data prediction model, and adjusting the imbalanced distribution of accuracy.

7. A multi-dimensional Weber parameter prediction method based on a neural network as claimed in claim 1, characterized in that: The training data set used for training the multidimensional data prediction model is finally constructed through data expansion and data cleaning based on small sample experimental data.

8. A multidimensional Weber parameter prediction system based on neural network, characterized in that: Including acquisition module and prediction module: An acquisition module is configured to: acquire multi-dimensional physical property operating condition parameters of the engine to be evaluated; The prediction module is configured to: input the multi-dimensional physical property operating condition parameters into the trained multi-dimensional data prediction model to obtain multi-dimensional Weber parameters for evaluating the combustion performance of the engine; Among them, the multidimensional data prediction model uses the Bayesian neural network as the underlying network, uses coverage and average interval width to quantify uncertainty, introduces it into the loss function, and constrains the model training process.

9. An electronic device, comprising: a memory for non-transitory storage of computer readable instructions; a processor for executing the computer-readable instructions; Wherein, when the computer-readable instructions are executed by the processor, the method according to any one of claims 1 to 7 is executed.

10. A storage medium, characterized in that: The computer-readable instructions are non-transitorily stored, wherein when the computer-readable instructions are executed by a computer, the method of any one of claims 1 to 7 is performed.

Images