A multi-source information fusion method for solving model prediction conflicts

Abstract

The application discloses a multi-source information fusion method for solving model prediction conflicts, characterized in that a plurality of single-source prediction models are constructed according to multi-source historical data in an industrial scene, basic probability distribution of the single-source prediction models is adaptively corrected in combination with operating conditions, and on this basis, evidence fusion rules are adopted to complete fusion of prediction results.

Description

Technical Field

[0001] This invention relates to the field of industrial data modeling and information fusion technology, specifically a multi-source information fusion method for resolving model prediction conflicts. Background Technology

[0002] In modern industrial production, key quality indicators and operating parameters have a significant impact on the safety, stability, and economic benefits of equipment. To achieve real-time monitoring of these indicators, various sensors and detection systems are typically deployed in industrial settings to collect process data, analytical data, and indirect measurement data from different information sources. In recent years, multi-source information fusion prediction methods have attracted much attention due to their ability to improve prediction accuracy and robustness. By integrating prediction results from different information sources and fully utilizing the complementarity of multi-source information, these methods can alleviate, to some extent, the problems of instability and insufficient accuracy in predictions using single models.

[0003] However, existing multi-source information fusion prediction methods still have shortcomings in practical industrial applications. On the one hand, prediction models constructed from different information sources exhibit significant differences in prediction performance under different operating conditions, and model reliability shows obvious operating condition dependence. Some fusion methods still use fixed weights or overall discount factors to process model prediction results, making it difficult to accurately characterize the dynamic characteristics of model reliability as operating conditions change. On the other hand, when dealing with multi-model prediction conflicts, some methods typically introduce global conflict discounting or conflict weight redistribution mechanisms to uniformly weaken or redistribute conflict quality. Such methods often ignore the structural characteristics of prediction uncertainty at the interval and operating condition levels, easily leading to excessive reduction of effective information and affecting the interpretability and stability of the fusion results.

[0004] Based on the above, this invention proposes a multi-source information fusion method to resolve model prediction conflicts. This method first constructs multiple single-source prediction models based on multi-source historical data. Then, it adaptively corrects the basic probability allocation of the single-source prediction models based on operational conditions. Finally, it employs evidence fusion rules to fuse the prediction results of multiple models, obtaining the fused prediction result for the target variable. Summary of the Invention

[0005] This invention addresses the problems existing in the background technology by proposing a multi-source information fusion method to resolve model prediction conflicts, thereby achieving the fusion of prediction results from multiple single-source prediction models. Specifically, it includes the following steps:

[0006] Step 1: Obtain multi-source historical samples and construct a dataset , For the i-th sub-information source model, Given the corresponding true values ​​of the samples, train n single-source information prediction models. .

[0007] Step 2: Based on historical sample data, classify the working conditions and use a clustering algorithm to form l working condition categories. .

[0008] Step 3: Divide the possible values ​​of the target variable into intervals and construct an identification framework.

[0009]

[0010] Where p is the number of intervals divided. Indicates different intervals.

[0011] Step 4: Based on the prediction error information of historical samples under each working condition category, calculate the reliability of the i-th single-source prediction model when the predicted value of the l-th working condition category is located in the k-th interval. The standard deviation of the prediction error for different models was calculated according to the operating condition category. ;

[0012] 1) Internal reliability Indicates the working condition category Next, the The predicted value of the single-source prediction model is located at the th The conditional probability that the true value also falls within a given interval is calculated as follows:

[0013]

[0014] Where N represents the number of samples, and y represents the sample size. The true value For the first The predicted values ​​of a single-source prediction model. To identify the first in the frame Each interval For the first Each working condition category.

[0015] Step 5: Use single-source prediction models respectively Treatment of Predicted Samples Output the prediction results of the target variable. Calculate the prediction uncertainty And construct the initial basic probability assignment :

[0016] 1) Prediction uncertainty The weighted average method is used to obtain the result. The specific calculation method is as follows:

[0017]

[0018] ,

[0019] in, For the first A single-source prediction model for samples The prediction uncertainty For the model In operating condition category The standard deviation of the prediction error For the first The weight of the contribution of the standard deviation of the prediction error to the calculation of the uncertainty value under each working condition category. Indicates the first Cluster centers for each work condition category and Representing samples respectively To the The Euclidean distance between the cluster centers of each work condition category, where L is the total number of work condition categories. This is the distance attenuation coefficient.

[0020] 2) Initial basic probability assignment The calculation method is as follows:

[0021]

[0022] in, This represents the k-th interval within the identification framework. The midpoint, For the sample predicted value, To predict uncertainty.

[0023] Step 6, based on internal reliability Initial basic probability assignment After making corrections, we obtain the corrected basic probability distribution. The calculation method is as follows:

[0024]

[0025] in, The fuzzy priority assignment coefficients are used to control the diffusion of uncertain mass to the union and uncertain set of adjacent intervals. The distribution ratio between them.

[0026] Step 7: Merge the basic probability assignments of each single-source prediction model processed in Step 6, and use the Dempster combination rule to apply the basic probability assignments. By performing fusion, the basic probability distribution of fusion is obtained. The formula is as follows:

[0027] ,

[0028] in, and This represents the focal element of the corrected basic probability assignment. To predict the conflict coefficient.

[0029] Step 8, Assigning basic fusion probabilities The solution is performed to obtain the fusion prediction result of the target variable. The calculation method is as follows:

[0030]

[0031] in, To identify the kth interval of the frame The midpoint value, It is the mean of the midpoint values ​​of all intervals.

[0032] Beneficial effects:

[0033] This invention discloses a multi-source information fusion method for resolving model prediction conflicts. The method first constructs multiple single-source prediction models based on multi-source historical data. Then, it adaptively corrects the basic probability allocation of the single-source prediction models based on operational conditions. Finally, it employs evidence fusion rules to fuse the multi-source prediction results, obtaining the fused prediction result for the target variable. This method requires no modification to the existing single-source prediction model structure, possesses good versatility and scalability, and can achieve high-precision and robust prediction of key quality indicators in various industrial scenarios, providing a reliable and effective technical means for the refined monitoring and intelligent optimization of industrial processes. Attached Figure Description

[0034] Figure 1 This is a flowchart of a multi-source information fusion method for resolving model prediction conflicts according to the present invention.

[0035] Figure 2 This is a diagram showing the internal reliability distribution of sub-information source model 1.

[0036] Figure 3 This is a diagram showing the internal reliability distribution of sub-information source model 2.

[0037] Figure 4 The probability density distribution of prediction errors for different sub-models and the fusion model on the test set.

[0038] Figure 5 A comparison chart of the conflict coefficients of prediction results before and after reliability correction on the test set. Detailed Implementation

[0039] The embodiments of the present invention are described in detail below. These embodiments are implemented based on the technical solution of the present invention, and provide detailed implementation methods and specific operation processes. However, the scope of protection of the present invention is not limited to the following embodiments.

[0040] This case study uses the industrial scenario of crude oil fractionation as an example. The sub-information sources are historical operating data of the actual production unit of a refinery and near-infrared spectral data of crude gasoline from the same period. Based on the two data samples, the final boiling point of crude gasoline is predicted, verifying the effectiveness of the multi-source information fusion method proposed in this invention for resolving model prediction conflicts.

[0041] The overall system flowchart of the present invention is as follows: Figure 1 As shown, the specific steps include:

[0042] 1) For the prediction of the final boiling point of crude gasoline, prediction models were constructed based on historical operating data of the production unit. And a prediction model built based on near-infrared spectral data. The model construction method will not be described in detail;

[0043] 2) Operating conditions are categorized based on historical sample data. The K-Means clustering algorithm is used to form multiple operating condition categories. Since operating condition information is mainly contained in the historical operating samples of the production unit, operating conditions are categorized only based on the historical operating data of the production unit. The number of operating condition categories is set to 3, denoted as follows: ;

[0044] 3) The final boiling point of the crude gasoline samples varied between 200℃ and 220℃, which was divided into 6 intervals to form the identification framework. The intervals are defined as follows:

[0045]

[0046] The midpoints of each interval are as follows:

[0047]

[0048] 4) Based on the prediction error information of historical samples under each working condition category, calculate the internal reliability of the single-source prediction model under different categories. and the standard deviation of the prediction error ,

[0049] For Model 1:

[0050]

[0051] For Model 2:

[0052]

[0053] Internal reliability of sub-information source model 1 Distribution as follows Figure 2 As shown, the internal reliability of sub-information source model 2 Distribution as follows Figure 3 As shown;

[0054] 5) Use single-source prediction models respectively Treatment of Predicted Samples Output the prediction results of the target variable. Calculate the prediction uncertainty And construct the initial basic probability assignment :

[0055] Here, we take the calculation of a single sample as an example. The Euclidean distance between the cluster centers is used as the basis for identifying the working condition, and its corresponding working condition category is... Using a weighted prediction uncertainty calculation method based on operating conditions, we obtain:

[0056] Model-1 predicts a temperature of 205.5℃, with an uncertainty of:

[0057] The Model-2 prediction is 210℃, with an uncertainty of:

[0058] An initial BPA was constructed using a Gaussian kernel function approach. The calculation results are shown in Table 1. It can be seen that there is significant conflict between the prediction results of the two sub-models at this point. The initial conflict coefficient is calculated as follows: .

[0059] Table 1 Initial BPA Calculation Values

[0060]

[0061] 6) According to the working conditions The internal reliability is adjusted based on the initial BPA, and a portion of the mass is allocated to the union of adjacent intervals and the uncertain set. The calculation results are shown in Table 2.

[0062] Table 2 Corrected BPA Calculation Values

[0063]

[0064] 7) The modified BPA of the two models is fused using the Dempster combination rule. The conflict degree after fusion is: .

[0065] The resulting BPA is:

[0066]

[0067] 8) The final predicted value is calculated using the interval midpoint weighted expectation method to obtain the fused prediction result:

[0068]

[0069] 9) Following the steps above, fuse the prediction results for all samples in the test set. The probability density distribution of the prediction errors of different sub-models and the fused model on the test set is as follows: Figure 4 As shown, a comparison of the conflict coefficients of the prediction results before and after reliability correction is presented. Figure 5 As shown.

[0070] It is evident that, compared to the individual sub-models, the prediction error distribution of the fusion model is more concentrated, with a higher probability density peak and a significantly reduced distribution width, indicating that the fusion prediction results have better stability and robustness across the entire test set. Furthermore, after introducing internal reliability correction, the overall conflict coefficient among the sub-models is significantly reduced, and the fluctuation amplitude of the conflict is effectively suppressed. This demonstrates that the proposed adaptive reliability correction mechanism can reasonably weaken unreliable evidence based on the differences in model prediction performance under different operating conditions, thereby improving the credibility and engineering applicability of multi-source information fusion prediction.

Claims

1. A multi-source information fusion method for resolving model prediction conflicts, characterized in that, Multiple single-source prediction models are constructed based on multi-source historical data in industrial scenarios. The basic probability allocation of the single-source prediction models is adaptively corrected for reliability by combining operating conditions. On this basis, evidence fusion rules are used to fuse the prediction results. The specific steps include the following: 1) Obtain multi-source historical samples and construct a dataset , For the i-th sub-information source model, Given the corresponding true values ​​of the samples, train n single-source information prediction models. ; 2) Based on historical sample data, operating conditions are divided, and a clustering algorithm is used to form l operating condition categories. ; 3) Divide the possible values ​​of the target variable into intervals and construct an identification framework. p is the number of intervals divided. Indicate different intervals; 4) Based on the prediction error information of historical samples under each working condition category, calculate the reliability of the i-th single-source prediction model when the predicted value of the l-th working condition category is located in the k-th interval. The standard deviation of the prediction error for different models was calculated according to the operating condition category. ; 5) Use the i-th single-source prediction model respectively Treatment of Predicted Samples Output the prediction results of the target variable. Calculate the prediction uncertainty And construct the initial basic probability assignment ; 6) Based on reliability Initial basic probability assignment After making corrections, we obtain the corrected basic probability distribution. ; 7) The basic probability assignments of each single-source prediction model processed in step 6) are fused to obtain the fused basic probability assignment. ; 8) Assignment of basic probability to fusion The solution is performed to obtain the fusion prediction results of the target variable.

2. The multi-source information fusion method for resolving model prediction conflicts according to claim 1, characterized in that, Internal reliability Indicates the working condition category Next, the The predicted value of the single-source prediction model is located at the th The conditional probability that the true value also falls within a given interval is calculated as follows: Where N represents the number of samples, and y represents the sample size. The true value For the first The predicted values ​​of a single-source prediction model. To identify the first in the frame Each interval For the first Each working condition category.

3. The multi-source information fusion method for resolving model prediction conflicts according to claim 1, characterized in that, Prediction uncertainty The weighted average method is used to obtain the result. The specific calculation method is as follows: ， in, For the first A single-source prediction model for samples The prediction uncertainty For the model In operating condition category The standard deviation of the prediction error For the first The weight of the contribution of the standard deviation of the prediction error to the calculation of the uncertainty value under each working condition category. Indicates the first Cluster centers for each work condition category and Representing samples respectively To the The Euclidean distance between the cluster centers of each work condition category, where L is the total number of work condition categories. This is the distance attenuation coefficient.

4. The multi-source information fusion method for resolving model prediction conflicts according to claim 1, characterized in that, Initial basic probability assignment The calculation method is as follows: in, This represents the k-th interval within the identification framework. The midpoint, For the sample predicted value, To predict uncertainty.

5. The multi-source information fusion method for resolving model prediction conflicts according to claim 1, characterized in that, Based on internal reliability, the first Initial basic probability assignment for a single-source prediction model After correction The calculation method is as follows: in, The fuzzy priority assignment coefficients are used to control the diffusion of uncertain mass to the union and uncertain set of adjacent intervals. The distribution ratio between them.

6. The multi-source information fusion method for resolving model prediction conflicts according to claim 1, characterized in that, Assigning basic probabilities using Dempster's combination rule The fusion process is as follows: ， in, and This represents the focal element of the corrected basic probability assignment. To predict the conflict coefficient.

7. The multi-source information fusion method for resolving model prediction conflicts according to claim 1, characterized in that, Basic probability assignment for fusion The method for obtaining the fusion prediction results through calculation is as follows: in, To identify the kth interval within the frame The midpoint value, It is the mean of the midpoint values ​​of all intervals.

Images