A high-performance adaptive complex industrial parameter prediction method based on regression application score

Abstract

The application discloses a kind of high-efficiency adaptive complex industrial parameter prediction methods based on regression application score, belong to industrial parameter prediction field.The application proposes the feature screening algorithm based on regression application score, the value of each feature is comprehensively evaluated by integrating three key indicators: hit rate within 3% error, determination coefficient R 2 And time-consuming score, and then the regression application score is obtained, the actual contribution of each feature to the model prediction performance can be accurately evaluated, and fast and effective feature selection is realized;Meanwhile, the way of iterative updating model is proposed, the model is trained using data for a period of time to predict data for a short period of time later, the model is retrained periodically using newly generated data to predict data for a period of time later, solving the problem of model timeliness, so that it can accurately predict for a long time.

Description

Technical Field

[0001] This invention relates to the field of industrial parameter prediction technology, and more specifically, to a high-performance adaptive method for predicting complex industrial parameters based on regression application scoring. Background Technology

[0002] In industrial settings, many critical process parameters require prediction. For example, in the steel industry, drum strength and ignition temperature are important factors affecting the quality of the final product. These key parameters play a crucial role in sinter quality, energy consumption, and output. However, many key parameters cannot be directly obtained through online measurement, thus requiring forecasting or soft sensing models to indirectly acquire this data. Establishing a mapping relationship between key parameters and process parameters can effectively solve this problem.

[0003] Existing modeling methods mainly include mechanistic models based on physical principles and data-driven intelligent models. However, due to the complex and ever-changing industrial production environment, the numerous process parameters involved, and the influence of factors such as equipment aging, and the fact that most models remain fixed after initial training and are difficult to adapt to these changes, the predictive accuracy of the models decreases significantly over time. Summary of the Invention

[0004] 1. The technical problem that the invention aims to solve

[0005] To address the problem that existing industrial parameter measurement models fail after long-term application, this invention aims to provide a high-efficiency adaptive method for predicting complex industrial parameters based on regression application scoring. This method can provide accurate predictions with long-term stability and alleviate computational burden.

[0006] 2. Technical Solution

[0007] To address the problem of traditional measurement models failing after long-term use in the industry, this invention innovatively proposes an iterative model update method. This involves training the model using data from a given period to predict data for a shorter subsequent timeframe, and then periodically retraining the model using newly generated data to predict data for that next period, instead of continuously using the same model after training. This method solves the problem of model timeliness, enabling long-term stable and accurate predictions. However, iteratively updating the training model requires frequent data collection and computation, placing a significant computational burden on the server and making it difficult to meet the requirements of real-time control of field equipment. To alleviate this computational burden, this invention further proposes a feature selection algorithm based on regression application scoring. This algorithm integrates three key indicators—namely, the hit rate within 3% error, the coefficient of determination R, and... 2The algorithm uses a combination of time consumption score and other metrics to comprehensively evaluate the value of each feature, thereby deriving a regression application score. By leveraging this score, the algorithm can accurately assess the actual contribution of each feature to the model's predictive performance, achieving fast and effective feature selection. This optimization strategy effectively overcomes the limitations of traditional feature selection algorithms. For example, features selected by the Pearson correlation algorithm and the Spearman correlation algorithm may be mathematically related, but they may not necessarily contribute positively during actual model training. The algorithm proposed in this invention not only eliminates useless or redundant features, reducing reliance on computational resources and significantly improving computational speed, but also significantly enhances the model's generalization ability and prediction accuracy, thereby improving the model's application effectiveness.

[0008] Specifically, this invention provides a high-performance adaptive method for predicting complex industrial parameters based on regression application scoring, the process of which is as follows:

[0009] Step 1: Data Acquisition and Preprocessing

[0010] (1) Data acquisition: Install the PLC acquisition device at the location of the equipment in the industry where data needs to be collected. For example, if it is necessary to collect the pressure of a pipeline valve, the PLC device is installed at the valve in the pipeline. Through the PLC acquisition device, it is ensured that the data can be collected quickly, accurately and stably.

[0011] (2) Data Preprocessing: After receiving the data, a data quality check is performed, especially to detect missing values. Once a missing value is found, it is filled with the value of the nearest neighbor data point to maintain the integrity and consistency of the dataset. Next, to ensure the reliability of the data, the Isolation Forest technique is used to perform outlier detection. An important parameter in building the Isolation Forest model is the number of decision trees in the forest. In this invention, the number of decision trees is set to 120 to ensure that the model has sufficient complexity and detection accuracy. The Isolation Forest model works by randomly splitting the data, and each split generates a path from the root node to the leaf node. Outliers are usually easier to isolate, so data points with shorter path lengths are assigned lower outlier scores, meaning they are more likely to be outliers. After scoring all data points for outliers, a threshold is set based on the scoring results to delete outliers. In industrial practice, a commonly used threshold is set to 0.01, meaning that approximately 1 outlier is deleted out of every 100 data points. This process helps to clean up the dataset and ensure the accuracy and reliability of subsequent model training. The Isolation Forest algorithm can initially remove outliers, but this is far from sufficient. This invention, through in-depth practical observation, reveals that in industrial environments, factors such as machine malfunctions can lead to data anomalies over a period of time. While the Isolation Forest algorithm effectively detects anomalous data when normal and abnormal conditions alternate, its performance is less than ideal when handling continuous anomalous data. Therefore, this invention further explores this issue based on in-depth analysis of the industrial process, discovering that certain key characteristics in the industrial process should not have values ​​of 0. Based on this standard, a deeper level of outlier cleanup was performed, deleting data samples that should not have a value of 0 but did. These measures aim to improve data quality and ensure the accuracy and reliability of the modeling.

[0012] Step 2: Feature Filtering

[0013] After data preprocessing, the next step is feature scoring. This invention introduces a regression application scoring method based on three key indicators: hit rate with an error within 3%, coefficient of determination R... 2 Value and time consumption score. Each metric is multiplied by its corresponding weight—hit rate multiplied by a, R. 2 The value is multiplied by b, and the time taken is multiplied by c; the sum of these weights is 1, i.e., a + b + c = 1. The specific values ​​of the weight coefficients a, b, and c need to be adjusted and determined based on the actual application scenario. In different application scenarios, the model's focus will be different. If prediction accuracy is the top priority, the hit rate coefficient a will be set higher; conversely, if the model's generalization ability or computational efficiency is more important, the hit rate coefficient a will be appropriately reduced, and R² will be increased. 2The coefficient 'b' represents the value of the regression application score, or the coefficient 'c' represents the time consumption score. Based on this principle, the values ​​of a, b, and c can be initially set to ensure they better reflect the needs and priorities of specific applications. Since the regression application score is an important indicator for evaluating changes in model performance, in different industrial scenarios, to further optimize these coefficients, it is necessary to use cross-validation techniques to test various combinations of coefficients, and finally select a set of coefficient values ​​that most sensitively reflects changes in model performance. For example, setting a to 0.5, b to 0.4, and c to 0.1, etc.

[0014] Because the scoring system introduced in this invention integrates the above three important factors, it effectively overcomes the limitation of traditional evaluation indicators in comprehensively and intuitively reflecting the performance of regression models. The calculation formula for the regression application score is shown in (1):

[0015] RAS = a * (Hit Rate) + b * R 2 +c*(Time Score) (1)

[0016] Where a, b, and c are three coefficients, Hit Rate is the hit rate within a 3% error, and R... 2 The coefficient of determination is calculated as shown in equation (2), and the time score is calculated as shown in equation (3).

[0017]

[0018] Where yi is the actual value of the industrial parameter. These are the actual average values ​​of industrial parameters. Here are the predicted values ​​for industrial parameters, where n is the total sample size and x is the value. i The time x is the time taken to train the model by deleting the i-th feature. max This is the time taken to train the model using all features.

[0019] The feature selection process in this invention is as follows: Figure 2 As shown, the feature selection algorithm based on regression application scores first trains the model using all features and records the regression application scores, which are then temporarily recorded as the baseline score. Next, the algorithm temporarily deletes one feature, retrains the model using the remaining features, and records a new regression application score. If, after temporarily excluding a feature, the model's regression application score is higher than the previous baseline score, it indicates that the feature is not essential for building the model and should be deleted; conversely, if the regression application score is lower than the previous baseline score, it indicates that the feature is an important feature required by the model, and deleting it would degrade model performance, so it should be retained. Through this process, the algorithm eliminates all features one by one, ultimately selecting the optimal feature set required for training the model.

[0020] Step 3: Iteratively update the trained model to predict the parameters for the next time step.

[0021] Finally, a random forest model is trained using the selected optimal feature set to predict the industrial parameters needed for the next time point. At set intervals, the algorithm collects recent datasets and retrains the model with this new data to optimize its ability to predict industrial parameters for the future.

[0022] The time interval for model updates and the size of the dataset used to train the model need to be set according to the specific industry. When the amount of training data is small, the model may underfit due to failing to learn enough rich data features, thus affecting its predictive performance. Conversely, if the amount of training data is too large, it will not only increase the computational burden but may also cause overfitting, reducing the model's ability to generalize to new data. Therefore, this invention first selects one day's worth of data as the initial training set for iterative updates and training of the model. After each training iteration, the generated model is used to predict the key quality indicators for the following day, and the regression application score and R-squared are recorded. 2 The hit rate and error are within a 3% range. Next, the training set size is gradually increased, using 2 days, 3 days...n days of data (n is not increased further once a suitable dataset size is found). That is, after each iteration, the model is updated, and the key quality indicators for the next day are predicted and the corresponding evaluation metrics are recorded. This aims to explore the impact of different training set sizes on model performance and find the optimal amount of training data. Given that the inventors performed multiple iterations on over 50,000 data points throughout the process, recording three evaluation metrics in each iteration, it is difficult to directly compare the performance of training sets of different lengths. Therefore, this invention calculates the model's average regression application score and average R-squared value for each training set. 2 The optimal amount of training data is determined by comprehensively evaluating the values ​​and the average hit rate with an error within 3%. This method helps overcome the analytical complexity caused by iterating over large amounts of data and provides a scientific basis for selecting the best training set size.

[0023] By regularly updating the model with the latest data, the effectiveness and predictive accuracy of the model are ensured, avoiding performance degradation due to data aging. This continuous learning and self-adjustment mechanism guarantees that the model can maintain high predictive accuracy over the long term and adapt to the ever-changing industrial production environment.

[0024] 3. Beneficial effects

[0025] The present invention has the following beneficial effects:

[0026] (1) When performing data preprocessing, the present invention fills in missing values ​​by using data collected from neighboring countries, which can better preserve the local features of the data; combined with the isolated forest algorithm to detect and delete outliers, the purity of the data is further improved; in addition, data that should not be zero is eliminated through process analysis, ensuring the rationality and accuracy of the data, thereby comprehensively improving the data quality.

[0027] (2) This invention proposes a feature selection algorithm based on regression application scoring, which integrates three factors for feature selection, effectively making up for the shortcomings of traditional evaluation indicators that cannot intuitively reflect the performance of regression models; by accurately selecting the features required for training the model, the algorithm not only reduces the computational pressure, but also significantly improves the overall performance of the model.

[0028] (3) This invention proposes a method for iteratively updating the parameters of the training model in the next nearby time period, which solves the problem of model timeliness and greatly improves the long-term stability of the model prediction accuracy. Attached Figure Description

[0029] Figure 1 This is a flowchart of the prediction method of the present invention;

[0030] Figure 2 This is a flowchart of feature selection in this invention;

[0031] Figure 3 This is a schematic diagram of the ignition furnace and its constituent pipes in the embodiment;

[0032] Figure 4 This is a schematic diagram of the prediction effect in the embodiment. The predicted value has a high degree of fit with the actual value, and good results have been achieved. Detailed Implementation

[0033] To further understand the content of this invention, a detailed description of the invention will be provided in conjunction with the accompanying drawings.

[0034] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0035] The present invention will be further described below with reference to embodiments.

[0036] Example

[0037] Combination Figures 1-4 As shown, the high-efficiency adaptive complex industrial parameter prediction method based on regression application scoring of the present invention takes the prediction of sintering ignition temperature as an example in an industrial scenario:

[0038] First, field data is collected via PLC. Missing values ​​are detected in the collected data, and the data is imputed using the values ​​of nearest neighbor data points to maintain the integrity and consistency of the dataset. Then, the Isolation Forest algorithm is used, with 120 decision trees in the forest, to detect outliers and remove the top 1% of data with the shortest paths. Combined with... Figure 3 As shown, during sintering ignition, the pressure before the valve, the pressure after the valve, the gas flow rate, the pressure at the front end of the main pipe, and the gas flow rate of the ignition furnace should not have values ​​of 0. If these features have values ​​of 0, it indicates that these data are abnormal and these data samples with values ​​of 0 need to be deleted.

[0039] Through the above data collection and preprocessing, a high-quality dataset is obtained. Then, a feature selection algorithm based on regression application scoring is used for feature selection. The feature selection process is as follows: Figure 2 As shown, the model used in the screening algorithm is random forest, and the features selected are shown in Table 1. Table 2 shows the features selected using the traditional Pearson correlation algorithm. Comparing Table 1 and Table 2, it can be seen that the feature screening algorithm based on regression application scoring of this invention requires only 6 features to be screened, which is significantly less than the number of features screened using the traditional Pearson correlation algorithm. This indicates that the algorithm proposed in this invention greatly reduces the computational burden.

[0040] Table 1. Input features selected based on regression application scoring.

[0041] Feature variable name Feature abbreviation Blast furnace gas regulating valve opening D02 Gas main back end pressure D07 Ignition furnace gas flow rate D09 Ignition furnace gas regulating valve opening D10 airflow D11 Current temperature of the ignition furnace D17

[0042] Table 2. Input features selected by the Pearson correlation algorithm

[0043]

[0044]

[0045] After using a feature selection algorithm based on regression application scoring to select the optimal feature combination from the sintering ignition data, a random forest regression prediction model was trained to predict the sintering temperature over a future period. Table 3 shows the various indicators of the predicted sintering ignition temperature for the next 1 to 5 days using the random forest model. Table 3 shows that the model's performance deteriorates over time.

[0046] Table 3: Changes in Model Prediction Performance

[0047] 1 day 2 days 3 days 4 days 5 days Regression application score 0.95 0.92 0.87 0.81 0.77 Hit rate within 3% of error 98％ 94％ 91％ 84％ 80％ <![CDATA[R 2 ]]> 0.96 0.91 0.85 0.81 0.78 Mean Absolute Error 6℃ 9℃ 14℃ 17℃ 22℃

[0048] To address the timeliness issue of models, this invention proposes a method for iteratively updating the model at regular intervals to predict sintering ignition temperatures over a future period. Through multiple experiments, this invention has found that using data from the most recent five days to train the model for predicting the sintering ignition temperature of the next day not only yields ideal model performance but also requires moderate computational resources. Specifically, every other day, the latest five days of data are collected to retrain the model, and then the newly trained model is used to predict the sintering ignition temperature of the following day. By continuously repeating this iterative update process, the model is ensured to reflect changes in the latest data in a timely manner, thereby maintaining and improving the accuracy and stability of the prediction. One example of the prediction results is shown below. Figure 3 As shown, the predicted values ​​have a high degree of fit with the actual values, achieving good results. Table 4 lists the evaluation indicators for predicting the sintering ignition temperature in the next 30 days using the traditional random forest model trained by iterative updating based on the Pearson correlation feature selection algorithm and the random forest model trained by iterative updating based on the feature selection algorithm proposed in this invention. By comparing the indicators in Table 4, it can be seen that the model trained using the features selected by the algorithm of this invention performs better in predicting the sintering ignition data in the next 30 days. Specifically, this model performs better in regression application score, hit rate within 3% error, R... 2 The model outperforms the model trained solely using the Pearson correlation feature selection algorithm in key metrics such as mean absolute error. This demonstrates that the algorithm of this invention not only effectively improves the model's accuracy but also enhances its generalization ability, i.e., its adaptability and accuracy on unseen data. Furthermore, comparing the metrics of the random forest model trained iteratively using the regression-based scoring feature selection algorithm (see Table 4) with those of the random forest model trained without iterative updates using the regression-based scoring feature selection algorithm (see Table 3), a significant improvement in all metrics is clearly observed in the iteratively updated training model. This result indicates that iteratively updating the training of the random forest model to predict the ignition temperature at the next time point can significantly improve the overall performance of the model and enhance the stability and reliability of the prediction.

[0049] Table 4 Comparison of the performance of two feature selection algorithms in predicting ignition temperature for the next 30 days when training random forest models.

[0050]

[0051] The present invention and its embodiments have been described above illustratively. This description is not restrictive and is merely one embodiment of the present invention, and is not actually limited thereto. Therefore, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the present invention, such designs should fall within the protection scope of the present invention.

Claims

1. A high-performance adaptive method for predicting complex industrial parameters based on regression application scoring, characterized in that: The process for predicting sintering ignition temperature parameters includes the following steps: Step 1, Data Acquisition and Preprocessing: Install the PLC acquisition device at the location to be tested, and preprocess the acquired sintering ignition data to fill in missing values ​​and remove outliers; The data preprocessing process is as follows: missing values ​​are filled using the values ​​of nearest neighbor data points, and outlier detection is performed on the sintering ignition data collected in step 1 using the isolated forest algorithm. After completing the outlier scoring of all data points, a threshold is set according to the scoring results to delete outliers. After deleting outliers using the isolated forest algorithm, the data is deeply cleaned to further delete data samples that should not have a value of 0 but have a value of 0. This part of the data samples that should not have a value of 0 includes the pressure value before the valve, the pressure value after the valve, the gas flow rate, the pressure at the front end of the main pipe, and the gas flow rate of the ignition furnace. Step 2, Feature Selection: Introduce a regression scoring method and combine key indicators such as hit rate, coefficient of determination, and time consumption within the 3% error range to select the optimal feature set required for training the model. The optimal feature set includes the blast furnace gas regulating valve opening characteristics, the gas main downstream pressure characteristics, the ignition furnace gas flow characteristics, the ignition furnace gas regulating valve opening characteristics, the air flow characteristics, and the current temperature characteristics of the ignition furnace. The calculation process of regression application score is as follows: Based on the key indicators of hit rate, coefficient of determination and time consumption score with an error within 3%, each indicator is multiplied by its corresponding weight. The regression application score calculation formula is shown in Equation (1): Where a, b, and c are the weights corresponding to hit rate, coefficient of determination, and time taken, respectively, a+b+c=1, and HitRate is the hit rate within 3% of the error; R 2 The coefficient of determination is calculated as shown in equation (2); the TimeScore is the time consumption score, and its calculation formula is shown in equation (3). in, These are actual values ​​of industrial parameters. These are the actual average values ​​of industrial parameters. These are predicted values ​​for industrial parameters. It is the total sample size. To delete the first The time required to train the model for each feature The time taken to train the model using all features; The process of selecting the optimal feature set required for training the model includes: first, training the model using all features and recording the regression application score, which is then used as the baseline score; next, temporarily deleting one feature, retraining the model using the remaining features, and recording the new regression application score; if the regression application score of the model is higher than the previous baseline score after temporarily excluding a feature, it indicates that the excluded feature is not necessary for building the model and should be deleted; conversely, if the regression application score is lower than the previous baseline score, it indicates that the excluded feature is an important feature required by the model, and its deletion would lead to a decrease in model performance, so it should be retained; through this process, all features are eliminated one by one, and finally the optimal feature set required for training the model is selected. Step 3: Predict the sintering ignition temperature at the next time point: Train a random forest model using the selected optimal feature set to predict the sintering ignition temperature at the next time point; and periodically update the random forest model using the latest data.

Images