A high-precision weighing method and system for a silo under weighing sensor failure

Abstract

This invention discloses a high-precision weighing method and system for weighing towers under load cell failure conditions, belonging to the field of industrial intelligent sensing technology. The method includes: acquiring load cell data from the weighing tower and preprocessing the data; modeling the load cell data using a nonlinear tree model, training the model to learn the nonlinear relationship between the load cell data and the total weight of the tower, and applying monotonic constraints to the target during model training; using the trained nonlinear tree model to predict new load cell data to obtain the total weight of the tower, and correcting the prediction results using a calibration method. This invention solves the problem of accurately predicting the total weight of a tower when key load cell data is missing. This invention utilizes machine learning technology to accurately calculate the total weight of the tower even when some load cells fail or data is lost.

Description

Technical Field

[0001] This invention relates to the field of industrial intelligent sensing technology, and in particular to a high-precision weighing method and system for weighing towers under load sensor failure. Background Technology

[0002] Material silos are typically large devices used for storing and measuring materials, and in automated production systems, their weighing sensors play a crucial role. Usually, multiple sensors are installed within the silo to monitor the total weight of the material. However, due to environmental factors, power outages, communication failures, sensor malfunctions, and other reasons, some sensor data may be lost, leading to inaccurate measurements of the silo's total weight.

[0003] Traditional solutions rely on manual inspection, repair, or replacement of faulty sensors. This approach not only increases maintenance costs but also, due to a lack of real-time capability, can lead to production stoppages or delays, thus impacting production efficiency and data traceability. Furthermore, existing data recovery methods, such as linear regression and interpolation, also have significant limitations. In the absence of critical sensor data, linear regression struggles to accurately predict the total weight of the silo, while the accuracy of interpolation methods is heavily influenced by the distribution of feeding and unloading data.

[0004] Therefore, how to accurately calculate the total weight of the tower using normal data from other sensors in the event of sensor failure has become an urgent problem to be solved in the field of automation control technology. Summary of the Invention

[0005] The purpose of this invention is to provide a high-precision weighing method and system for weighing towers under the condition of weighing sensor failure. By utilizing machine learning technology, the total weight of the tower can still be accurately calculated even when some weighing sensors fail or data is lost, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A high-precision weighing method for a material tower under load cell failure, comprising:

[0008] Acquire data from the weighing sensor of the silo and preprocess the data, including data standardization, missing value imputation and feature construction.

[0009] A nonlinear tree model is used to model the weighing sensor data of the tower. The nonlinear relationship between the weighing sensor data and the total weight of the tower is learned by training the model. Monotonic constraints are applied to the target during the model training process to ensure the monotonicity between specific features and prediction results.

[0010] The trained nonlinear tree model is used to predict the total weight of the tower from the new weighing sensor data. Based on the prediction results, a calibration method is used to correct the prediction results to improve the prediction accuracy.

[0011] Preferably, the data from the weighing sensors of the silo are preprocessed by performing the following operations:

[0012] A sliding window averaging algorithm is applied to each weighing sensor signal to remove short-term jitter.

[0013] By defining mapping relationships, the column names of the weighing sensor data of the pylon are standardized to ensure that the data format is consistent for subsequent model training and inference.

[0014] Missing data is handled by setting a reasonable filling scheme. For faulty sensors, the weighted average of two adjacent sensors is used as the first choice for filling. If adjacent data is also missing, the default value of 0 is used to fill the missing data to ensure the integrity of the dataset.

[0015] The S7sum feature is calculated on the load cell data of the silo, and the load cell data is summarized by the S7sum feature to ensure that other load cells can still provide effective information for total weight estimation in the event of failure of some load cells.

[0016] Preferably, the nonlinear tree model is the LightGBM model, which is trained using the gradient boosting tree algorithm. The nonlinear tree model is used for multivariate regression prediction, and the following operations are performed:

[0017] Multiple regression model: In model training, multiple weighing sensor data and S7sum aggregated data are selected as features, the target variable is the total weight of the tower, and the gradient boosting decision tree algorithm based on the LightGBM model captures the complex nonlinear relationships in the data;

[0018] Monotonic constraints: Monotonic constraints are set to ensure the monotonicity of certain features. For example, the total weight of the pylon increases monotonically with the increase of S7sum. Therefore, the model will impose a monotonically non-decreasing constraint on S7sum. During each incremental tree construction, the splitting direction is restricted according to this constraint, and branches opposite to the constraint are not allowed to be output. This is used to ensure that the output of the model conforms to physical reality and at the same time avoid unreasonable fluctuations in the model at extreme data points.

[0019] Preferably, the monotonic constraint is used to ensure that the sum of the weighing sensor data of the silo has a monotonically increasing relationship with the total weight of the silo.

[0020] Preferably, the feature construction includes calculating the sum of data from multiple tower weighing sensors and inputting it as a new feature into a nonlinear tree model for training.

[0021] Preferably, missing values ​​in the weighing sensor data of the material tower are filled by a default value or a weighted average of the weighing sensor data of adjacent material towers.

[0022] Preferably, the calibration method is an isotonic regression method, which is used to calibrate the prediction results to improve the accuracy of the model.

[0023] Preferably, the distribution of the total weight data of the material tower incorporates a sample weighting mechanism to improve the prediction accuracy of the model in scenarios with imbalanced or sparse data, including:

[0024] The key features of the weighing sensor of the silo are sorted according to their numerical values ​​and divided into several quantile intervals;

[0025] Based on the distribution density of the weighing sensor data of the tower in each quantile interval, the sample weights are dynamically allocated. Among them, the quantile intervals where the weighing sensor data of the tower are relatively concentrated are given higher weights, so that the model pays more attention to common working conditions. The intervals where the data distribution is sparse or where outliers are prone to occur are given lower weights to avoid the training process being overly affected by extreme values, so that the model can simultaneously take into account the prediction accuracy under the main working conditions and the robustness under abnormal working conditions.

[0026] During model training, a weighted squared error loss function is used, and the quantile weighting results are employed. The loss function is defined as follows:

[0027]

[0028] in, Indicates the number of samples. Indicates the first The true value of each sample Indicates the first The predicted value for each sample, This represents the weights calculated based on quantile intervals;

[0029] After the model training is completed, real tower weight data is obtained. and the corresponding prediction results Construct training sample pairs ( , Applying isotonic regression to the set of sample pairs, a monotonically increasing calibration function is obtained through training. .

[0030] According to another aspect of the present invention, a high-precision weighing system for a material tower under load cell failure is provided, for implementing the high-precision weighing method for a material tower under load cell failure as described above, comprising:

[0031] The data acquisition module is used to acquire data from the weighing sensors of the silo tower.

[0032] The data preprocessing module is used to standardize, fill in missing values, and construct features for the weighing sensor data of the pylon;

[0033] The model training module is used to model the weighing sensor data of the tower using a nonlinear tree model and apply monotonic constraints during the training process.

[0034] The prediction module is used to predict the total weight of the tower by using a trained nonlinear tree model to predict new sensor data.

[0035] The calibration module is used to calibrate the prediction results and improve the accuracy of the predictions.

[0036] Preferably, the data preprocessing module fills in the missing values ​​in the weighing sensor data of the material tower according to a preset strategy, and the filling value is the default value or the weighted average of the weighing sensor data of adjacent material towers.

[0037] Compared with the prior art, the beneficial effects of the present invention are:

[0038] This invention introduces monotonicity physical constraints into a data-driven machine learning model, forcing a monotonically increasing relationship between the model output (predicted total weight) and key input features (the sum of normal sensor data S7sum). This ensures that the prediction results strictly conform to the physical law that "as the weighing sensor reading increases, the total weight of the tower will inevitably increase," effectively avoiding erroneous predictions that defy common sense. Addressing the issue of uneven sample distribution across different weight ranges in tower weighing scenarios (few data for empty and full towers, and many data for intermediate states), a sample weight allocation strategy based on quantile intervals is designed. This guides the model to focus more on prediction accuracy under common operating conditions while reducing the interference of outliers on model training. A two-level optimization system combining in-model monotonic constraints and post-prediction calibration is constructed to further improve prediction accuracy. Attached Figure Description

[0039] Figure 1 This is a schematic diagram of the frame of the weighing system for the material tower of the present invention;

[0040] Figure 2 This is a flowchart of the model training process of the present invention;

[0041] Figure 3 This is a flowchart of the model calibration process of the present invention;

[0042] Figure 4 This is a flowchart of the real-time inference process of the model in this invention. Detailed Implementation

[0043] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0044] To address the challenges of accurately predicting the total weight of a material tower using linear regression methods in the absence of crucial weighing sensor data, and the significant impact of interpolation methods on the accuracy of feeding and unloading data distribution, please refer to [link to relevant documentation]. Figure 1-4 The following technical solution is provided in this embodiment:

[0045] A high-precision weighing system for use in a material tower under load sensor failure includes: a data acquisition module, a data preprocessing module, a model training module, a prediction module, and a calibration module.

[0046] The system comprises the following modules: a data acquisition module for acquiring data from the weighing sensors of the pylon; a data preprocessing module for standardizing, imputing missing values, and constructing features from the weighing sensor data, specifically imputing missing values ​​according to a preset strategy, with the imputed values ​​being either default values ​​or the weighted average of adjacent weighing sensor data; a model training module for modeling the weighing sensor data using a nonlinear tree model and applying monotonic constraints during training; a prediction module for predicting the total weight of the pylon using the trained nonlinear tree model on new sensor data; and a calibration module for calibrating the prediction results and improving the accuracy of the predictions.

[0047] It should be noted that data is acquired from the weighing sensors in the silo. Multiple weighing sensors (e.g., S1 to S8) are deployed within the silo system to measure the weight of the material inside the silo. The weighing sensor data is periodically transmitted to the data preprocessing module through the data acquisition module. The data acquisition module can be an IoT device connected to the sensors or a local control system, capable of reading the data from each weighing sensor in real time or at regular intervals and transmitting it to the central processing unit.

[0048] The data preprocessing module performs the following processing on the load cell data of the silo: 1) Standardization: Since load cells may use different units or ranges, the system needs to standardize all load cell data to make them have the same scale; 2) Missing value imputation: In the event of a load cell failure or data loss, the system will use a specific missing value imputation strategy. For example, for a faulty load cell (such as S6), the system will preferentially use the weighted average of two adjacent load cells to imput the missing value. If adjacent data is also missing, the system will fall back to the default value of 0 to imput the missing value, ensuring the integrity of the dataset; 3) Feature construction: The system calculates S7sum (the sum of multiple load cell data, usually including S1 to S5 and S7, S8), which is used as a new feature input into the model. S7sum can summarize the overall performance of multiple load cells and provide strong support for the inference of the total weight of the silo.

[0049] The preprocessed tower weighing sensor data was trained using a nonlinear tree model. The training steps were as follows: 1) Feature selection: S1 to S5, S7, S8, and S7sum were selected as input features, and the target variable was the total weight of the tower (i.e., the label value); 2) Data sorting: All feature data were sorted in ascending order according to S7sum; 3) Quantile calculation and weight allocation: Based on the overall data distribution, the data was divided into three quantile intervals: the first quantile (0%-10%), the second quantile (10%-90%), etc. The third quantile (90%-100%) is assigned a higher weight (e.g., 1.5) to intervals with relatively concentrated data distribution (such as the second quantile) and a lower weight (e.g., 0.8) to intervals with relatively sparse distribution (such as the first and third quantiles); 4) Model training: LightGBM is used to perform regression training on the selected features to establish a nonlinear relationship between the weighing sensor data and the total weight of the silo. During training, quantile weights are used to adjust the contribution of each data point to the loss function. For example, a weighted squared error loss function can be used. ,in, These are the quantile weights of sample i. It's a real label. These are model predictions.

[0050] During training, monotonic constraints are applied to ensure that the model output conforms to actual physical laws. For example, there is a monotonically increasing relationship between the S7sum feature and the total weight of the pylon, that is, as S7sum increases, the total weight of the pylon should increase. Therefore, the system applies a monotonically increasing constraint during training to ensure that when S7sum increases, the predicted total weight of the pylon also increases.

[0051] After the model training is completed, real tower weight data is obtained. and the corresponding prediction results Construct training sample pairs ( , Applying isotonic regression to the set of sample pairs, a monotonically increasing calibration function is obtained through training. .

[0052] In practical applications, load cells in the silo may malfunction, leading to the loss of some sensor data. To address this issue, the system is designed with a fault sensor handling mechanism: 1) Fault sensor detection: The system monitors the status of the load cells in real time. If any load cell data is found to be abnormal or missing, the system will automatically mark it as a fault sensor; 2) Fault sensor handling: For momentary fault sensors, the system uses default values ​​(such as the average value of adjacent sensors) to fill in the missing data, ensuring data integrity; while for sensors with long-term faults, a strategy of taking the weighted average of the readings of its adjacent sensors is used to fill in the missing data, ensuring that the model can still make effective predictions even when sensor data is missing.

[0053] After processing the weighing sensor data, the system inputs this data into the trained LightGBM model to predict the total weight of the silo. If some weighing sensor data is lost, the system will fill in the data and make a prediction. Finally, the model will output a predicted total weight value.

[0054] To further improve prediction accuracy, the generated isotonic regression function is used. By calibrating the prediction results, isotonic regression can correct inaccuracies caused by model errors, data biases, or other external factors, thereby making the final prediction results more accurate.

[0055] After the prediction is completed, the system outputs the predicted total weight of the tower along with the original data. The results are usually displayed in Excel file format, or the data is saved to a specified database table, or returned as a real-time data stream through an agreed interface.

[0056] This system can accurately calculate the total weight of the material tower in the event of partial failure of the weighing sensor or loss of data. It can be widely used in the automation control systems of industries such as chemical, feed, food, and mining, significantly improving the reliability and real-time performance of the data.

[0057] To better demonstrate the high-precision weighing process under load cell failure in a material tower, this embodiment provides a high-precision weighing method for load cell failure in a material tower, based on the aforementioned high-precision weighing system for load cell failure in a material tower, including:

[0058] Acquire data from the load cells of the silo, including data from multiple load cells.

[0059] The data from the weighing sensors of the silo are preprocessed, including data standardization, missing value imputation, and feature construction.

[0060] In this embodiment, feature construction includes calculating the sum of data from multiple tower weighing sensors and inputting it as a new feature into a nonlinear tree model for training.

[0061] In this embodiment, missing values ​​in the weighing sensor data of the silo are filled by the default value or the weighted average of the weighing sensor data of adjacent silos.

[0062] It should be noted that preprocessing the load cell data of the silo is the first step to ensure that the model can accurately calculate the total weight of the silo. The original load cell data usually contains noise, missing values ​​and different unit formats, and needs to be standardized and cleaned.

[0063] In this embodiment, the data from the weighing sensor of the silo tower is preprocessed by performing the following operations:

[0064] A sliding window averaging algorithm is applied to each weighing sensor signal to remove short-term jitter.

[0065] By defining mapping relationships (such as 'sensor1':'S1'), the column names of the weighing sensor data of the pylon are standardized to ensure that the data format is consistent for subsequent model training and inference.

[0066] Missing data is handled by setting a reasonable filling scheme. For faulty sensors, the weighted average of two adjacent sensors is used as the first choice for filling. If adjacent data is also missing, the default value of 0 is used to fill the missing data to ensure the integrity of the dataset.

[0067] The S7sum feature is calculated on the load cell data of the silo. S7sum is the weighted sum of multiple load cell data (usually S1 to S5 and S7, S8). This feature is used as an important input feature in model training. The S7sum feature can effectively summarize the load cell data and ensure that even if some load cells fail, other load cells can still provide effective information for total weight estimation.

[0068] It should be noted that the "fault sensor processing mechanism" of this invention differs from conventional data preprocessing steps. Data preprocessing aims to standardize and clean the original weighing sensor data to ensure the integrity of the model input data. The fault sensor processing mechanism, however, addresses anomalies or failures that occur during the operation of weighing sensors. Its goal is to ensure the reliability and stability of the system's predictions even when some or multiple weighing sensors fail. Specifically, the fault sensor processing mechanism of this invention includes the following steps: 1) Fault detection: The model can monitor the data stream of each weighing sensor. When an abnormal value range is detected or a reading shows a significant, illogical difference from adjacent weighing sensors, the weighing sensor is marked as faulty; 2) Missing value imputation: During prediction, if a short-term fault is detected in a weighing sensor, the model simulates missing data using a predefined imputation value (e.g., 0). If a weighing sensor is detected to be in a faulty state for an extended period (e.g., the fault time exceeds N sensor data acquisition cycles T), the model uses the weighted average of adjacent weighing sensors to imput the data, ensuring that the model can still make effective predictions even when weighing sensor data is abnormal.

[0069] A nonlinear tree model is used to model the load cell data of the silo. The nonlinear relationship between the load cell data and the total weight of the silo is learned by training the model. During the model training process, a monotonic constraint is applied to the target to ensure the monotonicity between specific features and the prediction results.

[0070] In this embodiment, the nonlinear tree model is the LightGBM model, which is trained using the gradient boosting tree algorithm. The use of a nonlinear tree model for multivariate regression prediction can handle highly nonlinear data and is suitable for complex tower weight prediction problems. The following operations are performed:

[0071] Multiple regression model: In model training, multiple weighing sensor data (such as S1 to S5 and S7 to S8) and S7sum summary data are selected as features. The target variable is the total weight of the tower. The gradient boosting decision tree algorithm based on the LightGBM model captures the complex nonlinear relationship in the data and has high computational efficiency.

[0072] Monotonic constraints: To ensure that the model's prediction results conform to physical laws, monotonic constraints are set to ensure the monotonicity of certain features. For example, the total weight of the pylon increases monotonically with the increase of S7sum. Therefore, the model applies a monotonically non-decreasing constraint to S7sum. During each incremental tree construction, the splitting direction is restricted according to this constraint, and branches opposite to the constraint are not allowed to be output. This is used to ensure that the model's output conforms to physical reality and at the same time avoid unreasonable fluctuations in the model at extreme data points.

[0073] In this embodiment, the monotonic constraint is used to ensure that the sum of the data from the weighing sensors of the silo has a monotonically increasing relationship with the total weight of the silo.

[0074] The trained nonlinear tree model is used to predict the total weight of the tower from the new weighing sensor data. Based on the prediction results, a calibration method is used to correct the prediction results to improve the prediction accuracy.

[0075] In this embodiment, the calibration method is isotonic regression, which is used to calibrate the prediction results to improve the accuracy of the model.

[0076] In this embodiment, a sample weighting mechanism is introduced to assess the distribution of the total weight data of the silo, in order to improve the prediction accuracy of the model in scenarios with imbalanced or sparse data, including:

[0077] The key features of the weighing sensor of the bulge (such as the sum feature S7sum) are sorted according to their numerical values ​​and divided into several quantile intervals;

[0078] Based on the distribution density of the weighing sensor data of the tower in each quantile interval, the sample weights are dynamically allocated. For quantile intervals where the weighing sensor data of the tower is relatively concentrated (such as the 10%-90% interval of S7sum), higher weights are assigned to make the model pay more attention to common working conditions. For intervals where the data distribution is sparse or where outliers are prone to occur, lower weights are assigned to avoid the training process being overly affected by extreme values, so that the model can simultaneously take into account the prediction accuracy under the main working conditions and the robustness under abnormal working conditions.

[0079] During model training, a weighted squared error loss function is used, and the quantile weighting results are employed. The loss function is defined as follows:

[0080]

[0081] in, Indicates the number of samples. Indicates the first The true value of each sample Indicates the first The predicted value for each sample, This represents the weights calculated based on quantile intervals;

[0082] After the model training is completed, real tower weight data is obtained. and the corresponding prediction results Construct training sample pairs ( , Applying isotonic regression to the set of sample pairs, a monotonically increasing calibration function is obtained through training. .

[0083] It should be noted that the total weight of the silo is calculated by using a trained nonlinear tree model to extrapolate from real-time weighing sensor data. The prediction process includes the following key steps: 1) Real-time extrapolation: When the silo system receives new weighing sensor data, it inputs this data into the trained model for prediction, calculating the total weight of the silo in real time; 2) Calibration and correction: Based on the prediction results, an isotonic regression calibration function is used. Further processing yields the corrected predicted values. Among them, calibration function It was pre-rehearsed.

[0084] This invention can process weighing sensor data in real time during the production process and output the predicted total weight of the silo. The result is output in the following manner:

[0085] Excel / Database Table Output: The total weight of the silo after model inference will be written to an Excel table or database table in real time, and operators can directly view the prediction results and perform subsequent operations.

[0086] Integration and Display: Supports seamless integration with existing production control systems, real-time monitoring of silo status, and automatic adjustments based on forecast results to improve production efficiency and accuracy.

[0087] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A high-precision weighing method for a material tower under load cell failure, characterized in that, include: Acquire data from the weighing sensor of the silo and preprocess the data, including data standardization, missing value imputation and feature construction. A nonlinear tree model is used to model the weighing sensor data of the tower. The nonlinear relationship between the weighing sensor data and the total weight of the tower is learned by training the model. Monotonic constraints are applied to the target during the model training process to ensure the monotonicity between specific features and prediction results. The trained nonlinear tree model is used to predict the new weighing sensor data to obtain the total weight of the tower. Based on the prediction results, a calibration method is used to correct the prediction results to improve the prediction accuracy. Preprocess the load cell data of the silo by performing the following operations: A sliding window averaging algorithm is applied to each weighing sensor signal to remove short-term jitter. By defining mapping relationships, the column names of the weighing sensor data of the pylon are standardized to ensure that the data format is consistent for subsequent model training and inference. Missing data is handled by setting a reasonable filling scheme. For faulty sensors, the weighted average of two adjacent sensors is used as the first choice for filling. If adjacent data is also missing, the default value of 0 is used to fill the missing data to ensure the integrity of the dataset. The sum of normal sensor data (S7sum) is calculated from the load cell data of the silo. The load cell data of the silo is summarized by the sum of normal sensor data (S7sum) to ensure that other load cells can still provide effective information for total weight estimation in the event of failure of some load cells.

2. The high-precision weighing method for a weighing tower under load sensor failure according to claim 1, characterized in that, The nonlinear tree model is the LightGBM model, which is trained using the gradient boosting tree algorithm. The nonlinear tree model is used for multivariate regression prediction, and the following operations are performed: Multiple regression model: In model training, multiple weighing sensor data and the sum of normal sensor data (S7sum) are selected as features, the target variable is the total weight of the tower, and the gradient boosting decision tree algorithm based on the LightGBM model captures the complex nonlinear relationships in the data. Monotonic constraints: Monotonic constraints are set to ensure the monotonicity of certain features. For example, the total weight of the pylon increases monotonically with the increase of the sum of normal sensor data S7sum. Therefore, the model will impose a monotonically non-decreasing constraint on the sum of normal sensor data S7sum. During each incremental tree construction, the splitting direction is restricted according to this constraint, and branches opposite to the constraint are not allowed to be output. This is used to ensure that the output of the model conforms to physical reality and at the same time avoid the model from generating unreasonable fluctuations at extreme data points.

3. The high-precision weighing method for a weighing tower under load sensor failure according to claim 1, characterized in that, The monotonic constraint is used to ensure that the sum of the data from the weighing sensors of the silo has a monotonically increasing relationship with the total weight of the silo.

4. The high-precision weighing method for a weighing tower under load cell failure according to claim 1, characterized in that, The feature construction includes calculating the sum of data from multiple weighing sensors in the tower, which is then used as a new feature input into a nonlinear tree model for training.

5. A high-precision weighing method for a material tower under load cell failure according to claim 1, characterized in that, Missing values ​​in the weighing sensor data of the hopper are filled by the default value or the weighted average of the weighing sensor data of adjacent hoppers.

6. A high-precision weighing method for a material tower under load cell failure according to claim 1, characterized in that, The calibration method is the isotonic regression method, which is used to calibrate the prediction results to improve the accuracy of the model.

7. A high-precision weighing method for a material tower under load cell failure according to claim 1, characterized in that, The distribution of the total weight data of the silo incorporates a sample weighting mechanism to improve the prediction accuracy of the model in scenarios with imbalanced or sparse data, including: The key features of the weighing sensor of the silo are sorted according to their numerical values ​​and divided into several quantile intervals; Based on the distribution density of the weighing sensor data of the tower in each quantile interval, the sample weights are dynamically allocated. Among them, the quantile intervals where the weighing sensor data of the tower are relatively concentrated are given higher weights, so that the model pays more attention to common working conditions. The intervals where the data distribution is sparse or where outliers are prone to occur are given lower weights to avoid the training process being overly affected by extreme values, so that the model can simultaneously take into account the prediction accuracy under the main working conditions and the robustness under abnormal working conditions.

8. A high-precision weighing system for a material tower under load sensor failure, used to implement the high-precision weighing method for a material tower under load sensor failure as described in any one of claims 1-7, characterized in that, include: The data acquisition module is used to acquire data from the weighing sensors of the silo tower. The data preprocessing module is used to standardize, fill in missing values, and construct features for the weighing sensor data of the pylon; The model training module is used to model the weighing sensor data of the tower using a nonlinear tree model and apply monotonic constraints during the training process. The prediction module is used to predict the total weight of the tower by using a trained nonlinear tree model to predict new sensor data. The calibration module is used to calibrate the prediction results and improve the accuracy of the predictions.

9. A high-precision weighing system for a material tower under load sensor failure according to claim 8, characterized in that, The data preprocessing module fills in the missing values ​​in the weighing sensor data of the material tower according to a preset strategy. The filling value is the default value or the weighted average of the weighing sensor data of adjacent material towers.

Images