Productivity management method based on wagon balance data and machine learning algorithm

Abstract

The invention discloses a capacity management method based on wagon balance data and a machine learning algorithm, and belongs to the technical field of industrial intelligent manufacturing. The method comprises the following steps: firstly, preprocessing collected original wagon balance data through a bilateral filtering noise reduction model and data cleaning; then, constructing multi-dimensional features from time, material and customer dimensions, and inputting the multi-dimensional features into an LSTM-Transform mixed demand prediction model for training so as to predict future material demands; and finally, based on the predicted demand, generating a productivity adjustment scheme through a productivity demand mapping model and a dynamic scheduling strategy. According to the method, the problems that traditional productivity management depends on artificial experience, the data dimension is single, prediction lags behind and the dynamic adjustment capability is weak are solved, data-driven accurate demand prediction and productivity dynamic optimization are realized, and the production efficiency and the inventory turnover rate are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of industrial intelligent manufacturing technology, specifically to a capacity management method based on weighbridge data and machine learning algorithms. Background Technology

[0002] In the production management of process manufacturing enterprises such as mining, metallurgy, and building materials, accurate capacity planning and efficient inventory control are key to improving operational efficiency. Currently, most enterprises still rely on manual experience combined with historical sales data for capacity forecasting and planning. This approach has significant drawbacks: First, the data dimension is limited, failing to fully utilize the high-frequency data reflecting actual material flow collected in real time by IoT devices such as weighbridges; second, forecasts are lagging, unable to respond promptly to rapid fluctuations in market demand, easily leading to insufficient capacity during peak seasons or inventory backlogs during off-seasons; third, dynamic adjustment capabilities are weak, production line scheduling lacks real-time data-driven support, making it difficult to cope with sudden changes in material supply or customer orders; finally, safety stock levels set based on experience are often unreasonable, resulting in excessive capital occupation or the risk of supply disruptions.

[0003] While some existing Enterprise Resource Planning (ERP) systems attempt to manage capacity, they typically fail to deeply integrate and explore the time-series characteristics of weighbridge data and its correlation with multi-dimensional business processes such as materials and customers. This hinders accurate demand forecasting and flexible capacity matching. Therefore, there is an urgent need in this field for a novel capacity management solution that deeply integrates real-time IoT data with intelligent algorithms. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a capacity management method, system and equipment based on weighbridge data and machine learning algorithms, so as to achieve accurate prediction of material demand and dynamic optimization scheduling of production capacity.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A capacity management method based on weighbridge data and machine learning algorithms includes the following steps: S1. Obtain the real-time material weighing data sequence collected by the weighbridge sensor. The material weighing data sequence is a sequence composed of timestamped weighbridge data. The material weighing data sequence is filtered by a bilateral filtering noise reduction model, and the filtered data is cleaned and normalized to generate a structured dataset. S2. Extract multi-dimensional features, including time, material, and customer dimensions, from the structured dataset and construct a multi-dimensional feature matrix; use the multi-dimensional feature matrix to train an LSTM-Transformer hybrid demand forecasting model to output material demand forecasts for a specified future period. S3. Based on the material demand forecast, calculate the total capacity demand within the planning period through the capacity demand mapping model, and generate a dynamic scheduling scheme for production line capacity based on the preset capacity adjustment threshold rules and optimization algorithm. S4. Push the dynamic scheduling plan for production line capacity to the production information management system to drive the production line to execute.

[0006] Furthermore, the weighbridge data in the material weighing data sequence is linked to the corresponding key business data, which includes material type, customer ID, production line equipment status, and inventory data.

[0007] Furthermore, the bilateral filtering noise reduction model adopts a state-space model. The state vector of the state-space model includes material weight, weight change rate, and weight acceleration. The process noise covariance matrix and observation noise covariance matrix used to suppress noise are adaptively adjusted through iterative update equations.

[0008] Furthermore, data cleaning and normalization include using the interquartile range method to detect and repair outliers in the filtered data, and performing standard deviation standardization to generate a structured dataset.

[0009] Furthermore, the time characteristics include rolling window statistics and periodic characteristics including year / month / week / day, holiday identifiers, and seasonal factors; the material characteristics include classification and continuous characteristics including material type, particle size grade, and grade index; and the customer characteristics include customer type, historical order frequency, and purchase volume fluctuation coefficient. Multidimensional features also include derived features; derived features include demand volatility and customer concentration index.

[0010] Furthermore, the LSTM-Transformer hybrid demand forecasting model includes: an input layer for inputting a multi-dimensional feature matrix, a bidirectional LSTM layer for capturing long-term and short-term dependencies in time series, a Transformer layer for modeling the correlation between different feature dimensions, and a fully connected layer for outputting material demand forecasts. The loss function of the LSTM-Transformer hybrid demand forecasting model is weighted mean square error, and hierarchical time series cross-validation is used during training.

[0011] Furthermore, the calculation formula for the capacity-demand mapping model is as follows: C p = ; in, C p This represents the total capacity demand within the planned period. T For the planned period, i An index representing a time sequence number. i= 1,2, ..., T , k An index representing the type of material. k = 1, 2, ..., m , D t+i，k For the predicted first t + i Monthly material k Material requirements, r k For materials k Processing conversion rate, or This represents the average operating efficiency of the production line.

[0012] Furthermore, the capacity adjustment threshold rules include: when the predicted total capacity demand growth rate is greater than the first threshold, a capacity expansion strategy is triggered; when the predicted total capacity demand decline rate is greater than the second threshold, a capacity contraction strategy is triggered.

[0013] Furthermore, the optimization algorithm is a genetic algorithm, and the objective function is: ; in, Minimize This indicates the ultimate goal of the entire optimization algorithm. C s To cover the costs of switching production, C i For inventory costs, C d To cover the cost of delayed delivery, α , β , c These are the weighting coefficients for production changeover costs, inventory costs, and delayed delivery costs, respectively.

[0014] Furthermore, this also includes: periodically calculating the prediction accuracy index of the LSTM-Transformer hybrid demand forecasting model and the capacity stability index of the production information management system, and incrementally training and dynamically adjusting the feature importance of the LSTM-Transformer hybrid demand forecasting model based on newly generated weighbridge data and actual production line capacity data.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. High prediction accuracy: This invention achieves high-precision prediction of material demand for the next few months by integrating high-frequency weighbridge data with multi-dimensional features and using an LSTM-Transformer hybrid model, which is far superior to traditional manual estimation.

[0016] 2. Strong dynamic response capability: This invention establishes a dynamic scheduling mechanism based on predicted demand, which can automatically trigger capacity adjustment according to demand fluctuations, reduce capacity volatility, and effectively respond to market changes.

[0017] 3. Inventory Management Optimization: This invention improves inventory turnover and reduces capital occupation and inventory costs through accurate demand forecasting and capacity matching.

[0018] 4. High degree of automation: This invention realizes full-process data-driven automation from data collection, processing, prediction to scheduling scheme generation, which greatly reduces the manual decision-making links and improves management efficiency.

[0019] 5. Strong systematicity: This invention constructs a complete closed-loop system of data-prediction-decision-execution-feedback, ensuring the continuous optimization and long-term effectiveness of the solution. Attached Figure Description

[0020] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. Obviously, the described embodiments are merely some embodiments of this invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0022] The terms "first," "second," etc., used in this invention are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, apparatus, product, or end that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or ends.

[0023] like Figure 1 As shown, the present invention provides a capacity management method based on weighbridge data and machine learning algorithms, comprising the following steps: S1. Obtain the material weighing data sequence collected in real time by the weighbridge sensor. The material weighing data sequence is a sequence composed of timestamped weighbridge data. The material weighing data sequence is filtered by a bilateral filtering noise reduction model, and the filtered data is cleaned and normalized to generate a structured dataset. S2. Extract multi-dimensional features, including time, material, and customer dimensions, from the structured dataset and construct a multi-dimensional feature matrix; use the multi-dimensional feature matrix to train an LSTM-Transformer hybrid demand forecasting model to output material demand forecasts for a specified future period. S3. Based on the material demand forecast, calculate the total capacity demand within the planning period through the capacity demand mapping model, and generate a dynamic scheduling scheme for production line capacity based on the preset capacity adjustment threshold rules and optimization algorithm. S4. Push the dynamic scheduling plan for production line capacity to the production information management system to drive the production line to execute.

[0024] This invention first preprocesses the collected weighbridge data using a bilateral filtering noise reduction model and data cleaning. Then, it constructs multi-dimensional features based on time, material, and customer dimensions, and inputs these features into an LSTM-Transformer hybrid demand forecasting model for training to predict future material demand. Finally, based on the predicted demand, it generates a dynamic capacity adjustment plan through a capacity demand mapping model and dynamic scheduling strategies (capacity adjustment threshold rules and optimization algorithms). This invention solves the problems of traditional capacity management, such as reliance on manual experience, limited data dimensions, predictive lag, and weak dynamic adjustment capabilities. It achieves data-driven, accurate demand forecasting and dynamic capacity optimization, significantly improving production efficiency and inventory turnover.

[0025] In some embodiments, the weighbridge data in the material weighing data sequence is associated with corresponding key business data, including material type, customer ID, production line equipment status, and inventory data. Linking the weighbridge data with key business data establishes a complex and dynamic relationship between time, materials, customers, capacity, and inventory, facilitating globally optimal prediction and decision-making.

[0026] In some embodiments, the bilateral filtering noise reduction model adopts a state-space model. The state vector of the state-space model includes material weight, weight change rate and weight acceleration. The process noise covariance matrix and observation noise covariance matrix used to suppress noise are adaptively adjusted through iterative update equations.

[0027] Specifically, the filtering process includes: 1.1. Convert the state vector x Defined as a three-dimensional vector [material weight, rate of change of weight, acceleration due to gravity], it is used to dynamically depict the material flow process.

[0028] 1.2 Observation matrix H Defined as [1, 0, 0], to extract weight observations from the state vector.

[0029] 1.3 Initialization process noise covariance matrix Q and observation noise covariance matrix R are used to characterize environmental equipment interference and weighbridge sensor measurement error, respectively.

[0030] 1.4 For each raw weight observation with a timestamp collected by the weighbridge z t (Weighbridge data), perform the following iterative calculations: a. Prediction sub-step: Based on the optimal state estimate from the previous time step, predict the current state estimate and error covariance: ; in, The current state. This refers to the state at the previous moment. A This is the state transition matrix, used to describe how weight, rate of change, and acceleration evolve over time. B The control input matrix is ​​used to quantify the impact of external forces or commands on changes in material weight. m t The control input vector is used to quantify known, non-random external control or instructions applied to the system; Let the error covariance be at the current moment. The error covariance of the previous time step. A T It is the state transition matrix A The transpose of , where Q is the process noise covariance matrix; b. Update sub-step: Using the current raw weight observations, perform optimal correction on the current state estimate and error covariance: ; in, This is the updated state at the current moment. The updated error covariance at the current time. The Kalman gain is obtained based on the observation noise covariance matrix R. H For the observation matrix, I It is the identity matrix. z t These are the original weight observations.

[0031] The status updated from the current moment The extracted weight component is the filtered data.

[0032] In filtering, the process noise covariance matrix Q and the observation noise covariance matrix R are adjusted to suppress noise. By adjusting Q and R, the behavior of the filter can be controlled.

[0033] In some embodiments, data cleaning and normalization include using the interquartile range method to detect and repair outliers in the filtered data, and performing standard deviation standardization to generate a structured dataset.

[0034] Specifically, data cleaning and normalization include: Outliers in weighbridge data are identified based on the interquartile range method, and abnormal values ​​in the data are repaired by linear interpolation of data from adjacent time points. Standard deviation standardization is performed on weighbridge data from different material types and customer dimensions to eliminate dimensional differences. The formula is as follows: ; in, X 'and X These are the data before and after standardization, respectively. m and s They are respectively X The arithmetic mean and standard deviation of all values ​​in the entire feature column.

[0035] In some embodiments, time features include rolling window statistics (such as the mean and variance of weighbridge data over the past 7 days and 30 days) and periodic features including year / month / week / day, holiday identifiers, and seasonal factors; material features include classification and continuous features including material type, particle size grade, and grade indicators; customer features include customer type (long-term / temporary), historical order frequency, and purchase volume fluctuation coefficient. Multidimensional features also include derived features; these include demand volatility and customer concentration index. The formula for calculating demand volatility is: ; in, Var d For demand volatility, or i For the first i Material processing volume per unit time This represents the average material handling rate. The customer concentration index is measured by the Herfindahl-Hirschman Index, which measures the proportion of purchases made by major customers.

[0036] In some embodiments, the LSTM-Transformer hybrid demand forecasting model includes: an input layer for inputting a multi-dimensional feature matrix, a bidirectional LSTM layer for capturing long-term and short-term dependencies in time series, a Transformer layer for modeling the correlation between different feature dimensions, and a fully connected layer for outputting material demand forecasts; the loss function of the LSTM-Transformer hybrid demand forecasting model is weighted mean square error, and hierarchical time series cross-validation is used during training.

[0037] The multi-dimensional feature matrix is ​​input through the input layer; then, the long-term and short-term dependencies of the time series are captured through the bidirectional LSTM layer, and the hidden state sequence is output; the self-attention mechanism of the Transformer layer is used to model the relationship between different feature dimensions and generate context-aware feature representations; finally, the fully connected layer is mapped to the material demand forecast values ​​for the next few months.

[0038] The formula for the weighted mean squared error is as follows (taking the forecast of material demand for the next 3 months as an example): ; in, Loss For weighted mean square error, The material demand forecast for the model. This represents the actual material requirements. This is the time weighting coefficient.

[0039] During training, hierarchical time series cross-validation is used to ensure the model's generalization ability across different materials and customer dimensions; and an early stopping mechanism is introduced to avoid overfitting and monitor the mean absolute percentage error metric on the validation set.

[0040] In some embodiments, the calculation formula for the capacity demand mapping model is as follows: C p = ; in, C p This represents the total capacity demand within the planned period. T For the planned period, i An index representing a time sequence number. i = 1,2, ..., T , k An index representing the type of material. k = 1, 2, ..., m , D t+i，k For the predicted first t + i Monthly material k Material requirements, r k For materials k Processing conversion rate, or This represents the average operating efficiency of the production line.

[0041] In some embodiments, the capacity adjustment threshold rules include: triggering a capacity increase strategy when the predicted total capacity demand growth rate is greater than a first threshold; and triggering a capacity contraction strategy when the predicted total capacity demand decline rate is greater than a second threshold. For example, when the predicted total capacity demand growth rate is >15%, a capacity increase strategy is triggered: increasing shifts, activating backup production lines, or outsourcing processing; when the predicted total capacity demand decline rate is >10%, a capacity contraction strategy is implemented: reducing shifts, optimizing equipment combinations, or switching to other materials.

[0042] In some embodiments, the optimization algorithm is a genetic algorithm, and the objective function is: ; in, Minimize This indicates the ultimate goal of the entire optimization algorithm. C s To cover the costs of switching production, C i For inventory costs, C d To cover the cost of delayed delivery, α , β , c These are the weighting coefficients for production changeover costs, inventory costs, and delayed delivery costs, respectively.

[0043] In some embodiments, the present invention further includes: periodically calculating the prediction accuracy index of the LSTM-Transformer hybrid demand forecasting model and the capacity stability index of the production information management system, and incrementally training and dynamically adjusting the feature importance of the LSTM-Transformer hybrid demand forecasting model based on newly generated weighbridge data and actual production line capacity data, so as to adapt to changes in mine material structure and customer demand. Example

[0044] This embodiment uses the capacity management of a mining company as an application scenario.

[0045] Raw data collection: The weighbridge sensor (accuracy ±0.1 tons, sampling frequency 5Hz) collects the material weighing data W(t) of transport vehicles in real time, i.e., weighbridge data, and records a timestamp T(t) accurate to the second. At the same time, the collected material weighing data W(t) is associated with the material type (such as iron ore, limestone), customer ID, production line equipment status, and inventory data.

[0046] Data preprocessing: A bilateral filtering denoising algorithm is applied to the material weighing data W(t). The state-space model of this algorithm has a state vector of [material weight, rate of change of weight, weight acceleration]. Through iterative filtering via prediction and update steps, the noise covariance is adaptively adjusted to output a smoothed weight sequence W'(t). Subsequently, the interquartile range (IQR) method is used to detect and repair outliers in W'(t), and all numerical features are standardized by standard deviation (Z-score) to generate a structured dataset D.

[0047] Feature engineering and model training: Construct multi-dimensional features from dataset D, including: Time characteristics: year, month, week, day, whether it is a holiday, seasonal factors, and the mean and variance of the weighing over the past 7 days and 30 days.

[0048] Material characteristics: material type, particle size grade, grade index.

[0049] Customer characteristics: customer type (long-term / temporary), historical order frequency, and purchase volume fluctuation coefficient.

[0050] Derivative characteristics: demand volatility, customer concentration index (Herfindahl index).

[0051] These features are concatenated into a feature matrix X. Using data from the past 36 months as the training set and demand for the next 3 months as the ground truth labels, an LSTM-Transformer hybrid demand prediction model is trained. This model first learns time dependencies through bidirectional LSTM layers, then learns the correlations between features through the Transformer's self-attention mechanism, and finally outputs predicted values ​​through fully connected layers. Weighted mean squared error (WMSE) is used as the loss function, and hierarchical time series cross-validation is used to ensure the model's generalization ability.

[0052] Demand Forecasting and Capacity Planning: The latest feature data is input into the trained model to obtain the demand forecasts for each material over the next three months. The total capacity demand for the next three months is calculated using the capacity-demand mapping model formula. Then, a dynamic scheduling strategy is applied: if the predicted total capacity demand growth rate is >15%, additional shifts are triggered or standby production lines are activated; if the predicted total capacity demand decline rate is >10%, fewer shifts are reduced. Simultaneously, a genetic algorithm is used to optimize the production schedule, with the objective function of minimizing the total cost of changeovers, inventory, and delayed delivery, generating specific dynamic capacity scheduling plans (such as monthly equipment uptime and shift arrangements).

[0053] Implementation and Feedback: The generated dynamic capacity scheduling plan will be pushed to the Production Management System (MES) for execution. At the end of each month for the next three months, the system will collect actual capacity and demand data, and calculate evaluation indicators such as Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE), capacity volatility, and inventory turnover rate. If model performance deteriorates or data distribution changes, incremental model training will be triggered, and the importance of updated features will be analyzed using Shapley sum and interpretation (SHAP) values ​​to achieve continuous model optimization.

[0054] Finally, it should be noted that the above embodiments are merely preferred embodiments of the present invention used to illustrate the technical solutions of the present invention, and are not intended to limit the invention, nor are they intended to limit the patent scope of the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. These modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention. That is to say, any changes or refinements made to the main design concept and spirit of the present invention that are not of substantial significance, but whose technical problems are still consistent with the present invention, should be included within the protection scope of the present invention. In addition, the direct or indirect application of the technical solutions of the present invention to other related technical fields are similarly included within the patent protection scope of the present invention.

Claims

1. A capacity management method based on weighbridge data and machine learning algorithms, characterized in that, Includes the following steps: S1. Obtain the material weighing data sequence collected in real time by the weighbridge sensor. The material weighing data sequence is a sequence composed of timestamped weighbridge data. The material weighing data sequence is filtered by a bilateral filtering noise reduction model, and the filtered data is cleaned and normalized to generate a structured dataset. S2. Extract multi-dimensional features, including time, material, and customer dimensions, from the structured dataset and construct a multi-dimensional feature matrix; An LSTM-Transformer hybrid demand forecasting model is trained using a multi-dimensional feature matrix to output material demand forecasts for a specified future period. S3. Based on the material demand forecast, calculate the total capacity demand within the planning period through the capacity demand mapping model, and generate a dynamic scheduling scheme for production line capacity based on the preset capacity adjustment threshold rules and optimization algorithm. S4. Push the dynamic scheduling plan for production line capacity to the production information management system to drive the production line to execute.

2. The capacity management method based on weighbridge data and machine learning algorithms according to claim 1, characterized in that, Link the weighbridge data in the material weighing data sequence to the corresponding key business data, which includes material type, customer ID, production line equipment status, and inventory data.

3. The capacity management method based on weighbridge data and machine learning algorithms according to claim 1, characterized in that, The bilateral filtering noise reduction model adopts a state-space model. The state vector of the state-space model includes material weight, weight change rate and weight acceleration. The process noise covariance matrix and observation noise covariance matrix are adaptively adjusted through iterative update equations to suppress noise.

4. The capacity management method based on weighbridge data and machine learning algorithms according to claim 1, characterized in that, Data cleaning and normalization include using the interquartile range method to detect and repair outliers in the filtered data, and performing standard deviation standardization to generate a structured dataset.

5. The capacity management method based on weighbridge data and machine learning algorithms according to claim 1, characterized in that, Time characteristics include rolling window statistics and periodic characteristics including year / month / week / day, holiday identifiers, and seasonal factors; material characteristics include classification and continuous characteristics including material type, particle size grade, and grade index; customer characteristics include customer type, historical order frequency, and purchase volume fluctuation coefficient. Multidimensional features also include derived features; Derivative features include demand volatility and customer concentration index.

6. The capacity management method based on weighbridge data and machine learning algorithms according to claim 1, characterized in that, The LSTM-Transformer hybrid demand forecasting model includes: an input layer for inputting a multi-dimensional feature matrix, a bidirectional LSTM layer for capturing long-term and short-term dependencies in time series, a Transformer layer for modeling the correlation between different feature dimensions, and a fully connected layer for outputting material demand forecasts. The loss function of the LSTM-Transformer hybrid demand forecasting model is weighted mean square error, and hierarchical time series cross-validation is used during training.

7. The capacity management method based on weighbridge data and machine learning algorithms according to claim 1, characterized in that, The calculation formula for the capacity-demand mapping model is as follows: C p = ； in, C p This represents the total capacity demand within the planned period. T For the planned period, i An index representing a time sequence number. i = 1, 2,..., T , k An index representing the type of material. k = 1, 2, ..., m , D t+i，k For the predicted first t + i Monthly material k Material requirements r k For materials k Processing conversion rate, η This represents the average operating efficiency of the production line.

8. The capacity management method based on weighbridge data and machine learning algorithms according to claim 1, characterized in that, The capacity adjustment threshold rules include: when the predicted total capacity demand growth rate is greater than the first threshold, a capacity expansion strategy is triggered; when the predicted total capacity demand decline rate is greater than the second threshold, a capacity contraction strategy is triggered.

9. A capacity management method based on weighbridge data and machine learning algorithms according to claim 1, characterized in that, The optimization algorithm is a genetic algorithm, and the objective function is: ； in, Minimize This indicates the ultimate goal of the entire optimization algorithm. C s To cover the cost of switching production, C i For inventory costs, C d To cover the cost of delayed delivery, α , β , γ These are the weighting coefficients for production changeover costs, inventory costs, and delayed delivery costs, respectively.

10. A capacity management method based on weighbridge data and machine learning algorithms according to claim 1, characterized in that, Also includes: The prediction accuracy index of the LSTM-Transformer hybrid demand forecasting model and the capacity stability index of the production information management system are calculated periodically. Based on the newly generated weighbridge data and the actual capacity data of the production line, the LSTM-Transformer hybrid demand forecasting model is incrementally trained and the feature importance is dynamically adjusted.

Images