Multi-scale feature mixed production business process remaining execution time prediction method
Through the integration of multi-scale feature mixing and multiple machine learning models, the problem of insufficient flexibility and accuracy of production business process prediction methods in the existing technology is solved, and comprehensive and accurate prediction of production business processes is achieved, supporting more efficient production management.
Patent Information
- Application Number
- CN202410965424.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-18
- Publication Date
- 2024-11-19
AI Technical Summary
The existing production business process remaining time prediction methods lack flexibility and accuracy, making them difficult to adapt to complex and changeable production environments, and ignores the multi-scale characteristics of production business processes, resulting in inconsistent performance on different time scales and insufficient generalization capabilities.
Using a multi-scale feature mixing method, features from seconds to year level are extracted from the event log of the production business process, and the ARIMA, RNN, decision tree, Holt-Winters, XGBoost, CNN, LSTM, and 1D-CNN models are used for prediction. , and integrate the results of different scale models through random forest algorithms to enhance generalization and anti-noise performance.
It realizes more accurate and robust production business process remaining execution time prediction, supports more efficient production scheduling and resource management, and improves production efficiency and market response speed.
Smart Images

Figure CN118982094A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of business process management and relates to a method for predicting the remaining execution time of a business process. Background Art
[0002] In today's rapidly developing industrial environment, production efficiency and resource optimization have become key factors in improving competitiveness. The prediction of the remaining execution time of production business processes is an important part of achieving this goal. It can help enterprises better carry out production scheduling, inventory management and decision support. With the development of information technology, more and more enterprises have begun to use historical data to predict the future state of production processes. Traditional methods for predicting the remaining time of production business processes rely on experience and rules. Although this method is effective in some cases, it often lacks flexibility and accuracy. Since most of them are based on heuristic rules set by humans, these models are difficult to adapt to the complexity and variability of the production process. In addition, traditional methods find it difficult to capture nonlinear relationships and implicit patterns in the data, which limits their application in dynamic and unpredictable production environments. With the development of big data and machine learning technology, data-driven prediction methods have begun to be widely used. These methods predict future production activities by analyzing patterns and trends in historical data. Machine learning models, especially deep learning models, provide new solutions for time series prediction of production business processes due to their powerful feature extraction capabilities and modeling capabilities for complex data structures.
[0003] Although existing forecasting methods can provide valuable insights in some cases, most existing methods only focus on a single time scale, such as analyzing data at the daily or monthly level. This single-scale analysis ignores the multi-scale characteristics of production business processes. For example, at the second or minute level, the production process may exhibit completely different behavioral patterns than at the daily or weekly level. In addition, existing methods may perform well on specific datasets, but have insufficient generalization capabilities in new or different production environments. The characteristics of production business processes often manifest differently at different time scales. For example, the second scale may capture instantaneous failures of machines or rapid production fluctuations, while the monthly scale may reflect seasonal changes or long-term trends. Therefore, in order to fully understand the production process and make accurate remaining execution time predictions, a method that can integrate the characteristics of different time scales is needed to fully capture the complexity of the production business process. This method is able to process and analyze multi-scale data and extract valuable information from it. In addition, this method has good generalization capabilities, can work stably in different production environments and conditions, and adapt to the uncertainty and dynamic changes in the production process. Summary of the invention
[0004] The present invention provides a method for predicting the remaining execution time of a production business process with a multi-scale feature mixture, aiming to improve the accuracy and robustness of the prediction by comprehensively considering the features of different time scales. The method first extracts the features of the second, minute, hour, day, week, month, quarter and year levels from the event log of the production business process, and then predicts the features of each scale using ARIMA, RNN, decision tree, Holt-Winters, XGBoost, CNN, LSTM and 1D-CNN models respectively. Then, the results from the prediction models of different scales are integrated using the random forest algorithm to obtain a comprehensive and integrated prediction output. The random forest enhances the generalization ability and anti-noise performance of the model by constructing multiple decision trees and integrating their prediction results. The present invention can provide enterprises with accurate and real-time prediction of the remaining execution time of production business processes, support more efficient production scheduling and resource management, thereby improving production efficiency and market response speed.
[0005] The method for predicting the remaining execution time of a production business process with multi-scale feature mixing includes the following steps:
[0006] Step 1) Input the production business process event log, and extract second-level, minute-level, hour-level, day-level, week-level, month-level, quarter-level, and year-level features from the event log;
[0007] Step 2) Predict the remaining execution time of the production business process based on second-level features. Second-level features reflect the high-frequency changes in the production business process, which are usually sudden and random. Therefore, the Autoregressive Integrated Moving Average (ARIMA) model is introduced to capture the short-term fluctuations in the second-level features and predict the remaining execution time of the production business process in seconds. The ARIMA model combines three components: autoregressive (AR), integrated (I), and moving average (MA). Its basic form can be expressed as ARIMA (p, d, q) = AR (p) + I (d) + MA (q), where AR (p) represents the autoregressive model, I (d) represents the difference model, MA (q) represents the moving average model, p is the order of the autoregressive term, which represents the number of linear relationships between the current value and the previous value, d is the order of the difference, which is used to convert the non-stationary time series into a stationary time series, and q is the order of the moving average term, which represents the linear combination of the current prediction error and the previous error;
[0008] Step 3) Predict the remaining execution time of the production business process based on minute-level features. Minute-level features reflect changes in the production process in a shorter period of time. These changes may contain short-term trends and periodic patterns, but compared with the second-level scale, minute-level data focuses more on capturing the continuity and stability of the production process. Therefore, the recurrent neural network (RNN) is introduced to use its memory ability when processing sequence data to predict the remaining execution time of the production business process at the minute level. The basic structure of RNN can be expressed as:
[0009] h t =σ(W·h t-1 +U·x t +b)
[0010]
[0011] Among them, h t is the hidden state at time step t, x t is the input at time step t, is the predicted remaining execution time output at time step t, W, U, V are weight matrices, b and c are bias terms, and σ is the activation function ReLU;
[0012] Step 4) Predict the remaining execution time of the production business process based on hourly features. Hourly features usually show strong regularity and periodicity. For this reason, a decision tree model with intuitiveness, strong explanatory power and automatic feature selection capability is introduced to predict the remaining execution time of the production business process at the hourly level. The decision tree consists of nodes and branches. Each internal node represents a decision rule on a feature, each branch represents a result of the decision rule, and each leaf node represents the final prediction result. The decision tree construction process is as follows:
[0013] Feature selection: select an optimal feature for segmentation, which can best distinguish the data;
[0014] Split the dataset: Split the dataset into subsets based on the threshold of the selected feature;
[0015] Constructing decision nodes of the tree: recursively apply feature selection and segmentation steps on the segmented subsets to construct a decision tree;
[0016] Stopping condition: Stop the growth of the tree when a certain stopping condition is reached, such as reaching the maximum depth, all data points belonging to the same category, or the classification error is below a certain threshold;
[0017] The evaluation index of the decision tree is information gain (IG), which is used to measure the reduction in uncertainty before and after using feature segmentation data. The specific formula is:
[0018]
[0019]
[0020] Where S is the event log of the production business process, A is the hourly feature, Entropy(S) is the entropy of the event log S, indicating the uncertainty of the data set, S v is the data subset separated by the value v of feature A, |S| and |S v | respectively event log S and subset S v The number of samples in , c is the length of the event log S, and p(i) is the probability distribution of each case i in the event log;
[0021] Step 5) Predict the remaining execution time of the production business process based on daily features. The characteristics of the daily scale help to reveal the long-term trends and seasonal patterns in the production business process, which are usually affected by factors such as weekdays, weekends, and holidays. Therefore, the Holt-Winters model suitable for seasonal time series prediction is introduced to predict the day and remaining execution time of the production business process. The Holt-Winters model consists of three parts: level (L), trend (T) and seasonality (S), and the calculation methods are as follows:
[0022] S t =(1-α)·S t-s +α·(L t +T t )
[0023] T t =(1-β)·T t-1 +β·(S t +S t-s )
[0024] L t =(1-γ)·(L t-1 +T t-1 )+γ·S t
[0025] Among them, S t is the seasonal component at time t, T t is the trend component at time point t, L t is the horizontal component at time t, α, β, γ are smoothing parameters whose values are between 0 and 1, and s is the length of the seasonal cycle;
[0026] Step 6) Predict the remaining execution time of the production business process based on weekly features. The weekly scale features reflect the medium-term trends and cyclical changes in the production process, and a powerful model is needed to capture complex nonlinear relationships and long-term dependencies. In this regard, the XGBoost model is introduced to predict the remaining execution time of the production business process at the weekly level. It consists of multiple decision trees, and each tree i has a prediction function f i (x), the final prediction result is the sum of all tree prediction results:
[0027]
[0028] Where n is the number of trees. The loss function Obj(F) of the XGBoost model is as follows:
[0029]
[0030] Among them, L is the loss between the predicted value and the actual value, is the mean square error, Ω(fi) is the regularization term, including the number of leaf nodes of the tree and the L2 norm of the tree weight;
[0031] Step 7) Predict the remaining execution time of the production business process based on monthly features. At the monthly scale, the characteristics of the production business process usually show more macro trends and seasonal patterns. For this reason, a convolutional neural network (CNN) is introduced to predict the remaining time of the production business process at the monthly level. It captures local dependencies and patterns in event logs through convolution operations. The model structure of CNN is as follows:
[0032] Input layer: receives the monthly features of event logs;
[0033] Convolutional layer: Use convolution kernel sliding to extract local features on event logs:
[0034]
[0035] Among them, f is the convolution kernel, x is the input month-level feature, * represents the convolution operation, t is the time point, and k is the convolution kernel size;
[0036] Activation function: Use the ReLU function to help CNN learn more complex patterns:
[0037] Pooling layer: Reduces the spatial dimension of features, extracts the main features, and uses the maximum pooling operation to select the maximum value in the input features;
[0038] Fully connected layer: Converts the output of the convolutional layer and the pooling layer into the final prediction result:
[0039] Z=W·h+b
[0040] Among them, W is the weight matrix, b is the bias, h is the output of the previous layer, and z is the output of the fully connected layer;
[0041] Step 8) Predict the remaining execution time of the production business process based on quarterly features. Quarterly-scale data reflects the performance of the production business process over a longer time span, including seasonal changes, long-term trends, and possible cyclical patterns. To this end, a long short-term memory network (LSTM) that can learn long-term dependencies and effectively capture dynamic characteristics in time series data is introduced to predict the remaining execution time of the production business process at the quarterly level. The LSTM network consists of multiple LSTM units, each of which can contain an input gate, a forget gate, a cell state, and an output gate. These gates control the flow of information, allowing the network to remember or forget certain information, as follows:
[0042] Forget Gate:
[0043] f t =δ(W f ·[h t-1 ,x t ]+b f )
[0044] Among them, f t is the activation value of the forget gate, W f and b f is the weight and bias of the forget gate, [h t-1 ,x t ] is the concatenation of the previous hidden state and the current input, σ is the sigmoid activation function;
[0045] Input Gate:
[0046] i t =δ(W i ·[h t-1 ,x t ]+b i )
[0047]
[0048] Among them, i t is the activation value of the input gate, is the candidate cell state, W i , W C and b i 、b C are the weights and biases of the input gate and candidate cell states, respectively;
[0049] Unit status update:
[0050]
[0051] Among them, C t is the current cell state, C t-1 is the cell state at the previous time step;
[0052] Output Gate:
[0053] o t =δ(W o ·[h t-1 ,x t ]+b o )
[0054] h t =o t *tanh(C t )
[0055] Among them, t is the activation value of the output gate, h t is the current hidden state;
[0056] Step 9) Predict the remaining execution time of the production business process based on annual-level features. Time series data at the annual scale usually show obvious annual periodicity and long-term trends, which are crucial for long-term planning and strategic decision-making. To this end, a one-dimensional convolutional neural network (1D-CNN) is introduced to predict the remaining execution time of the production business process at the annual scale. It can capture local patterns and long-term dependencies in the time series through the convolution layer. 1D-CNN consists of an input layer, a convolution layer, an activation function, a pooling layer, and a fully connected layer. The important parts are described as follows:
[0057] Convolution operation:
[0058]
[0059] Where f is the convolution kernel, x is the input annual feature, * represents the convolution operation, t is the time point, and k is the convolution kernel size;
[0060] Activation function: Use ReLU function;
[0061] Pooling operation: Use the maximum pooling operation to select the maximum value in the input feature;
[0062] Fully connected layer:
[0063] Z=W·h+b
[0064] Among them, W is the weight matrix, b is the bias, h is the output of the previous layer, and z is the output of the fully connected layer;
[0065] Step 10) Combine the prediction results at different time scales to obtain a more comprehensive and accurate prediction of the remaining execution time of the production business process. For this, the random forest method is used to improve the accuracy and robustness of the prediction by building multiple decision trees and combining their results.
[0066] Further, the step 1) is composed of the following steps:
[0067] Step 1.1) Extract the second-level features of the event log. The second-level scale focuses on the real-time status and instantaneous changes of the production business process. The extracted features include:
[0068] Event interval: the time interval between two consecutive events;
[0069] Instantaneous output: the amount of production per second;
[0070] Step 1.2) Extract minute-level features of event logs. The minute-level scale provides a more macro perspective than the second-level scale and is suitable for analyzing the execution of jobs in production business processes. The extracted features include:
[0071] Productivity: average output per minute;
[0072] Job completion time: the number of minutes required to complete a job;
[0073] Quality control indicators: scrap rate or defective rate per minute;
[0074] Step 1.3) Extract hourly features of the event log. Hourly features are helpful for analyzing the production efficiency and resource consumption of work shifts in the production business process. The extracted features include:
[0075] Shift output: total output of each work shift;
[0076] Human resource utilization: productivity of employees per hour;
[0077] Step 1.4) Extract the daily features of the event log. The daily features reveal key indicators of production operations, including:
[0078] Daily output: total output per day;
[0079] Scrap rate: the proportion of scrap in daily production;
[0080] Step 1.5) Extract the weekly features of the event log. The weekly features can reflect the periodic changes in the production business process, including:
[0081] Weekly production forecast: weekly production forecast;
[0082] Demand forecasting: forecast demand for the next week based on historical data;
[0083] Step 1.6) Extract monthly features of the event log. Monthly features are used to analyze long-term trends and seasonal fluctuations in production capacity, including:
[0084] Production capacity forecast: forecast the production capacity for the next month based on historical data;
[0085] Monthly market demand forecast: forecast the market demand for the next month based on historical data;
[0086] Step 1.7) Extract quarterly features of the event log. Quarterly features are helpful for long-term planning. The extracted features include:
[0087] Quarterly market demand forecast: forecast the market demand for the next quarter based on historical data;
[0088] Long-term resource planning: forecasting resource requirements based on quarterly data;
[0089] Step 1.8) Extract annual features of event logs. Annual features are the highest level of time scale and are suitable for enterprise strategic planning and long-term decision-making. The extracted features include:
[0090] Annual production volume: the total production volume for the entire year;
[0091] Annual employee attendance; the number of attendances for each employee throughout the year;
[0092] Further, the step 2) consists of the following steps:
[0093] Step 2.1) Autoregressive component. The ARIMA model is based on autoregression, that is, the value at the current time point is related to the values at several previous time points. The autoregressive order p means that the current time point is related to the values at the previous p time points. The calculation formula is as follows:
[0094] X t =c+φ1X t-1 +φ2X t-2 +···+φ p X t-p +ε t
[0095] Among them, X t is the observed value at time point t, c is the constant term, φ is the autoregressive coefficient, ∈ t is the error term;
[0096] Step 2.2) Difference component, in order to eliminate the non-stationarity of time series data, the ARIMA model usually needs to perform a difference operation, that is, convert the original data into difference data to eliminate the influence of trends and seasonality. The difference order d represents the number of times the time series data needs to be differentiated, and the calculation formula is as follows:
[0097]
[0098] Among them, X t is the observed value at time point t, ▽ represents the difference operator, and θ is the difference coefficient;
[0099] Step 2.3) Moving average component, the ARIMA model is based on moving average, that is, the value at the current time point is related to the error of several previous time points. The moving average order q means that the error at the current time point is related to the error of the previous q time points. The calculation formula is as follows:
[0100] ε t =μ+θ1ε t-1 +θ2ε t-2 +···+θ q ε t-q
[0101] Where μ is the mean, θ is the sliding average coefficient, ∈ t is the error term;
[0102] Further, the step 10) is composed of the following steps:
[0103] Step 10.1) The components of the random forest are as follows:
[0104] Decision Tree: Random forest consists of multiple decision trees, each of which is built independently;
[0105] Training process: Each decision tree uses a different data subset during training. These subsets are obtained through bootstrap sampling. Bootstrap sampling is a sampling method with replacement. Multiple training sets are extracted from the original data set. The size of each training set is the same as the original data set.
[0106] Feature selection: At each split node of each tree, a portion of features is randomly selected to search for the best split point;
[0107] Prediction: Each decision tree will give a prediction result Final prediction result of random forest is the average of all decision tree predictions:
[0108]
[0109] Where J is the number of decision trees;
[0110] Step 10.2) Use random forest to integrate prediction results of different scales to provide more accurate and robust prediction of remaining execution time of production business processes:
[0111] Prediction result collection: Collect prediction results from the time series prediction models of different scales trained in steps 2 to 9;
[0112] Feature conversion: first adjust the prediction results of all models to the same time scale, that is, convert all prediction results into minute averages, then normalize or standardize the prediction results to eliminate the dimensional effects of features of different scales, and finally use these features as inputs to the random forest model;
[0113] Random forest training: From the collected multi-scale prediction results, select a part of the data as the training set of the random forest, and the remaining data as the test data set. Use the training set data to initialize the random forest model and start the training process. During the training process, each decision tree will independently learn how to predict the remaining execution time from the multi-scale features;
[0114] Prediction and integration: First, multi-scale feature extraction is applied to the test dataset to generate the same feature set as that used in training. These features are then input into the trained random forest model to obtain the prediction results of each decision tree. Finally, the predictions of all decision trees are combined through an ensemble method using a voting mechanism to obtain the final prediction result.
[0115] The technical concept of the present invention is: first, extract the second-level, minute-level, hour-level, day-level, week-level, month-level, quarter-level and year-level features from the event log of the production business process, and then use ARIMA, RNN, decision tree, Holt-Winters, XGBoost, CNN, LSTM, 1D-CNN models to predict the features of each scale. Then, the random forest algorithm is used to integrate the results from the prediction models of different scales to obtain a comprehensive and integrated prediction output. The random forest enhances the generalization ability and anti-noise performance of the model by constructing multiple decision trees and integrating their prediction results. The present invention can provide enterprises with accurate and real-time prediction of the remaining execution time of production business processes, support more efficient production scheduling and resource management, thereby improving production efficiency and market response speed.
[0116] The advantages of the present invention are as follows: by integrating multi-scale features from seconds to years, it is possible to comprehensively capture the micro and macro changes in the production business process and provide a more comprehensive remaining execution time prediction result; by utilizing multiple prediction models such as ARIMA, RNN, decision tree, Holt-Winters, XGBoost, CNN, LSTM, 1D-CNN, etc., it is possible to accurately capture the complex patterns and long-term dependencies in time series data; by integrating the prediction results of multiple models through the random forest algorithm, the present invention can reduce the overfitting problem that may exist in a single model and improve the stability and reliability of the prediction results. BRIEF DESCRIPTION OF THE DRAWINGS
[0117] Figure 1 It is the overall flow chart of the present invention;
[0118] Figure 2 This is an example diagram of an event log for a new drug production business process;
[0119] Figure 3 This is an example diagram of the ARIMA model;
[0120] Figure 4 This is an example diagram of the RNN model;
[0121] Figure 5 This is an example diagram of a decision tree;
[0122] Figure 6 This is an example diagram of the Holt-Winters model;
[0123] Figure 7 This is an example diagram of the XGBoost model;
[0124] Figure 8 This is an example diagram of the CNN model;
[0125] Fig. 9 This is an example diagram of the LSTM model;
[0126] Fig.10 This is an example diagram of the 1D-CNN model;
[0127] Fig.11 An example graph showing the remaining execution time prediction using a random forest algorithm to ensemble multiple models. DETAILED DESCRIPTION
[0128] The present invention will be further described below in conjunction with the accompanying drawings.
[0129] Reference Figure 1
[0130] The method for predicting the remaining execution time of a production business process with multi-scale feature mixing includes the following steps:
[0131] (1) Input the production business process event log and extract features at the second, minute, hour, day, week, month, quarter, and year levels from the event log;
[0132] (2) Prediction of the remaining execution time of production business processes based on second-level features. Second-level features reflect the high-frequency changes in production business processes, which are usually sudden and random. Therefore, the Autoregressive Integrated Moving Average (ARIMA) model is introduced to capture the short-term fluctuations in second-level features and predict the remaining execution time of production business processes in seconds. The ARIMA model combines three components: autoregressive (AR), integrated (I), and moving average (MA). Its basic form can be expressed as ARIMA(p, d, q) = AR(p) + I(d) + MA(q), where AR(p) represents the autoregressive model, I(d) represents the difference model, MA(q) represents the moving average model, p is the order of the autoregressive term, which represents the number of linear relationships between the current value and the previous value, d is the order of the difference, which is used to convert the non-stationary time series into a stationary time series, and q is the order of the moving average term, which represents the linear combination of the current prediction error and the previous error.
[0133] (3) Prediction of the remaining execution time of the production business process based on minute-level features. Minute-level features reflect changes in the production process over a shorter period of time. These changes may contain short-term trends and periodic patterns, but compared with the second-level scale, minute-level data focuses more on capturing the continuity and stability of the production process. Therefore, a recurrent neural network (RNN) is introduced to predict the remaining execution time of the production business process at the minute level by using its memory ability when processing sequence data. The basic structure of RNN can be expressed as:
[0134] h t =σ(W·h t-1 +U·x t +b)
[0135]
[0136] Among them, h t is the hidden state at time step t, x t is the input at time step t, is the predicted remaining execution time output at time step t, W, U, V are weight matrices, b and c are bias terms, and σ is the activation function ReLU;
[0137] (4) Prediction of the remaining execution time of production business processes based on hourly features. Hourly features usually show strong regularity and periodicity. Therefore, a decision tree model with intuitiveness, strong explanatory power and automatic feature selection capability is introduced to predict the remaining execution time of production business processes at the hourly level. The decision tree consists of nodes and branches. Each internal node represents a decision rule on a feature, each branch represents a result of the decision rule, and each leaf node represents the final prediction result.
[0138] (5) Prediction of the remaining execution time of the production business process based on daily features. The characteristics of the daily scale help to reveal the long-term trends and seasonal patterns in the production business process, which are usually affected by factors such as weekdays, weekends, and holidays. Therefore, the Holt-Winters model suitable for seasonal time series prediction is introduced to predict the day and remaining execution time of the production business process. The Holt-Winters model consists of three parts: level (L), trend (T) and seasonality (S), and the calculation methods are as follows:
[0139] S t =(1-α)·S t-s +α·(L t +T t )
[0140] T t =(1-β)·T t-1 +β·(S t +S t-s )
[0141] L t =(1-γ)·(L t-1 +T t-1 )+γ·S t
[0142] Among them, S t is the seasonal component at time t, T t is the trend component at time point t, L t is the horizontal component at time t, α, β, γ are smoothing parameters whose values are between 0 and 1, and s is the length of the seasonal cycle;
[0143] (6) Prediction of the remaining execution time of production business processes based on weekly features. The weekly scale features reflect the medium-term trends and cyclical changes in the production process, and a powerful model is needed to capture complex nonlinear relationships and long-term dependencies. To this end, the XGBoost model is introduced to predict the remaining execution time of production business processes at the weekly level. It consists of multiple decision trees, and each tree i has a prediction function f i (x), the final prediction result is the sum of all tree prediction results:
[0144]
[0145] Where n is the number of trees. The loss function Obj(F) of the XGBoost model is as follows:
[0146]
[0147] Among them, L is the loss between the predicted value and the actual value, is the mean square error, Ω(fi) is the regularization term, including the number of leaf nodes of the tree and the L2 norm of the tree weight;
[0148] (7) Prediction of the remaining execution time of production business processes based on monthly features. At the monthly scale, the characteristics of production business processes usually show more macro trends and seasonal patterns. For this reason, convolutional neural networks (CNNs) are introduced to predict the remaining execution time of production business processes at the monthly level. Convolutional neural networks capture local dependencies and patterns in event logs through convolution operations. The model structure of CNN is as follows:
[0149] Input layer: receives the monthly features of event logs;
[0150] Convolutional layer: Use convolution kernel sliding to extract local features on event logs:
[0151]
[0152] Among them, f is the convolution kernel, x is the input month-level feature, * represents the convolution operation, t is the time point, and k is the convolution kernel size;
[0153] Activation function: Use the ReLU function to help CNN learn more complex patterns:
[0154] Pooling layer: Reduces the spatial dimension of features, extracts the main features, and uses the maximum pooling operation to select the maximum value in the input features;
[0155] Fully connected layer: Converts the output of the convolutional layer and the pooling layer into the final prediction result:
[0156] Z=W·h+b
[0157] Among them, W is the weight matrix, b is the bias, h is the output of the previous layer, and z is the output of the fully connected layer;
[0158] (8) Prediction of the remaining execution time of production business processes based on quarterly characteristics. Quarterly-scale data reflects the performance of production business processes over a longer time span, including seasonal changes, long-term trends, and possible cyclical patterns. To this end, a long short-term memory network (LSTM) that can learn long-term dependencies and effectively capture the dynamic characteristics of time series data is introduced to predict the remaining execution time of production business processes at the quarterly level. The LSTM network consists of multiple LSTM units, each of which can contain an input gate, a forget gate, a cell state, and an output gate. These gates control the flow of information, allowing the network to remember or forget certain information, as follows:
[0159] Forget Gate:
[0160] f t =δ(W f ·[h t-1 ,x t ]+b f )
[0161] Among them, f t is the activation value of the forget gate, W f and b f is the weight and bias of the forget gate, [h t-1 ,x t ] is the concatenation of the previous hidden state and the current input, σ is the sigmoid activation function;
[0162] Input Gate:
[0163] i t =δ(W i ·[h t-1 ,x t ]+b i )
[0164]
[0165] Among them, i t is the activation value of the input gate, is the candidate cell state, W i , W C and b i 、b C are the weights and biases of the input gate and candidate cell states, respectively;
[0166] Unit status update:
[0167]
[0168] Among them, C t is the current cell state, C t-1 is the cell state at the previous time step;
[0169] Output Gate:
[0170] o t =δ(W o ·[h t-1 ,x t ]+b o )
[0171] h t =o t *tanh(C t )
[0172] Among them, t is the activation value of the output gate, h t is the current hidden state;
[0173] (9) Prediction of the remaining execution time of production business processes based on annual-level features. Time series data at the annual scale usually show obvious annual periodicity and long-term trends, which are crucial for long-term planning and strategic decision-making. To this end, a one-dimensional convolutional neural network (1D-CNN) is introduced to predict the remaining execution time of production business processes at the annual scale. It can capture local patterns and long-term dependencies in time series through convolutional layers. 1D-CNN consists of an input layer, a convolutional layer, an activation function, a pooling layer, and a fully connected layer. The important parts are described as follows:
[0174] Convolution operation:
[0175]
[0176] Where f is the convolution kernel, x is the input annual feature, * represents the convolution operation, t is the time point, and k is the convolution kernel size;
[0177] Activation function: Use ReLU function;
[0178] Pooling operation: Use the maximum pooling operation to select the maximum value in the input feature;
[0179] Fully connected layer:
[0180] Z=W·h+b
[0181] Among them, W is the weight matrix, b is the bias, h is the output of the previous layer, and z is the output of the fully connected layer;
[0182] (10) Combine the prediction results at different time scales to obtain a more comprehensive and accurate prediction of the remaining execution time of the production business process. For this purpose, the random forest method is used to improve the accuracy and robustness of the prediction by constructing multiple decision trees and combining their results.
[0183] Further, the step (1) is composed of the following steps:
[0184] (1.1) Extract the second-level features of the event log. The second-level scale focuses on the real-time status and instantaneous changes of the production business process. The extracted features include:
[0185] Event interval: the time interval between two consecutive events;
[0186] Instantaneous output: the amount of production per second;
[0187] (1.2) Extract minute-level features from event logs. The minute-level scale provides a more macroscopic perspective than the second-level scale and is suitable for analyzing the execution of jobs in production business processes. The extracted features include:
[0188] Productivity: average output per minute;
[0189] Job completion time: the number of minutes required to complete a job;
[0190] Quality control indicators: scrap rate or defective rate per minute;
[0191] (1.3) Extract hourly features from event logs. Hourly features help analyze the production efficiency and resource consumption of work shifts in production business processes. The extracted features include:
[0192] Shift output: total output of each work shift;
[0193] Human resource utilization: productivity of employees per hour;
[0194] (1.4) Extracting daily features from event logs. Daily features reveal key indicators of production operations, including:
[0195] Daily output: total output per day;
[0196] Scrap rate: the proportion of scrap in daily production;
[0197] (1.5) Extract weekly features from event logs. Weekly features can reflect the periodic changes in production business processes, including:
[0198] Weekly production forecast: weekly production forecast;
[0199] Demand forecasting: forecast demand for the next week based on historical data;
[0200] (1.6) Extract monthly features from event logs. Monthly features are used to analyze long-term trends and seasonal fluctuations in production capacity, including:
[0201] Production capacity forecast: forecast the production capacity for the next month based on historical data;
[0202] Monthly market demand forecast: forecast the market demand for the next month based on historical data;
[0203] (1.7) Extract quarterly features from event logs. Quarterly features are helpful for long-term planning. The extracted features include:
[0204] Quarterly market demand forecast: forecast the market demand for the next quarter based on historical data;
[0205] Long-term resource planning: forecasting resource requirements based on quarterly data;
[0206] (1.8) Extract annual features from event logs. Annual features are the highest level of time scale and are suitable for enterprise strategic planning and long-term decision-making. The extracted features include:
[0207] Annual production volume: the total production volume for the entire year;
[0208] Annual employee attendance; the number of attendances for each employee throughout the year;
[0209] Further, the step (2) is composed of the following steps:
[0210] (2.1) Autoregressive component. The ARIMA model is based on autoregression, that is, the value at the current time point is related to the values at several previous time points. The autoregressive order p means that the current time point is related to the values at the previous p time points. The calculation formula is as follows:
[0211] X t =c+φ1X t-1 +φ2X t-2 +···+φ p X t-p +ε t
[0212] Among them, X t is the observed value at time point t, c is the constant term, φ is the autoregressive coefficient, ∈ t is the error term;
[0213] (2.2) Difference component. In order to eliminate the non-stationarity of time series data, the ARIMA model usually needs to perform a difference operation, that is, convert the original data into difference data to eliminate the influence of trends and seasonality. The difference order d represents the number of times the time series data needs to be differentiated. The calculation formula is as follows:
[0214]
[0215] Among them, X t is the observed value at time point t, ▽ represents the difference operator, and θ is the difference coefficient;
[0216] (2.3) Moving average component. The ARIMA model is based on moving average, that is, the value at the current time point is related to the error at several previous time points. The moving average order q means that the error at the current time point is related to the error at the previous q time points. The calculation formula is as follows:
[0217] ε t =μ+θ1ε t-1 +θ2ε t-2 +···+θ q ε t-q
[0218] Where μ is the mean, θ is the sliding average coefficient, ∈ t is the error term;
[0219] Further, the step (10) is composed of the following steps:
[0220] (10.1) The components of a random forest are as follows:
[0221] Decision Tree: Random forest consists of multiple decision trees, each of which is built independently;
[0222] Training process: Each decision tree uses a different data subset during training. These subsets are obtained through bootstrap sampling. Bootstrap sampling is a sampling method with replacement. Multiple training sets are extracted from the original data set. The size of each training set is the same as the original data set.
[0223] Feature selection: At each split node of each tree, a portion of features is randomly selected to search for the best split point;
[0224] Prediction: Each decision tree will give a prediction result Final prediction result of random forest is the average of all decision tree predictions:
[0225]
[0226] Where J is the number of decision trees;
[0227] (10.2) Use random forest to integrate prediction results of different scales to provide more accurate and robust prediction of remaining execution time of production business processes:
[0228] Prediction result collection: Collect prediction results from the time series prediction models of different scales trained in steps 2 to 9;
[0229] Feature conversion: first adjust the prediction results of all models to the same time scale, that is, convert all prediction results into minute averages, then normalize or standardize the prediction results to eliminate the dimensional effects of features of different scales, and finally use these features as inputs to the random forest model;
[0230] Random forest training: From the collected multi-scale prediction results, select a part of the data as the training set of the random forest, and the remaining data as the test data set. Use the training set data to initialize the random forest model and start the training process. During the training process, each decision tree will independently learn how to predict the remaining execution time from the multi-scale features;
[0231] Prediction and integration: First, multi-scale feature extraction is applied to the test dataset to generate the same feature set as that used in training. These features are then input into the trained random forest model to obtain the prediction results of each decision tree. Finally, the predictions of all decision trees are combined through an ensemble method using a voting mechanism to obtain the final prediction result.
[0232] Attached Figure 2 This is an example of an event log for a pharmaceutical company's new drug production business process. The first line is the header of the event log, including case ID, activity name, start time, end time, and required resources. Each entry in the event log corresponds to an event instance in a production activity. For example, case ID is 1, activity name is "raw material preparation", start time is 8:00, end time is 9:00, and required resources are "raw material R1, worker W1", which is an event instance. Events with the same case ID belong to the same case. For example, the entries corresponding to case ID 1 and activity names are "raw material preparation" (A), "chemical reaction" (B), "filtration" (C), "purification" (D), "quality inspection" (E), "packaging" (F), and "final inspection" (G) belong to the same case. The features extracted for case 1 in the above event log are as follows:
[0233] Second-level features:
[0234] Event interval: the time interval between two consecutive events.
[0235] Event interval (s) = (end time - start time) / 3600
[0236] Instantaneous output: the amount of production per second, that is
[0237] Instantaneous output = output / event interval (s)
[0238] Minute-level features:
[0239] Productivity: average output per minute, i.e.
[0240] Productivity = Output / ((End Time - Start Time) / 60)
[0241] Job completion time: the number of minutes required to complete a job, i.e.
[0242] Job completion time = (end time - start time) / 60
[0243] Quality control indicators: scrap rate or defective rate per minute, that is,
[0244] Quality control index = (output 原料准备 -output 最终检验 ) / ((End time 最终检验 -Start time 原料准备 ) / 60)Hourly features:
[0245] Shift output: total output per work shift, i.e. output;
[0246] Human resource utilization: the productivity of employees per hour, i.e.
[0247] Human resource utilization = output / (end time - start time)
[0248] Day-level features:
[0249] Daily output: the total output per day, i.e. the output of the final inspection;
[0250] Scrap rate: the proportion of scrap in daily production, i.e.
[0251] Scrap rate = (output 原料准备 -output 最终检验 ) / output 原料准备
[0252] Weekly features:
[0253] Weekly production forecast: weekly production forecast, that is, weekly production forecast = output × 7;
[0254] Weekly demand forecast: forecast demand for the next week based on historical data, that is, take the average of historical weekly order volumes.
[0255] Moon Level Features:
[0256] Production capacity forecast: Based on historical data, the production capacity for the next month is forecasted. A month is calculated as 30 days, that is,
[0257] Production capacity forecast = output × 30
[0258] Monthly market demand forecast: forecast the market demand for the next month based on historical data, that is, monthly market demand forecast = (weekly order volume / 7) × 30
[0259] Quarter Level Features:
[0260] Quarterly market demand forecast: forecast the market demand for the next quarter based on historical data, i.e.
[0261] Quarterly market demand forecast = (weekly order volume / 7) × 30 × 3
[0262] Long-term resource planning: forecasting resource requirements based on quarterly data, i.e.
[0263] Long-term resource planning = output 原料准备 ×30×3
[0264] Year-level features:
[0265] Annual production volume: the total production volume for the entire year, that is, the sum of the output of the final inspection activities every day throughout the year;
[0266] Annual employee attendance; the attendance number of each employee throughout the year, that is, the attendance number of each employee in the execution resources every day throughout the year.
[0267] Attached Figure 3 This is an example diagram of the ARIMA model modeling process. Step 1: Take the event log of the production business process as input data. These data are usually in time series format, recording key events and time points in the production process. Step 2: Perform a stationarity test on the input event log. The stationarity test method is ADF (Augmented Dickey-Fuller), and the purpose of the test is to determine whether the event log has constant statistical characteristics. Step 3: If the stationarity test shows that the data is non-stationary, perform a difference operation on the data. The difference operation aims to eliminate the trend and seasonal components in the time series and make it reach a stationary state. The order of difference (d) is the minimum number of differences required to make the series stationary. Step 4: Re-test the stationarity of the time series after difference to confirm whether it is stationary. Step 5: If the time series is stationary, perform a white noise test to determine whether the series is a random process. If the white noise test is passed, it means that the model has fully captured the statistical characteristics of the data and the modeling can be ended. Step 6: If the white noise test fails, p (the order of the autoregressive term) and q (the order of the moving average term) need to be estimated together with the ARIMA model. Step 7: After determining the parameters of the ARIMA model, re-test the model residuals for white noise. If the residuals are white noise, it means that the model is able to fit the data well. Step 8: Once the white noise test is passed, the ARIMA model is established. Use this model to predict the remaining execution time of production business processes in seconds.
[0268] Attached Figure 4 This is an example diagram of the RNN model. The hidden state at time t is as follows:
[0269] h t =f(Ux t +Wh t-1 +b)
[0270] Among them, h t 、h t-1 represents the hidden state at the current time t and the previous time t-1, f is the activation function ReLU, x t represents the event log input at the current time t, U and W represent the model parameters, and b is the bias term. The output at time t is as follows:
[0271] o t =Vh t +c
[0272] Among them, t represents the model output at the current time t, h t represents the hidden state at the current time t, V represents the model parameters, c is the bias term, and the predicted output at time t is as follows:
[0273] y t =f(o t )
[0274] Among them, function f represents the regression function. The loss function L at time t t as follows:
[0275]
[0276] in, is the predicted remaining execution time, y t Represents the remaining execution time label at the current moment. The overall loss function L of the RNN model is as follows:
[0277]
[0278] Where T is the number of cases in the event log.
[0279] Attached Figure 5 This is an example diagram of a decision tree. The decision tree construction process is as follows:
[0280] Feature selection: select an optimal feature for segmentation, which can best distinguish the data;
[0281] Split the dataset: Split the dataset into subsets based on the threshold of the selected feature;
[0282] Constructing decision nodes of the tree: recursively apply feature selection and segmentation steps on the segmented subsets to construct a decision tree;
[0283] Stopping condition: Stop the growth of the tree when a certain stopping condition is reached, such as reaching the maximum depth, all data points belonging to the same category, or the classification error is below a certain threshold;
[0284] The evaluation index of the decision tree is information gain (IG), which is used to measure the reduction in uncertainty before and after using feature segmentation data. The specific formula is:
[0285]
[0286] Where S is the event log of the production business process, A is the hourly feature, Entropy(S) is the entropy of the event log S, indicating the uncertainty of the data set, S v is the data subset separated by the value v of feature A, |S| and |S v | respectively event log S and subset S v The number of samples in , c is the length of the event log S, and p(i) is the probability distribution of each case i in the event log;
[0287] Attached Figure 6 This is an example diagram of the Holt-Winters model. Step 1: Use the event log of the production business process as input data. This data is usually in time series format, recording key events and time points in the production process. Step 2: Determine the initial baseline value L0, the initial trend value B0, and the initial seasonal factor S0. Step 3: Select three smoothing parameters α, β, γ, which control the degree of smoothing of the baseline, trend, and seasonal factors respectively. Step 4: For each new data point Y in the time series t , iteratively calculated according to the following equation:
[0288] Baseline update equation: L t =α·(Y t -S t-m )+(1-α)·(L t-1 +B t-1 )
[0289] Trend update equation: B t =β·(L t -L t-1 )+(1-β)·B t-1
[0290] Seasonal factor update equation: S t =γ·(Y t -L t-1 -B t-1)+(1-γ)·S t-m
[0291] Here, m is the length of the seasonal cycle and t is the current time point or observation point.
[0292] Step 5: Use the updated parameters to predict future values. The prediction formula depends on whether you are using an additive or multiplicative model:
[0293] Additive Model:
[0294] Multiplicative Model:
[0295] where h is the lead step size for prediction.
[0296] Finally, the performance of the model is evaluated by comparing the predicted values with the actual values. Indicators such as mean square error (MSE) or root mean square error (RMSE) can be used. Depending on the results of the model evaluation, it may be necessary to return to step 2 to adjust the smoothing parameters and re-iterate the calculation to improve the prediction performance of the model.
[0297] Attached Figure 7 This is an example diagram of the XGBoost model. The XGBoost model process mainly has the following key steps:
[0298] Initialization: Set the initial prediction value, usually 0 or the mean value of the target variable.
[0299] Iterative training: Enter a loop and build a decision tree each time.
[0300] Calculate residuals: For each data point, calculate the residual between the predicted value and the actual value.
[0301] Build Tree: Use the residual information to build a new decision tree.
[0302] Update the model: Multiply the new tree by the learning rate and add it to the existing model to update the model's predictions.
[0303] Performance evaluation: Evaluate the performance of the model on the validation set.
[0304] Early stopping method: If the performance of the model does not improve in consecutive iterations, the iteration is stopped.
[0305] Output prediction: Output the remaining execution time prediction result of the final model.
[0306] Attached Figure 8 This is an example diagram of a CNN model. The simplified CNN model is as follows:
[0307] Input layer: This is the starting point of the model, taking event logs as input data.
[0308] Convolutional layer: Performs a convolution operation, using a set of learnable convolution kernels (or filters) to slide over the input data to extract features. A convolutional layer can have multiple convolution kernels, each of which can extract different features of the input data. Convolution operation:
[0309] C(i,j)=(I*K)(i,j)=∑ m ∑ n I(im,jn)·K(m,n)
[0310] Among them, (i,j) is the position of the output feature map. (m,n) is the moving position of the convolution kernel.
[0311] Activation layer: Immediately following the convolutional layer, a ReLU (Rectified Linear Unit) activation function is usually used, which introduces nonlinearity and allows the network to learn complex features.
[0312] Pooling layer: Perform downsampling operations to reduce the spatial size of the data (for example, using maximum pooling or average pooling). This helps reduce the amount of computation, prevents overfitting, and extracts more abstract features. The mathematical expression of the pooling operation is as follows:
[0313] P(i,j)=max m,n∈window I(im,jn)
[0314] Among them, P(i,j) is the value of the output feature map at position (i,j), I is the input feature map, m and n are the relative position indexes within the window, and window refers to a local area defined on I for pooling operation.
[0315] Fully connected layer: Usually located after the convolution and pooling layers, it expands the features from the spatial dimension into a one-dimensional feature vector. The fully connected layer can learn the complex relationship between features.
[0316] Output layer: The last layer of the model, which outputs the remaining execution time prediction result of the final model.
[0317] Attached Fig. 9 This is an example diagram of the LSTM model. A brief description and core idea of the LSTM network are as follows:
[0318] Forget Gate: The forget gate is responsible for deciding which information is no longer needed and should be forgotten from the cell state. It makes decisions by comparing the current input with the hidden state of the previous time step.
[0319] Input Gate: The input gate consists of two parts: a sigmoid layer that determines which values will be updated, and a tanh layer that creates a vector of new candidate values that will be added to the state to update the cell state. The mathematical expression is as follows:
[0320] i t =σ(W i ·[h t-1 ,x t ]+b f )
[0321]
[0322] Among them, x t is the input at time step t, h t-1 is the hidden state at the previous time step, i t is the output value of the input gate, is the candidate cell state, W and b are weight and bias parameters corresponding to different gating mechanisms, σ is usually a sigmoid function, which is used to output a value between 0 and 1, indicating the degree of opening of the gate, and tanh is a hyperbolic tangent function, which is used to output a value between -1 and 1, indicating the degree of activation of the state.
[0323] Updating the cell state: The cell state is the memory part of the LSTM network, which carries information about the observed input sequence. The cell state is updated at each time step by discarding old information through the forget gate and adding new information through the input gate.
[0324] Output Gate: The output gate determines what the output of the next hidden state should be. It determines what information is useful based on the current cell state and input, and passes this information to the next time step. Its mathematical expression is as follows:
[0325] o t =σ(W o ·[h t-1 ,x t ]+b o )
[0326] Among them, t is the output value of the output gate.
[0327] Current hidden state: The current hidden state is calculated based on the decision of the output gate and the updated cell state. This state will be passed to the next time step, or used for the output of the current time step.
[0328] Attached Fig.10 This is an example diagram of a 1D-CNN model. The 1D-CNN model is mainly composed of the following parts:
[0329] Input Layer: The input layer is designed to receive raw data sequences.
[0330] Convolutional layer: applies a one-dimensional convolution kernel to extract local features in the data.
[0331] y t =f(∑ k w k ·x t-k +b)
[0332] Among them, f is the activation function, w k is the convolution kernel weight and b is the bias.
[0333] Activation layer: ReLU activation function is usually used to introduce nonlinearity.
[0334] Pooling layer (optional): Use pooling operations to reduce feature dimensions and extract key information.
[0335] Fully connected layers: pass and combine features from convolutional and pooling layers to prepare for the final prediction.
[0336] Since the research problem is a regression task, the evaluation indicators can be MSE, RMSE or MAE. Taking MSE as an example, its calculation method is:
[0337]
[0338] Where N is the number of samples, y i is the true value, is the predicted value.
[0339] Attached Fig.11This is an example diagram of using the random forest algorithm to integrate multiple models for the remaining execution time prediction. First, the event log is input into multiple models such as ARIMA, RNN, decision tree, Holt-Winters, XGBoost, CNN, LSTM and 1D-CNN to learn the features of different time scales and train them, and collect their prediction results. Then, the prediction results of all models are uniformly adjusted to the same time scale, specifically by converting them into minute averages. In order to eliminate the dimensional influence of features of different scales, these prediction results are further normalized or standardized; then, a part of the data collected from the multi-scale prediction results is selected to form the training set of the random forest. The random forest model is initialized with these training set data and trained. During the training process, each decision tree in the forest independently learns how to use these multi-scale features to predict the remaining execution time; finally, after the model training is completed, multi-scale feature extraction is performed on the new test data set to generate a set of features that are the same as the training stage. These features are input into the trained random forest model, and each decision tree will give its prediction results. Through the ensemble method of the voting mechanism, the predictions of all decision trees are combined to obtain the final prediction result. This process improves the accuracy and robustness of the prediction by integrating the prediction results of multiple models.
[0340] The contents described in the embodiments of this specification are merely an enumeration of the implementation forms of the inventive concept. The protection scope of the present invention should not be regarded as limited to the specific forms described in the embodiments. The protection scope of the present invention also extends to the equivalent technical means that can be conceived by those skilled in the art based on the inventive concept.
Claims
1. A method for predicting the remaining execution time of a production business process based on a multi-scale feature mixture, characterized in that: The following steps are involved: 1) Input the production business process event log and extract the second-level, minute-level, hour-level, day-level, week-level, month-level, quarter-level and year-level features from the event log; 2) Prediction of the remaining execution time of production business processes based on second-level features: The autoregressive integrated moving average ARIMA model is introduced to capture short-term fluctuations in second-level features and predict the remaining execution time of production business processes in seconds. The ARIMA model combines three components: autoregressive AR, differential I, and moving average MA. Its basic form is expressed as ARIMA(p, d, q) = AR(p) + I(d) + MA(q), where AR(p) represents the autoregressive model, I(d) represents the differential model, MA(q) represents the moving average model, p is the order of the autoregressive term, which represents the number of linear relationships between the current value and the previous value, d is the order of the differential, which is used to convert the non-stationary time series into a stationary time series, and q is the order of the moving average term, which represents the linear combination of the current prediction error and the previous error. 3) Prediction of the remaining execution time of production business processes based on minute-level features: Introducing the recursive neural network (RNN) and using its memory ability when processing sequence data to predict the remaining execution time of production business processes at the minute level; The basic structure of RNN is expressed as: h t =σ(W·h t-1 +U·x t +b) Among them, h t is the hidden state at time step t, x t is the input at time step t, is the predicted remaining execution time output at time step t, W, U, V are weight matrices, b and c are bias terms, and σ is the activation function ReLU; 4) Prediction of the remaining execution time of production business processes based on hourly features: A decision tree model is introduced to predict the remaining execution time of production business processes at the hourly level. The decision tree consists of nodes and branches. Each internal node represents a decision rule on a feature, each branch represents a result of the decision rule, and each leaf node represents the final prediction result. The decision tree construction process is as follows: Feature selection: select an optimal feature for segmentation, which can best distinguish the data; Split the dataset: Split the dataset into subsets based on the threshold of the selected feature; Constructing decision nodes of the tree: recursively apply feature selection and segmentation steps on the segmented subsets to construct a decision tree; Stop condition: Stop the growth of the tree when a certain stop condition is reached; The evaluation index of the decision tree is information gain IG, which is used to measure the reduction in uncertainty before and after using feature segmentation data. The specific formula is: Where S is the event log of the production business process, A is the hourly feature, Entropy(S) is the entropy of the event log S, indicating the uncertainty of the data set, S v is the data subset separated by the value v of feature A, |S| and |S v | respectively event log S and subset S v The number of samples in , c is the length of the event log S, p(i) is the probability distribution of each case i in the event log, and Values(A) is the total number of subsets obtained after segmentation according to feature A; 5) Prediction of the remaining execution time of production business processes based on daily characteristics: The Holt-Winters model is introduced to predict the daily and remaining execution time of production business processes; the Holt-Winters model consists of three parts: level L, trend T and seasonality S, and the calculation methods are as follows: S t =(1-α)·S t-s +α·(L t +T t ) T t =(1-β)·T t-1 +β·(S t +S t-s ) L t =(1-γ)·(L t-1 +T t-1 )+γ·S t Among them, S t is the seasonal component at time t, T t is the trend component at time point t, L t is the horizontal component at time point t, α, β, and γ are smoothing parameters, and their values are between 0 and 1; 6) Prediction of the remaining execution time of production business processes based on weekly features: The XGBoost model is introduced to predict the remaining execution time of production business processes at the weekly level. It consists of multiple decision trees, and each tree i has a prediction function f i (x), the final prediction result is the sum of all tree prediction results: Where n is the number of trees; The loss function Obj(F) of the XGBoost model is as follows: Among them, L is the loss between the predicted value and the actual value, is the mean square error, Ω(fi) is the regularization term, including the number of leaf nodes in the tree and the L2 norm of the tree weight, y i is the true value of the i-th sample; 7) Prediction of remaining execution time of production business process based on monthly features: Convolutional neural network (CNN) is introduced to predict the remaining execution time of production business process at the monthly level. It captures local dependencies and patterns in event logs through convolution operations. The model structure of CNN is as follows: Input layer: receives the monthly features of event logs; Convolutional layer: Use convolution kernel sliding to extract local features on event logs: Among them, f is the convolution kernel, x is the input month-level feature, * represents the convolution operation, t is the time point, k is the convolution kernel size, and f k represents a convolution kernel of size k; Activation function: Use the ReLU function to help CNN learn more complex patterns: Pooling layer: Reduces the spatial dimension of features, extracts the main features, and uses the maximum pooling operation to select the maximum value in the input features; Fully connected layer: Converts the output of the convolutional layer and the pooling layer into the final prediction result: Z=W·h+b Among them, W is the weight matrix, b is the bias, h is the output of the previous layer, and z is the output of the fully connected layer; 8) Prediction of the remaining execution time of production business processes based on quarterly features: Long short-term memory network LSTM is introduced to predict the remaining execution time of production business processes at the quarterly level; the LSTM network consists of multiple LSTM units, each of which contains an input gate, a forget gate, a cell state, and an output gate; these gates control the flow of information, allowing the network to remember or forget certain information, as follows: Forget Gate: f t =δ(W f ·[h t-1 ,x t ]+b f ) Among them, f t is the activation value of the forget gate, W f and b f is the weight and bias of the forget gate, [h t-1 ,x t ] is the concatenation of the previous hidden state and the current input, σ is the sigmoid activation function; Input Gate: i t =δ(W i ·[h t-1 ,x t ]+b i ) Among them, i t is the activation value of the input gate, is the candidate cell state, W i , W C and b i 、b C are the weights and biases of the input gate and candidate cell states, respectively; Unit status update: Among them, C t is the current cell state, C t-1 is the cell state of the previous time step, * represents the element-by-element multiplication operation; Output Gate: o t =δ(W o ·[h t-1 ,x t ]+b o ) h t =o t *tanh(C t ) Among them, t is the activation value of the output gate, h t is the current hidden state, W o and b o is the weight matrix and bias term of the output gate; 9) Prediction of the remaining execution time of production business processes based on annual features: A one-dimensional convolutional neural network 1D-CNN is introduced to predict the remaining execution time of production business processes on an annual scale. The convolutional layer can capture local patterns and long-term dependencies in the time series. The 1D-CNN consists of an input layer, a convolutional layer, an activation function, a pooling layer, and a fully connected layer, as follows: Convolution operation: Where f is the convolution kernel, x is the input annual feature, * represents the convolution operation, t is the time point, and k is the convolution kernel size; Activation function: Use ReLU function; Pooling operation: Use the maximum pooling operation to select the maximum value in the input feature; Fully connected layer: Z=W·h+b Among them, W is the weight matrix, b is the bias, h is the output of the previous layer, and z is the output of the fully connected layer; 10) Combine the prediction results at different time scales to obtain a more comprehensive and accurate prediction of the remaining execution time of the production business process. For this, the random forest method is used to improve the accuracy and robustness of the prediction by building multiple decision trees and combining their results.
2. The method for predicting the remaining execution time of a production business process with multi-scale feature mixing according to claim 1, characterized in that: Step 1) comprises the following steps: 1.1) Extract the second-level features of the event log. The extracted features include: Event interval: the time interval between two consecutive events; Instantaneous output: the amount of production per second; 1.2) Extract minute-level features of event logs. The extracted features include: Productivity: average output per minute; Job completion time: the number of minutes required to complete a job; Quality control indicators: scrap rate or defective rate per minute; 1.3) Extract hourly features of event logs. The extracted features include: Shift output: total output of each work shift; Human resource utilization: productivity of employees per hour; 1.4) Extract the day-level features of the event log. The extracted features include: Daily output: total output per day; Scrap rate: the proportion of scrap in daily production; 1.5) Extract weekly features of event logs. The extracted features include: Weekly production forecast: weekly production forecast; Demand forecasting: forecast demand for the next week based on historical data; 1.6) Extract monthly features of event logs. The extracted features include: Production capacity forecast: forecast the production capacity for the next month based on historical data; Monthly market demand forecast: forecast the market demand for the next month based on historical data; 1.7) Extract quarterly features of event logs. The extracted features include: Quarterly market demand forecast: forecast the market demand for the next quarter based on historical data; Long-term resource planning: forecasting resource requirements based on quarterly data; 1.8) Extract annual features of event logs. The extracted features include: Annual production volume: the total production volume for the entire year; Annual employee attendance; the number of attendances for each employee throughout the year.
3. The method for predicting the remaining execution time of a production business process with multi-scale feature mixing according to claim 1, characterized in that: Step 2) comprises the following steps: 2.1) Autoregressive component. The ARIMA model is based on autoregression, that is, the value at the current time point is related to the values at several previous time points; the autoregressive order p means that the current time point is related to the values at the previous p time points. The calculation formula is as follows: X t =c+φ1X t-1 +φ2X t-2 +···+φ p X t-p +e t Among them, X t is the observed value at time point t, c is the constant term, φ is the autoregressive coefficient, ∈ t is the error term; 2.2) Difference component. In order to eliminate the non-stationarity of time series data, the ARIMA model usually needs to perform a difference operation, that is, convert the original data into difference data to eliminate the influence of trends and seasonality. The difference order d represents the number of times the time series data needs to be differentiated. The calculation formula is as follows: Among them, X t is the observed value at time point t, represents the difference operator, θ is the difference coefficient; 2.3) Moving average component. The ARIMA model is based on moving average, that is, the value at the current time point is related to the error of several previous time points; the moving average order q means that the error at the current time point is related to the error of the previous q time points. The calculation formula is as follows: e t =μ+θ1ε t-1 +θ2ε t-2 +···+θ q e t-q Where μ is the mean, θ is the sliding average coefficient, ∈ t is the error term.
4. The method for predicting the remaining execution time of a production business process with multi-scale feature mixing according to claim 1, characterized in that: Step 10) The following steps are involved: 10.1) The components of a random forest are as follows: Decision Tree: Random forest consists of multiple decision trees, each of which is built independently; Training process: Each decision tree uses a different data subset during training. These subsets are obtained through bootstrap sampling. Bootstrap sampling is a sampling method with replacement. Multiple training sets are extracted from the original data set. The size of each training set is the same as the original data set. Feature selection: At each split node of each tree, a portion of features is randomly selected to search for the best split point; Prediction: Each decision tree will give a prediction result Final prediction result of random forest is the average of all decision tree predictions: in, J is the number of decision trees; 10.2) Use random forest to integrate prediction results of different scales to provide more accurate and robust prediction of remaining execution time of production business processes: Prediction result collection: collect prediction results from the time series prediction models of different scales trained in step 2)-step 9); Feature conversion: first adjust the prediction results of all models to the same time scale, that is, convert all prediction results into minute averages, then normalize or standardize the prediction results to eliminate the dimensional effects of features of different scales, and finally use these features as inputs to the random forest model; Random forest training: From the collected multi-scale prediction results, select a part of the data as the training set of the random forest, and the remaining data as the test data set. Use the training set data to initialize the random forest model and start the training process. During the training process, each decision tree will independently learn how to predict the remaining execution time from the multi-scale features; Prediction and integration: First, multi-scale feature extraction is applied to the test dataset to generate the same feature set as that used in training. These features are then input into the trained random forest model to obtain the prediction results of each decision tree. Finally, the predictions of all decision trees are combined through an ensemble method using a voting mechanism to obtain the final prediction result.
Citation Information
Cited By
Basin flood disaster early prediction method based on machine learning
CN120409822A
Electroencephalogram signal processing method and system based on motion artifact prediction
CN120549516A
