Intelligent production scheduling method and system based on multi-model fusion

Abstract

The invention provides an intelligent production scheduling method and system based on multi-model fusion in the technical field of machine yield prediction, and the method comprises the steps: S1, obtaining a large amount of historical operation data, at least including the number, date, operation duration, rotating speed and density of a machine, of the machine, and carrying out the preprocessing of each historical operation data, and constructing a data set; s2, based on the multi-layer perceptron, a gradient lifting regression module, a decision tree module and a fusion output module, creating a machine operation duration prediction model; s3, training the machine operation time length prediction model through the data set; and S4, predicting the operation duration of each machine through the trained machine operation duration prediction model to obtain a prediction result, obtaining work order information, machine information and production scheduling priority of each machine, and performing intelligent production scheduling based on the prediction result, the work order information, the machine information and the production scheduling priority. The method has the advantages that the accuracy and robustness of machine production scheduling are greatly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of machine output prediction, and in particular to an intelligent production scheduling method and system based on multi-model fusion. Background Art

[0002] With the rapid development of intelligent technology, machines (mechanical equipment used for processing and manufacturing) generate a large amount of production data during operation, such as machine equipment ID, machine operation time, machine speed, product density, etc. In order to better adapt to the production needs of enterprises and improve their production efficiency, the need to use intelligent scheduling technology to schedule the production of machines has emerged. Intelligent scheduling technology is a tool that uses artificial intelligence, algorithm optimization and data analysis technology to solve task allocation and resource scheduling problems in complex production environments. Traditional intelligent scheduling technologies include:

[0003] 1. Rule-driven intelligent scheduling technology, which uses predefined rules and logic to optimize production task allocation and resource scheduling, is relatively basic but still plays an important role in many practical applications, especially in production scenarios with limited resources, simple constraints or clear rules.

[0004] 2. Using heuristic rules and experience to optimize production scheduling is widely used in intelligent production scheduling scenarios in actual production due to its high computational efficiency and strong flexibility, especially in multi-constraint and multi-objective optimization problems. For example, the greedy algorithm is an algorithm design strategy based on local optimal selection. In the process of problem solving, it always chooses the option that seems to be the best in the current situation (i.e., greedy selection), and tries to reach the global optimum through a series of local optimal decisions.

[0005] 3. Mathematical programming methods, by establishing mathematical models and using solvers to optimize production scheduling problems, such as linear programming (LP), integer programming (IP), and mixed integer linear programming (MILP); mathematical programming methods are widely used in the field of intelligent production scheduling, especially in the fields of manufacturing, logistics and supply chain management, energy scheduling, etc.; for example, in the manufacturing field, mathematical programming methods can be used to optimize the allocation and processing sequence of tasks in the production workshop, solve the flexible job shop scheduling problem (FJSP), and determine the production time and machine allocation of different batches of products through batch production scheduling.

[0006] 4. Artificial neural network (ANN) and deep learning (DL) technology, that is, to achieve production scheduling optimization through training models in solving complex nonlinear problems, optimization and prediction in dynamic environments, etc.; for example, in production scheduling optimization, features can be extracted from a large amount of historical data to optimize task allocation and resource scheduling, and convolutional neural networks (CNN) can be used to extract production equipment status features, combined with fully connected layers to optimize task scheduling, and combined with deep reinforcement learning (DRL) to optimize multi-stage production scheduling.

[0007] 5. Reinforcement Learning (RL), with its ability of self-learning and optimization in dynamic and uncertain environments, is used to solve complex production scheduling problems, such as multi-objective optimization, real-time scheduling, resource allocation, etc.; for example, the Multi-Agent Reinforcement Learning (MARL) method uses multiple agents to collaborate or compete to complete production scheduling tasks. In terms of distributed production scheduling, each agent represents different production units and collaborates to achieve the global optimum, and is used for task allocation and resource coordination among multiple factories to achieve resource sharing optimization.

[0008] However, traditional intelligent production scheduling technology usually relies on empirical data and fixed rules, and makes estimates by analyzing historical operating time data (such as mean, median or distribution characteristics). However, it has poor adaptability to data fluctuations. In a high-mix, low-batch production mode, it is difficult to cope with complex and changing production environments, such as equipment status fluctuations, changes in material properties, and the need for urgent orders to be inserted into production. Deep learning-based prediction methods include time series models, deep neural networks, and multimodal fusion models. Although they have strong modeling capabilities, they require a large amount of data support, have high training costs and are time-consuming. In actual enterprise applications, due to the limited dimension of data features, deep learning models are difficult to fully learn effective features, which in turn affects the accuracy and robustness of production scheduling.

[0009] Therefore, how to provide an intelligent production scheduling method and system based on multi-model fusion to improve the accuracy and robustness of machine production scheduling has become a technical problem that needs to be solved urgently. Summary of the invention

[0010] The technical problem to be solved by the present invention is to provide an intelligent production scheduling method and system based on multi-model fusion, so as to improve the accuracy and robustness of machine production scheduling.

[0011] In a first aspect, the present invention provides an intelligent production scheduling method based on multi-model fusion, comprising the following steps:

[0012] Step S1, obtaining a large amount of historical operation data of the machine, including at least the machine number, date, operation time, rotation speed and density, converting the format of each of the historical operation data from the JSON format to the DataFrame format, performing an outlier removal operation on each of the historical operation data after the format conversion, and then converting the encoding format of the machine number in each of the historical operation data into One-Hot encoding, performing a mean-standard deviation standardization operation on the rotation speed and density, thereby completing the preprocessing of each of the historical operation data and constructing a data set;

[0013] Step S2, creating a machine operation time prediction model based on a multi-layer perceptron, a gradient boosting regression module, a decision tree module and a fusion output module;

[0014] The loss function of the machine operation time prediction model is:

[0015]

[0016] Where L represents the loss value of the loss function of the machine running time prediction model; α j represents a learnable parameter; L 1 Represents the loss function of the multi-layer perceptron, L 2 represents the loss function of the gradient boosting regression module, L 3 Represents the loss function of the decision tree module, L 1 , L 2 , L 3 All are mean square error functions;

[0017] Step S3, training a machine operation time prediction model using the data set;

[0018] Step S4: Use the trained machine operating time prediction model to predict the operating time of each machine to obtain the operating time prediction result, obtain the work order information, machine information and production scheduling priority of each machine, and perform intelligent production scheduling based on the operating time prediction result, work order information, machine information and production scheduling priority.

[0019] Furthermore, in step S1, the formula for the mean-standard deviation standardization operation is:

[0020]

[0021] in, Indicates the speed after standardized operation; Z 1 Indicates the speed before standardization operation; u 1 Represents the average value of the speed; σ 1 represents the standard deviation of the rotation speed; represents the density after standardization; Z 2 represents the density before standardization; u 2 represents the average value of density; σ 2 Represents the standard deviation of the density.

[0022] Furthermore, in step S2, the multilayer perceptron includes an input layer, a plurality of hidden layers and an output layer, and each of the input layer, the hidden layer and the output layer is composed of a plurality of stacked fully connected layers; the input layer is used to input operation data, the hidden layer is used to perform feature operation on the operation data to obtain a first prediction result, and the output layer is used to output the first prediction result;

[0023] The formula of the fully connected layer is:

[0024] X (l) =(W (l) X (l-1) +b (l) );

[0025] Among them, X (l) represents the output of the lth fully connected layer; X (l-1) represents the output of the l-1th fully connected layer, and is also the input of the lth fully connected layer. The input of the 0th fully connected layer is the running data; W (l) represents the weight matrix of the lth fully connected layer; b (l) Represents the bias vector of the lth fully connected layer;

[0026] The gradient boosting regression module constructs a weak prediction model and iterates the weak prediction model. During the iteration, a new weak prediction model is fitted according to the negative gradient direction of the loss function of the gradient boosting regression module, and the prediction values ​​of all the iterated weak prediction models are weighted summed to obtain a second prediction result;

[0027] The iterative formula of the weak prediction model is:

[0028] F m (x) = F m-1 (x)+v·h m (x);

[0029] Among them, F m (x) represents the second prediction result of the mth iteration; F m-1 (x) represents the second prediction result of the m-1th iteration; v represents the learning rate; h m (x) represents the predicted value of the mth weak prediction model; x represents the operating data;

[0030] The formula for the negative ladder direction is:

[0031] r m =y i -F m-1 (x i );

[0032] Among them, r m Represents the residual, that is, the negative gradient direction; yi represents the true value corresponding to the i-th running data; F m-1 (x i ) represents the i-th running data x i Input the second prediction result obtained by the gradient boosting regression module after m-1 iterations;

[0033] The calculation formula of the second prediction result is:

[0034]

[0035] Wherein, F(x) represents the second prediction result; M represents the total number of weak prediction models;

[0036] The decision tree module splits the features in the input operation data based on information gain, and then performs regression analysis to output a third prediction result; the features include machine number, date, operation time, rotation speed and density;

[0037] The calculation formula of the information gain is:

[0038]

[0039] Where IG represents information gain; H() represents entropy function; S represents the data set containing running data; S i represents the i-th subset in the data set; n represents the total number of subsets; m~ represents the total number of feature types; p k Represents the probability of the k-th feature in S.

[0040] Furthermore, in step S2, the fusion output module is used to perform weighted fusion on the first prediction result, the second prediction result and the third prediction result output by the multilayer perceptron, the gradient boosting regression module and the decision tree module, and output the running time prediction result, and the formula is:

[0041]

[0042] Among them, prediction represents the running time prediction result; w j Represents weight; model_predict 1 Indicates the first prediction result; model_predict 2 Indicates the second prediction result; model_predict 3 Represents the third prediction result.

[0043] Furthermore, the step S3 is specifically as follows:

[0044] Perform a ripple addition operation on the speed and density in the dataset:

[0045]

[0046] in, Indicates the speed after adding fluctuations; represents the speed before adding fluctuations; Δ[Z 1 ] i Indicates the i-th speed fluctuation value, and its value range is (-1,1); represents the density after adding fluctuations; represents the density before adding fluctuations; Δ[Z 2 ] i Indicates the i-th density fluctuation value, ranging from (-20, 20);

[0047] Training a machine operation time prediction model using the data set after the fluctuation addition operation;

[0048] The step S4 is specifically as follows:

[0049] The running time prediction model of the machine after training is used to perform real-time prediction of the running time of each machine to obtain the running time prediction result, and obtain the work order information, machine information and production scheduling priority of each machine;

[0050] The work order information at least includes the expected completion time; the machine information at least includes the number of idle machines, the idle machine numbers and the total production output; the production scheduling priority is idle machine priority, occupied machine priority, short remaining time for the machine head priority, high output priority or short expected occupancy time priority;

[0051] The estimated total output is calculated based on the running time prediction results, work order information, machine information and production scheduling priority, and intelligent production scheduling is performed based on the estimated total output and the scheduled total output.

[0052] In a second aspect, the present invention provides an intelligent production scheduling system based on multi-model fusion, comprising the following modules:

[0053] A data set construction module is used to obtain a large amount of historical operation data of the machine, including at least the machine number, date, operation time, rotation speed and density, convert the format of each of the historical operation data from JSON format to DataFrame format, perform an outlier removal operation on each of the historical operation data after format conversion, and then convert the encoding format of the machine number in each of the historical operation data into One-Hot encoding, perform a mean-standard deviation standardization operation on the rotation speed and density, and then complete the preprocessing of each of the historical operation data and construct a data set;

[0054] A machine operation time prediction model creation module is used to create a machine operation time prediction model based on a multi-layer perceptron, a gradient boosting regression module, a decision tree module, and a fusion output module;

[0055] The loss function of the machine operation time prediction model is:

[0056]

[0057] Where L represents the loss value of the loss function of the machine running time prediction model; α j represents a learnable parameter; L 1 Represents the loss function of the multi-layer perceptron, L 2 represents the loss function of the gradient boosting regression module, L 3 Represents the loss function of the decision tree module, L 1 , L 2 , L 3 All are mean square error functions;

[0058] A machine operation time prediction model training module is used to train the machine operation time prediction model using the data set;

[0059] The intelligent production scheduling module is used to predict the operating time of each machine through the trained machine operating time prediction model to obtain the operating time prediction result, obtain the work order information, machine information and production scheduling priority of each machine, and perform intelligent production scheduling based on the operating time prediction result, work order information, machine information and production scheduling priority.

[0060] Furthermore, in the data set construction module, the formula for the mean-standard deviation standardization operation is:

[0061]

[0062] in, Indicates the speed after standardized operation; Z 1 Indicates the speed before standardization operation; u 1 Represents the average value of the speed; σ 1 Indicates the standard deviation of the speed; Z 2 scaled represents the density after standardization; Z 2 represents the density before standardization; u 2 represents the average value of density; σ 2 Represents the standard deviation of the density.

[0063] Furthermore, in the machine operation time prediction model creation module, the multilayer perceptron includes an input layer, a plurality of hidden layers and an output layer, and each of the input layer, the hidden layer and the output layer is composed of a plurality of fully connected layers stacked; the input layer is used for inputting operation data, the hidden layer is used for performing feature operation on the operation data to obtain a first prediction result, and the output layer is used for outputting the first prediction result;

[0064] The formula of the fully connected layer is:

[0065] X (l) =(W (l) X (l-1) +b (l) );

[0066] Among them, X (l) represents the output of the lth fully connected layer; X (l-1) represents the output of the l-1th fully connected layer, and is also the input of the lth fully connected layer. The input of the 0th fully connected layer is the running data; W (l) represents the weight matrix of the lth fully connected layer; b (l) Represents the bias vector of the lth fully connected layer;

[0067] The gradient boosting regression module constructs a weak prediction model and iterates the weak prediction model. During the iteration, a new weak prediction model is fitted according to the negative gradient direction of the loss function of the gradient boosting regression module, and the prediction values ​​of all the iterated weak prediction models are weighted summed to obtain a second prediction result;

[0068] The iterative formula of the weak prediction model is:

[0069] F m (x) = F m-1 (x)+v·h m (x);

[0070] Among them, F m (x) represents the second prediction result of the mth iteration; F m-1 (x) represents the second prediction result of the m-1th iteration; v represents the learning rate; h m (x) represents the predicted value of the mth weak prediction model; x represents the operating data;

[0071] The formula for the negative ladder direction is:

[0072] r m =y i -F m-1 (x i );

[0073] Among them, r mRepresents the residual, that is, the negative gradient direction; y i represents the true value corresponding to the i-th running data; F m-1 (x i ) represents the i-th running data x i Input the second prediction result obtained by the gradient boosting regression module after m-1 iterations;

[0074] The calculation formula of the second prediction result is:

[0075]

[0076] Wherein, F(x) represents the second prediction result; M represents the total number of weak prediction models;

[0077] The decision tree module splits the features in the input operation data based on information gain, and then performs regression analysis to output a third prediction result; the features include machine number, date, operation time, rotation speed and density;

[0078] The calculation formula of the information gain is:

[0079]

[0080] Where IG represents information gain; H() represents entropy function; S represents the data set containing running data; S i represents the i-th subset in the data set; n represents the total number of subsets; m~ represents the total number of feature types; p k Represents the probability of the k-th feature in S.

[0081] Furthermore, in the machine operation time prediction model creation module, the fusion output module is used to weightedly fuse the first prediction result, the second prediction result and the third prediction result output by the multilayer perceptron, the gradient boosting regression module and the decision tree module, and output the operation time prediction result, and the formula is:

[0082]

[0083] Among them, prediction represents the running time prediction result; w j Represents weight; model_predict 1 Indicates the first prediction result; model_predict 2 Indicates the second prediction result; model_predict 3 Represents the third prediction result.

[0084] Furthermore, the machine operation time prediction model training module is specifically used for:

[0085] Perform a ripple addition operation on the speed and density in the dataset:

[0086]

[0087] in, Indicates the speed after adding fluctuations; represents the speed before adding fluctuations; Δ[Z 1 ] i Indicates the i-th speed fluctuation value, and its value range is (-1,1); represents the density after adding fluctuations; represents the density before adding fluctuations; Δ[Z 2 ] i Indicates the i-th density fluctuation value, ranging from (-20, 20);

[0088] Training a machine operation time prediction model using the data set after the fluctuation addition operation;

[0089] The intelligent production scheduling module is specifically used for:

[0090] The running time prediction model of the machine after training is used to perform real-time prediction of the running time of each machine to obtain the running time prediction result, and obtain the work order information, machine information and production scheduling priority of each machine;

[0091] The work order information at least includes the expected completion time; the machine information at least includes the number of idle machines, the idle machine numbers and the total production output; the production scheduling priority is idle machine priority, occupied machine priority, short remaining time for the machine head priority, high output priority or short expected occupancy time priority;

[0092] The estimated total output is calculated based on the running time prediction results, work order information, machine information and production scheduling priority, and intelligent production scheduling is performed based on the estimated total output and the scheduled total output.

[0093] The advantages of the present invention are:

[0094] 1. Obtain a large amount of historical operation data of the machine, including at least the machine number, date, operation time, rotation speed and density, convert the format of each historical operation data from JSON format to DataFrame format, perform outlier removal operation on each historical operation data after format conversion, and then convert the encoding format of the machine number in each historical operation data into One-Hot encoding, perform mean-standard deviation standardization operation on the rotation speed and density, and then complete the preprocessing of each historical operation data and construct a data set; then create a machine operation time prediction model based on a multi-layer perceptron, a gradient boosting regression module, a decision tree module and a fusion output module, and train the machine operation time prediction model through the data set; finally, predict the operation time of each machine through the trained machine operation time prediction model to obtain the operation time prediction result, obtain the work order information, machine information and scheduling priority of each machine, and perform intelligent scheduling based on the operation time prediction result, work order information, machine information and scheduling priority; multi-layer perceptron, gradient boosting regression module, decision tree module and fusion output module, and train the machine operation time prediction model through the data set; finally, predict the operation time of each machine through the trained machine operation time prediction model, obtain the work order information, machine information and scheduling priority of each machine, and perform intelligent scheduling based on the operation time prediction result, work order information, machine information and scheduling priority; The perceptron is used to perform feature operations on the operating data to obtain the first prediction result; the gradient boosting regression module is used to construct a weak prediction model and iterate it. During the iteration process, a new weak prediction model is fitted according to the negative gradient direction of the loss function, and the prediction values ​​of all iterated weak prediction models are weighted and summed to obtain the second prediction result; the decision tree module is used to split the features in the input operating data according to the information gain, and then perform regression analysis to output the third prediction result; the fusion output module is used to weightedly fuse the first prediction result, the second prediction result and the third prediction result to output the operating time prediction result; that is, by fusing the multi-layer perceptron (MLP), the gradient boosting regression module (GB), and the decision tree module (DT) to predict the operating time of the machine, effectively combining the advantages of the fusion multi-layer perceptron, the gradient boosting regression module and the decision tree module, the accuracy and robustness of the operating time prediction are improved, and then the operating time prediction results, work order information, machine information and production scheduling priority are combined for intelligent scheduling, which ultimately greatly improves the accuracy and robustness of machine scheduling.

[0095] 2. By performing fluctuation addition operations on the rotation speed and density in the data set before training the machine operation time prediction model, the possible fluctuations in the machine production process (such as changes in the machine operating status, fluctuations in production speed, etc.) are simulated. Such fluctuations are added to the data set, making the operation time prediction results closer to the actual situation, enhancing the stability and reliability of the operation time prediction results, and further improving the accuracy and robustness of machine scheduling.

[0096] 3. By setting the production scheduling priority to idle machine priority, occupied machine priority, short remaining time for disk head priority, high output priority or short estimated occupancy time priority, and combining the running time prediction results, work order information, machine information and production scheduling priority, intelligent production scheduling is performed. That is, a greedy algorithm is used to achieve intelligent production scheduling according to the customized production scheduling priority. Under the premise of maximizing resource utilization and production efficiency, the production scheduling process is efficient and reasonable. In addition, intelligent production scheduling can effectively reduce machine waiting time, balance production load, and significantly improve production efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0097] The present invention will be further described below in conjunction with embodiments with reference to the accompanying drawings.

[0098] Figure 1 It is a flow chart of an intelligent production scheduling method based on multi-model fusion of the present invention.

[0099] Figure 2 It is a structural schematic diagram of an intelligent production scheduling system based on multi-model fusion in the present invention. DETAILED DESCRIPTION

[0100] The technical solution in the embodiments of the present application has the following overall idea: the operation time of the machine is predicted by integrating the machine operation time prediction model of the multi-layer perceptron, the gradient boosting regression module and the decision tree module, and the advantages of the multi-layer perceptron, the gradient boosting regression module and the decision tree module are effectively combined to improve the accuracy and robustness of the operation time prediction, and then the operation time prediction results, work order information, machine information and production scheduling priority are combined to perform intelligent scheduling, thereby improving the accuracy and robustness of machine scheduling.

[0101] Please refer to Figure 1 to Figure 2 As shown, a preferred embodiment of the intelligent production scheduling method based on multi-model fusion of the present invention includes the following steps:

[0102] Step S1, obtaining a large amount of historical operation data of the machine, including at least the machine number, date, operation time, speed and density, converting the format of each of the historical operation data from JSON format to DataFrame format for analysis and processing, performing an outlier removal operation on each of the historical operation data after format conversion to improve data quality, and then converting the encoding format of the machine number in each of the historical operation data to One-Hot encoding, performing a mean-standard deviation standardization operation on the speed and density to eliminate the dimensional differences between different features, thereby completing the preprocessing of each of the historical operation data and constructing a data set;

[0103] One-Hot encoding is an encoding method that converts categorical data (categorical variables) into numerical form. It is widely used in machine learning and deep learning. The core idea is to map each category into a one-hot vector (One-Hot Vector), in which only one position is 1 and the rest are 0. The advantages of using One-Hot encoding are that it can eliminate the size relationship between categories, adapt to machine learning algorithms, avoid information loss, and has strong flexibility, so that the model can process classified machine numbers and facilitate distinction during model training.

[0104] Step S2, creating a machine operation time prediction model based on a multi-layer perceptron (MLP), a gradient boosting regression module (GB), a decision tree module (DT) and a fusion output module;

[0105] The loss function of the machine operation time prediction model is:

[0106]

[0107] Where L represents the loss value of the loss function of the machine running time prediction model; α j represents a learnable parameter; L 1 Represents the loss function of the multi-layer perceptron, L 2 represents the loss function of the gradient boosting regression module, L 3 Represents the loss function of the decision tree module, L 1 , L 2 , L 3 All are mean square error functions;

[0108] The prediction of machine operation time is of great significance in modern manufacturing, especially in the era of intelligent manufacturing and Industry 4.0. The demand behind it mainly comes from optimizing production efficiency, reducing operating costs and improving equipment utilization. In the production process, the machine operation time not only affects the formulation of production plans, but also directly relates to the realization of production goals and the accuracy of order delivery time. The accurate prediction of machine operation time can not only improve the scientific nature of production scheduling, but also help enterprises optimize equipment maintenance plans and reduce unplanned downtime, while providing important support for intelligent scheduling systems and dynamic capacity management.

[0109] The formula of the mean square error function is:

[0110]

[0111] Among them, y i Indicates actual value; represents the average predicted value;

[0112] Step S3, training a machine operation time prediction model using the data set;

[0113] Step S4: Use the trained machine operating time prediction model to predict the operating time of each machine to obtain the operating time prediction result, obtain the work order information, machine information and production scheduling priority of each machine, and perform intelligent production scheduling based on the operating time prediction result, work order information, machine information and production scheduling priority.

[0114] Under the condition of a limited number of features, the present invention makes predictions by integrating the advantages of different machine learning models (multi-layer perceptron, gradient boosting regression module, decision tree module), which effectively improves the prediction performance; the present invention also designs a scheme to simulate the fluctuation of machine operation (perform fluctuation addition operation), which can truly simulate the fluctuation characteristics in the actual production environment, thereby enhancing the applicability and robustness of the machine operation time prediction model; in terms of intelligent production scheduling, the present invention supports the greedy algorithm of user-defined production scheduling priority, so as to flexibly adjust the production scheduling strategy to adapt to different production needs. In summary, the present invention not only significantly improves the accuracy and robustness of machine operation time prediction, but also provides efficient decision support for intelligent production by optimizing the production scheduling plan.

[0115] In step S1, the formula for the mean-standard deviation standardization operation is:

[0116]

[0117] in, Indicates the speed after standardized operation; Z 1 Indicates the speed before standardization operation; u 1 Represents the average value of the speed; σ 1 represents the standard deviation of the rotation speed; represents the density after standardization; Z 2 represents the density before standardization; u 2 represents the average value of density; σ 2 Represents the standard deviation of the density.

[0118] In step S2, the multilayer perceptron includes an input layer, a plurality of hidden layers, and an output layer, and each of the input layer, the hidden layer, and the output layer is composed of a plurality of fully connected layers stacked; the input layer is used for inputting operation data, the hidden layer is used for performing feature operation on the operation data to obtain a first prediction result, and the output layer is used for outputting the first prediction result;

[0119] The formula of the fully connected layer is:

[0120] X(l) =(W (l) X (l-1) +b (l) );

[0121] Among them, X (l) represents the output of the lth fully connected layer; X (l-1) represents the output of the l-1th fully connected layer, and is also the input of the lth fully connected layer. The input of the 0th fully connected layer is the running data; W (l) represents the weight matrix of the lth fully connected layer; b (l) Represents the bias vector of the lth fully connected layer;

[0122] The gradient boosting regression module constructs a weak prediction model and iterates the weak prediction model. During the iteration, a new weak prediction model is fitted according to the negative gradient direction of the loss function of the gradient boosting regression module, and the prediction values ​​of all iterated weak prediction models are weighted summed to obtain a second prediction result; the weak prediction model is preferably a decision tree model;

[0123] The iterative formula of the weak prediction model is:

[0124] F m (x) = F m-1 (x)+v·h m (x);

[0125] Among them, F m (x) represents the second prediction result of the mth iteration; F m-1 (x) represents the second prediction result of the m-1th iteration; v represents the learning rate (Step Size), which is used to control the contribution of each tree; h m (x) represents the predicted value of the mth weak prediction model; x represents the operating data;

[0126] The formula for the negative ladder direction is:

[0127] r m =y i -F m-1 (x i );

[0128] Among them, r m Represents the residual, that is, the negative gradient direction; y i represents the true value corresponding to the i-th running data; F m-1 (x i ) represents the i-th running data x i Input the second prediction result obtained by the gradient boosting regression module after m-1 iterations;

[0129] The calculation formula of the second prediction result is:

[0130]

[0131] Wherein, F(x) represents the second prediction result; M represents the total number of weak prediction models;

[0132] The decision tree module splits the features in the input operation data based on information gain, and then performs regression analysis to output the third prediction result; the features include machine number, date, operation time, speed and density; the decision tree is a tree-structured machine learning model used for classification and regression tasks. Its basic idea is to divide the data set into smaller subsets, and the corresponding decision trees are gradually constructed; the decision tree is split based on maximizing information gain, which measures the degree of reduction of uncertainty before and after the division;

[0133] The calculation formula of the information gain is:

[0134]

[0135] Where IG represents information gain; H() represents entropy function, H(S) represents entropy before splitting; S represents the data set containing running data; S i represents the i-th subset in the data set, H(S i ) represents the entropy of each subset; n represents the total number of subsets; Indicates the total number of feature types; p k Represents the probability of the k-th feature in S. By calculating the difference in entropy before and after the data set is divided, the feature with the largest information gain is selected for division, thereby reducing the confusion of the data.

[0136] In step S2, the fusion output module is used to perform weighted fusion on the first prediction result, the second prediction result and the third prediction result output by the multilayer perceptron, the gradient boosting regression module and the decision tree module, and output the running time prediction result, and the formula is:

[0137]

[0138] Among them, prediction represents the running time prediction result; w j Represents weight; model_predict 1 Indicates the first prediction result; model_predict 2 Indicates the second prediction result; model_predict 3 Represents the third prediction result.

[0139]

[0140] in, It represents the prediction result of the j-th model (multilayer perceptron, gradient boosting regression module or decision tree module) after the i-th fluctuation of the running data (feature); n represents the number of simulated fluctuations of the running data.

[0141] The step S3 is specifically as follows:

[0142] Perform a ripple addition operation on the speed and density in the dataset:

[0143]

[0144] in, Indicates the speed after adding fluctuations; represents the speed before adding fluctuations; Δ[Z 1 ] i Indicates the i-th speed fluctuation value, and its value range is (-1,1); represents the density after adding fluctuations; represents the density before adding fluctuations; Δ[Z 2 ] i Indicates the i-th density fluctuation value, ranging from (-20, 20);

[0145] Training a machine operation time prediction model using the data set after the fluctuation addition operation;

[0146] Since the input density and rotation speed are user-defined, they are usually set to integers; however, in the actual production process, due to individual differences in machines, the statistical density and rotation speed often differ from the set values, which makes there a certain deviation between the actual production environment and the theoretical setting; in order to simulate the fluctuations in these actual environments, the present invention proposes a simple and effective data perturbation processing method, so that the disturbed data can more truly reflect the actual production situation and improve the prediction effect in actual production.

[0147] That is, firstly, the speed and density are processed for fluctuations, and the characteristic data after fluctuations are input into the model for training; since the result of a single fluctuation simulation may deviate greatly from the actual situation, the result is input into the fusion output module after n iterations in the specific implementation; the model is tuned through continuous iterative training. During the training stage, the model gradually adjusts the parameters through the back propagation and gradient update mechanism, and learns the influence of characteristic fluctuations on the running time of the machine.

[0148] The step S4 is specifically as follows:

[0149] The running time prediction model of the machine after training is used to perform real-time prediction of the running time of each machine to obtain the running time prediction result, and obtain the work order information, machine information and production scheduling priority of each machine;

[0150] The work order information at least includes expected completion time, average rotation speed, average density, single width, planned output, gram weight and enterprise credit code; the machine information at least includes machine name, machine status, door width, expected occupancy time, remaining time of head, number of idle machines, idle machine number and total production output; the production scheduling priority is idle machine priority, occupied machine priority, short remaining time of head, high output priority or short expected occupancy time priority;

[0151] The estimated total output is calculated based on the running time prediction results, work order information, machine information and production scheduling priority, and intelligent production scheduling is performed based on the estimated total output and the scheduled total output, that is, production scheduling is performed if the estimated total output is less than the scheduled total output.

[0152] Since the machine may be affected by various factors during the actual production process, such as equipment failure, operating errors, etc., which may cause the expected completion time to deviate from the actual situation, it is necessary to monitor the operation status of the machine in real time and make dynamic adjustments based on the current production progress and historical data. The present invention can accurately correct the expected completion time by integrating the runtime prediction results with the actual production status, thereby improving the accuracy and reliability of the production plan.

[0153] In the process of intelligent production scheduling, the machine operation time prediction model will predict the daily operation time of the machine according to different strategies, and calculate the predicted daily output of each machine by combining the work order information and machine information. Then, the user can customize the production scheduling priority as an input parameter in the production scheduling process, for example, setting the production scheduling priority as: idle machine > occupied machine > machine with short estimated occupation time > machine with short remaining time of head. First, determine whether the idle machine can complete the expected production scheduling demand: if it can, directly arrange the idle machine for production scheduling; if not, consider using the occupied machine for production scheduling; when considering the occupied machine, first consider whether the machine with short estimated occupation time can complete the production scheduling; if not, then consider whether the machine with short remaining time of head can complete the production scheduling. The production scheduling priority plan of the occupied machine can be customized according to user needs. If the total output is expected to reach the expected output, the machine will be arranged for production scheduling; otherwise, it will prompt that the production scheduling target cannot be completed under the current conditions. After the production scheduling is completed, the generated machine data and prediction results are returned to the front end to provide users with detailed production scheduling plans and related information.

[0154] In specific implementation, it also supports real-time data processing and feedback mechanisms. For additional information about the machine, such as the estimated completion time and the remaining time of the equipment, the production schedule can be adjusted dynamically in real time based on these data. At the same time, during the forecasting process, if the transmission data is found to be abnormal (such as inconsistent equipment status, missing data, etc.), an error message will be returned and the process will be terminated. This exception handling mechanism enhances the robustness of the system.

[0155] A preferred embodiment of an intelligent production scheduling system based on multi-model fusion of the present invention includes the following modules:

[0156] A data set construction module is used to obtain a large amount of historical operation data of the machine, including at least the machine number, date, operation time, speed and density, convert the format of each of the historical operation data from JSON format to DataFrame format for analysis and processing, perform an outlier removal operation on each of the historical operation data after format conversion to improve data quality, and then convert the encoding format of the machine number in each of the historical operation data to One-Hot encoding, perform a mean-standard deviation standardization operation on the speed and density to eliminate the dimensional differences between different features, thereby completing the preprocessing of each of the historical operation data and constructing a data set;

[0157] One-Hot encoding is an encoding method that converts categorical data (categorical variables) into numerical form. It is widely used in machine learning and deep learning. The core idea is to map each category into a one-hot vector (One-Hot Vector), in which only one position is 1 and the rest are 0. The advantages of using One-Hot encoding are that it can eliminate the size relationship between categories, adapt to machine learning algorithms, avoid information loss, and has strong flexibility, so that the model can process classified machine numbers and facilitate distinction during model training.

[0158] A machine operation time prediction model creation module is used to create a machine operation time prediction model based on a multi-layer perceptron (MLP), a gradient boosting regression module (GB), a decision tree module (DT) and a fusion output module;

[0159] The loss function of the machine operation time prediction model is:

[0160]

[0161] Where L represents the loss value of the loss function of the machine running time prediction model; α j represents a learnable parameter; L 1 Represents the loss function of the multi-layer perceptron, L 2 represents the loss function of the gradient boosting regression module, L3 Represents the loss function of the decision tree module, L 1 , L 2 , L 3 All are mean square error functions;

[0162] The prediction of machine operation time is of great significance in modern manufacturing, especially in the era of intelligent manufacturing and Industry 4.0. The demand behind it mainly comes from optimizing production efficiency, reducing operating costs and improving equipment utilization. In the production process, the machine operation time not only affects the formulation of production plans, but also directly relates to the realization of production goals and the accuracy of order delivery time. The accurate prediction of machine operation time can not only improve the scientific nature of production scheduling, but also help enterprises optimize equipment maintenance plans and reduce unplanned downtime, while providing important support for intelligent scheduling systems and dynamic capacity management.

[0163] The formula of the mean square error function is:

[0164]

[0165] Among them, y i Indicates actual value; represents the average predicted value;

[0166] A machine operation time prediction model training module is used to train the machine operation time prediction model using the data set;

[0167] The intelligent production scheduling module is used to predict the operating time of each machine through the trained machine operating time prediction model to obtain the operating time prediction result, obtain the work order information, machine information and production scheduling priority of each machine, and perform intelligent production scheduling based on the operating time prediction result, work order information, machine information and production scheduling priority.

[0168] Under the condition of a limited number of features, the present invention makes predictions by integrating the advantages of different machine learning models (multi-layer perceptron, gradient boosting regression module, decision tree module), which effectively improves the prediction performance; the present invention also designs a scheme to simulate the fluctuation of machine operation (perform fluctuation addition operation), which can truly simulate the fluctuation characteristics in the actual production environment, thereby enhancing the applicability and robustness of the machine operation time prediction model; in terms of intelligent production scheduling, the present invention supports the greedy algorithm of user-defined production scheduling priority, so as to flexibly adjust the production scheduling strategy to adapt to different production needs. In summary, the present invention not only significantly improves the accuracy and robustness of machine operation time prediction, but also provides efficient decision support for intelligent production by optimizing the production scheduling plan.

[0169] In the data set construction module, the formula for the mean-standard deviation standardization operation is:

[0170]

[0171] in, Indicates the speed after standardized operation; Z 1 Indicates the speed before standardization operation; u 1 Represents the average value of the speed; σ 1 represents the standard deviation of the rotation speed; represents the density after standardization; Z 2 represents the density before standardization; u 2 represents the average value of density; σ 2 Represents the standard deviation of the density.

[0172] In the machine operation time prediction model creation module, the multilayer perceptron includes an input layer, a plurality of hidden layers and an output layer, and each of the input layer, the hidden layer and the output layer is composed of a plurality of fully connected layers stacked; the input layer is used for inputting operation data, the hidden layer is used for performing feature operation on the operation data to obtain a first prediction result, and the output layer is used for outputting the first prediction result;

[0173] The formula of the fully connected layer is:

[0174] X (l) =(W (l) X (l-1) +b (l) );

[0175] Among them, X (l) represents the output of the lth fully connected layer; X (l-1) represents the output of the l-1th fully connected layer, and is also the input of the lth fully connected layer. The input of the 0th fully connected layer is the running data; W (l) represents the weight matrix of the lth fully connected layer; b (l) Represents the bias vector of the lth fully connected layer;

[0176] The gradient boosting regression module constructs a weak prediction model and iterates the weak prediction model. During the iteration, a new weak prediction model is fitted according to the negative gradient direction of the loss function of the gradient boosting regression module, and the prediction values ​​of all iterated weak prediction models are weighted summed to obtain a second prediction result; the weak prediction model is preferably a decision tree model;

[0177] The iterative formula of the weak prediction model is:

[0178] F m (x) = F m-1(x)+v·h m (x);

[0179] Among them, F m (x) represents the second prediction result of the mth iteration; F m-1 (x) represents the second prediction result of the m-1th iteration; v represents the learning rate (Step Size), which is used to control the contribution of each tree; h m (x) represents the predicted value of the mth weak prediction model; x represents the operating data;

[0180] The formula for the negative ladder direction is:

[0181] r m =y i -F m-1 (x i );

[0182] Among them, r m Represents the residual, that is, the negative gradient direction; y i represents the true value corresponding to the i-th running data; F m-1 (x i ) represents the i-th running data x i Input the second prediction result obtained by the gradient boosting regression module after m-1 iterations;

[0183] The calculation formula of the second prediction result is:

[0184]

[0185] Wherein, F(x) represents the second prediction result; M represents the total number of weak prediction models;

[0186] The decision tree module splits the features in the input operation data based on information gain, and then performs regression analysis to output the third prediction result; the features include machine number, date, operation time, speed and density; the decision tree is a tree-structured machine learning model used for classification and regression tasks. Its basic idea is to divide the data set into smaller subsets, and the corresponding decision trees are gradually constructed; the decision tree is split based on maximizing information gain, which measures the degree of reduction of uncertainty before and after the division;

[0187] The calculation formula of the information gain is:

[0188]

[0189] Where IG represents information gain; H() represents entropy function, H(S) represents entropy before splitting; S represents the data set containing running data; S irepresents the i-th subset in the data set, H(S i ) represents the entropy of each subset; n represents the total number of subsets; Indicates the total number of feature types; p k Represents the probability of the k-th feature in S. By calculating the difference in entropy before and after the data set is divided, the feature with the largest information gain is selected for division, thereby reducing the confusion of the data.

[0190] In the machine operation time prediction model creation module, the fusion output module is used to perform weighted fusion on the first prediction result, the second prediction result and the third prediction result output by the multilayer perceptron, the gradient boosting regression module and the decision tree module, and output the operation time prediction result, and the formula is:

[0191]

[0192] Among them, prediction represents the running time prediction result; w j Represents weight; model_predict 1 Indicates the first prediction result; model_predict 2 Indicates the second prediction result; model_predict 3 Represents the third prediction result.

[0193]

[0194] in, It represents the prediction result of the j-th model (multilayer perceptron, gradient boosting regression module or decision tree module) after the i-th fluctuation of the running data (feature); n represents the number of simulated fluctuations of the running data.

[0195] The machine operation time prediction model training module is specifically used for:

[0196] Perform a ripple addition operation on the speed and density in the dataset:

[0197]

[0198] in, Indicates the speed after adding fluctuations; represents the speed before adding fluctuations; Δ[Z 1 ] i Indicates the i-th speed fluctuation value, and its value range is (-1,1); represents the density after adding fluctuations; represents the density before adding fluctuations; Δ[Z 2 ] i Indicates the i-th density fluctuation value, ranging from (-20, 20);

[0199] Training a machine operation time prediction model using the data set after the fluctuation addition operation;

[0200] Since the input density and rotation speed are user-defined, they are usually set to integers; however, in the actual production process, due to individual differences in machines, the statistical density and rotation speed often differ from the set values, which makes there a certain deviation between the actual production environment and the theoretical setting; in order to simulate the fluctuations in these actual environments, the present invention proposes a simple and effective data perturbation processing method, so that the disturbed data can more truly reflect the actual production situation and improve the prediction effect in actual production.

[0201] That is, firstly, the speed and density are processed for fluctuations, and the characteristic data after fluctuations are input into the model for training; since the result of a single fluctuation simulation may deviate greatly from the actual situation, the result is input into the fusion output module after n iterations in the specific implementation; the model is tuned through continuous iterative training. During the training stage, the model gradually adjusts the parameters through the back propagation and gradient update mechanism, and learns the influence of characteristic fluctuations on the running time of the machine.

[0202] The intelligent production scheduling module is specifically used for:

[0203] The running time prediction model of the machine after training is used to perform real-time prediction of the running time of each machine to obtain the running time prediction result, and obtain the work order information, machine information and production scheduling priority of each machine;

[0204] The work order information at least includes expected completion time, average rotation speed, average density, single width, planned output, gram weight and enterprise credit code; the machine information at least includes machine name, machine status, door width, expected occupancy time, remaining time of head, number of idle machines, idle machine number and total production output; the production scheduling priority is idle machine priority, occupied machine priority, short remaining time of head, high output priority or short expected occupancy time priority;

[0205] The estimated total output is calculated based on the running time prediction results, work order information, machine information and production scheduling priority, and intelligent production scheduling is performed based on the estimated total output and the scheduled total output, that is, production scheduling is performed if the estimated total output is less than the scheduled total output.

[0206] Since the machine may be affected by various factors during the actual production process, such as equipment failure, operating errors, etc., which may cause the expected completion time to deviate from the actual situation, it is necessary to monitor the operation status of the machine in real time and make dynamic adjustments based on the current production progress and historical data. The present invention can accurately correct the expected completion time by integrating the runtime prediction results with the actual production status, thereby improving the accuracy and reliability of the production plan.

[0207] In the process of intelligent production scheduling, the machine operation time prediction model will predict the daily operation time of the machine according to different strategies, and calculate the predicted daily output of each machine by combining the work order information and machine information. Then, the user can customize the production scheduling priority as an input parameter in the production scheduling process, for example, setting the production scheduling priority as: idle machine > occupied machine > machine with short estimated occupation time > machine with short remaining time of head. First, determine whether the idle machine can complete the expected production scheduling demand: if it can, directly arrange the idle machine for production scheduling; if not, consider using the occupied machine for production scheduling; when considering the occupied machine, first consider whether the machine with short estimated occupation time can complete the production scheduling; if not, then consider whether the machine with short remaining time of head can complete the production scheduling. The production scheduling priority plan of the occupied machine can be customized according to user needs. If the total output is expected to reach the expected output, the machine will be arranged for production scheduling; otherwise, it will prompt that the production scheduling target cannot be completed under the current conditions. After the production scheduling is completed, the generated machine data and prediction results are returned to the front end to provide users with detailed production scheduling plans and related information.

[0208] In specific implementation, it also supports real-time data processing and feedback mechanisms. For additional information about the machine, such as the estimated completion time and the remaining time of the equipment, the production schedule can be adjusted dynamically in real time based on these data. At the same time, during the forecasting process, if the transmission data is found to be abnormal (such as inconsistent equipment status, missing data, etc.), an error message will be returned and the process will be terminated. This exception handling mechanism enhances the robustness of the system.

[0209] In summary, the advantages of the present invention are:

[0210] 1. Obtain a large amount of historical operation data of the machine, including at least the machine number, date, operation time, rotation speed and density, convert the format of each historical operation data from JSON format to DataFrame format, perform outlier removal operation on each historical operation data after format conversion, and then convert the encoding format of the machine number in each historical operation data into One-Hot encoding, perform mean-standard deviation standardization operation on the rotation speed and density, and then complete the preprocessing of each historical operation data and construct a data set; then create a machine operation time prediction model based on a multi-layer perceptron, a gradient boosting regression module, a decision tree module and a fusion output module, and train the machine operation time prediction model through the data set; finally, predict the operation time of each machine through the trained machine operation time prediction model to obtain the operation time prediction result, obtain the work order information, machine information and scheduling priority of each machine, and perform intelligent scheduling based on the operation time prediction result, work order information, machine information and scheduling priority; multi-layer perceptron, gradient boosting regression module, decision tree module and fusion output module, and train the machine operation time prediction model through the data set; finally, predict the operation time of each machine through the trained machine operation time prediction model, obtain the work order information, machine information and scheduling priority of each machine, and perform intelligent scheduling based on the operation time prediction result, work order information, machine information and scheduling priority; The perceptron is used to perform feature operations on the operating data to obtain the first prediction result; the gradient boosting regression module is used to construct a weak prediction model and iterate it. During the iteration process, a new weak prediction model is fitted according to the negative gradient direction of the loss function, and the prediction values ​​of all iterated weak prediction models are weighted and summed to obtain the second prediction result; the decision tree module is used to split the features in the input operating data according to the information gain, and then perform regression analysis to output the third prediction result; the fusion output module is used to weightedly fuse the first prediction result, the second prediction result and the third prediction result to output the operating time prediction result; that is, by fusing the multi-layer perceptron (MLP), the gradient boosting regression module (GB), and the decision tree module (DT) to predict the operating time of the machine, effectively combining the advantages of the fusion multi-layer perceptron, the gradient boosting regression module and the decision tree module, the accuracy and robustness of the operating time prediction are improved, and then the operating time prediction results, work order information, machine information and production scheduling priority are combined for intelligent scheduling, which ultimately greatly improves the accuracy and robustness of machine scheduling.

[0211] 2. By performing fluctuation addition operations on the rotation speed and density in the data set before training the machine operation time prediction model, the possible fluctuations in the machine production process (such as changes in the machine operating status, fluctuations in production speed, etc.) are simulated. Such fluctuations are added to the data set, making the operation time prediction results closer to the actual situation, enhancing the stability and reliability of the operation time prediction results, and further improving the accuracy and robustness of machine scheduling.

[0212] 3. By setting the production scheduling priority to idle machine priority, occupied machine priority, short remaining time for disk head priority, high output priority or short estimated occupancy time priority, and combining the running time prediction results, work order information, machine information and production scheduling priority, intelligent production scheduling is performed. That is, a greedy algorithm is used to achieve intelligent production scheduling according to the customized production scheduling priority. Under the premise of maximizing resource utilization and production efficiency, the production scheduling process is efficient and reasonable. In addition, intelligent production scheduling can effectively reduce machine waiting time, balance production load, and significantly improve production efficiency.

[0213] Although the specific implementation modes of the present invention are described above, those skilled in the art should understand that the specific implementation modes described are only illustrative and are not intended to limit the scope of the present invention. Equivalent modifications and changes made by those skilled in the art in accordance with the spirit of the present invention should be included in the scope of protection of the claims of the present invention.

Claims

1. An intelligent production scheduling method based on multi-model fusion, characterized by: The steps include: Step S1, obtaining a large amount of historical operation data of the machine, including at least the machine number, date, operation time, rotation speed and density, converting the format of each of the historical operation data from the JSON format to the DataFrame format, performing an outlier removal operation on each of the historical operation data after the format conversion, and then converting the encoding format of the machine number in each of the historical operation data into One-Hot encoding, performing a mean-standard deviation standardization operation on the rotation speed and density, thereby completing the preprocessing of each of the historical operation data and constructing a data set; Step S2, creating a machine operation time prediction model based on a multi-layer perceptron, a gradient boosting regression module, a decision tree module and a fusion output module; The loss function of the machine operation time prediction model is: Where L represents the loss value of the loss function of the machine running time prediction model; α j Represents learnable parameters; L1 represents the loss function of the multilayer perceptron, L2 represents the loss function of the gradient boosting regression module, and L3 represents the loss function of the decision tree module. L1, L2, and L3 are all mean square error functions; Step S3, training a machine operation time prediction model using the data set; Step S4: Use the trained machine operating time prediction model to predict the operating time of each machine to obtain the operating time prediction result, obtain the work order information, machine information and production scheduling priority of each machine, and perform intelligent production scheduling based on the operating time prediction result, work order information, machine information and production scheduling priority.

2. The intelligent production scheduling method based on multi-model fusion according to claim 1, characterized in that: In step S1, the formula for the mean-standard deviation standardization operation is: in, represents the speed after the standardization operation; Z1 represents the speed before the standardization operation; u1 represents the average speed; σ1 represents the standard deviation of the speed; Z2 scaled represents the density after standardization; Z2 represents the density before standardization; u2 represents the average value of density; σ2 represents the standard deviation of density.

3. The intelligent production scheduling method based on multi-model fusion according to claim 1, characterized in that: In step S2, the multilayer perceptron includes an input layer, a plurality of hidden layers and an output layer, and each of the input layer, the hidden layer and the output layer is composed of a plurality of stacked fully connected layers; the input layer is used to input operation data, the hidden layer is used to perform feature operation on the operation data to obtain a first prediction result, and the output layer is used to output the first prediction result; The formula of the fully connected layer is: X (l) =(W (l) X (l-1) +b (l) ); Among them, X (l) represents the output of the lth fully connected layer; X (l-1) represents the output of the l-1th fully connected layer, and is also the input of the lth fully connected layer. The input of the 0th fully connected layer is the running data; W (l) represents the weight matrix of the lth fully connected layer; b (l) Represents the bias vector of the lth fully connected layer; The gradient boosting regression module constructs a weak prediction model and iterates the weak prediction model. During the iteration, a new weak prediction model is fitted according to the negative gradient direction of the loss function of the gradient boosting regression module, and the prediction values ​​of all the iterated weak prediction models are weighted summed to obtain a second prediction result; The iterative formula of the weak prediction model is: F m (x)=F m-1 (x)+v·h m (x); Among them, F m (x) represents the second prediction result of the mth iteration; F m-1 (x) represents the second prediction result of the m-1th iteration; v represents the learning rate; h m (x) represents the predicted value of the mth weak prediction model; x represents the operating data; The formula for the negative ladder direction is: r m =y i -F m-1 (x i ); Among them, r m Represents the residual, that is, the negative gradient direction; y i represents the true value corresponding to the i-th running data; F m-1 (x i ) represents the i-th running data x i Input the second prediction result obtained by the gradient boosting regression module after m-1 iterations; The calculation formula of the second prediction result is: Wherein, F(x) represents the second prediction result; M represents the total number of weak prediction models; The decision tree module splits the features in the input operation data based on information gain, and then performs regression analysis to output a third prediction result; the features include machine number, date, operation time, rotation speed and density; The calculation formula of the information gain is: Where IG represents information gain; H() represents entropy function; S represents the data set containing running data; S i represents the i-th subset in the data set; n represents the total number of subsets; Indicates the total number of feature types; p k Represents the probability of the k-th feature in S.

4. The intelligent production scheduling method based on multi-model fusion according to claim 1, characterized in that: In step S2, the fusion output module is used to perform weighted fusion on the first prediction result, the second prediction result and the third prediction result output by the multilayer perceptron, the gradient boosting regression module and the decision tree module, and output the running time prediction result, and the formula is: Among them, prediction represents the running time prediction result; w j Represents weight; model_predict 1 Indicates the first prediction result; model_predict 2 Indicates the second prediction result; model_predict 3 Represents the third prediction result.

5. The intelligent production scheduling method based on multi-model fusion according to claim 1, characterized in that: The step S3 is specifically as follows: Perform a ripple addition operation on the speed and density in the dataset: in, Indicates the speed after adding fluctuations; Indicates the speed before adding fluctuations; Δ[Z1] i Indicates the i-th speed fluctuation value, and its value range is (-1,1); represents the density after adding fluctuations; represents the density before adding fluctuations; Δ[Z2] i Indicates the i-th density fluctuation value, ranging from (-20, 20); Training a machine operation time prediction model using the data set after the fluctuation addition operation; The step S4 is specifically as follows: The running time prediction model of the machine after training is used to perform real-time prediction of the running time of each machine to obtain the running time prediction result, and obtain the work order information, machine information and production scheduling priority of each machine; The work order information at least includes the expected completion time; the machine information at least includes the number of idle machines, the idle machine numbers and the total production output; the production scheduling priority is idle machine priority, occupied machine priority, short remaining time for the machine head priority, high output priority or short expected occupancy time priority; The estimated total output is calculated based on the running time prediction results, work order information, machine information and production scheduling priority, and intelligent production scheduling is performed based on the estimated total output and the scheduled total output.

6. An intelligent production scheduling system based on multi-model fusion, characterized by: Includes the following modules: A data set construction module is used to obtain a large amount of historical operation data of the machine, including at least the machine number, date, operation time, rotation speed and density, convert the format of each of the historical operation data from JSON format to DataFrame format, perform an outlier removal operation on each of the historical operation data after format conversion, and then convert the encoding format of the machine number in each of the historical operation data into One-Hot encoding, perform a mean-standard deviation standardization operation on the rotation speed and density, and then complete the preprocessing of each of the historical operation data and construct a data set; A machine operation time prediction model creation module is used to create a machine operation time prediction model based on a multi-layer perceptron, a gradient boosting regression module, a decision tree module, and a fusion output module; The loss function of the machine operation time prediction model is: Where L represents the loss value of the loss function of the machine running time prediction model; α j Represents learnable parameters; L1 represents the loss function of the multilayer perceptron, L2 represents the loss function of the gradient boosting regression module, and L3 represents the loss function of the decision tree module. L1, L2, and L3 are all mean square error functions; A machine operation time prediction model training module is used to train the machine operation time prediction model using the data set; The intelligent production scheduling module is used to predict the operating time of each machine through the trained machine operating time prediction model to obtain the operating time prediction result, obtain the work order information, machine information and production scheduling priority of each machine, and perform intelligent production scheduling based on the operating time prediction result, work order information, machine information and production scheduling priority.

7. The intelligent production scheduling system based on multi-model fusion according to claim 6, characterized in that: In the data set construction module, the formula for the mean-standard deviation standardization operation is: in, represents the speed after the standardization operation; Z1 represents the speed before the standardization operation; u1 represents the average speed; σ1 represents the standard deviation of the speed; represents the density after standardization; Z2 represents the density before standardization; u2 represents the average value of density; σ2 represents the standard deviation of density.

8. The intelligent production scheduling system based on multi-model fusion according to claim 6, characterized in that: In the machine operation time prediction model creation module, the multilayer perceptron includes an input layer, a plurality of hidden layers and an output layer, and each of the input layer, the hidden layer and the output layer is composed of a plurality of stacked fully connected layers; the input layer is used for inputting operation data, the hidden layer is used for performing feature operation on the operation data to obtain a first prediction result, and the output layer is used for outputting the first prediction result; The formula of the fully connected layer is: X (l) =(W (l) X (l-1) +b (l) ); Among them, X (l) represents the output of the lth fully connected layer; X (l-1) represents the output of the l-1th fully connected layer, and is also the input of the lth fully connected layer. The input of the 0th fully connected layer is the running data; W (l) represents the weight matrix of the lth fully connected layer; b (l) Represents the bias vector of the lth fully connected layer; The gradient boosting regression module constructs a weak prediction model and iterates the weak prediction model. During the iteration, a new weak prediction model is fitted according to the negative gradient direction of the loss function of the gradient boosting regression module, and the prediction values ​​of all the iterated weak prediction models are weighted summed to obtain a second prediction result; The iterative formula of the weak prediction model is: F m (x)=F m-1 (x)+v·h m (x); Among them, F m (x) represents the second prediction result of the mth iteration; F m-1 (x) represents the second prediction result of the m-1th iteration; v represents the learning rate; h m (x) represents the predicted value of the mth weak prediction model; x represents the operating data; The formula for the negative ladder direction is: r m =y i -F m-1 (x i ); Among them, r m Represents the residual, that is, the negative gradient direction; y i represents the true value corresponding to the i-th running data; F m-1 (x i ) represents the i-th running data x i Input the second prediction result obtained by the gradient boosting regression module after m-1 iterations; The calculation formula of the second prediction result is: Wherein, F(x) represents the second prediction result; M represents the total number of weak prediction models; The decision tree module splits the features in the input operation data based on information gain, and then performs regression analysis to output a third prediction result; the features include machine number, date, operation time, rotation speed and density; The calculation formula of the information gain is: Where IG represents information gain; H() represents entropy function; S represents the data set containing running data; S i represents the i-th subset in the data set; n represents the total number of subsets; Indicates the total number of feature types; p k Represents the probability of the k-th feature in S.

9. The intelligent production scheduling system based on multi-model fusion according to claim 6, characterized in that: In the machine operation time prediction model creation module, the fusion output module is used to perform weighted fusion on the first prediction result, the second prediction result and the third prediction result output by the multilayer perceptron, the gradient boosting regression module and the decision tree module, and output the operation time prediction result, and the formula is: Among them, prediction represents the running time prediction result; w j Represents weight; model_predict 1 Indicates the first prediction result; model_predict 2 Indicates the second prediction result; model_predict 3 Represents the third prediction result.

10. The intelligent production scheduling system based on multi-model fusion according to claim 6, characterized in that: The machine operation time prediction model training module is specifically used for: Perform a ripple addition operation on the speed and density in the dataset: in, Indicates the speed after adding fluctuations; Indicates the speed before adding fluctuations; Δ[Z1] i Indicates the i-th speed fluctuation value, and its value range is (-1,1); represents the density after adding fluctuations; represents the density before adding fluctuations; Δ[Z2] i Indicates the i-th density fluctuation value, ranging from (-20, 20); Training a machine operation time prediction model using the data set after the fluctuation addition operation; The intelligent production scheduling module is specifically used for: The running time prediction model of the machine after training is used to perform real-time prediction of the running time of each machine to obtain the running time prediction result, and obtain the work order information, machine information and production scheduling priority of each machine; The work order information at least includes the expected completion time; the machine information at least includes the number of idle machines, the idle machine numbers and the total production output; the production scheduling priority is idle machine priority, occupied machine priority, short remaining time for the machine head priority, high output priority or short expected occupancy time priority; The estimated total output is calculated based on the running time prediction results, work order information, machine information and production scheduling priority, and intelligent production scheduling is performed based on the estimated total output and the scheduled total output.

Images