CN118114836A - Method and system for predicting and scheduling productivity based on artificial intelligence - Google Patents

Method and system for predicting and scheduling productivity based on artificial intelligence Download PDF

Info

Publication number
CN118114836A
CN118114836A CN202410329818.1A CN202410329818A CN118114836A CN 118114836 A CN118114836 A CN 118114836A CN 202410329818 A CN202410329818 A CN 202410329818A CN 118114836 A CN118114836 A CN 118114836A
Authority
CN
China
Prior art keywords
time
denotes
hidden state
gate
objective function
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202410329818.1A
Other languages
Chinese (zh)
Inventor
于邦齐
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Individual
Original Assignee
Individual
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Individual filed Critical Individual
Priority to CN202410329818.1A priority Critical patent/CN118114836A/en
Publication of CN118114836A publication Critical patent/CN118114836A/en
Pending legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/004—Artificial life, i.e. computing arrangements simulating life
    • G06N3/006—Artificial life, i.e. computing arrangements simulating life based on simulated virtual individual or collective life forms, e.g. social simulations or particle swarm optimisation [PSO]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G06N3/0442—Recurrent networks, e.g. Hopfield networks characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/048—Activation functions
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
    • G06Q50/04—Manufacturing

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Business, Economics & Management (AREA)
  • Software Systems (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Evolutionary Computation (AREA)
  • Computational Linguistics (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Biophysics (AREA)
  • Mathematical Physics (AREA)
  • Biomedical Technology (AREA)
  • Artificial Intelligence (AREA)
  • Data Mining & Analysis (AREA)
  • Economics (AREA)
  • Human Resources & Organizations (AREA)
  • Strategic Management (AREA)
  • Tourism & Hospitality (AREA)
  • Marketing (AREA)
  • General Business, Economics & Management (AREA)
  • Primary Health Care (AREA)
  • Manufacturing & Machinery (AREA)
  • Development Economics (AREA)
  • Game Theory and Decision Science (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Operations Research (AREA)
  • Quality & Reliability (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention provides a productivity prediction and scheduling method and system based on artificial intelligence, which relate to the field of data processing, collect historical production data, construct a productivity prediction model, receive the historical production data through an input layer of the productivity prediction model, analyze the historical production data based on a time-dependent capturing layer, capture a time-dependent relationship and obtain a time-dependent result; flowing the time dependent result into a feature focusing layer, obtaining key features by calculating focus attention scores, and screening the key features through a hidden state analysis layer to obtain productivity prediction results by a gating mechanism; based on the productivity prediction result, combining to-be-completed order data to construct a first, a second and a third objective functions; constructing a comprehensive objective function based on the first, second and third objective functions; and solving the comprehensive objective function through a multi-objective optimization algorithm to obtain a scheduling optimal solution.

Description

Method and system for predicting and scheduling productivity based on artificial intelligence
Technical Field
The invention relates to a data processing technology, in particular to an artificial intelligence-based productivity prediction and scheduling method and system.
Background
The industrial development is faster and faster, the productivity of industrial products is a certain index of a factory, the past productivity data and the related data of influencing factors are utilized to process the data and mine the influence degree between the data, a certain method is utilized to predict the future productivity, the operation department of the factory can better judge and make decisions, and the staff number, equipment maintenance, material control and the like of the factory can be adjusted according to the decisions so as to obtain ideal profitability.
The invention discloses a CN201911224534.1, belongs to the field of digital factory productivity prediction, and particularly relates to a method, a system and a device for predicting the production productivity of a digital factory, aiming at solving the problems that the model in the prior art is single and is not easy to generalize, the multi-layer difference relation between data and a prediction target cannot be explored, and the accuracy of a prediction result is low. The method comprises the following steps: acquiring production capacity data in a preset time period, and extracting and expanding feature vectors through feature engineering; dividing the feature vector set into a set training test set by a K-fold cross validation method; and obtaining the predicted capacity of the digital factory through the capacity prediction model. According to the invention, a plurality of base models which are better in production capacity prediction of a digital factory are integrated by adopting a GBDT-Stacking method, characteristics which are more reliable in production capacity prediction of the factory are constructed through characteristic engineering, a model training test set is divided through a K-fold cross validation method, and the model prediction accuracy is high, fitting is not easy to pass, generalization is easy, and robustness is good;
In summary, the prior art scheme has some defects that only one algorithm and one model are adopted in prediction, so that multi-layer difference relation between complex data and a prediction target cannot be explored, more accurate prediction cannot be made, integration is lacked, and model accuracy is not high.
Disclosure of Invention
The embodiment of the invention provides an artificial intelligence-based productivity prediction and scheduling method and system, which at least can solve part of problems in the prior art.
In a first aspect of an embodiment of the present invention,
Providing collected historical production data, cleaning the historical production data, constructing a productivity prediction model, receiving the historical production data through an input layer of the productivity prediction model, analyzing the historical production data based on a time-dependent capturing layer of the productivity prediction model, capturing a time-dependent relationship, and obtaining a time-dependent result;
The time dependent result flows into a feature focusing layer of the productivity prediction model, key features are obtained through calculation of focusing attention scores, and the key features are screened out through a gating mechanism by a hidden state analysis layer;
based on the productivity prediction result, combining to-be-completed order data, and constructing a first objective function by taking the minimum total completion time as a target; constructing a second objective function with the minimum total cost as a target; constructing a third objective function by taking the minimum production task completion time in advance as a target; constructing a comprehensive objective function based on the first objective function, the second objective function and the third objective function; and solving the comprehensive objective function through a preset multi-objective optimization algorithm to obtain a scheduling optimal solution.
In an alternative embodiment of the present invention,
Analyzing the historical production data based on a time-dependent capturing layer of the productivity prediction model, capturing a time-dependent relationship, and obtaining a time-dependent result, wherein the method comprises the following steps:
the time-dependent capture layer is constructed based on a long-short-time memory model;
Setting an initial hidden state and an initial unit state, extracting time-series data based on the historical production data, and performing an operation on each time step of the time-series data:
Calculating forgetting gate output, input gate output and candidate value vectors through forgetting gate and input gate, updating unit states by utilizing the forgetting gate output, the input gate output and the candidate value vectors, calculating output gate output through the output gate, calculating a hidden state of the time step by combining the output gate output based on the result of updating the unit states, and transmitting the hidden state to the next time step;
repeating the operation until all time steps of the time series data are calculated;
Setting jump connection according to the preset interval step number between the hidden states, expanding the hidden states for each time step, introducing jump weights for each jump connection, combining the jump weights based on the expanded hidden states, and obtaining a time dependent result through fusion weighting.
In an alternative embodiment of the present invention,
Operating on each time step of the time series data, comprising:
The time series data is operated through forget gate, input gate, output gate and cell state update, and the formula is as follows:
Wherein f t denotes a forgetting gate output at time step t, σ denotes a sigmoid activation function, W f denotes a weight of the forgetting gate, h t-1 denotes a hidden state of the previous time step, x t denotes an input of the current time step, b f denotes a bias of the forgetting gate, i t denotes an input gate output at time step t, W i denotes a weight of the input gate, b i denotes a bias of the input gate, C' t denotes a candidate value vector at time step t, W C denotes a candidate value vector weight, b C denotes a bias of the candidate value vector, C t denotes a cell state at time step t, C t-1 denotes a cell state of the previous time step, o t denotes an output of the output gate at time step t, W o denotes a weight of the output gate, b o denotes a bias of the output gate, and h t denotes a hidden state at current time step t.
In an alternative embodiment of the present invention,
Setting jump connection according to a preset interval step number between hidden states, expanding the hidden states for each time step, introducing jump weights for each jump connection, combining the jump weights based on the expanded hidden states, and obtaining a time dependent result through fusion weighting, wherein the jump connection comprises the following steps:
For each time step the hidden state is extended, the formula is as follows:
hext＝[ht;ht-skip;...;ht-n×skip]；
Wherein hex t represents an extended hidden state of the current time step t, h t represents a hidden state of the current time step t, h t-skip represents a hidden state after skip of skip steps, and n×skip represents n skip intervals;
a hopping weight is introduced for each of the hopping connections, the formula of which is as follows:
Wskip·ht-skip；
wherein W skip represents a weight matrix of the jump connection;
The time dependent result is obtained by fusion weighting, and the formula is as follows:
hcbt＝Wh·hext；
Wherein hcb t represents the time-dependent result of the current time step t, and W h represents the fusion weight matrix.
In an alternative embodiment of the present invention,
And flowing the time dependent result into a feature focusing layer of the productivity prediction model, and obtaining key features by calculating a focusing attention score, wherein the key features comprise:
the characteristic focusing layer is constructed based on a self-attention mechanism;
carrying out head separation processing on the time dependent result, and dividing the time dependent result into a plurality of heads;
Performing linear transformation on each head, calculating a focus attention score by using the scaled dot product attention, and normalizing the focus attention score by using a softmax activation function;
After weighted summation, all heads are combined in series, and the combined output is subjected to linear transformation to obtain key characteristics.
In an alternative embodiment of the present invention,
Further comprises:
performing linear transformation on each head to obtain a query Q matrix, a key K matrix and a value V matrix of the head, wherein the formula is as follows:
Wherein H represents an input hidden state sequence, Q k represents a query Q matrix of a kth head, WQ k represents a weight matrix of a query operation of the kth head on the input hidden state sequence, K k represents a key K matrix of the kth head, WK k represents a weight matrix of a key operation of the kth head on the input hidden state sequence, V k represents a value V matrix of the kth head, WV k represents a weight matrix of a value operation of the kth head on the input hidden state sequence;
The focus attention score is calculated using the scaled dot product attention, with the formula:
wherein, Represents the attention score of the kth head for the sequence position t and t ', t represents the time step corresponding to the query operation, t' represents the time step corresponding to the key operation,/>Representing a transpose of the key K matrix of the kth head, d k representing the dimension of the query and key;
Normalized and weighted summed using a softmax activation function, the formula is as follows:
wherein, Representing the result of normalization of focus attention score,/>Representing the output of the kth header;
the outputs of all the heads are combined as follows:
Where M t represents the combined output, concat () function represents the series connection;
based on the combined outputs, a linear transformation is performed, which is shown as follows:
H′t＝WM·Mt；
Where H' t represents the output vector and W M represents the weight matrix of the linear transformation.
In an alternative embodiment of the present invention,
Screening the yield prediction result by the key features through a gate control mechanism through a hidden state analysis layer comprises the following steps:
the hidden state analysis layer is constructed based on a gating circulating unit;
And taking the key features output by the feature focusing layer as input data of the hidden state analysis layer, and operating based on each time step:
Sequentially passing through an update gate and a reset gate, calculating update gate output and reset gate output, creating a candidate hidden state based on the reset gate output, and combining the candidate hidden state based on the update gate output to form an update hidden state;
repeating the operation until all time steps of the input data are calculated, and obtaining a productivity prediction result.
In an alternative embodiment of the present invention,
Performing operations on a per time step basis includes:
Calculating an update gate output and a reset gate output, creating a candidate hidden state, and updating the hidden state, wherein the formula is as follows:
Wherein z t denotes the update gate output at time step t, σ denotes the sigmoid activation function, W z denotes the weight of the update gate, v t-1 denotes the hidden state of the previous time step, y t denotes the input of the current time step, b z denotes the bias of the update gate, r t denotes the reset gate output at time step t, W r denotes the weight of the reset gate, b r denotes the bias of the reset gate, v' t denotes the candidate hidden state, W denotes the weight matrix of the candidate hidden state, b denotes the bias of the candidate hidden state, and v t denotes the updated hidden state.
In an alternative embodiment of the present invention,
Constructing a composite objective function based on the first, second, and third objective functions includes:
based on the productivity prediction model, a productivity prediction result is obtained, and the to-be-completed order data is combined to construct a comprehensive objective function, wherein the formula is as follows:
Wherein ω1 represents the weight of the total completion time, N represents the total number of production lines, i represents the ith production line, t ei represents the production end time of the ith production line, ω2 represents the weight of the total cost, F i represents the cost of the ith production line, ω3 represents the weight of the task completion time in advance, M represents the total number of orders, j represents the j order, t odj represents the latest completion time of the j order, t oej represents the actual completion time of the j order, c j represents the completion of the j order, c j is-1 when t odj＜toej is the completion of the out-of-date, c j is 1 when t odj≥toej is the completion of the in advance.
In an alternative embodiment of the present invention,
Solving the comprehensive objective function through a preset multi-objective optimization algorithm to obtain a scheduling optimal solution, wherein the step of obtaining the scheduling optimal solution comprises the following steps:
randomly generating particle swarms, wherein each particle represents a production scheme, initializing the position and the speed of the particle, and setting the initial temperature, the cooling rate and the disturbance range;
evaluating the production scheduling scheme corresponding to the particles, calculating a fitness value by combining the comprehensive objective function, and starting updating iteration of the particle swarm:
Updating the individual optimal position of each particle, simultaneously updating the global optimal position of the particle group, updating the speed of the particle according to the individual optimal position and the global optimal position, and updating the position of the particle according to the updated speed;
After the updating iteration, perturbing the production scheduling scheme corresponding to the particles within the perturbation range to generate a new production scheduling scheme, calculating the fitness value of the new production scheduling scheme, if the fitness value is optimal, accepting the new production scheduling scheme, otherwise, determining whether to accept the new production scheduling scheme according to the temperature and the annealing probability;
And reducing the temperature according to a preset cooling rate, restarting updating iteration and disturbance of the particle swarm until the particle swarm meets a preset iteration threshold, evaluating the fitness value of the final production scheme, and selecting the highest fitness value as the optimal production solution.
In a second aspect of an embodiment of the present invention,
The first unit is used for collecting historical production data, cleaning the historical production data, constructing a productivity prediction model, receiving the historical production data through an input layer of the productivity prediction model, analyzing the historical production data based on a time-dependent capturing layer of the productivity prediction model, capturing a time-dependent relationship, and obtaining a time-dependent result;
the second unit is used for flowing the time dependent result into a feature focusing layer of the productivity prediction model, obtaining key features by calculating focusing attention scores, and screening the key features through a gating mechanism by a hidden state analysis layer to obtain productivity prediction results;
A third unit, configured to construct a first objective function based on the productivity prediction result and combining to-be-completed order data, with minimum total completion time as a target; constructing a second objective function with the minimum total cost as a target; constructing a third objective function by taking the minimum production task completion time in advance as a target; constructing a comprehensive objective function based on the first objective function, the second objective function and the third objective function; and solving the comprehensive objective function through a preset multi-objective optimization algorithm to obtain a scheduling optimal solution.
In a third aspect of an embodiment of the present invention,
There is provided an electronic device including:
A processor;
a memory for storing processor-executable instructions;
wherein the processor is configured to invoke the instructions stored in the memory to perform the method described previously.
In a fourth aspect of an embodiment of the present invention,
There is provided a computer readable storage medium having stored thereon computer program instructions which, when executed by a processor, implement the method as described above.
In the embodiment of the invention, the capacity can be more comprehensively understood and predicted by integrating different models; the time dependency capture layer is designed, so that the time correlation in the historical data can be better understood and utilized, and the model can be more flexibly adapted to the change and trend of different time periods; through the arrangement of jump connection, the transfer efficiency of information between hidden states is improved, the gradient vanishing problem is relieved, and the model is more effective in processing long-sequence data; jump weight and weighted fusion are introduced, so that the model can more flexibly weight information of different time steps when a time dependent result is obtained, and the attention of the model to key time points in historical data is improved; by calculating the focusing attention score, the model can extract key features from the time dependent results, so that the focusing degree of the model on the key features is improved, and the important time dependent relation in productivity prediction can be accurately captured; through a multi-head self-attention mechanism, the model can learn different representation subspaces in parallel, and the modeling capacity of the model on complex relations is improved; by modeling based on the gating loop unit, time series data can be better modeled, including capturing time dependence, and the gating mechanism allows the model to dynamically adjust the weight of information according to input and previous hidden states, helping to more accurately represent patterns in the time series; due to the effects of the update gate and the reset gate, the model can effectively process long-term dependence, and is important to the scene needing to consider long-term trend in tasks such as capacity prediction; the combination of different multi-objective optimization algorithms utilizes the global searching capability of the particle swarm optimization algorithm and the local searching and escaping local optimal capability of the simulated annealing algorithm; the two algorithms are combined, so that the search space can be enlarged, the possibility of finding the globally optimal solution or the non-inferior solution is improved, the characteristics of the problem are better matched, and the adaptability of the algorithm is improved.
Drawings
FIG. 1 is a flow chart of a method for capacity prediction and scheduling based on artificial intelligence according to an embodiment of the invention;
FIG. 2 is a schematic diagram of a system for capacity prediction and scheduling based on artificial intelligence according to an embodiment of the present invention.
Detailed Description
For the purpose of making the objects, technical solutions and advantages of the embodiments of the present invention more apparent, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention, and it is apparent that the described embodiments are only some embodiments of the present invention, not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
The technical scheme of the invention is described in detail below by specific examples. The following embodiments may be combined with each other, and some embodiments may not be repeated for the same or similar concepts or processes.
FIG. 1 is a schematic flow chart of a method for capacity prediction and scheduling based on artificial intelligence according to an embodiment of the invention, as shown in FIG. 1, the method includes:
S101, collecting historical production data, cleaning the historical production data, constructing a productivity prediction model, receiving the historical production data through an input layer of the productivity prediction model, analyzing the historical production data based on a time-dependent capturing layer of the productivity prediction model, capturing a time-dependent relationship, and obtaining a time-dependent result;
Collecting historical production data, wherein the historical production data comprises the operation time length of a production line, the output quantity, the production cost, the order execution time length of the production line and the like, cleaning the historical production data, improving the quality of the production data, complementing the missing value, correcting the abnormal value, deleting the repeated value, further checking a time point mark of the data after data standardization, unifying time data formats, sequentially transmitting the time data formats in a time step mode, and analyzing the processed historical production data in a time sequence according to a specific time step;
Constructing a comprehensive productivity prediction model, and dividing the productivity prediction model into a plurality of data processing layers for further training and prediction;
the input layer receives the history production data after the cleaning treatment, transmits the history production data to the next step for operation, captures the time dependence relationship through the time dependence capturing layer, and enables the model to accurately give a corresponding time dependence result after receiving the new history production data through repeated training of the history production data.
In this embodiment, integrating different models can more fully understand and predict capacity; the time dependency capture layer is designed, so that the time correlation in the historical data can be better understood and utilized, and the model can be more flexibly adapted to the change and trend of different time periods.
In an alternative embodiment, the analyzing the historical production data based on the time-dependent capturing layer of the productivity prediction model captures the time-dependent relationship to obtain the time-dependent result, including:
the time-dependent capture layer is constructed based on a long-short-time memory model;
Setting an initial hidden state and an initial unit state, extracting time-series data based on the historical production data, and performing an operation on each time step of the time-series data:
Calculating forgetting gate output, input gate output and candidate value vectors through forgetting gate and input gate, updating unit states by utilizing the forgetting gate output, the input gate output and the candidate value vectors, calculating output gate output through the output gate, calculating a hidden state of the time step by combining the output gate output based on the result of updating the unit states, and transmitting the hidden state to the next time step;
repeating the operation until all time steps of the time series data are calculated;
Setting jump connection according to the preset interval step number between the hidden states, expanding the hidden states for each time step, introducing jump weights for each jump connection, combining the jump weights based on the expanded hidden states, and obtaining a time dependent result through fusion weighting.
The forgetting gate is specifically used for determining which information in the memory state of the previous time step should be reserved and which information should be forgotten, and the updating of the memory unit is affected by calculating the output of the forgetting gate, wherein the output value of the forgetting gate is between 0 and 1, and 0 represents complete forgetting and 1 represents complete reservation;
The input gate is specifically used for controlling the influence of new input information on the memory state of the current time step, and the new information is fused into the memory unit by calculating the output of the input gate;
the candidate value vector is specifically generated by input information of the current time step, is a vector containing potential new information and is used for updating the cell state and providing an updated value to be selected;
Setting an initial hidden state and an initial memory unit state, which are usually zero or preset values, starting to operate each time step of the historical production data, calculating forgetting gate output by utilizing the input of the current time step and the hidden state of the previous time step, and determining the historical information to be reserved; calculating input gate output through the operation of the input of the current time step and the hidden state of the previous time step, and determining new information to be introduced into the memory unit; calculating candidate value vectors by utilizing the input of the current time step and the hidden state of the previous time step to obtain a potential memory state update; updating the state of the memory unit by utilizing the forgetting gate output, the input gate output and the candidate value vector; calculating output gate output based on the updated memory unit state, calculating the hidden state of the current time step, and transmitting the current hidden state to the next time step for circulation in sequence;
Jump connection is arranged between hidden states according to a preset interval step number, so that information flow is promoted, wherein the interval step number is usually 2 steps, 3 steps, 4 steps and the like; introducing jump weights for each jump connection, weighting the extended hidden state according to the jump weights, extending the hidden state of each time step, and introducing flexibility for information transfer between different time steps; based on the extended hidden state and the jump weight, a final time dependent result is obtained in a weighted fusion mode.
In an alternative embodiment, each time step of the time series data is operated on, comprising:
The time series data is operated through forget gate, input gate, output gate and cell state update, and the formula is as follows:
Wherein f t denotes a forgetting gate output at time step t, σ denotes a sigmoid activation function, W f denotes a weight of the forgetting gate, h t-1 denotes a hidden state of the previous time step, x t denotes an input of the current time step, b f denotes a bias of the forgetting gate, i t denotes an input gate output at time step t, W i denotes a weight of the input gate, b i denotes a bias of the input gate, C' t denotes a candidate value vector at time step t, W C denotes a candidate value vector weight, b C denotes a bias of the candidate value vector, C t denotes a cell state at time step t, C t-1 denotes a cell state of the previous time step, o t denotes an output of the output gate at time step t, W o denotes a weight of the output gate, b o denotes a bias of the output gate, and h t denotes a hidden state at current time step t.
In an alternative embodiment, setting jump connection according to a preset number of steps at intervals between hidden states, expanding the hidden states for each time step, introducing jump weights for each jump connection, and obtaining time dependent results by fusion weighting based on the expanded hidden states and combining the jump weights, wherein the method comprises the following steps:
For each time step the hidden state is extended, the formula is as follows:
hext＝[ht;ht-skip;...;ht-n×skip]；
Wherein hex t represents an extended hidden state of the current time step t, h t represents a hidden state of the current time step t, h t-skip represents a hidden state after skip of skip steps, and n×skip represents n skip intervals;
a hopping weight is introduced for each of the hopping connections, the formula of which is as follows:
Wskip·ht-skip；
wherein W skip represents a weight matrix of the jump connection;
The time dependent result is obtained by fusion weighting, and the formula is as follows:
hcbt＝Wh·hext；
Wherein hcb t represents the time-dependent result of the current time step t, and W h represents the fusion weight matrix.
In the embodiment, through the time dependency capturing layer, the time dependency relationship in the historical production data is better modeled, so that long-term and short-term time dependency can be captured, and the modeling capability of time sequence data is improved; through the arrangement of jump connection, the transfer efficiency of information between hidden states is improved, the gradient vanishing problem is relieved, and the model is more effective in processing long-sequence data; jump weight and weighted fusion are introduced, so that the model can more flexibly weight information of different time steps when a time dependent result is obtained, and the attention of the model to key time points in historical data is improved; through the operation steps, the model can comprehensively understand the historical production data, better capture key modes and trends in the time sequence, and is beneficial to improving the accuracy and the robustness of productivity prediction.
S102, flowing the time dependent result into a feature focusing layer of the productivity prediction model, obtaining key features by calculating focusing attention scores, and screening the key features through a hidden state analysis layer to obtain productivity prediction results through a gating mechanism;
the dimension of the time-dependent result is matched with the expected input dimension of the characteristic focusing layer, the time-dependent result is used as the input of the characteristic focusing layer, linear projection is carried out on the time-dependent result, the initial representation of the attention score is obtained, and the activation function is applied to normalize the score, so that the focusing attention score is obtained; applying the focus attention score to the time dependent results, and obtaining key features by carrying out weighted summation on the results of each time step;
And transmitting the obtained key features to a hidden state analysis layer, screening the key features by using a gating mechanism, further extracting and strengthening information related to productivity, obtaining a productivity prediction result by outputting the gating mechanism, and transmitting the output of the hidden state analysis layer to an output layer of a productivity prediction model to obtain a final productivity prediction result.
In the embodiment, by calculating the focus attention score, the model can extract key features from the time dependent results, so that the focusing degree of the model on the key features is improved, and the important time dependent relation in productivity prediction can be accurately captured; the focusing attention score is utilized to weight and sum the key features, so that information which has important roles in productivity prediction is enhanced, the expression of the model on the key features is enhanced, and the sensitivity of the model on productivity influence factors is improved; the key features are screened through a gating mechanism at the hidden state analysis layer, so that the model is helped to selectively pay attention to information contributing to the result in productivity prediction, the screening and learning capabilities of the model in a complex time sequence are improved, and the key features can be better understood and utilized;
In an alternative embodiment, flowing the time dependent results into a feature focusing layer of the capacity prediction model, obtaining key features by calculating focus attention scores, comprises:
the characteristic focusing layer is constructed based on a self-attention mechanism;
carrying out head separation processing on the time dependent result, and dividing the time dependent result into a plurality of heads;
Performing linear transformation on each head, calculating a focus attention score by using the scaled dot product attention, and normalizing the focus attention score by using a softmax activation function;
After weighted summation, all heads are combined in series, and the combined output is subjected to linear transformation to obtain key characteristics.
The heads in the head separation process specifically refer to a parallel attention mechanism module, and each head is focused on learning and capturing different attention points and features in an input sequence; decomposing the input into a plurality of heads, wherein each head independently learns different representation subspaces, and finally fusing information of the subspaces to obtain more comprehensive representation;
Dividing the time dependent result into a plurality of heads, dividing the input data according to the number of the heads, so that each head can independently process a subset; performing linear transformation on the input of each head, multiplying the input of each head with a corresponding weight matrix to obtain a linearly transformed representation, and creating different representation subspaces;
For each head's linearly transformed input, computing a focus attention score using scaled dot product attention, including computing a matrix of queries Q, keys K, values V, and performing focus attention score computation by dot product and scaling; performing an activation function operation on the attention score to obtain attention distribution, and ensuring that the attention weight sum of each time step is 1;
And carrying out weighted summation on the value matrix by using the focusing attention score to obtain the attention output of each head, combining the outputs of all heads in series to form a comprehensive representation, and carrying out linear transformation on the combined outputs to obtain the final key characteristics.
In an alternative embodiment, further comprising:
performing linear transformation on each head to obtain a query Q matrix, a key K matrix and a value V matrix of the head, wherein the formula is as follows:
Wherein H represents an input hidden state sequence, Q k represents a query Q matrix of a kth head, WQ k represents a weight matrix of a query operation of the kth head on the input hidden state sequence, K k represents a key K matrix of the kth head, WK k represents a weight matrix of a key operation of the kth head on the input hidden state sequence, V k represents a value V matrix of the kth head, WV k represents a weight matrix of a value operation of the kth head on the input hidden state sequence;
The focus attention score is calculated using the scaled dot product attention, with the formula:
wherein, Represents the attention score of the kth head for the sequence position t and t ', t represents the time step corresponding to the query operation, t' represents the time step corresponding to the key operation,/>Representing a transpose of the key K matrix of the kth head, d k representing the dimension of the query and key;
Normalized and weighted summed using a softmax activation function, the formula is as follows:
wherein, Representing the result of normalization of focus attention score,/>Representing the output of the kth header;
the outputs of all the heads are combined as follows:
Where M t represents the combined output, concat () function represents the series connection;
based on the combined outputs, a linear transformation is performed, which is shown as follows:
H′t＝WM·Mt；
Where H' t represents the output vector and W M represents the weight matrix of the linear transformation.
In the embodiment, the characteristic focusing layer adopts a self-attention mechanism, so that important characteristics are highlighted by calculating the focusing attention score on the basis of the head separation processing, key information is extracted, and the sensitivity of the model to an input sequence is improved; through a multi-head self-attention mechanism, the model can learn different representation subspaces in parallel, and the modeling capacity of the model on complex relations is improved; the outputs of the heads are combined and subjected to linear transformation, so that the learned characteristics of the heads are fused, and more comprehensive representation is obtained; the scaled dot product attention is introduced in the attention calculation, normalization is carried out through the operation of an activation function, and the nonlinear relation is effectively modeled, so that the model can be more flexibly adapted to the importance of different parts in the sequence.
In an alternative embodiment, screening the yield prediction result by the key feature through a gating mechanism through a hidden state parsing layer includes:
the hidden state analysis layer is constructed based on a gating circulating unit;
And taking the key features output by the feature focusing layer as input data of the hidden state analysis layer, and operating based on each time step:
Sequentially passing through an update gate and a reset gate, calculating update gate output and reset gate output, creating a candidate hidden state based on the reset gate output, and combining the candidate hidden state based on the update gate output to form an update hidden state;
repeating the operation until all time steps of the input data are calculated, and obtaining a productivity prediction result.
The update gate is specifically used for controlling whether to bring new information into the candidate hidden state, when the output of the update gate approaches 1, the model considers the new information more, and when the output approaches 0, the old information is reserved more;
the reset gate is specifically used for determining whether to forget the past hidden state, and when the output of the reset gate is close to 1, the model forgets more past information; while the output approaches 0, the model retains more past information;
Setting an initial hidden state and an initial unit state by taking key features output by a feature focusing layer as input data of a hidden state analysis layer, preparing for calculation of a gating circulation unit, and sequentially performing the following operations for each time step of the input data:
Calculating the output of an update gate of the current time step by using the input data and the hidden state of the previous time step;
Calculating the reset gate output of the current time step by utilizing the input data and the hidden state of the previous time step;
Creating a candidate hidden state based on the output of the reset gate, wherein the candidate hidden state contains the information of the current time step, and calculating the updated hidden state of the current time step by utilizing the output of the update gate and the candidate hidden state and combining the hidden state of the previous time step; and (3) circularly operating until all time steps of input data are calculated, and obtaining a productivity prediction result.
In an alternative embodiment, the operations include, on a per time step basis:
Calculating an update gate output and a reset gate output, creating a candidate hidden state, and updating the hidden state, wherein the formula is as follows:
Wherein z t denotes the update gate output at time step t, σ denotes the sigmoid activation function, W z denotes the weight of the update gate, v t-1 denotes the hidden state of the previous time step, y t denotes the input of the current time step, b z denotes the bias of the update gate, r t denotes the reset gate output at time step t, W r denotes the weight of the reset gate, b r denotes the bias of the reset gate, v' t denotes the candidate hidden state, W denotes the weight matrix of the candidate hidden state, b denotes the bias of the candidate hidden state, and v t denotes the updated hidden state.
In the embodiment, by modeling based on the gating loop unit, time series data can be better modeled, including capturing time dependence, and the gating mechanism allows the model to dynamically adjust the weight of the information according to the input and previous hidden states, helping to more accurately represent the patterns in the time series; the number of the parameters of the gating circulation unit is small, so that the gradient disappearance problem is reduced, the model parameters are better optimized, and the training efficiency is improved; the model can dynamically forget or retain the past information through the regulation and control of the update gate and the reset gate, is better suitable for the input of different time steps, can effectively process long-term dependence due to the effect of the update gate and the reset gate, and is important for the scene needing to consider long-term trend in tasks such as productivity prediction and the like.
S103, based on the productivity prediction result, combining to-be-completed order data, and constructing a first objective function by taking the minimum total completion time as a target; constructing a second objective function with the minimum total cost as a target; constructing a third objective function by taking the minimum production task completion time in advance as a target; constructing a comprehensive objective function based on the first objective function, the second objective function and the third objective function; solving the comprehensive objective function through a preset multi-objective optimization algorithm to obtain a scheduling optimal solution;
Acquiring a capacity prediction result of each production line in a future period of time by using the capacity prediction model, and analyzing order data to be completed, wherein the order data comprises the production quantity of orders, the predicted completion time, required resources and the like;
Constructing a first objective function based on the capacity forecast and the order information, so as to minimize the total completion time; constructing a second objective function by combining the information such as the production line cost and the like, so that the total cost is minimized; constructing a third objective function according to the order cut-off time and the actual completion time, so that the task completion time is minimized in advance;
The priorities of the three objective functions are determined according to the current production environment and strategic targets. For example, in the case of a shortage of funds, the minimum total cost is a primary consideration; constructing a comprehensive objective function, combining the three objective functions in a weighted manner, wherein the weight distribution depends on the relative importance of each objective;
And generating an initial solution set by adopting a multi-objective optimization algorithm established based on a particle swarm optimization algorithm and a simulated annealing algorithm, and finally enabling a comprehensive objective function to tend to an optimal solution by iterative computation and generation of new particles, and selecting a production scheduling scheme with highest fitness from all non-inferior solutions as the production scheduling optimal solution.
In the embodiment, the multi-objective optimization algorithm can search and find non-inferior solutions in multiple objective dimensions, so as to realize global optimization, and help ensure that the production scheduling scheme has better performance under different objectives, rather than being excellent in a certain aspect only; the combination of different multi-objective optimization algorithms utilizes the global searching capability of the particle swarm optimization algorithm and the local searching and escaping local optimal capability of the simulated annealing algorithm; the two algorithms are combined, so that the search space can be enlarged, the possibility of finding the globally optimal solution or the non-inferior solution is improved, the characteristics of the problem are better matched, and the adaptability of the algorithm is improved.
In an alternative embodiment, constructing the integrated objective function based on the first objective function, the second objective function, and the third objective function includes:
based on the productivity prediction model, a productivity prediction result is obtained, and the to-be-completed order data is combined to construct a comprehensive objective function, wherein the formula is as follows:
Wherein ω1 represents the weight of the total completion time, N represents the total number of production lines, i represents the ith production line, t ei represents the production end time of the ith production line, ω2 represents the weight of the total cost, F i represents the cost of the ith production line, ω3 represents the weight of the task completion time in advance, M represents the total number of orders, j represents the jth order, t odj represents the latest completion time of the jth order, t oej represents the actual completion time of the jth order, c j represents the completion of the jth order, c j is 1 when t odj＜toej is completed over time, c j is 1 when t odj≥toej is completed in advance.
In the comprehensive objective function, the aim is to make the production end time of all production lines as average as possible by minimizing the maximum production end time, thereby being beneficial to improving the utilization rate of the whole production line and reducing the influence of a single bottleneck production line on the whole production progress; by minimizing the total cost, the goal is to reduce the production cost under limited resources, by optimizing the utilization of resources, reducing the energy consumption, optimizing the production plan, and the like; by minimizing the total cost of the task completion time in advance, the goal is to ensure that the production task can be completed on time, thereby helping to improve customer satisfaction and reduce the risk of out-of-date delivery; by adjusting the weight parameters, the priorities among the three sub-targets are balanced, and reasonable weight setting can be better adapted to different production environments and strategic targets so as to realize more flexible optimization.
In an alternative embodiment, solving the comprehensive objective function through a preset multi-objective optimization algorithm to obtain a production optimal solution includes:
randomly generating particle swarms, wherein each particle represents a production scheme, initializing the position and the speed of the particle, and setting the initial temperature, the cooling rate and the disturbance range;
evaluating the production scheduling scheme corresponding to the particles, calculating a fitness value by combining the comprehensive objective function, and starting updating iteration of the particle swarm:
Updating the individual optimal position of each particle, simultaneously updating the global optimal position of the particle group, updating the speed of the particle according to the individual optimal position and the global optimal position, and updating the position of the particle according to the updated speed;
After the updating iteration, perturbing the production scheduling scheme corresponding to the particles within the perturbation range to generate a new production scheduling scheme, calculating the fitness value of the new production scheduling scheme, if the fitness value is optimal, accepting the new production scheduling scheme, otherwise, determining whether to accept the new production scheduling scheme according to the temperature and the annealing probability;
And reducing the temperature according to a preset cooling rate, restarting updating iteration and disturbance of the particle swarm until the particle swarm meets a preset iteration threshold, evaluating the fitness value of the final production scheme, and selecting the highest fitness value as the optimal production solution.
Randomly generating a group of particles in a set of all production schemes, wherein each particle represents one possible production scheme, and simultaneously setting an initial temperature, a cooling rate and a disturbance range for influencing adjustment of simulated annealing and the amplitude of disturbance;
Evaluating a production scheduling scheme corresponding to each particle, calculating a fitness value by using a comprehensive objective function, measuring the performance of each particle, updating the optimal position of each particle by comparing the fitness values, updating the global optimal position of the whole particle group, and updating the speed and the position of each particle to enable the particles to be effectively searched;
After each time of updating, each particle is subjected to simulated annealing operation, and a current production scheduling scheme is disturbed to generate a new production scheduling scheme; if the fitness value of the new scheduling scheme is higher, receiving the new scheduling scheme; if the fitness value is low, determining whether to accept a new production scheduling scheme according to the current temperature and the annealing probability; reducing the temperature according to the initialized cooling rate;
repeating the iteration of the position and the speed of the particles and the simulated annealing operation of each particle until the fitness value meets a preset value or the iteration times meet a preset iteration threshold value, and ending the loop iteration;
and (3) comprehensively evaluating the final scheduling scheme, calculating the fitness value by using a comprehensive objective function, ensuring that the final scheduling scheme has good performance under a plurality of targets, and selecting the scheduling scheme with the highest fitness value from all particles as a final scheduling optimal solution.
In this embodiment, the mutual influence and combination of the particle swarm optimization algorithm and the simulated annealing algorithm can find a balance between the global and the local, so that faster global convergence and finer local search can be realized, the optimization efficiency is improved, the optimal solution is ensured to be more comprehensively found in the search space, the diversity of the solution is increased, the situation of sinking into the local optimal solution is avoided, the robustness of the algorithm is improved, the adaptability of the overall algorithm can be improved, and the optimization requirements of different stages are more suitable.
FIG. 2 is a schematic structural diagram of a system for capacity prediction and scheduling based on artificial intelligence according to an embodiment of the present invention, as shown in FIG. 2, the system includes:
The first unit is used for collecting historical production data, cleaning the historical production data, constructing a productivity prediction model, receiving the historical production data through an input layer of the productivity prediction model, analyzing the historical production data based on a time-dependent capturing layer of the productivity prediction model, capturing a time-dependent relationship, and obtaining a time-dependent result;
the second unit is used for flowing the time dependent result into a feature focusing layer of the productivity prediction model, obtaining key features by calculating focusing attention scores, and screening the key features through a gating mechanism by a hidden state analysis layer to obtain productivity prediction results;
A third unit, configured to construct a first objective function based on the productivity prediction result and combining to-be-completed order data, with minimum total completion time as a target; constructing a second objective function with the minimum total cost as a target; constructing a third objective function by taking the minimum production task completion time in advance as a target; constructing a comprehensive objective function based on the first objective function, the second objective function and the third objective function; and solving the comprehensive objective function through a preset multi-objective optimization algorithm to obtain a scheduling optimal solution.
In a third aspect of an embodiment of the present invention,
There is provided an electronic device including:
A processor;
a memory for storing processor-executable instructions;
wherein the processor is configured to invoke the instructions stored in the memory to perform the method described previously.
In a fourth aspect of an embodiment of the present invention,
There is provided a computer readable storage medium having stored thereon computer program instructions which, when executed by a processor, implement the method as described above.
The present invention may be a method, apparatus, system, and/or computer program product. The computer program product may include a computer readable storage medium having computer readable program instructions embodied thereon for performing various aspects of the present invention.
Finally, it should be noted that: the above embodiments are only for illustrating the technical solution of the present invention, and not for limiting the same; although the invention has been described in detail with reference to the foregoing embodiments, it will be understood by those of ordinary skill in the art that: the technical scheme described in the foregoing embodiments can be modified or some or all of the technical features thereof can be replaced by equivalents; such modifications and substitutions do not depart from the spirit of the invention.

Claims (13)

1. A method for capacity prediction and scheduling based on artificial intelligence, comprising:
Collecting historical production data, cleaning the historical production data, constructing a productivity prediction model, receiving the historical production data through an input layer of the productivity prediction model, analyzing the historical production data based on a time-dependent capturing layer of the productivity prediction model, capturing a time-dependent relationship, and obtaining a time-dependent result;
The time dependent result flows into a feature focusing layer of the productivity prediction model, key features are obtained through calculation of focusing attention scores, and the key features are screened out through a gating mechanism by a hidden state analysis layer;
based on the productivity prediction result, combining to-be-completed order data, and constructing a first objective function by taking the minimum total completion time as a target; constructing a second objective function with the minimum total cost as a target; constructing a third objective function by taking the minimum production task completion time in advance as a target; constructing a comprehensive objective function based on the first objective function, the second objective function and the third objective function; and solving the comprehensive objective function through a preset multi-objective optimization algorithm to obtain a scheduling optimal solution.
2. The method of claim 1, wherein resolving the historical production data based on a time-dependent capture layer of the capacity prediction model captures time-dependent relationships to obtain time-dependent results, comprising:
the time-dependent capture layer is constructed based on a long-short-time memory model;
Setting an initial hidden state and an initial unit state, extracting time-series data based on the historical production data, and performing an operation on each time step of the time-series data:
Calculating forgetting gate output, input gate output and candidate value vectors through forgetting gate and input gate, updating unit states by utilizing the forgetting gate output, the input gate output and the candidate value vectors, calculating output gate output through the output gate, calculating a hidden state of the time step by combining the output gate output based on the result of updating the unit states, and transmitting the hidden state to the next time step;
repeating the operation until all time steps of the time series data are calculated;
Setting jump connection according to the preset interval step number between the hidden states, expanding the hidden states for each time step, introducing jump weights for each jump connection, combining the jump weights based on the expanded hidden states, and obtaining a time dependent result through fusion weighting.
3. The method of claim 2, wherein operating each time step of the time series data comprises:
The time series data is operated through forget gate, input gate, output gate and cell state update, and the formula is as follows:
Wherein f t denotes a forgetting gate output at time step t, σ denotes a sigmoid activation function, W f denotes a weight of the forgetting gate, h t-1 denotes a hidden state of the previous time step, x t denotes an input of the current time step, b f denotes a bias of the forgetting gate, i t denotes an input gate output at time step t, W i denotes a weight of the input gate, b i denotes a bias of the input gate, C' t denotes a candidate value vector at time step t, W C denotes a candidate value vector weight, b C denotes a bias of the candidate value vector, C t denotes a cell state at time step t, C t-1 denotes a cell state of the previous time step, o t denotes an output of the output gate at time step t, W o denotes a weight of the output gate, b o denotes a bias of the output gate, and h t denotes a hidden state at current time step t.
4. A method according to claim 3, wherein jump connections are set between hidden states according to a preset number of steps of intervals, and the hidden states are expanded for each time step, jump weights are introduced for each jump connection, and based on the expanded hidden states, time dependent results are obtained by fusion weights in combination with the jump weights, comprising:
For each time step the hidden state is extended, the formula is as follows:
hext＝[ht;ht-skip;...;ht-n×skip]；
Wherein hex t represents an extended hidden state of the current time step t, h t represents a hidden state of the current time step t, h t-skip represents a hidden state after skip of skip steps, and n×skip represents n skip intervals;
a hopping weight is introduced for each of the hopping connections, the formula of which is as follows:
Wskip·ht-skip；
wherein W skip represents a weight matrix of the jump connection;
The time dependent result is obtained by fusion weighting, and the formula is as follows:
hcbt＝Wh·hext；
Wherein hcb t represents the time-dependent result of the current time step t, and W h represents the fusion weight matrix.
5. The method of claim 1, wherein flowing the time dependent results into a feature focusing layer of the capacity prediction model, obtaining key features by calculating focus attention scores, comprises:
the characteristic focusing layer is constructed based on a self-attention mechanism;
carrying out head separation processing on the time dependent result, and dividing the time dependent result into a plurality of heads;
Performing linear transformation on each head, calculating a focus attention score by using the scaled dot product attention, and normalizing the focus attention score by using a softmax activation function;
After weighted summation, all heads are combined in series, and the combined output is subjected to linear transformation to obtain key characteristics.
6. The method as recited in claim 5, further comprising:
performing linear transformation on each head to obtain a query Q matrix, a key K matrix and a value V matrix of the head, wherein the formula is as follows:
Wherein H represents an input hidden state sequence, Q k represents a query Q matrix of a kth head, WQ k represents a weight matrix of a query operation of the kth head on the input hidden state sequence, K k represents a key K matrix of the kth head, WK k represents a weight matrix of a key operation of the kth head on the input hidden state sequence, V k represents a value V matrix of the kth head, WV k represents a weight matrix of a value operation of the kth head on the input hidden state sequence;
The focus attention score is calculated using the scaled dot product attention, with the formula:
wherein, Represents the attention score of the kth head for the sequence position t and t ', t represents the time step corresponding to the query operation, t' represents the time step corresponding to the key operation,/>Transpose of key K matrix representing the kth head, d k represents query, key dimension:
Normalized and weighted summed using a softmax activation function, the formula is as follows:
wherein, Representing the result of normalization of focus attention score,/>Representing the output of the kth header;
the outputs of all the heads are combined as follows:
Where M t represents the combined output, concat () function represents the series connection;
based on the combined outputs, a linear transformation is performed, which is shown as follows:
H′t＝WM·Mt；
Where H' t represents the output vector and W M represents the weight matrix of the linear transformation.
7. The method of claim 1, wherein screening the key features through a gating mechanism via a hidden state parsing layer comprises:
the hidden state analysis layer is constructed based on a gating circulating unit;
And taking the key features output by the feature focusing layer as input data of the hidden state analysis layer, and operating based on each time step:
Sequentially passing through an update gate and a reset gate, calculating update gate output and reset gate output, creating a candidate hidden state based on the reset gate output, and combining the candidate hidden state based on the update gate output to form an update hidden state;
repeating the operation until all time steps of the input data are calculated, and obtaining a productivity prediction result.
8. The method of claim 7, wherein operating on a per time step basis comprises:
Calculating an update gate output and a reset gate output, creating a candidate hidden state, and updating the hidden state, wherein the formula is as follows:
Wherein z t denotes the update gate output at time step t, σ denotes the sigmoid activation function, W z denotes the weight of the update gate, v t-1 denotes the hidden state of the previous time step, y t denotes the input of the current time step, b z denotes the bias of the update gate, r t denotes the reset gate output at time step t, W r denotes the weight of the reset gate, b r denotes the bias of the reset gate, v' t denotes the candidate hidden state, W denotes the weight matrix of the candidate hidden state, b denotes the bias of the candidate hidden state, and v t denotes the updated hidden state.
9. The method of claim 1, wherein constructing a composite objective function based on the first objective function, the second objective function, and the third objective function comprises:
based on the productivity prediction model, a productivity prediction result is obtained, and the to-be-completed order data is combined to construct a comprehensive objective function, wherein the formula is as follows:
Wherein ω1 represents the weight of the total completion time, N represents the total number of production lines, i represents the ith production line, t ei represents the production end time of the ith production line, ω2 represents the weight of the total cost, F i represents the cost of the ith production line, ω3 represents the weight of the task completion time in advance, M represents the total number of orders, j represents the jth order, t odj represents the latest completion time of the jth order, t oej represents the actual completion time of the jth order, c j represents the completion of the jth order, c j is-1 when t odj＜toej is completed over time, c j is 1 when t odj≥toej is completed in advance.
10. The method of claim 1, wherein solving the integrated objective function by a preset multi-objective optimization algorithm to obtain a yield-optimal solution comprises:
randomly generating particle swarms, wherein each particle represents a production scheme, initializing the position and the speed of the particle, and setting the initial temperature, the cooling rate and the disturbance range;
evaluating the production scheduling scheme corresponding to the particles, calculating a fitness value by combining the comprehensive objective function, and starting updating iteration of the particle swarm:
Updating the individual optimal position of each particle, simultaneously updating the global optimal position of the particle group, updating the speed of the particle according to the individual optimal position and the global optimal position, and updating the position of the particle according to the updated speed;
After the updating iteration, perturbing the production scheduling scheme corresponding to the particles within the perturbation range to generate a new production scheduling scheme, calculating the fitness value of the new production scheduling scheme, if the fitness value is optimal, accepting the new production scheduling scheme, otherwise, determining whether to accept the new production scheduling scheme according to the temperature and the annealing probability;
And reducing the temperature according to a preset cooling rate, restarting updating iteration and disturbance of the particle swarm until the particle swarm meets a preset iteration threshold, evaluating the fitness value of the final production scheme, and selecting the highest fitness value as the optimal production solution.
11. A system for capacity prediction and scheduling based on artificial intelligence, comprising:
The first unit is used for collecting historical production data, cleaning the historical production data, constructing a productivity prediction model, receiving the historical production data through an input layer of the productivity prediction model, analyzing the historical production data based on a time-dependent capturing layer of the productivity prediction model, capturing a time-dependent relationship, and obtaining a time-dependent result;
the second unit is used for flowing the time dependent result into a feature focusing layer of the productivity prediction model, obtaining key features by calculating focusing attention scores, and screening the key features through a gating mechanism by a hidden state analysis layer to obtain productivity prediction results;
A third unit, configured to construct a first objective function based on the productivity prediction result and combining to-be-completed order data, with minimum total completion time as a target; constructing a second objective function with the minimum total cost as a target; constructing a third objective function by taking the minimum production task completion time in advance as a target; constructing a comprehensive objective function based on the first objective function, the second objective function and the third objective function; and solving the comprehensive objective function through a preset multi-objective optimization algorithm to obtain a scheduling optimal solution.
12. An electronic device, comprising:
A processor;
a memory for storing processor-executable instructions;
wherein the processor is configured to invoke the instructions stored in the memory to perform the method of any of claims 1 to 10.
13. A computer readable storage medium having stored thereon computer program instructions, which when executed by a processor, implement the method of any of claims 1 to 10.
CN202410329818.1A 2024-03-21 2024-03-21 Method and system for predicting and scheduling productivity based on artificial intelligence Pending CN118114836A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202410329818.1A CN118114836A (en) 2024-03-21 2024-03-21 Method and system for predicting and scheduling productivity based on artificial intelligence

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202410329818.1A CN118114836A (en) 2024-03-21 2024-03-21 Method and system for predicting and scheduling productivity based on artificial intelligence

Publications (1)

Publication Number Publication Date
CN118114836A true CN118114836A (en) 2024-05-31

Family

ID=91208359

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202410329818.1A Pending CN118114836A (en) 2024-03-21 2024-03-21 Method and system for predicting and scheduling productivity based on artificial intelligence

Country Status (1)

Country Link
CN (1) CN118114836A (en)

Similar Documents

Publication Publication Date Title
Huang et al. Addressing the loss-metric mismatch with adaptive loss alignment
Aydilek et al. A novel hybrid approach to estimating missing values in databases using k-nearest neighbors and neural networks
Sheta Software effort estimation and stock market prediction using takagi-sugeno fuzzy models
Dowlatshahi et al. Training spiking neurons with gravitational search algorithm for data classification
Yue et al. Effective, efficient and robust neural architecture search
CN117455208B (en) Injection molding production scheduling optimization method and system based on artificial intelligence
Ma et al. Feature information prediction algorithm for dynamic multi-objective optimization problems
Stankovic et al. Forecasting ethereum price by tuned long short-term memory model
Shi et al. Handling uncertainty in financial decision making: a clustering estimation of distribution algorithm with simplified simulation
Wei et al. K-NN based neuro-fuzzy system for time series prediction
Yadav et al. Software reliability prediction and optimization using machine learning algorithms: A review
Al-Laham et al. Parameter identification of a multilayer perceptron neural network using an optimized salp swarm algorithm
Memarian et al. Smart optimization with PPCR modeling in the presence of missing data, time delay and model-plant mismatch
Zhang et al. Promptst: Prompt-enhanced spatio-temporal multi-attribute prediction
Parsa et al. Accurate and accelerated neuromorphic network design leveraging a Bayesian hyperparameter pareto optimization approach
Lo Predicting software reliability with support vector machines
CN117437507A (en) Prejudice evaluation method for evaluating image recognition model
Zheng et al. Data-driven optimization based on random forest surrogate
Wang et al. Time series prediction with incomplete dataset based on deep bidirectional echo state network
CN118114836A (en) Method and system for predicting and scheduling productivity based on artificial intelligence
Janikova et al. Prediction of production line performance using neural networks
Sun et al. A decision variable classification strategy based on the degree of environmental change for dynamic multiobjective optimization
Pisa et al. Transfer Learning Suitability Metric for ANN-based Industrial Controllers
Jiang et al. A Cooperative Silicon Content Dynamic Prediction Method With Variable Time Delay Estimation in the Blast Furnace Ironmaking Process
Gholamnezhad et al. A model-based many-objective evolutionary algorithm with multiple reference vectors

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination