Batching strategy optimization method for reserve pool computing network based on parameterized channel

Abstract

The application relates to a feeding strategy optimization method based on a reserve pool computing network with parameter channels, which is applied to a multiple entry manufacturing system, and the method comprises the following steps: obtaining training bifurcation parameters and a training set, training the reserve pool computing network by taking the training bifurcation parameters as the input of a bifurcation parameter channel of the reserve pool computing network with parameter channels and taking the training set as the input of other channels of the reserve pool computing network, obtaining the trained reserve pool computing network, obtaining a sample to be predicted and a prediction bifurcation parameter, inputting the trained reserve pool computing network for prediction, obtaining a production line production state evaluation index corresponding to the prediction bifurcation parameter, selecting a bifurcation parameter corresponding to an optimal production state evaluation index as a target feeding strategy, and realizing feeding strategy optimization. Compared with the prior art, the application has the advantages that the prediction is close to the actual situation, and the production capacity can be effectively optimized.

Description

Technical Field

[0001] This invention relates to the field of manufacturing and processing control technology, and in particular to a method for optimizing feeding strategies based on a storage pool computing network with parameter channels. Background Technology

[0002] Material feeding control, situated at the forefront of the multi-entry manufacturing system scheduling framework, is the input to the manufacturing system. It determines when, how many, and what type of workpieces are fed into the system, maintaining the production line's workpiece load at the target level and influencing other types of scheduling to maximize the manufacturing system's production capacity and optimize its performance. Compared to workpiece scheduling, material feeding control has a more significant impact on the manufacturing system's operational performance and product performance indicators, playing a crucial role in improving the overall performance of the manufacturing system.

[0003] Current research on material feeding strategies for multi-entry manufacturing systems, both domestically and internationally, mainly focuses on two aspects: commonly used material feeding control strategies and improved material feeding control strategies. Commonly used material feeding strategies can be further divided into open-loop and closed-loop strategies based on whether they consider production line information. Open-loop material feeding strategies feed materials at a pre-set rate based on forecasts and demand, which is simple and easy to implement. However, because the feeding strategy cannot be changed according to actual production conditions, it cannot achieve good control results in actual production. Traditional closed-loop material feeding strategies establish a threshold based on a specific method. When this value on the production line is below the threshold, material feeding is initiated; otherwise, feeding is stopped, maintaining this value on the production line at a constant level. Compared to open-loop strategies, closed-loop strategies utilize production line information and heuristic methods when making material feeding decisions. They can dynamically change the feeding plan based on system disturbances, resulting in better performance and relative stability. However, since the thresholds determined by each strategy are not natural constants of the system, they must be derived separately for each target system, and cannot be dynamically adjusted based on the actual production line status information. In summary, neither of the two commonly used material feeding control strategies fully considers the dynamic evolution of the production line status, and thus has a certain degree of one-sidedness.

[0004] The improved feeding strategy has improved some important performance indicators compared with the commonly used feeding strategy. However, the complex decision-making mechanism often requires a lot of calculation time when it is applied to actual production, which affects the efficiency of feeding decision-making and makes it difficult to implement in practice. Moreover, when the production environment changes, new rule parameters or rule combinations need to be regenerated, which requires a lot of simulation calculation time, which contradicts the practicality.

[0005] In summary, the production status of a manufacturing system undergoes dynamic evolution, but current material feeding strategies lack a comprehensive assessment from a global perspective of the impact of the dynamic evolution of the material feeding strategy on the operational status and performance of the manufacturing system. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the existing technology by providing a feeding strategy optimization method for multi-entry manufacturing systems based on a reservoir computing network with parameter channels. By associating the feeding strategy with the actual operating conditions of the manufacturing system, the feeding strategy parameters are optimized and controlled. This method can effectively track the actual operating conditions of the manufacturing system, perform reasonable chaotic behavior analysis, and determine the feeding parameters to maximize the production capacity of the manufacturing system, thereby enabling the feeding strategy to play a better role on the production line.

[0007] The objective of this invention can be achieved through the following technical solutions:

[0008] This invention provides a feeding strategy optimization method based on a reservoir computing network with parameter channels, applicable to a multi-entry manufacturing system. The method includes the following steps:

[0009] Obtain training bifurcation parameters and training set. By using the training bifurcation parameters as input to the bifurcation parameter channels of the reservoir computing network with parameter channels, and using the training set as input to other channels of the reservoir computing network, train the reservoir computing network to obtain the trained reservoir computing network.

[0010] The sample to be predicted and the bifurcation parameters for prediction are obtained and input into the trained reserve pool computing network for prediction. The production status evaluation index of the production line corresponding to the bifurcation parameters for prediction is obtained. The bifurcation parameter corresponding to the optimal production status evaluation index is selected as the target feeding strategy to achieve feeding strategy optimization.

[0011] Among them, the bifurcation parameters used for training and the bifurcation parameters used for prediction are the same type of parameter whose values ​​are not necessarily the same.

[0012] As a preferred technical solution, the training bifurcation parameters and prediction bifurcation parameters are one or more of the following: material feeding quantity, order arrival order, material feeding time interval, material feeding quantity, material feeding order, equipment scheduling rules, and order delivery period. The production line production status evaluation index is any one or more of the following: daily processing steps, equipment utilization rate, daily processing steps, inventory level, processing cycle, work-in-process quantity, and equipment queue length.

[0013] As a preferred technical solution, the acquisition of the training set or the sample to be predicted includes the following steps:

[0014] Using Anylogic simulation software, we simulated production scenarios corresponding to different bifurcation parameters to obtain training samples or samples to be predicted, including real-time status information of the production line.

[0015] The training set is obtained based on multiple training samples.

[0016] As a preferred technical solution, the reservoir computing network includes an input layer, a hidden layer, and an output layer connected in sequence. The input layer is an m-dimensional column vector, the hidden layer includes n nodes, the input layer includes an m-1 dimensional state vector and a 1-dimensional parameter channel vector, the input data of each dimension of the state vector is connected to n / (m-1) hidden layer nodes, the data of the parameter channel vector is connected to each hidden layer node, and the output layer has one fewer node than the input layer.

[0017] As a preferred technical solution, the process of training the reservoir computing network includes the following steps:

[0018] Obtain the hidden layer weight connection matrix and input matrix, and select the optimal hyperparameters for the reserve pool computation network;

[0019] Based on the optimal hyperparameters, hidden layer weight connection matrix, and input matrix, the output matrix of the reservoir computing network is obtained, and the reservoir computing network is trained based on the output matrix.

[0020] As a preferred technical solution, the process of obtaining the hidden layer weight connection matrix and the input matrix includes the following steps:

[0021] Determine the size and sparsity of the reservoir, randomly generate a connection matrix, and obtain the hidden layer weight connection matrix that gives the reservoir echo state characteristics by scaling based on the connection matrix and the scaling factor.

[0022] Based on the preset continuous probability distribution, obtain the scaled input matrix.

[0023] As a preferred technical solution, the hyperparameters include ridge regression coefficients, spectral radius of the hidden layer weight connection matrix, average degree of the reservoir computational network, reservoir input cell size, and leakage rate.

[0024] As a preferred technical solution, the output matrix is ​​calculated using the following formula:

[0025]

[0026] Among them, U, R and These are the input time series after discarding transient states, the normalized hidden layer state sequence of the reservoir, and the sequence constructed based on R, respectively. The odd-numbered row elements of the vector R are the same as the odd-numbered row elements of the reservoir state vector R, and the even-numbered row elements are the squares of the corresponding even-numbered row elements of R. I is an n×n identity matrix, and β is a hyperparameter.

[0027] As a preferred technical solution, the process of inputting the sample to be predicted and the bifurcation parameters for prediction into the trained reservoir computing network includes the following steps:

[0028] Using any training sample from the training set as the initial condition for the preheating reservoir computing network, the sample to be predicted and the bifurcation parameters for prediction are input into the trained reservoir computing network for prediction.

[0029] As a preferred technical solution, the training of the reservoir computing network uses root mean square error to calculate the closeness between the predicted output and the expected output.

[0030] Compared with the prior art, the present invention has the following advantages:

[0031] (1) Make the impact of parameter changes on system state clearer: Considering the impact of the dynamic evolution of the material feeding strategy of the multi-entry manufacturing system on the evolution of the operating state and performance of the manufacturing system, a reserve pool computing network including bifurcation parameter channels is constructed. Since the operating state of the multi-entry manufacturing system is affected by many factors, by setting bifurcation parameter channels to predict the production status evaluation index of the production line corresponding to different parameter values, it is possible to clearly reflect the impact of different values ​​of bifurcation parameters on the system of each factor. By determining the value of the bifurcation parameters, the system performance can be optimized.

[0032] (2) Fast training speed of the reserve pool computing network: The Anylogic simulation software is used to simulate the production scenarios corresponding to different bifurcation parameters, and training samples including real-time status information of the production line are obtained. The sample set composed of the training samples is used to train the reserve pool computing network. There is no need to obtain the real-time status of the production line from the actual multi-factory manufacturing system, and the solution is easy to implement.

[0033] (3) The feeding strategy can play a better role on the production line: This method can effectively track the actual working conditions of the manufacturing system, conduct reasonable chaotic behavior analysis, and determine the feeding parameters to maximize the production capacity of the manufacturing system, so that the feeding strategy can play a better role on the production line. Attached Figure Description

[0034] Figure 1 This is a flowchart of the feeding strategy optimization method based on the storage pool computing network with parameter channels in Example 1;

[0035] Figure 2 Flowchart for training the computational network for the reservoir;

[0036] Figure 3 This is a schematic diagram of the computational network structure for the reservoir.

[0037] Figure 4 A flowchart for optimizing material feeding methods in a multi-entry manufacturing system;

[0038] Figure 5This is a comparison chart of simulation results showing the changes in the number of processing steps per day when the feeding quantity is 8 pieces per 8 hours;

[0039] Figure 6 This is a comparison chart of simulation results showing the changes in the number of processing steps per day when the feeding quantity is 12 pieces every eight hours;

[0040] Figure 7 This is a comparison chart of simulation results showing the changes in the number of daily processing steps when the material feeding quantity is 20 pieces every eight hours;

[0041] Figure 8 This is a comparison chart of simulation results showing the change in the number of processing steps per day when the material feeding quantity is 40 pieces per eight hours;

[0042] Figure 9 A comparison chart of simulation results showing the change in the number of processing steps per day when the feeding quantity is 50 pieces every eight hours;

[0043] Figure 10 This is a comparison chart of simulation results showing the changes in the number of processing steps per day when the feeding quantity is 60 pieces every eight hours;

[0044] Figure 11 The figure shows the simulation results of the daily processing steps of the production line changing with the amount of material fed. Detailed Implementation

[0045] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0046] Example 1

[0047] like Figure 1 The present embodiment provides a feeding strategy optimization method based on a storage pool computing network with parameter channels, including the following steps:

[0048] Step S1: Use production line data to train the reserve pool network to establish a reserve pool calculation network that takes input bifurcation parameters and multiple real-time states as inputs and outputs corresponding quantity prediction states.

[0049] Step S2: Optimize the material feeding strategy of the multi-entry manufacturing system using the trained reserve pool computing network and production line status data.

[0050] like Figure 3As shown, the pooled computation network with parameter channels consists of m input layer nodes, n hidden layer nodes, and l output layer nodes, where l = m-1. The input layer is an m-dimensional column vector containing m-1 dimensional state data and 1-dimensional parameter channel data. The (m-1) dimensional state data is evenly distributed across the n nodes of the hidden layer; that is, each dimension of input data is connected to n / (m-1) hidden layer nodes. The parameter channel data is connected to every node in the hidden layer. Each hidden layer node is connected to all output layer nodes.

[0051] like Figure 2 As shown, the training method for a reservoir-computing network with an additional input parameter channel includes the following steps:

[0052] Step S11, obtain the dataset;

[0053] Step S12: Determine the size and sparsity of the reserve pool, and initialize the reserve pool input matrix and the hidden layer weight connection matrix;

[0054] Step S13: Input the bifurcation parameters and training dataset into the reservoir computing network to obtain hyperparameters;

[0055] Step S14: Input the obtained hyperparameters, bifurcation parameters and training dataset into the reservoir computing network to obtain the output matrix of the reservoir computing network.

[0056] A reservoir-computing network with an additional input parameter channel adds a parameter channel to an existing reservoir-computing network. Data from this parameter channel is input to the reservoir network simultaneously with data from the input channel. The input data for this parameter channel is a controllable parameter of interest that affects system state changes, such as the amount of material fed into the system that influences the daily processing steps in the preferred technical solution. The training method for the reservoir-computing network with the parameter channel is described below, where the input matrix W... in element w ij The gain between input node i and hidden layer node j is represented by the following formula (1):

[0057]

[0058] In the following formula (2), the hidden layer weight connection matrix W r element w jkThis represents the gain between hidden layer neuron j and hidden layer neuron k. The hidden layer weight connection matrix W. r It is a large sparse square matrix of n×n dimensions, where the non-zero elements represent which neurons are connected, the direction of the connection, and the weight.

[0059]

[0060] In the following formula (3), the output matrix W out element w kt This represents the gain between hidden layer neuron k and output node t.

[0061]

[0062] In this embodiment, the input matrix W of the reservoir computing network with parameter channels in And the hidden layer weight connection matrix W r No iterative training with learning data is required; the weights of both are randomly generated from a specific distribution and kept fixed, and the output matrix W... out The only matrix that needs to be trained for the network in the reservoir calculation.

[0063] Let the training set have N t There are N samples, and the length of the prediction data is N. p The input matrix is ​​U, the output matrix is ​​V, and the hidden layer state matrix is ​​R, which can be represented by the following formulas (4)-(6):

[0064]

[0065]

[0066]

[0067] Step 1: Initialization Phase

[0068] Determine the size and sparsity of the reserve pool, and randomly generate the hidden layer weight connection matrix W0. Scale the matrix using the scaling factor ρ according to the following formula (7) to obtain W. r To ensure the stability of the reservoir network and its echo state property (ESP), ρ(W0) is called the spectroscopic radius (SR) of the internal connectivity weights of the reservoir, which is the hidden layer weight connectivity matrix W. r The eigenvalue with the largest absolute value.

[0069]

[0070] The input matrix W is generated according to a certain continuous probability distribution. inFor example, matrix W in the preferred technical solution in The elements are randomly generated in a uniform distribution of [-σ,σ], and are scaled using hyperparameters to ensure that the state of the reservoir is far from the saturation region.

[0071] The internal state of the reserve pool can be initialized arbitrarily, usually set to 0.

[0072] Step 2: Determine hyperparameters

[0073] A set of hyperparameter values ​​is randomly generated within the range and used for the calculations in the following steps.

[0074] In this embodiment, the activation function of the hidden layer is assumed to be the hyperbolic tangent function. Then, the state transition of the reservoir is described by the mapping function of the following formula (8).

[0075]

[0076] Define a new vector Its odd-numbered row elements are the same as the odd-numbered row elements of r, and its even-numbered row elements are the squares of the corresponding even-numbered row elements of r. To overcome the influence of the initial transient, it is assumed that the internal state variables are collected from time m until time P, then W out The calculation can be performed using the following formula (9).

[0077]

[0078] Among them, U, R and These are the input time series after discarding transient states, the normalized hidden layer state series of the reservoir, and the sequence constructed based on R, respectively. I is an n×n identity matrix, and β is a hyperparameter.

[0079] The input data vector u(t) is replaced by the output vector v(t) and input into the reservoir, making the entire reservoir network a closed-loop, self-evolving dynamic system. v(t) is mapped to v(t+dt) according to the following formulas (10)-(12) to obtain the current W. out Predicted time series

[0080]

[0081]

[0082]

[0083] The goal of training the pooling network is to make its output V as close as possible to the desired output V. desiredIn this embodiment, the root mean square error (RMSE) is used to calculate the degree to which the predicted output approximates the expected output and to evaluate the prediction capability of the reservoir network, as shown in the following formulas (13) and (14).

[0084] min||W out RV|| 2 (13)

[0085]

[0086] This can be summarized by the following formula (15).

[0087] W out =(M -1 ×T) T (15)

[0088] Where M represents the inputs u1(t), u2(t), ..., u N (t), where t = m, m+1, ..., P is a (P-m+1)×N dimensional matrix; T is a (P-m+1)×1 column matrix composed of the output v(t).

[0089] Finally, the optimal values ​​for the seven hyperparameters of the reservoir computing network were determined, and the set of hyperparameter values ​​with the best performance was used as the hyperparameters of the reservoir network.

[0090] Step 3: Training Phase

[0091] Using the obtained hyperparameters, calculate the output matrix W according to formulas (8) and (9). out .

[0092] Step 4: Prediction Phase

[0093] Finally, the established reservoir network can be used to perform chaotic behavior learning analysis on the target system, predicting the system's behavior under bifurcation parameters different from the training values. The obtained output matrix W is then used... out The output v(t) of the corresponding input u(t) is predicted according to the following formulas (16) and (17).

[0094] r(t+1)=(1-Δ)r(t)+αtanh[W r ·r(t)+W in u(t+1)] (16)

[0095] v(t) = W out ·r(t) (17)

[0096] In this embodiment, the process of optimizing the material feeding strategy of the multi-entry manufacturing system using the trained reservoir computing network and the state data of the production line includes the following steps:

[0097] Step S21: Select different bifurcation parameters to perform simulation and obtain samples;

[0098] Step S22: Input the acquired samples into the storage pool computing network with parameter channels for training and learning to obtain the optimized feeding strategy.

[0099] The process of optimizing the material feeding strategy of a multi-entry manufacturing system using a trained pooled computing network and production line status data includes the following steps:

[0100] Step 1: Select different bifurcation parameters and fixed performance indicators to conduct simulations and obtain samples;

[0101] Step 2: The bifurcation parameters and the real-time status corresponding to the fixed performance index in the sample are used as inputs to the reserve pool computing network with parameter channels, and the real-time status corresponding to the fixed performance index is selected as the output of the reserve pool computing network.

[0102] Step 3: Train and learn the parameters of the reservoir computing network based on the samples, the input and the output. The parameters include the input matrix, the hidden layer weight connection matrix, the output matrix and seven hyperparameters.

[0103] Step 4: Select bifurcation parameter values ​​that are different from those in the sample to perform simulation to obtain the dataset to be predicted;

[0104] Step 5: Input the bifurcation parameter values ​​and the dataset to be predicted into the trained reserve pool computing network, and use any one of the datasets used for training as the initial condition for restarting / warming up the reserve pool network to predict the production status.

[0105] Step 6: By comparing the changes in the production status of the production line under different bifurcation parameters, the material feeding strategy can be optimized.

[0106] like Figure 4 As shown, the material feeding strategy with fixed feeding time and fixed feeding quantity is selected as the optimization object, and the daily processing steps of the production line are selected as the performance index. The process of optimizing the material feeding strategy of the multi-entry manufacturing system using the trained reserve pool computing network and the state data of the production line specifically includes the following steps:

[0107] Step 1, Sample Collection

[0108] This embodiment focuses on optimizing a feeding strategy with fixed feeding intervals and fixed feeding quantities. Different simulation scenarios are designed, involving modifications to order information to change the feeding quantity. Simulations are performed on a model built using the Anylogic simulation software platform, recording real-time production line status information and performance indicators, specifically the daily processing steps. Real-time status information includes the quantity of work-in-process, daily processing steps (MOV), and the utilization rate of bottleneck equipment. These data records will serve as the training set.

[0109] Step 2: Learning Process

[0110] We select the input and output data for establishing the storage pool computational network. Here, we choose the material feeding quantity as the input for the bifurcation parameter channel, real-time daily processing steps data as other input channel data, and the predicted daily processing steps as the output.

[0111] Then, the algorithm code for calculating the reservoir network with parameter channels was implemented in MATLAB software, and sample data was added to the MATLAB code. The W value of the reservoir network was recorded through MATLAB simulation. in W r W out And seven hyperparameters. Among them, W in W r It is generated randomly, while W out Based on formulas 8 and 9, the meanings of the seven hyperparameters are shown in Table 1.

[0112] Table 1. Seven Hyperparameters of the Reservoir Network

[0113]

[0114]

[0115] Through the steps above, all parameters in the reservoir computing network with parameter channels have been determined.

[0116] Step 3: Model Application

[0117] Once the parameters of the reservoir network are determined, the feeding strategy can be optimized using the trained reservoir computing network.

[0118] A short-time simulation was conducted using a different number of feed samples than those used for training to obtain the dataset to be predicted.

[0119] By inputting the input quantity and the corresponding dataset recording short-term production status into the trained reserve pool computing network, and using any data from the training dataset as the initial condition for restarting / warming up the reserve pool network, the production status of the production line under a certain input quantity can be predicted.

[0120] By comparing the changes in the daily processing steps of the production line under different material input quantities, the material input quantity that maximizes the daily processing steps is selected as the material input strategy, thereby achieving the goal of optimizing the material input strategy.

[0121] In this embodiment, the material feeding interval is 8 hours, and the number of bifurcation parameters used for training is m=3, with values ​​of p=8, 16, and 24 (pieces). For each bifurcation parameter, each simulation lasts 2000 days, and the number of processing steps of the manufacturing system within that hour is recorded at one-hour intervals, resulting in a total of 3*47999 data points. For each bifurcation parameter, a segment N_t=2000 (approximately 833 days) is randomly selected from its corresponding 47999 data points for training, with the first wa=50 data points (approximately 2 days) serving as a warm-up period.

[0122] The simulation results are shown in Table 2.

[0123] Table 2 Simulation data of processing steps per hour

[0124]

[0125]

[0126] The feed quantity and training dataset are input into a reservoir network with parameter channels. A training reservoir computational network is built using MATLAB, and the W value of the reservoir network is recorded through MATLAB simulation. in W r W out And seven hyperparameters. The values ​​of the seven hyperparameters of the reservoir network recorded by MATLAB are shown in Table 3.

[0127] Table 3 Hyperparameter values

[0128] Hyperparameter symbols Value β <![CDATA[3.5077×10 -5 ]]> ρ 0.1384 <k> < / k> 50 W_in_a 1.9549 α 0.4579 <![CDATA[p b ]]> 8 <![CDATA[k p ]]> 0.1

[0129] After establishing the storage pool computational network, MATLAB simulations were run to compare the advantages and disadvantages of different feeding strategies with varying feeding quantities. This embodiment selected feeding quantities of 8, 12, 20, 40, 50, and 60 for prediction, and the results are attached. Figure 5-10 As shown

[0130] Figure 5-10 This indicates that the actual number of processing steps per hour fluctuates around the value predicted by the reserve pool, demonstrating that the trained reserve pool can perform chaotic behavior analysis and prediction of the hourly processing steps of the manufacturing system under a given input quantity. Figures 5 to 10 The number of processing steps per hour increased from 35 to 40 and eventually stabilized at around 45, indicating that the maximum capacity of the production line is 45 processing steps per hour. Figure 8 and Figure 9This indicates that the production line capacity is mobilized to its maximum and reaches a critical state when the amount of raw materials fed is between 40 and 50.

[0131] To facilitate observation of the critical point at which production capacity reaches saturation, we conducted production simulations with multiple material input quantities in the simulation system. Considering that the production line will enter a stable processing state approximately five days after startup, and that the daily processing steps will not change significantly after entering this stable state, we set the simulation period to 100 days to simplify statistics. To eliminate the impact of fluctuations in data during the production line's preheating phase, we selected the median of each material input quantity for plotting. The resulting material input quantity - daily processing steps curve is shown below. Figure 11 As shown.

[0132] Figure 11 This indicates that the inflection point at which production capacity reaches saturation occurs around the input quantity p = 45, at which point the daily processing steps are approximately 1090. According to... Figure 8 When p=40, the production line processes approximately 43 steps per hour. 43*24=1032. According to... Figure 9 When p=10, p=50, and p=60, the number of processing steps per hour on the production line is 45, 45*24=1080. The calculated prediction results of the reserve pool are consistent with the actual operation results of the simulated production line.

[0133] The method provided by this invention is based on a reserve pool calculation algorithm with parameter channels. It constructs a multi-entry manufacturing system material feeding strategy optimization model that takes into account the dynamic evolution of the feeding strategy and its impact on the operating status and performance of the manufacturing system. This model can effectively track the actual working conditions of the manufacturing system, perform reasonable chaotic behavior analysis, and determine the feeding parameters to maximize the production capacity of the manufacturing system, so that the feeding strategy can play a better role on the production line.

[0134] Example 2

[0135] Compared to Example 1, the bifurcation parameter in this example is the order in which orders arrive.

[0136] Example 3

[0137] Compared to Example 1, the bifurcation parameters in this example are the order of order arrival and the quantity of materials fed. Correspondingly, the number of parameter channels in the storage pool calculation network is 2. Similarly, the number of bifurcation parameters can be adjusted as needed.

[0138] Example 4

[0139] This embodiment provides an electronic device, including: one or more processors and a memory, wherein the memory stores one or more programs, the one or more programs including instructions for executing the feeding strategy optimization method based on a storage pool computing network with parameter channels as described in any of Embodiments 1-3.

[0140] Example 5

[0141] This embodiment provides a computer-readable storage medium including one or more programs executable by one or more processors of an electronic device, the one or more programs including instructions for performing a feeding strategy optimization method based on a reservoir computing network with parameter channels as described in any of Embodiments 1-3.

[0142] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for optimizing feeding strategies based on a storage pool computational network with parameter channels, characterized in that, Applied to a multi-entry manufacturing system, the method includes the following steps: Obtain training bifurcation parameters and training set. By using the training bifurcation parameters as input to the bifurcation parameter channels of the reservoir computing network with parameter channels, and using the training set as input to other channels of the reservoir computing network, train the reservoir computing network to obtain the trained reservoir computing network. The sample to be predicted and the bifurcation parameters for prediction are obtained and input into the trained reserve pool computing network for prediction. The production status evaluation index of the production line corresponding to the bifurcation parameters for prediction is obtained. The bifurcation parameter corresponding to the optimal production status evaluation index is selected as the target feeding strategy to achieve feeding strategy optimization. The aforementioned reservoir computing network comprises an input layer, a hidden layer, and an output layer connected sequentially. The input layer is an m-dimensional column vector, and the hidden layer comprises n nodes. The input layer includes an m-1 dimensional state vector and a 1-dimensional parameter channel vector. The input data for each dimension of the state vector is connected to n / (m-1) hidden layer nodes, and the data in the parameter channel vector is connected to each hidden layer node. The output layer has one fewer node than the input layer. The process of training the reservoir computing network includes the following steps: Obtain the hidden layer weight connection matrix and input matrix, and select the optimal hyperparameters for the reserve pool computation network; Based on the optimal hyperparameters, hidden layer weight connection matrix, and input matrix, the output matrix of the reservoir computing network is obtained, and the reservoir computing network is trained based on the output matrix. The process of obtaining the hidden layer weight connection matrix and the input matrix includes the following steps: Determine the size and sparsity of the reservoir, randomly generate a connection matrix, and obtain the hidden layer weight connection matrix that gives the reservoir echo state characteristics by scaling based on the connection matrix and the scaling factor. Based on a preset continuous probability distribution, obtain the scaled input matrix. The hyperparameters mentioned include ridge regression coefficients, spectral radius of the hidden layer weight connection matrix, average degree of the reservoir computation network, reservoir input cell size, and leakage rate. The output matrix is ​​calculated using the following formula: in, , and These are, respectively, the input time series after discarding transient states, the normalized reservoir hidden layer state series, and the sequence of states after discarding transient states. Sequences and vectors constructed from the basics. Odd-numbered row elements and the state vector of the reservoir The odd-numbered rows have the same elements, and the even-numbered rows have the same elements. The square of the corresponding even-numbered row element, yes The identity matrix is ​​β, where β is a hyperparameter.

2. The feeding strategy optimization method based on a storage pool computing network with parameter channels according to claim 1, characterized in that, The training bifurcation parameters and prediction bifurcation parameters are one or more of the following: material feeding quantity, order arrival order, material feeding time interval, material feeding quantity, material feeding order, equipment scheduling rules, and order delivery date. The production line production status evaluation indicators are one or more of the following: daily processing steps, equipment utilization rate, daily processing steps, inventory level, processing cycle, work-in-process quantity, and equipment queue length.

3. The feeding strategy optimization method based on a storage pool computing network with parameter channels according to claim 1, characterized in that, The acquisition of the training set or the sample to be predicted includes the following steps: Using Anylogic simulation software, we simulated production scenarios corresponding to different bifurcation parameters to obtain training samples or samples to be predicted, including real-time status information of the production line. The training set is obtained based on multiple training samples.

4. The feeding strategy optimization method based on a storage pool computing network with parameter channels according to claim 1, characterized in that, The process of inputting the sample to be predicted and the bifurcation parameters for prediction into the trained reservoir computing network includes the following steps: Using any training sample from the training set as the initial condition for the preheating reservoir computing network, the sample to be predicted and the bifurcation parameters for prediction are input into the trained reservoir computing network for prediction.

5. The feeding strategy optimization method based on a storage pool computing network with parameter channels according to claim 1, characterized in that, The training of the aforementioned reservoir computing network uses root mean square error to calculate the closeness between the predicted output and the expected output.

Images