Multi-dimensional data prediction method and system for discrete manufacturing workshop

Abstract

A multi-dimensional data prediction method and system for a discrete manufacturing workshop. The method comprises the following steps: obtaining an original data set; using a DBSCAN algorithm to perform clustering on the original data set, so as to obtain a plurality of data clusters, and establishing corresponding labels for the data clusters; searching for association relationships between the labels, so as to constitute data knowledge; deleting redundant data from all the data knowledge, so as to form a knowledge network; obtaining original data which is cleaned in step S1 and corresponds to labels included in the data knowledge; initializing the weight of a BP network structure, so as to obtain an initial model; obtaining a final prediction model; and acquiring original data of discrete manufacturing to be processed, and inputting same into the prediction model, and the prediction model outputting a prediction result. By means of a qualitative cognition method and a quantitative and qualitative two-stage cognition method, the method solves the problems of a large data volume and large dimensions of a discrete manufacturing workshop in a divide-and-rule mode, reduces the calculation cost, and can efficiently mine valuable information required for prediction.

Description

A multidimensional data prediction method and system for discrete manufacturing workshops Technical Field

[0001] The present invention relates to a data prediction method and system, and in particular to a multidimensional data prediction method and system for a discrete manufacturing workshop. Background Art

[0002] Currently, most manufacturing companies have completed the transition from manual management to information-based management, accumulating massive amounts of data related to processing, quality, equipment, and personnel during the manufacturing process. Faced with this constantly generated, high-dimensional manufacturing data, how to extract valuable information from it to provide accurate guidance for daily manufacturing operations, thereby reducing quality, energy consumption, and operating costs during the manufacturing process and improving production efficiency, has become a pressing issue.

[0003] Discrete manufacturing, in particular, features complex and diverse business scenarios, with varying production processes and methods. This generates large volumes of high-dimensional data. Currently, most approaches to mining useful information directly utilize deep learning methods to build neural network models for production factor forecasting. However, high-dimensional data can lead to redundant network nodes, increasing computational costs. Summary of the Invention

[0004] Purpose of the invention: The first purpose of the present invention is to provide a multidimensional data prediction method for discrete manufacturing workshops that can reduce computing costs and efficiently mine valuable information.

[0005] A second object of the present invention is to provide a multi-dimensional data prediction system for discrete manufacturing workshops.

[0006] Technical solution: The present invention discloses a multidimensional data prediction method for a discrete manufacturing workshop, comprising the following steps:

[0007] S1: Collect and clean the raw data of discrete manufacturing to obtain the original data set; the raw data includes product quality index data, equipment energy consumption index data, and operating cost data in the discrete manufacturing process;

[0008] S2: Use the DBSCAN algorithm to cluster the original data set, obtain multiple data clusters, and establish corresponding labels for the data clusters;

[0009] S3: Establish qualitative cognition and use the Apriori algorithm to search for the association between each label. Two or more labels are associated to form a piece of data knowledge.

[0010] S4: Delete all redundant data in the data knowledge, extract the associations between each piece of data knowledge and construct a knowledge network to complete qualitative cognition;

[0011] S5: Traverse each piece of data knowledge contained in the knowledge network in step S4, obtain the original data cleaned in step S1 corresponding to the label contained in the data knowledge, and use this part of the original data as the performance indicator sample;

[0012] S6: Using the DMTLBO-BP network model, construct the original BP network structure, initialize the weight of the BP network structure according to the performance index sample, and obtain the initial model;

[0013] S7: Train the initial model and update the weight of the initial model to obtain the final prediction model;

[0014] S8: Obtain performance indicator samples corresponding to the original data of discrete manufacturing to be processed, input them into the prediction model, and the prediction model outputs the prediction results.

[0015] Furthermore, the BP network structure in step S6 includes an input layer, two hidden layers and an output layer, wherein the number of neurons in the input layer is N m , and N m The total number of data knowledge obtained in step S3 is equal to the number of neurons in hidden layer 1. i , the number of neurons in hidden layer 2 is N k , the number of neurons in the output layer is N j , and N j The number of conclusions obtained by associating with the Apriori algorithm in step S3 is the same.

[0016] Furthermore, the steps of initializing the weights of the BP network structure according to the performance indicator samples in step S6 are as follows:

[0017] S6-1: Use real number coding to encode the performance indicator samples to obtain a coding array; the coding formula is as follows: L = N m ×(N k +N i )+N j ×(N k +N i )+(N k +N i +N j )

[0018] Where L is the code length;

[0019] S6-2: Define the encoding array as a class group, determine the class group size as L, and the number of iterations as a;

[0020] S6-3: Randomly divide the class into b groups and select the individual with the best fitness in each group as the teacher. The fitness is defined as the performance;

[0021] S6-4: Utilize the achievement differences between teachers and students in the class group to enable teachers to impart knowledge to students;

[0022] S6-5: Calculate the average scores of b groups, define the top half of the students in each group as excellent students, and the rest as poor students;

[0023] S6-6: Implement self-learning strategies for good students and mutual learning strategies for poor students;

[0024] S6-7: Each group conducts collaborative communication optimization, and selects the top k best individuals in the group based on their performance to replace the last k worst individuals in the group. This process is repeated h times until the difference between the current performance of each group and the performance of the previous iteration is less than a preset value. The calculation formula for k is: k = Gen / C;

[0025] Where Gen represents the current number of iterations, and C is a constant;

[0026] S6-8: Repeat S6-4 to S6-7 until the number of iterations is a;

[0027] S6-9: Find the individual with the best performance in the class group. The encoding of this individual is the initial weight of the BP network structure.

[0028] Furthermore, the fitness calculation formula in step S6-3 is:

[0029] where N j is the number of neurons in the output layer, d q is the expected output of the kth neuron in the output layer, y q is the actual output of the kth neuron in the output layer.

[0030] Furthermore, the calculation formula for imparting knowledge in step S6-4 is: D_M=r i ×(M taecher -TF×M mean )

[0031] where r i =rand(0,1) represents the learning step, TF = [1+rand(0,1)] represents the teaching factor, M taecher Indicates teacher performance, M mean Indicates contemporary achievements;

[0032] Students move closer to teachers according to D_M formula, X′ i,j =X i,j +D_M

[0033] where X′ i,jrepresents the performance of the i-th generation student j after teaching, X i,j represents the performance of student j of generation i before teaching.

[0034] Furthermore, the mutual learning strategy of the eugenics in step S6-6 means that the eugenics conducts N in-depth explorations on itself, and the calculation formula for the number of explorations N is as follows:

[0035] Among them F i Indicates the performance of the i-th individual in the group; F max Indicates the best score in the group, F min Indicates the worst result in the group; UL is the upper limit of the number of explorations, which can be set according to the actual situation; LL is the lower limit of the number of explorations, which can be set according to the actual situation;

[0036] The mutual learning strategy for poor students means that poor students learn from other better individuals through mutual learning between groups. p Randomly select x that is better than q To learn, the learning formula is as follows: X′ i,p =X i,p +r i (X i,p -X i,q ),F(X i,p )≤F(X i,q ) X′ i,p =X i,p +r i (X i,p -X i,q ),F(X i,p )>F(X i,q )

[0037] where X′ i,p represents individual x p The results after teaching, X i,p Represents individual x p In the results before teaching, when F(X i,p ) is better than F(X i,q ) accepts the new individual, otherwise maintains the old individual.

[0038] Furthermore, in step S7, the initial model is trained and the weight of the initial model is updated to obtain the final prediction model, which includes the following steps:

[0039] S7-1: Calculate the input value of the hidden layer 1 of the BP network structure;

[0040] S7-2: Calculate the output value of hidden layer 1 based on the input value of hidden layer 1;

[0041] S7-3: Calculate the input value of hidden layer 2 based on the output value of hidden layer 1;

[0042] S7-4: Calculate the output value of hidden layer 2 based on the input value of hidden layer 2;

[0043] S7-5: Calculate the input value of the output layer based on the output value of the hidden layer 2;

[0044] S7-6: Calculate the output value of the output layer based on the input value of the output layer;

[0045] S7-7: Calculate the back propagation error of the BP network structure based on the output value of the output layer;

[0046] S7-8: Update the weight value of the BP network structure according to the back propagation error;

[0047] S7-9: Set the number of iterations of the weight value of the BP network structure. If the number of iterations is not reached, repeat steps S7-1 to S7-8. If the number of iterations is reached, output the initial model of the current weight value. The initial model is the final prediction model.

[0048] Furthermore, the calculation formula in step S7-1 is:

[0049] in represents the data matrix composed of performance indicator samples, N is the distance from the input layer to the hidden layer 1 i ×N m Weight matrix;

[0050] The calculation formula in step S7-2 is:

[0051] Among them, g(z) represents the activation function,

[0052] The calculation formula in step S7-3 is:

[0053] in Represents N k ×N i The weight matrix of

[0054] The calculation formula in step S7-4 is:

[0055] The calculation formula in step S7-5 is:

[0056] in Represents N j ×N k The weight matrix of

[0057] The calculation formula in step S7-6 is:

[0058] The back propagation error in step S7-7 includes the output layer error and the hidden layer error, where the calculation formula of the output layer error is:

[0059] in Represents the actual output value matrix;

[0060] The calculation formula of the hidden layer error is:

[0061] in Represents the transpose of the weight matrix between the hidden layer and the output layer;

[0062] The calculation formula of the weight value in step S7-8 is:

[0063] in represents the expected output matrix, represents the expected output matrix, represents the expected output matrix, Represents the weight value of the previous iteration, ɑ represents the learning rate, ɑ∈(0,1).

[0064] Based on the same inventive concept, the present invention also discloses a multidimensional data prediction system for a discrete manufacturing workshop, comprising:

[0065] The data processing module is used to clean, cluster, label, and associate the raw data of discrete manufacturing in order to obtain the performance indicator module;

[0066] The model building module uses the DMTLBO-BP network model to build the original BP network structure, initializes the weight of the BP network structure according to the performance index sample, and obtains the initial model;

[0067] The prediction model module trains the initial model and updates the weights of the initial model to obtain the final prediction model.

[0068] Furthermore, the data processing module includes:

[0069] The data input submodule acquires the raw data of discrete manufacturing and cleans it to obtain the original data set;

[0070] The clustering classification submodule is used to cluster the original data set using the DBSCAN algorithm, obtain multiple data clusters, and establish corresponding labels for the data clusters;

[0071] The qualitative cognitive submodule uses the Apriori algorithm to search for the association between each label, and two or more labels are associated to form a piece of data knowledge;

[0072] The knowledge network construction submodule deletes redundant data from all data knowledge in the qualitative cognitive module, extracts the associations between each piece of data knowledge and constructs a knowledge network;

[0073] The data reconstruction submodule obtains the cleaned original data in the data input submodule corresponding to the labels contained in the network knowledge in the knowledge network construction submodule, and uses this part of the original data as a performance indicator sample.

[0074] Beneficial effects: Compared with the existing technology, the present invention has the following significant advantages: the present invention solves the problem of large data volume and high dimensionality in discrete manufacturing workshops in a divide-and-conquer manner through qualitative cognition and a two-stage quantitative and qualitative cognition method, reduces the computing cost, and the prediction model of the present invention can efficiently mine the valuable information required for prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] FIG1 is a flow chart of the method of the present invention;

[0076] FIG2 is a flow chart of training an initial model of the present invention;

[0077] FIG3 is a structural diagram of a BP network structure of the present invention;

[0078] FIG4 is a schematic diagram of the discrete manufacturing process of the present invention;

[0079] FIG5 is a flow chart of the system of the present invention. DETAILED DESCRIPTION

[0080] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0081] Example 1

[0082] The multidimensional data prediction method for a discrete manufacturing workshop according to the present invention, as shown in FIG1 and FIG3, comprises the following steps:

[0083] S1: Collect and clean raw discrete manufacturing data to form a raw data set. Raw data includes product quality indicators, equipment energy consumption indicators, and operating cost data from the discrete manufacturing process. Data cleaning involves removing duplicate values ​​and completing missing values ​​in the raw data.

[0084] S2: Preprocess the original data set and cluster it using the DBSCAN algorithm to obtain multiple data clusters, and then establish corresponding labels for the data clusters. Using the DBSCAN algorithm, in the original data set containing t indicator data, define the radius and the minimum number of minpts. Randomly select an indicator data with ξ as the radius to see if it contains the minimum number of minpts. If it meets the requirements, the indicator data is considered the core point. Then, all corresponding indicator data are found as new clusters. The cycle continues in this way until all indicator data in the original data set has been processed. Take the device energy consumption indicator data as an example: the device energy consumption data includes temperature, pressure, etc.; construct similar labels as shown below: device ① temperature ① pressure ①, device ① temperature ① pressure ②, device ① temperature ② pressure ①, and other similar labels.

[0085] S3: Establish qualitative cognition and use the Apriori algorithm to search for the association between each label. Two or more labels that are associated constitute a piece of data knowledge. As shown in Figure 4, take the label corresponding to the product quality indicator data as an example: Use the Apriori association algorithm to scan the labels related to the product quality indicators in step S2 into a data set, accumulate the count of each item, and mine the items that meet the minimum support. Find the set L1 of frequent 1-item sets, product 1 measurement 1 quality is excellent; L1 is used to find the set L2 of frequent 2-item sets, that is, the association rules of two different process measurements, and L2 is used to find L3, that is, the association rules of three process measurements, until no frequent n-item sets can be found. Establish qualitative cognition and establish a mathematical model for product quality indicator data. Among them, P i and K represent the conditions and conclusions of the corresponding labels, respectively. For example It means that conclusion K can be drawn from conditions P1 and P2. For example, if product 1 measurement 1 has excellent quality + product 1 measurement 2 has excellent quality + ..... product 1 measurement n has excellent quality, it can be concluded that product 1 has excellent quality.

[0086] S4: Delete all redundant data in the data knowledge, then extract the association between each piece of cleaned data knowledge and build a knowledge network to complete qualitative cognition. Deleting redundant data refers to the following two situations: ① If there are two pieces of knowledge and and satisfy Knowledge It is knowledge If there is redundancy, delete ②If there are two pieces of knowledge and and satisfy Knowledge It is knowledge If there is redundancy, delete

[0087] S5: Traverse each piece of data knowledge contained in the knowledge network in step S4, obtain the raw data cleaned in step S1 corresponding to the labels contained in the data knowledge, and use this raw data as the performance indicator sample. This step and subsequent steps involve quantitative data cognition. This embodiment uses qualitative cognition and a two-stage quantitative and qualitative cognition method to solve the problem of large data volumes and high dimensionality in discrete manufacturing workshops in a divide-and-conquer approach, reducing computational costs and efficiently mining valuable information required for prediction.

[0088] S6: Establish the DMTLBO-BP network model as the original model, construct the original BP network structure, initialize the weight of the BP network structure according to the performance index sample, and obtain the preliminary model. The BP network structure includes an input layer, two hidden layers and an output layer. The neurons in the input layer are N m , and N m The total number of data knowledge obtained in step S3 is equal to N m and The number of characteristic elements is equal to that of the hidden layer 1, and the number of neurons in the hidden layer 1 is N i , the number of neurons in hidden layer 2 is N k , the number of neurons in the output layer is N j , and N j The same number of conclusions obtained by the Apriori algorithm in step S3, that is, N j It is equal to K in step S4, as shown in Figure 3.

[0089] The steps of initializing the weights of the BP network structure according to the performance indicator samples in step S6 are as follows:

[0090] S6-1: Use real number coding to encode the performance indicator samples to obtain a coding array. The coding method is: L = N m ×(N k +N i )+N j ×(N k +N i )+(N k +N i +N j )

[0091] Where L is the code length.

[0092] S6-2: Define the encoding array as a class group, determine the class group size as L, and the number of iterations as a. The number of iterations a can be set according to the situation.

[0093] S6-3: Randomly divide the class into b groups and select the individual with the best fitness in each group as the teacher. The fitness is defined as the score. The fitness calculation formula is:

[0094] where N j is the number of neurons in the output layer, d q is the expected output of the kth neuron in the output layer, y q is the actual output of the kth neuron in the output layer.

[0095] S6-4: Utilize the differences between teachers and students in a class group to enable teachers to impart knowledge to students in the class group. The calculation formula for imparting knowledge is: D_M=r i ×(M taecher -TF×M mean )

[0096] where r i =rand(0,1) represents the learning step, TF = [1+rand(0,1)] represents the teaching factor, M taecher Indicates teacher performance, M mean Indicates contemporary achievements;

[0097] Students move closer to teachers according to D_M formula, X′ i,j =X i,j +D_M, where X′ i,j represents the performance of the i-th generation student j after teaching, X i,j represents the performance of student j of generation i before teaching.

[0098] S6-5: Calculate the average scores of b groups, define the top half of the students in each group as excellent students, and the remaining students as poor students.

[0099] S6-6: Implement the self-learning strategy for top students and the mutual learning strategy for bottom students. The self-learning strategy for top students means that top students improve their performance by conducting N in-depth explorations of themselves. The calculation formula for the number of explorations N is as follows:

[0100] Among them F i Indicates the performance of the i-th individual in the group; F max Indicates the best score in the group, F min Indicates the worst result in the group; UL is the upper limit of the number of explorations, which can be set according to the actual situation; LL is the lower limit of the number of explorations, which can be set according to the actual situation;

[0101] The mutual learning strategy for poor students means that poor students learn from other better individuals through mutual learning between groups. pRandomly select x that is better than q To learn, the learning formula is as follows: X′ i,p =X i,p +r i (X i,p -X i,q ),F(X i,p )≤F(X i,q ) X′ i,p =X i,p +r i (X i,p -X i,q ),F(X i,p )>F(X i,q )

[0102] where X′ i,p Represents individual x p The results after teaching, X i,p Represents individual x p In the results before teaching, when F(X i,p ) is better than F(X i,q ) accepts the new individual, otherwise maintains the old individual.

[0103] S6-7: Each group conducts collaborative communication optimization, selecting the top k best individuals in the group based on their scores to replace the last k worst individuals in the group. This process is repeated h times until the difference between the current score and the score of the previous iteration is less than the preset value, indicating that the scores of each group have reached a preset stable state. The formula for calculating the number of students k participating in the inter-group replacement communication is: k = Gen / C,

[0104] Where Gen represents the current number of iterations, C is a constant, and C can be set according to actual conditions.

[0105] S6-8: Determine whether the number of iterations meets the conditions, and repeat S6-4 to S6-7 until the number of iterations a is reached.

[0106] S6-9: Initialize the initial weights of the initial model and find the individual with the best performance in the class group. The encoding of this individual is the initial weight of the BP network structure.

[0107] S7: Train the initial model and update the weights of the initial model to obtain the final prediction model. This includes the following steps:

[0108] S7-1: Calculate the input value of the hidden layer 1 of the BP network structure. The calculation formula is as follows:

[0109] in represents the data matrix composed of performance indicator samples, N is the distance from the input layer to the hidden layer 1 i ×N m Weight matrix.

[0110] S7-2: Calculate the output value of hidden layer 1 based on the input value of hidden layer 1. The calculation formula is:

[0111] Among them, g(z) represents the activation function,

[0112] S7-3: Calculate the input value of hidden layer 2 based on the output value of hidden layer 1. The calculation formula is:

[0113] in, Represents N k ×N i The weight matrix, where N i Represents the number of neurons in hidden layer 1, N k Represents the number of neurons in hidden layer 2.

[0114] S7-4: Calculate the output value of hidden layer 2 based on the input value of hidden layer 2. The calculation formula is:

[0115] S7-5: Calculate the input value of the output layer based on the output value of hidden layer 2. The calculation formula is:

[0116] in Represents N j ×N k The weight matrix, where N j Represents the number of neurons in the output layer, that is, the number of label categories of the prediction factors (product quality index data, equipment energy consumption index data, and operating cost data), N k Represents the number of neurons in hidden layer 2.

[0117] S7-6: Calculate the output value of the output layer based on the input value of the output layer. The calculation formula is:

[0118] Among them, g(z) represents the activation function.

[0119] S7-7: Calculate the back propagation error of the BP network structure based on the output value of the output layer. The back propagation error includes the output layer error and the hidden layer error. The calculation formula of the output layer error is:

[0120] in Represents the actual output value matrix.

[0121] The calculation formula of the hidden layer error is:

[0122] in Represents the transpose of the weight matrix between the hidden layer and the output layer.

[0123] S7-8: Update the weight value of the BP network structure based on the back propagation error. The calculation formula is:

[0124] in represents the expected output matrix, represents the expected output matrix, represents the expected output matrix, Represents the weight value of the previous iteration, α represents the learning rate, α∈(0,1).

[0125] S7-9: Set the number of iterations of the weight value of the BP network structure. If the number of iterations is not reached, repeat steps S7-1 to S7-8. After the number of iterations is reached, output the initial model of the current weight value. The initial model is the final prediction model.

[0126] S8: Obtain performance indicator samples corresponding to the original data of discrete manufacturing to be processed, input them into the prediction model, and the prediction model outputs the prediction results.

[0127] Example 2

[0128] The multidimensional data prediction system for a discrete manufacturing workshop described in the present invention, as shown in FIG2 , includes a data processing module, a model building module, and a prediction model module.

[0129] The data processing module is used to clean, cluster, label, and associate the raw data of discrete manufacturing to obtain the performance indicator module. The data processing module includes a data input submodule, a clustering and classification submodule, a qualitative recognition submodule, a knowledge network construction submodule, and a data reconstruction submodule.

[0130] The data input submodule acquires and cleans raw discrete manufacturing data to create a raw data set. This raw data includes product quality indicators, equipment energy consumption indicators, and operating cost data from the discrete manufacturing process. Data cleaning involves deleting duplicate values ​​and completing missing values ​​in the raw data.

[0131] The clustering and classification submodule uses the DBSCAN algorithm to cluster the original dataset, obtaining multiple data clusters and assigning corresponding labels to these clusters. Using the DBSCAN algorithm, within an original dataset containing t indicators, a radius and a minimum number of minpts are defined. A randomly selected indicator, with a radius of ξ, is observed to see if it contains the minimum number of minpts. If it does, it is designated as the core point. All corresponding indicators are then identified as new clusters, and the process continues in this manner until all indicators in the original dataset have been processed. For example, device energy consumption data includes temperature, pressure, and other parameters. Similar labels are constructed, such as: Device ① Temperature ① Pressure ①, Device ① Temperature ① Pressure ②, Device ① Temperature ② Pressure ①, and so on.

[0132] The qualitative cognition submodule uses the Apriori algorithm to search for the association between each label, and two or more related labels constitute a piece of data knowledge. Take the label corresponding to the product quality indicator data as an example: using the Apriori association relationship algorithm, the labels related to the product quality indicators in the clustering classification submodule are scanned into a data set, the count of each item is accumulated, and the items that meet the minimum support are mined to find the set L1 of frequent 1-item sets, product 1 measurement 1 quality is excellent; L1 is used to find the set L2 of frequent 2-item sets, that is, the association rules of two different process measurements, and L2 is used to find L3, that is, the association rules of three process measurements, until no more frequent n-item sets can be found to establish qualitative cognition and establish a mathematical model for product quality indicator data. Among them, P i and K represent the conditions and conclusions of the corresponding labels, respectively. For example It means that conclusion K can be drawn from conditions P1 and P2. For example, if product 1 measurement 1 has excellent quality + product 1 measurement 2 has excellent quality + ..... product 1 measurement n has excellent quality, it can be concluded that product 1 has excellent quality.

[0133] The knowledge network construction submodule deletes the redundant data in all the data knowledge in the qualitative cognitive module, and then extracts the relationship between each piece of data knowledge and constructs a knowledge network. Deleting redundant data refers to the following two situations: ① If there are two pieces of knowledge and and satisfy Knowledge It is knowledge If there is redundancy, delete ②If there are two pieces of knowledge and and satisfy Knowledge It is knowledge If there is redundancy, delete

[0134] The data reconstruction submodule obtains the cleaned original data from the data input submodule corresponding to the data knowledge contained in the knowledge network in the knowledge network construction submodule, and uses this part of the original data as a performance indicator sample.

[0135] The model building module uses the DMTLBO-BP network model to build the original BP network structure, initializes the weight of the BP network structure based on the performance index sample, and obtains the initial model. The BP network structure includes an input layer, two hidden layers and an output layer. The neurons in the input layer are N m , and N m The total number of data knowledge obtained by the qualitative cognitive submodule is equal to N m and The number of characteristic elements is equal to that of the hidden layer 1, and the number of neurons in the hidden layer 1 is N i , the number of neurons in hidden layer 2 is N k , the number of neurons in the output layer is N j , and N j The same number of conclusions as those obtained by the Apriori algorithm in the qualitative cognitive submodule, that is, N j Equal to K in the knowledge network construction submodule.

[0136] Use real number coding to encode the performance index samples and obtain the encoding array. The encoding method is: L = N m ×(N k +N i )+N j ×(N k +N i )+(N k +N i +N j )

[0137] Where L is the code length.

[0138] Define the encoding array as a class group, determine the class group size as L, and the number of iterations as a. The number of iterations a can be set according to the situation.

[0139] The class population is randomly divided into b groups, and the individual with the best fitness in each group is selected as the teacher. The fitness is defined as the score. The fitness calculation formula is:

[0140] where N j is the number of neurons in the output layer, d q is the expected output of the kth neuron in the output layer, y q is the actual output of the kth neuron in the output layer.

[0141] By utilizing the differences between teachers and students in a class group, teachers can impart knowledge to students in the class group. The calculation formula for imparting knowledge is: D_M=r i ×(M taecher -TF×M mean )

[0142] where r i =rand(0,1) represents the learning step, TF = [1+rand(0,1)] represents the teaching factor, M taecher Indicates teacher performance, M mean Indicates contemporary achievements;

[0143] Students move closer to teachers according to D_M formula, X′ i,j =X i,j +D_M, where X′ i,j represents the performance of the i-th generation student j after teaching, X i,j represents the performance of student j of generation i before teaching.

[0144] Calculate the average scores of b groups, define the top half of the students in each group as excellent students, and the remaining students as poor students.

[0145] Implement the self-study strategy for top students and the mutual learning strategy for bottom students. The self-study strategy for top students means that top students improve their performance by conducting N in-depth explorations of themselves. The calculation formula for the number of explorations N is as follows:

[0146] Among them F i Indicates the performance of the i-th individual in the group; F max Indicates the best score in the group, F min Indicates the worst result in the group; UL is the upper limit of the number of explorations, which can be set according to the actual situation; LL is the lower limit of the number of explorations, which can be set according to the actual situation;

[0147] The mutual learning strategy for poor students means that poor students learn from other better individuals through mutual learning between groups. p Randomly select x that is better than q To learn, the learning formula is as follows: X′ i,p =X i,p +r i (X i,p -X i,q ),F(X i,p )≤F(X i,q ) X′ i,p =X i,p +r i (X i,p-X i,q ),F(X i,p )>F(X i,q )

[0148] where X′ i,p Represents individual x p The results after teaching, X i,p Represents individual x p In the results before teaching, when F(X i,p ) is better than F(X i,q ) accepts the new individual, otherwise maintains the old individual.

[0149] Each group conducts collaborative communication optimization, and selects the top k best individuals in the group based on their scores to replace the last k worst individuals in the group. This process is repeated h times until the difference between the current score and the score of the previous iteration is less than the preset value, that is, the scores of each group tend to the preset stable state. The formula for calculating the number of students k participating in the inter-group replacement exchange is: k = Gen / C,

[0150] Where Gen represents the current number of iterations, C is a constant, and C can be set according to actual conditions.

[0151] Determine whether the number of iterations meets the conditions and repeat the iteration until it reaches a times.

[0152] Initialize the initial weights of the initial model and find the individual with the best performance in the class group. The encoding of this individual is the initial weight of the BP network structure.

[0153] The prediction model module trains the initial model and updates the weights of the initial model to obtain the final prediction model.

[0154] Calculate the input value of the hidden layer 1 of the BP network structure. The calculation formula is as follows:

[0155] in represents the data matrix composed of performance indicator samples, N is the distance from the input layer to the hidden layer 1 i ×N m Weight matrix.

[0156] The output value of hidden layer 1 is calculated based on the input value of hidden layer 1. The calculation formula is:

[0157] Among them, g(z) represents the activation function,

[0158] The input value of hidden layer 2 is calculated based on the output value of hidden layer 1. The calculation formula is:

[0159] in, Represents N k ×N i The weight matrix, where N i Represents the number of neurons in hidden layer 1, N k Represents the number of neurons in hidden layer 2.

[0160] The output value of hidden layer 2 is calculated based on the input value of hidden layer 2. The calculation formula is:

[0161] The input value of the output layer is calculated based on the output value of the hidden layer 2. The calculation formula is:

[0162] in Represents N j ×N k The weight matrix, where N j Represents the number of neurons in the output layer, that is, the number of label categories of the prediction factors (product quality index data, equipment energy consumption index data, and operating cost data), N k Represents the number of neurons in hidden layer 2.

[0163] The output value of the output layer is calculated based on the input value of the output layer. The calculation formula is:

[0164] Among them, g(z) represents the activation function.

[0165] The back propagation error of the BP network structure is calculated based on the output value of the output layer. The back propagation error includes the output layer error and the hidden layer error. The calculation formula of the output layer error is:

[0166] in Represents the actual output value matrix.

[0167] The calculation formula of the hidden layer error is:

[0168] in Represents the transpose of the weight matrix between the hidden layer and the output layer.

[0169] The weight value of the BP network structure is updated according to the back propagation error. The calculation formula is:

[0170] in represents the expected output matrix, represents the expected output matrix, represents the expected output matrix, Represents the weight value of the previous iteration, ɑ represents the learning rate, ɑ∈(0,1).

[0171] The number of iterations of the weight value of the BP network structure is set. When the number of iterations is reached, the initial model of the current weight value is output. The initial model is the final prediction model.

[0172] When using the multidimensional data prediction system of the discrete manufacturing workshop, the original data of the discrete manufacturing to be processed is input into the data processing module to obtain the corresponding performance index module. The performance index module is input into the prediction module as input data, and the prediction module can output the prediction result.

Claims

1. A multidimensional data prediction method for a discrete manufacturing workshop, characterized by: The following steps are included: S1: Collect and clean the raw data of discrete manufacturing to obtain the original data set; the raw data includes product quality index data, equipment energy consumption index data, and operating cost data in the discrete manufacturing process; S2: Use the DBSCAN algorithm to cluster the original data set, obtain multiple data clusters, and establish corresponding labels for the data clusters; S3: Establish qualitative cognition and use the Apriori algorithm to search for the association between each label. Two or more labels are associated to form a piece of data knowledge. S4: Delete all redundant data in the data knowledge, extract the associations between each piece of data knowledge and construct a knowledge network to complete qualitative cognition; S5: Traverse each piece of data knowledge contained in the knowledge network in step S4, obtain the original data cleaned in step S1 corresponding to the label contained in the data knowledge, and use this part of the original data as the performance indicator sample; S6: Using the DMTLBO-BP network model, construct the original BP network structure, initialize the weight of the BP network structure according to the performance index sample, and obtain the initial model; The BP network structure in step S6 includes an input layer, two hidden layers and an output layer, wherein the number of neurons in the input layer is N. m , and N m The total number of data knowledge obtained in step S3 is equal to the number of neurons in hidden layer 1. i , the number of neurons in hidden layer 2 is N k , the number of neurons in the output layer is N j , and N j The same number of conclusions as those obtained by the Apriori algorithm in step S3; S6-1: Use real number coding to encode the performance indicator samples to obtain a coding array; the coding formula is as follows: L=N m ×(N k +N i )+N j ×(N k +N i )+(N k +N i +N j ) Where L is the code length; S6-2: Define the encoding array as a class group, determine the class group size as L, and the number of iterations as a; S6-3: Randomly divide the class into b groups and select the individual with the best fitness in each group as the teacher. The fitness is defined as the performance; The fitness calculation formula is: where N j is the number of neurons in the output layer, d q is the expected output of the kth neuron in the output layer, y q is the actual output of the kth neuron in the output layer; S6-4: Utilize the achievement differences between teachers and students in the class group to enable teachers to impart knowledge to students; The calculation formula for imparting knowledge is: D_M=r i ×(M taecher -TF×M mean ) where r i =rand(0,1) represents the learning step, TF = [1+rand(0,1)] represents the teaching factor, M taecher Indicates teacher performance, M mean Indicates contemporary achievements; Students move closer to teachers according to the D_M method, X′ i,j =X i,j +D_M where X′ i,j represents the performance of the i-th generation student j after teaching, X i,j represents the performance of student j in generation i before teaching; S6-5: Calculate the average scores of b groups, define the top half of the students in each group as excellent students, and the rest as poor students; S6-6: Implement self-learning strategies for good students and mutual learning strategies for poor students; The mutual learning strategy of eugenics means that eugenics conducts N in-depth explorations of itself. The calculation formula for the number of explorations N is as follows: Among them F i Indicates the score of the i-th individual in the group; F max Indicates the best score in the group, F min Indicates the worst result in the group; UL is the upper limit of the number of explorations, which can be set according to the actual situation; LL is the lower limit of the number of explorations, which can be set according to the actual situation; The mutual learning strategy for poor students means that poor students learn from other better individuals through mutual learning between groups. p Randomly select x that is better than q To learn, the learning formula is as follows: X′ i,p =X i,p +r i (X i,p -X i,q ),F(X i,p )≤F(X i,q ) X′ i,p =X i,p +r i (X i,p -X i,q ),F(X i,p )>F(X i,q ) where X′ i,p Represents individual x p The results after teaching, X i,p Represents individual x p In the results before teaching, when F(X i,p ) is better than F(X i,q ) accept the new individual, otherwise maintain the old individual; S6-7: Each group conducts collaborative communication optimization, and selects the top k best individuals in the group based on their scores to replace the last k worst individuals in the group. This process is repeated h times until the difference between the current score of each group and the score of the previous iteration is less than the preset value. The calculation formula for k is: k = Gen / C; Where Gen represents the current number of iterations, and C is a constant; S6-8: Repeat S6-4 to S6-7 until the number of iterations is a; S6-9: Find the individual with the best performance in the class group. The encoding of this individual is the initial weight of the BP network structure. S7: Train the initial model and update the weight of the initial model to obtain the final prediction model; S7-1: Calculate the input value of the hidden layer 1 of the BP network structure; The calculation formula is: in represents the data matrix composed of performance indicator samples, N is the distance from the input layer to the hidden layer 1 i ×N m Weight matrix; S7-2: Calculate the output value of hidden layer 1 based on the input value of hidden layer 1; The calculation formula is: Among them, g(z) represents the activation function, S7-3: Calculate the input value of hidden layer 2 based on the output value of hidden layer 1; The calculation formula is: in Represents N k ×N i The weight matrix of S7-4: Calculate the output value of hidden layer 2 based on the input value of hidden layer 2; The calculation formula is: S7-5: Calculate the input value of the output layer based on the output value of the hidden layer 2; The calculation formula is: in Represents N j ×N k The weight matrix of S7-6: Calculate the output value of the output layer based on the input value of the output layer; The calculation formula is: S7-7: Calculate the back propagation error of the BP network structure based on the output value of the output layer; The back propagation error includes the output layer error and the hidden layer error, where the calculation formula of the output layer error is: in Represents the actual output value matrix; The calculation formula of the hidden layer error is: in Represents the transpose of the weight matrix between the hidden layer and the output layer; S7-8: Update the weight value of the BP network structure according to the back propagation error; The calculation formula of the weight value is: in represents the expected output matrix, represents the expected output matrix, represents the expected output matrix, Represents the weight value of the previous iteration, ɑ represents the learning rate, α∈(0,1); S7-9: Set the number of iterations of the weight value of the BP network structure. If the number of iterations is not reached, repeat steps S7-1 to S7-8. When the number of iterations is reached, output the initial model of the current weight value. The initial model is the final prediction model. S8: Obtain performance indicator samples corresponding to the original data of discrete manufacturing to be processed, input them into the prediction model, and the prediction model outputs the prediction results.

2. A multidimensional data prediction system for discrete manufacturing workshops, characterized by: include: The data processing module is used to clean, cluster, label, and associate the raw data of discrete manufacturing in order to obtain the performance indicator module; The data processing module includes: The data input submodule acquires the raw data of discrete manufacturing and cleans it to obtain the original data set; The clustering classification submodule is used to cluster the original data set using the DBSCAN algorithm, obtain multiple data clusters, and establish corresponding labels for the data clusters; The qualitative cognitive submodule uses the Apriori algorithm to search for the association between each label, and two or more labels are associated to form a piece of data knowledge; The knowledge network construction submodule deletes redundant data from all data knowledge in the qualitative cognitive module, extracts the associations between each piece of data knowledge and constructs a knowledge network; The data reconstruction submodule obtains the cleaned raw data from the data input submodule corresponding to the labels contained in the network knowledge in the knowledge network construction submodule, and uses this part of the raw data as a performance indicator sample; The model building module uses the DMTLBO-BP network model to build the original BP network structure, initializes the weight of the BP network structure according to the performance index sample, and obtains the initial model; The BP network structure in the model building module includes an input layer, two hidden layers and an output layer, where the number of neurons in the input layer is N. m , and N m The total number of data knowledge obtained in step S3 is equal to the number of neurons in hidden layer 1. i , the number of neurons in hidden layer 2 is N k , the number of neurons in the output layer is N j , and N j The same number of conclusions as those obtained by the Apriori algorithm in step S3; Initializing the weights of the BP network structure based on the performance index sample to obtain the initial model includes the following steps: I-1: Use real number coding to encode the performance indicator samples to obtain a coding array; the coding formula is as follows: L=N m ×(N k +N i )+N j ×(N k +N i )+(N k +N i +N j ) Where L is the code length; I-2: Define the encoding array as a class group, determine the class group size as L, and the number of iterations as a; I-3: Randomly divide the class into b groups and select the individual with the best fitness in each group as the teacher. The fitness is defined as the grade. The fitness calculation formula is: where N j is the number of neurons in the output layer, d q is the expected output of the kth neuron in the output layer, y q is the actual output of the kth neuron in the output layer; I-4: Utilize the achievement differences between teachers and students in the class group to enable teachers to impart knowledge to students; The calculation formula for imparting knowledge is: D_M=r i ×(M taecher -TF×M mean ) where r i =rand(0,1) represents the learning step, TF = [1+rand(0,1)] represents the teaching factor, M taecher Indicates teacher performance, M mean Indicates contemporary achievements; Students move closer to teachers according to the D_M method, X′ i,j =X i,j +D_M where X′ i,j represents the performance of the i-th generation student j after teaching, X i,j represents the performance of student j in generation i before teaching; I-5: Calculate the average scores of b groups, define the top half of the students in each group as excellent students, and the rest as poor students; I-6: Implement self-learning strategies for top students and mutual learning strategies for bottom students; The mutual learning strategy of eugenics means that eugenics conducts N in-depth explorations of itself. The calculation formula for the number of explorations N is as follows: Among them F i Indicates the score of the i-th individual in the group; F max Indicates the best score in the group, F min Indicates the worst result in the group; UL is the upper limit of the number of explorations, which can be set according to the actual situation; LL is the lower limit of the number of explorations, which can be set according to the actual situation; The mutual learning strategy for poor students means that poor students learn from other better individuals through mutual learning between groups. p Randomly select x that is better than q To learn, the learning formula is as follows: X′ i,p =X i,p +r i (X i,p -X i,q ),F(X i,p )≤F(X i,q ) X′ i,p =X i,p +r i (X i,p -X i,q ),F(X i,p )>F(X i,q ) where X′ i,p Represents individual x p The results after teaching, X i,p Represents individual x p In the results before teaching, when F(X i,p ) is better than F(X i,q ) accept the new individual, otherwise maintain the old individual; I-7: Each group conducts collaborative communication optimization. The top k best individuals in the group are selected based on their scores to replace the last k worst individuals in the group. This process is repeated h times until the difference between the current score of each group and the score of the previous iteration is less than the preset value. The calculation formula for k is: k = Gen / C; Where Gen represents the current number of iterations, and C is a constant; I-8: Repeat S6-4 to S6-7 until the number of iterations is a; I-9: Find the best individual in the class. The encoding of this individual is the initial weight of the BP network structure. Obtaining the final prediction model includes the following steps: II-1: Calculate the input value of hidden layer 1 of BP network structure; The calculation formula is: in represents the data matrix composed of performance indicator samples, N is the distance from the input layer to the hidden layer 1 i ×N m Weight matrix; II-2: Calculate the output value of hidden layer 1 based on the input value of hidden layer 1; The calculation formula is: Among them, g(z) represents the activation function, II-3: Calculate the input value of hidden layer 2 based on the output value of hidden layer 1; The calculation formula is: in Represents N k ×N i The weight matrix of II-4: Calculate the output value of hidden layer 2 based on the input value of hidden layer 2; The calculation formula is: II-5: Calculate the input value of the output layer based on the output value of hidden layer 2; The calculation formula is: in Represents N j ×N k The weight matrix of II-6: Calculate the output value of the output layer based on the input value of the output layer; The calculation formula is: II-7: Calculate the back propagation error of the BP network structure based on the output value of the output layer; The back propagation error includes the output layer error and the hidden layer error, where the calculation formula of the output layer error is: in Represents the actual output value matrix; The calculation formula of the hidden layer error is: in Represents the transpose of the weight matrix between the hidden layer and the output layer; II-8: Update the weight value of the BP network structure according to the back propagation error; The calculation formula of the weight value is: in represents the expected output matrix, represents the expected output matrix, represents the expected output matrix, Represents the weight value of the previous iteration, α represents the learning rate, α∈(0,1); II-9: Set the number of iterations of the weight value of the BP network structure. If the number of iterations is not reached, repeat steps S7-1 to S7-8. After the number of iterations is reached, output the initial model of the current weight value. The initial model is the final prediction model.

Images