A work time prediction method based on process clustering

Abstract

The application discloses a kind of time prediction methods based on process clustering, it is related to the research field of time series prediction.The time of processing target process of specific target team is predicted, which provides reference basis for the selection of scheduling scheme and the evaluation of production cycle.According to the clustering results of the process, the different types of process are fitted with the corresponding work time data, and the learning curve model associated with the cumulative processing quantity of the team, the personnel shift system and the team size is obtained respectively.After determining the target process and the target team, the target process category is calculated according to the clustering results, and the three-factor learning curve of the target team under the task of this category is used to predict the target work time.

Description

TECHNICAL FIELD

[0001] The present application relates to the research field of time series prediction, and more particularly, to a process clustering method and a time series prediction method based on learning curve. BACKGROUND

[0002] In most electronic product workshops, product processing needs to go through multiple complex processes. Such processes need to be completed by manual and machine processing, and the processing time is closely related to factors such as the degree of human participation, the proficiency of personnel operation, the health status of personnel, the mood status, and the working environment. Due to factors such as large personnel mobility in the workshop and different employee learning efficiencies, it is difficult to accurately predict the processing time. At present, most electronic workshops mainly use the way of manual rough estimation to formulate working hours, and the scheduling scheme and production cycle evaluation made based on such inaccurate working hour prediction will have a large deviation from the actual situation, thereby affecting the grasp of the production rhythm.

[0003] The existing working hour prediction research for semi-manual processes is mostly limited to predicting one or a type of process, which is not suitable for the actual electronic product precision workshop where there are many types of processes, and the operation difficulty and employee participation degree are different. SUMMARY

[0004] The purpose of the present application is to provide a working hour prediction method for electronic product precision semi-manual and semi-automatic production workshops, and to provide a reference for arranging production schemes and estimating production cycles in the workshop.

[0005] The present application proposes a working hour prediction method for electronic product precision semi-manual and semi-automatic production workshops. The method first realizes the clustering of processes by analyzing process characteristics. Then, considering the cumulative processing quantity of the team, the personnel shift system, and the number of team members, the learning efficiency is comprehensively evaluated, and a learning curve model associated with the three factors is derived from the traditional CD production function. Through collecting and regression analyzing the relevant historical working hour data of each type of process, the coefficients in each type of learning curve model are solved, and a corresponding learning curve model database is established. Combined with the process clustering results and the three-factor learning curve model specific to this category, the working hour prediction of the newly arrived process processed by the specified team can be realized.

[0006] Thus, the technical scheme of the present application is a working hour prediction method based on process clustering, which comprises:

[0007] Step 1: clustering processes based on self-organizing feature mapping neural network;

[0008] Step 1.1: collecting the names and process characteristics of all processing processes from enterprise CAPP and ERP systems, and quantifying the characteristics;

[0009] Step 1.2: All the characteristics of the process are arranged into an n-dimensional vector form and normalized as input samples of the SOFM network, and the SOFM network represents a self-organizing feature map network;

[0010] Step 1.3: The SOFM network is designed according to the characteristics of the input data, including the number of input layer neuron nodes, the number of output layer nodes, the learning rate and the number of learning times;

[0011] Step 1.4: The SOFM network determines the winning node and the dominant neighborhood by calculating the distance between all neuron nodes and input samples, and adjusts the weight of the neuron nodes in the dominant neighborhood; when the learning rate of the winning node is less than a fixed value, the SOFM network training process is ended, and the clustering result of the input n-dimensional vector is obtained;

[0012] Step 2: Derive and fit a three-factor learning curve model to predict the target process processing time;

[0013] Step 2.1: Derive a learning curve model associated with the cumulative processing quantity of the team, the personnel shift system and the number of team members from the Cobb-Douglas production function;

[0014] Step 2.2: Collect and process the time data of various types of processes processed by different teams, and perform regression analysis for different types of processes respectively to solve the coefficients in the learning curve model, obtain the specific learning curve model for each type of process, and store in the database;

[0015] Step 2.3: Determine the process to be processed, classify it according to the SOFM network, and obtain the corresponding learning curve model from the database;

[0016] Step 2.4: Determine the operation team of the process, and substitute the cumulative processing quantity of the team, the personnel shift system and the number of team members into the learning curve model of the corresponding type to obtain the target time.

[0017] Further, the specific method of 1.3 is:

[0018] Step 1.3.1: Input and output layer node number setting;

[0019] The input layer node number is set to 5, and the output layer is finally determined to be a two-dimensional grid structure composed of 80 five-dimensional nodes, 8*10;

[0020] Step 1.3.2: Output layer neuron initialization;

[0021] Random small numbers in the interval [0, 1] are used to assign initial values to each neuron component of the output layer, and the output layer neuron initialization is obtained j = 1, 2, …, m, m is the number of output layer neurons;

[0022] Step 1.3.3: weight initialization;

[0023] The connection weights {w ij , i = 1,..., n, j = 1,..., m} between n input nodes and m output nodes are initialized with random numbers in the interval [0, 1], and the weight vector between input node i and output node j is denoted as W j = [W 1j , W 2j ,..., W nj ];

[0024] Step 1.3.4: winner neighborhood setting;

[0025] During the initial training, the winner neighborhood is set as an 8*8 grid centered on the winning neuron, and as the number of iterations increases, the side length d of the winner neighborhood n j* (t) gradually decreases according to formula (2);

[0026] d(t) = d(0)(1 - t / t n ) (2)

[0027] where r(0) is the initial side length of the winner neighborhood, t is the current iteration number, and t n is the total number of iterations;

[0028] Step 1.3.5: setting the learning rate of the winning node;

[0029] During the initial training, the learning rate a(0) is set to 0.9, and as the number of iterations t increases, the learning rate a(t) gradually decreases according to formula (3);

[0030] a(t) = a(0)(1 - t / t n ) (3)

[0031] where a(0) is the initial learning rate, t is the current iteration number, and t n is the total number of iterations.

[0032] The specific method of step 1.4 is:

[0033] Step 1.4.1: similarity comparison between input vector and all competitive layer neurons;

[0034] Select the input vector and the competitive layer neuron According to formula (4), calculate the Euclidean distance between them, and the shorter the Euclidean distance, the more similar the input vector is to the neuron;

[0035]

[0036] Step 1.4.2: Select the winning neuron;

[0037] The input vector is compared with all neurons in the competition layer according to formula (4), and the neuron with the highest similarity is selected as the winning neuron according to formula (5)

[0038]

[0039] Step 1.4.3: Select the winning neighborhood and adjust the weights of the neurons in the winning neighborhood;

[0040] Calculate d(t) as the side length of the square winning neighborhood at time t according to formula (2), and determine the current winning neighborhood N j* (t) with the current winning neuron as the center;

[0041] Calculate the connection weights between all nodes and the input vector according to formula (6), wherein the weights of the nodes not belonging to the winning neighborhood N j* (t) do not change;

[0042]

[0043] In the formula, the learning rate a(t) is related to the distance between the competition node and the winning neuron, and the greater the distance from the winning neuron, the smaller a(t) is; after t iterations, the kth layer node in the winning neighborhood of the winning neuron j* is represented by The distance from all nodes in each layer to the winning node is the same, and the distance between each layer node and the winning node increases with the increase of the layer level;

[0044] According to formula (3), the learning rate a * of the winning node j j* is calculated, and the learning rate of the kth layer node in the winning neighborhood N j* changes according to formula (7), and the farther the node is from the winning node, the lower the learning rate is;

[0045]

[0046] In the formula, K is the maximum layer level in the current winning neighborhood;

[0047] Step 1.4.4: End the training and obtain the clustering result of the input procedure vector;

[0048] The learning ends when the learning rate of the winning node is less than the set threshold or the number of iterations reaches the set number, and the mapping position of all input procedure vectors in the output layer two-dimensional grid structure after training is obtained, thereby realizing the clustering division of the procedure.

[0049] Further, the step 2.2 various types of process learning curve model y=kx f(l，s) .

[0050] The present application selects process characteristics by on-site investigation of semi-manual and semi-automatic production workshops, uses SOFM network to cluster processes, separately fits specific learning curves for each type of process, so that the curve fitting result is more detailed and accurate, and more in line with the actual situation of various types of processes in the semi-manual and semi-automatic production workshop of electronic product precision parts, and the operation difficulty and employee participation are different. At the same time, clustering multiple processes into one category can also avoid the problem of insufficient processing time data of a single process. In addition, the present application associates the learning curve model with three factors of team cumulative processing quantity, personnel shift system and team size, so that the differences between individual processing teams can be considered when predicting processing time, making the prediction result more accurate, and also helping to reasonably arrange and distribute the processing teams in subsequent production scheduling and other related work. BRIEF DESCRIPTION OF DRAWINGS

[0051] Figure 1 The core step schematic diagram of the present application is a process clustering-based processing time prediction method;

[0052] Figure 2 The work flow chart of the present application is a process clustering-based processing time prediction method;

[0053] Figure 3 The clustering algorithm flow chart in the specific embodiment of the present application;

[0054] Figure 4 The winning neighborhood level division example diagram in the present application;

[0055] Figure 5 The three-factor learning curve model processing time prediction flow chart of the present application;

[0056] Figure 6 Comparison curve diagram of actual processing time and processing time prediction result of thermoacoustic welding 80. DETAILED DESCRIPTION

[0057] The present application will be described in further detail below in conjunction with the accompanying drawings:

[0058] In conjunction with Figure 1 A process clustering-based processing time prediction method includes the following steps:

[0059] S0FM network training phase:

[0060] Step 1, collect all the relevant characteristic information of the processing process from the enterprise CAPP, ERP system, and arrange it as the input vector of the SOFM network;

[0061] 1.1 Process feature vector generation

[0062] The evaluation index of the machining process includes three quantitative indexes and two qualitative indexes. The quantitative indexes include the number of process operators, the process belonging to the reporting production line, and the process version, wherein the reporting production line includes the component pre-integration line, the component integration line, and the SIP integration line. The qualitative indexes include the process automation degree and the employee operation difficulty, which are quantified by adopting the five-point system scoring. A set of five-dimensional feature vectors is generated for each process according to the quantified process features as the training set of the SOM network, and the vector form is

[0063] X = [X1, X2, X3, X4, X5], wherein X1 represents the number of process operators, X2 represents the process belonging to the reporting production line, X3 represents the process version number, X4 represents the process automation degree score, and X5 represents the employee operation process difficulty score.

[0064] 1.2 Process feature data collection

[0065] The reporting examples are derived from the APS system of the factory, and the data is screened and cleaned to obtain the quantitative feature data of all processes. Through on-site investigation and employee visits, the qualitative feature data of all processes is obtained.

[0066] 1.3 Data normalization processing

[0067] According to formula (1), each component X K of all process feature vectors is normalized and mapped to the interval [0, 1] to obtain n is the number of training set feature vectors.

[0068]

[0069] Step 2, SOFM network design: The input and output layer node numbers, output layer neurons, initial weights, winning neighborhood, and learning rate of the SOFM network are set.

[0070] 2.1 Input and output layer node number setting

[0071] After on-site investigation of the semi-automated production workshop, the characteristics of all processes are analyzed and screened to determine three quantitative indexes and two qualitative indexes. After completing the data collection of all process feature indexes, the data is screened and cleaned, and a set of five-dimensional feature vectors is generated for each process according to the quantified process features, and the vector form is X = [X1, X2, X3, X4, X5]. Each component X KAll are normalized to map to the interval [0, 1], and get the process feature matrix of size m*5. Since the input process feature vector is a five-dimensional vector, the number of input layer nodes is set to 5; after calculation and adjustment, the output layer is finally determined as an 8*10 two-dimensional grid structure composed of 80 five-dimensional nodes.

[0072] 2.2 Output layer neuron initialization

[0073] Assign initial values to each neuron component of the output layer with random numbers in the interval [0, 1], and get m is the number of output layer neurons.

[0074] 2.3 Weight initialization

[0075] Assign initial values to the connection weights {w ij , i = 1,..., n, j = 1,..., m} between n input nodes and m output nodes with random numbers in the interval [0, 1]. The weight vector between output node j and input node is represented by W j = [W 1j , W 2j ,..., W nj ].

[0076] 2.4 Winner neighborhood setting

[0077] During initial training, the winner neighborhood is set to an 8*8 grid centered on the winning neuron, and as the number of iterations increases, the side length d of the winner neighborhood N j* (t) gradually decreases according to formula (2).

[0078] d(t) = d(0)(1-t / t n ) (2)

[0079] Where: r(0) is the initial winner neighborhood side length; t is the current iteration number; t n is the total iteration number.

[0080] 2.5 Winner node learning rate setting

[0081] During initial training, the learning rate a(0) is set to 0.9, and as the number of iterations t increases, the learning rate a(t) gradually decreases according to formula (3).

[0082] a(t) = a(0)(1-t / t n ) (3)

[0083] Where: a(0) is the initial learning rate; t is the current iteration number; t n is the total iteration number.

[0084] Step 3, randomly select an input vector as the input layer of SOFM, find the neuron with the shortest distance as the winning node by calculating the distance between it and each competitive layer neuron.

[0085] 3.1 Similarity comparison between input vector and all competitive layer neurons

[0086] Select input vector and competitive layer neurons According to formula (4), calculate the Euclidean distance between them, the shorter the Euclidean distance, the more similar the input vector and the neuron.

[0087]

[0088] 3.2 Select the winning neuron

[0089] After similarity comparison between input vector and all neurons in competitive layer according to formula (4), select the neuron with the highest similarity as the winning neuron according to formula (5)

[0090]

[0091] 3.3 Select the winning neighborhood and adjust the node weight of neurons in the winning neighborhood

[0092] According to formula (2), calculate d(t) as the side length of the square winning neighborhood at time t, and determine the current winning neighborhood N j* (t) with the current winning neuron as the center.

[0093] According to formula (6), calculate the connection weight between all nodes and input vector, where the node weight outside the winning neighborhood N j* (t) does not change.

[0094]

[0095] In the formula, the learning rate a(t) is related to the distance between the competitive node and the winning neuron, and the larger the distance from the winning neuron, the smaller a(t). In the winning neighborhood N j* (t), the nodes are divided into distance levels in the order of Figure 4 .

[0096] Figure 4 Take the winning neighborhood with a side length of 5 nodes as an example to show the hierarchical division method in the winning neighborhood. After t iterations, in the winning neighborhood of the winning neuron j*, the kth level node is represented by The winning node is the node with the shortest distance to the input node. The distance from each node in each layer to the winning node is the same, and the distance between each layer node and the winning node increases with the increase of the layer level.

[0097] The winning node j can be calculated according to formula (3) * The learning rate a of each node in the winning neighborhood N j* The learning rate of the kth layer node in the winning neighborhood N j* varies according to formula (7), and the farther the node is from the winning node, the lower the learning rate.

[0098]

[0099] where K is the maximum layer level in the current winning neighborhood.

[0100] 3.4 End of training and obtain input process vector clustering results

[0101] When the learning rate of the winning node is less than 0.1 or the number of iterations reaches 10,000, the learning is ended, and the mapping positions of all input process vectors in the output layer two-dimensional grid structure after the training is stable are obtained, thereby realizing the clustering division of the processes.

[0102] Regression solution learning curve:

[0103] Step 1, three-factor learning curve model derivation

[0104] In the background of the electronic product precision part production workshop, the processing team is only composed of employees of the same shift, and all employees participating in processing are divided into two types of formal employees and outsourcing employees. The salary of the formal employee is settled fixedly every month, and the amount of salary is also a fixed value, which has nothing to do with the number of processed workpieces; the salary settlement mode of the outsourcing employee is daily settlement, and the salary is paid according to the piece, that is, the more the number of processed workpieces, the higher the salary.

[0105] This patent considers three factors of team cumulative processing quantity, personnel shift and team number to comprehensively judge the learning efficiency, determines the learning curve model associated with the three factors, and derives the three-factor learning curve model based on the traditional CD production function. The traditional CD production function is in the form of formula (8):

[0106] Y=AK α L β (8)

[0107] Wherein, Y represents output, A is efficiency coefficient, which is the reaction of generalized technical progress, K, L represent capital investment and labor investment, a, b represent the output elasticity of capital and labor, that is, when labor and capital investment increase by 1%, the output increases by a% and b% on average respectively. Labor investment L is determined by the team cumulative processing quantity and the team number, therefore, labor investment L can be decomposed according to formula (9).

[0108] L=HQ a N b (9)

[0109] Wherein, H represents the correction coefficient, Q represents the team cumulative processing quantity per unit time, N represents the team number, a, b represent the influence of cumulative processing quantity and team number on labor investment respectively, that is, when the team cumulative processing quantity per unit time increases by 1%, the labor investment increases by a%; when the team number increases by 1%, the labor investment increases by b%.

[0110] In the electronic product precision part production workshop targeted by the patent, capital investment only includes labor capital, that is, with the increase of output, the change of capital stock is related to labor investment, which can be specifically represented by formula (10):

[0111] K=uL λ (10)

[0112] Wherein, u is the correction coefficient, and l represents the influence coefficient of labor investment on capital investment. Formula (9) and formula (10) are brought into formula (8), and the correction coefficient part is uniformly represented by k, so that formula (11) can be obtained:

[0113] Y=kQ a(αλ+β) N b(αλ+β) (11)

[0114] Let p=a( l a+ b), q=b( l a+ b), and take the logarithm to obtain formula (12):

[0115] lnY=lnk+plnQ+qlnN (12)

[0116] Unit working hour represents the average working hour of each process, which can be represented by formula (13):

[0117] t=T / Y (13)

[0118] Wherein, T is total working hour. Formula (13) is brought into formula (12) to obtain formula (14):

[0119] lnt=lnT-lnk-plnQ-qlnN (14)

[0120] Let u = lnT - lnk, then formula (14) can be expressed as:

[0121] lnt=u-plnQ-qlnN (15)

[0122] In actual production, formal employees and outsourced employees belong to different human resource laborers, and their impact on labor input is also different. The derivation process of coefficients u, p, and q in formula (15) is all related to labor input. Therefore, fitting the same working hour data of employees in different shifts will yield different coefficients. Let M represent the two types of employees, where M=1 represents formal employees and M=0 represents outsourced employees, and we can obtain formula (16).

[0123] lnt=M(u1-p1ln Q-q1lnN)+(1-M)(u2-p2ln Q-q2lnN) (16)

[0124] Formula (16) is the logarithmic form of the derived three-factor learning curve model. In this step, under the background of the production of precision electronic parts, the labor input and capital input in CD are decomposed, and then the total output and unit working hours are converted, finally obtaining the learning curve model related to the cumulative processing quantity of the team, the shift system of personnel and the number of team members.

[0125] Step 2: Construction of the work time prediction system

[0126] Regression analysis was performed on the time data of each category of process divided by the SOFM network, the coefficients in the learning curve model were solved, the specific learning curve model of each type of process was obtained, and a database of corresponding information was established.

[0127] 2.1 Collection and Analysis of Working Hour Data

[0128] For each process type obtained by SOFM network clustering, historical working hours data of multiple processing teams under various cumulative processing times are collected and abnormal data information is removed. Based on this information, regression analysis is performed on formula (16) to obtain the values ​​of coefficients u, p, q1, and q2 under each category.

[0129] 2.2 Constructing a database of work hour prediction information

[0130] Create a learning curve data table and a team processing information data table in the database. In the learning curve table, the process category number corresponds one-to-one with the corresponding learning curve; in the team processing information table, each data entry consists of the team number, the team's cumulative processing times for each type of process, the team's shift system, and the team's number of members.

[0131] Step 3: Conduct time forecasting based on process type and operating team.

[0132] After the process to be processed and the processing team operating the process are determined, the processing time of the process is predicted according to the following steps.

[0133] 3.1 Classification of the process to be predicted according to the SOFM model

[0134] The feature vector of the process is obtained, normalized according to formula (1), and the distance between the vector and each category vector is calculated according to formula (4). The process is classified into the category with the closest distance.

[0135] 3.2 Determination of the learning curve model of the process to be predicted

[0136] The number of times the team processes the process, the personnel shift of the team, and the number of personnel are extracted from the database team processing information table. The learning curve model of the category to which the process to be predicted belongs is extracted from the learning curve table. The target processing time is calculated by combining the learning curve model and the team information.

[0137] The processing time prediction method proposed in this patent is verified by taking an actual production workshop of complex and precise electronic products as an example. The process characteristics in Table 1 are used as evaluation indicators, and each process in the processing of microwave devices of electronic products is used as a classification evaluation object. The data samples shown in Table 2 are established, wherein the process consists of process name and process number.

[0138] Table 1 Process evaluation feature selection

[0139]

[0140] Table 2 Classification and evaluation sample of processing process

[0141]

[0142]

[0143] According to the data samples provided in Table 2, after normalization, they are input into the SOFM network for training. After calculation and adjustment, the classification number is finally taken as 6, and the clustering results are shown in Table 3.

[0144] Table 3 Process clustering results

[0145]

[0146] The processing time data information of each process in Table 3 is counted. Due to the large amount of processing time information data, only part of the processing time data samples are shown here. Taking process category 2 as an example, part of the processing time data information is shown in Table 4 and Table 5.

[0147] Table 4 Part of the outsourcing team processing time data information of process category 2

[0148]

[0149]

[0150] Table 5 Process Category 2 Part Formal Team Man-hour Data Information

[0151]

[0152]

[0153]

[0154] According to the logarithmic form of three-factor learning curve:

[0155] Int = M(u1-p1lnQ-q1lnN) + (1-M)(u2-p2lnQ-q2lnN)

[0156] The cumulative processing quantity Q, team number N and processing man-hour t of each team of processing process category 2 are taken logarithm respectively, and then the processed outsourcing team processing data and formal team processing data are analyzed by multiple regression analysis by using spss software, and the regression coefficients are obtained:

[0157] u1 = 3.685; u2 = 2.987; p1 = 1.100; p2 = 1.077; q1 = 0.524; q2 = 0.516

[0158] Therefore, the logarithmic form of the three-factor learning curve model of process category 2 is:

[0159] Int = M(3.685-1.100lnQ-0.524lnN) + (1-M)(2.987-1.077lnQ-0.516lnN)

[0160] The three-factor learning curve model of process category 2 is transformed as:

[0161] t = 39.845MQ -1.100 N -0.524 + 19.826(1-M)Q -1.077 N -0.516

[0162] Similarly, the three-factor learning curve models of the other five categories can be obtained through data collection and regression analysis, and they are stored in the learning curve table in the database corresponding to the process category, and the team processing information table is updated to obtain the database of work time prediction information. Randomly select the process hot sound welding 80, and select different processing teams to process this process. The work time prediction model proposed in this patent is used to predict the work time, and the prediction results are compared with the traditional learning curve prediction results and the actual work time results. The results are shown in the figure.

[0163] Figure 6 The team 1238 is an outsourcing team with 1 person and a cumulative processing quantity of 2; the team 3086 is an outsourcing team with 3 people and a cumulative processing quantity of 1; the team 2375 is a formal team with 2 people and a cumulative processing quantity of 3; the team 1041 is a formal team with 1 person and a cumulative processing quantity of 2; and the team 1529 is a formal team with 3 people and a cumulative processing quantity of 2. According to the three-factor learning curve model, the work time prediction information database is updated. Figure 6 The data can calculate that the prediction error of the work time prediction model proposed in this patent is between-14% and 7%, while the prediction error of the traditional learning curve model is between-56% and 64%. Compared with the calculation results of the three-factor learning curve model, the prediction error is larger. It is shown that the traditional learning curve model is not suitable for the production background of the semi-automated workshop of electronic product precision parts. The work time prediction results of the three-factor learning curve model considering the cumulative processing quantity of the team, the personnel shift system and the team size are more in line with the actual situation.

Claims

1. A time prediction method based on process clustering, the method comprising: Step 1: Cluster the processes based on a self-organizing feature map neural network; Step 1.1: Collect the names and technological characteristics of all processing steps from the enterprise's CAPP and ERP systems, and quantify these characteristics; Step 1.2: Organize the features of all processes into an n-dimensional vector form and normalize them to serve as input samples for the SOFM network, which stands for Self-Organizing Feature Mapping Network. Step 1.3: Design the SOFM network based on the characteristics of the input data, including the number of neurons in the input layer, the number of nodes in the output layer, the learning rate, and the number of learning iterations; Step 1.4: The SOFM network determines the winning node and dominant neighborhood by calculating the distance between all neuron nodes and the input sample, and adjusts the weights of the neuron nodes in the dominant neighborhood; when the learning rate of the winning node is less than a fixed value, the SOFM network training process ends and the clustering result of the input n-dimensional vector is obtained; Step 2: Derive and fit a three-factor learning curve model to predict the processing time of the target process; Step 2.1: Derive the learning curve model relating to three factors—cumulative processing quantity of the team, shift system, and team size—from the Cobb-Douglas production function; The traditional CD production function is in the form of formula (8): （8） Where Y represents output, A is the efficiency coefficient, which reflects generalized technological progress, and K and L represent capital input and labor input, respectively. , This represents the output elasticity of capital and labor, i.e., when labor and capital inputs increase by 1%, output increases by an average of [missing value]. and The labor input L is determined by the cumulative processing quantity of the team and the number of team members. Therefore, the labor input L is decomposed according to formula (9). （9） Where H represents the correction coefficient, Q represents the cumulative processing quantity of the team per unit time, N represents the number of team members, and a and b represent the impact of the cumulative processing quantity and the number of team members on labor input, respectively. That is, when the cumulative processing quantity of the team increases by 1% per unit time, the labor input increases by a%; when the number of team members increases by 1%, the labor input increases by b%. In the production workshop for precision electronic components, capital investment only includes labor capital. That is, as output increases, the change in capital stock is related to labor input, as specifically expressed by formula (10): （10） in, For correction factor, The coefficient representing the influence of labor input on capital input; substituting formulas (9) and (10) into formula (8), and uniformly representing the correction coefficient with k, we can obtain: （11） make , And taking the logarithm, we get formula (12): (12) The unit working time represents the average working time for each process, expressed as formula (13): （13） Where T is the total working hours; substituting formula (13) into formula (12) yields formula (14): （14） make Then formula (14) can be expressed as: (15) In actual production, formal employees and outsourced employees belong to different human resources workers, and their impact on labor input is also different. The derivation process of coefficients u, p, and q in formula (15) is related to labor input. Therefore, fitting the same working hour data of different shifts will yield different coefficients. Represent the two types of employees with M, where M=1 represents formal employees and M=0 represents outsourced employees, and formula (16) can be obtained. （16） Step 2.2: Collect and process the time data of various processes from multiple different teams, perform regression analysis for different types of processes, solve the coefficients in the learning curve model, obtain the specific learning curve model for each type of process, and store it in the database; Step 2.3: Determine the processes to be processed, classify them according to the SOFM network, and obtain the corresponding learning curve models from the database; Step 2.4: Determine the operation team for this process, and substitute the team's cumulative processing quantity, shift system, and team size into the learning curve model to obtain the target working hours.

2. The time prediction method based on process clustering as described in claim 1, characterized in that, The specific method described in 1.3 is as follows: Step 1.3.1: Setting the number of input / output layer nodes; The number of nodes in the input layer is set to 5, and the output layer is finally determined to be an 8*10 two-dimensional mesh structure composed of 80 five-dimensional nodes; Step 1.3.2: Initialize the output layer neurons; Initialize each neuron component in the output layer with a random decimal number in the interval [0, 1], and obtain m is the number of neurons in the output layer; Step 1.3.3: Weight initialization; Use random decimals within the interval [0, 1] as the connection weights between n input nodes and m output nodes. Assign initial values, and use the weight vector between output node j and input node j. express; Step 1.3.4: Setting up the winning neighborhood; During initial training, the winning neighborhood is set to an 8x8 grid centered on the winning neuron. As the number of iterations increases, the winning neighborhood... The side length d gradually decreases according to formula (2); （2） In the formula: The initial winning neighborhood side length; This represents the current iteration number; This represents the total number of iterations. Step 1.3.5: Setting the learning rate for the winning node; During initial training, the learning rate Set to 0.9, as the number of iterations... The increase in learning rate It gradually decreases according to formula (3); （3） In the formula: The initial learning rate; This represents the current iteration number; This represents the total number of iterations.

3. The time prediction method based on process clustering as described in claim 1, characterized in that, The specific method for step 1.4 is as follows: Step 1.4.1: Compare the similarity between the input vector and all neurons in the competing layers; Selecting the input vector Competing layer neurons The Euclidean distance between the two is calculated according to formula (4). The shorter the Euclidean distance, the more similar the input vector is to the neuron. （4） Step 1.4.2: Select the winning neuron; input vector After comparing the similarity with all neurons in the competing layer according to formula (4), the neuron with the highest similarity is selected as the winning neuron according to formula (5). ; （5） Step 1.4.3: Select the winning neighborhood and adjust the weights of the neurons within the winning neighborhood; Calculate according to formula (2) As the side length of the winning neighborhood of the square at time t, and with the current winning neuron Determine the current winning neighborhood for the center. ; Calculate the connection weights between all nodes and the input vector according to formula (6), where the nodes do not belong to the winning neighborhood. The node weights within remain unchanged; ； In the formula, the learning rate The distance between the competing node and the winning neuron is related; the greater the distance, the better. The smaller; After t iterations, the winning neuron In the winning neighborhood, the k-th layer node uses This means that all nodes in each layer are equidistant from the winning node, and the distance between each node and the winning node increases with the number of layers. Calculate the winning node according to formula (3) learning rate Winning Neighborhood The learning rate of the k-th layer node varies according to formula (7), and the learning rate of the node that is further away from the winning node is lower. ， (7) In the formula, K is the maximum number of levels in the current winning neighborhood; Step 1.4.4: End training and obtain the clustering results of the input process vectors; Learning ends when the learning rate of the winning node is less than a set threshold or the number of iterations reaches a set number. This results in the mapping positions of all input process vectors in the output layer's two-dimensional grid structure after training has stabilized, thereby achieving the clustering and partitioning of processes.

Images