Single-piece customized product assembly process progress prediction method

Abstract

The invention provides a single-piece customized product assembly process progress prediction method, which comprises the steps of calculating the time lower bound of all procedures of single-piece customized product assembly based on domain knowledge; performing offline training of a data-driven completion time deviation prediction model based on the time lower bound of all procedures; and deploying the trained prediction model in an actual production environment, inputting real-time manufacturing data into the prediction model, and predicting the completion time of a next process in a recursion mode. According to the invention, accurate prediction of the completion time of all the uncompleted processes in a single-piece customization production environment can be realized; the problems that in the prior art, under different production scenes, generalization performance is poor, a complex and changeable single-piece customization environment is difficult to adapt, accurate prediction of the production progress is difficult to achieve, and the completion time of an intermediate process is difficult to predict are solved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent manufacturing technology, and in particular to a method for predicting the progress of a single-piece customized product assembly process. Background Art

[0002] Due to the development of personalized needs and customer requirements, multi-variety small batch production, or even single-piece custom production (OKP) has become increasingly popular. OKP is a manufacturing paradigm that can produce customized products according to customer requirements and is widely used in the manufacturing of heavy industrial products, such as ships, steam turbines, boilers, and large industrial equipment. The characteristics of OKP can be summarized as follows: highly customized, product design and manufacturing based on customer orders, product design changes almost with every new order; one-time successful production, due to the high unit price, most customer orders only contain one sample, adopt a production mode without spare parts, and there are no prototypes in the entire development cycle; there is a large uncertainty in production control. Due to customized manufacturing, OKP production has poor stability, low production and process specialization, and unexpected interference in production control often occurs, affecting the normal execution of production plans and even leading to delayed delivery.

[0003] Due to the high degree of customization and one-time successful production characteristics of OKP, the process data of the product is usually estimated by the process department by modifying the data of historical models. Some process data, such as working hours, are inaccurate. Secondly, in actual OKP production, unexpected interference time often occurs, such as machine failure, logistics delays and quality rework. Inaccurate process data and unexpected interference often affect normal operation plans. How to accurately judge the current production status, predict future production progress, and make timely adjustments before uncertain factors cause serious impacts are the core issues of OKP production control.

[0004] Existing methods for production schedule prediction can be roughly divided into three categories: analytical methods, simulation methods, and data-driven methods. The analytical method simplifies the manufacturing system into a random process model and constructs a queuing model, Markov chain model, or other mathematical models to predict the production schedule. The application scenarios of analytical methods are mostly limited to simple systems with strict assumptions, and are suitable for some uncomplicated and ideal scenarios in the hope of establishing an accurate mathematical model of the artificial manufacturing system. However, as manufacturing becomes more and more complex, it becomes very difficult to systematically describe the real manufacturing process and establish an accurate analytical model.

[0005] The simulation method establishes a simulation model that is close to the actual workshop operation according to the actual situation, and performs a large number of repeated simulations to obtain a sufficient number of random samples to achieve the prediction of production progress. For the increasingly complex and multi-disturbance OKP manufacturing system, it is difficult to establish a simulation model that can accurately simulate the actual operating state. At the same time, due to the increasing complexity of product processing and the increasing size of manufacturing systems, simulation analysis requires a long computing time and huge computing resources, which limits the application of simulation methods in real-time production scenarios.

[0006] With the development of artificial intelligence technology, data-driven workshop production schedule prediction methods have gradually attracted the interest of researchers. The data-driven method first analyzes the high-dimensional features related to the prediction target, then uses historical manufacturing data to train the prediction model, and finally inputs the corresponding real-time data into the model for prediction. This method does not need to know the complex relationship between the input features, and has high computational efficiency and prediction accuracy in practical applications.

[0007] However, current research on production schedule prediction based on data-driven methods usually models it as a pure machine learning problem, and ignores the prospect of adding domain knowledge to the prediction model by designing more sophisticated algorithms and features to achieve better fitting effects, resulting in difficulty in achieving satisfactory generalization performance in complex and changeable OKP production scenarios. In addition, most existing studies predict remaining time, product completion time, cycle time, etc., and only focus on the completion time of the last process, which cannot be used to predict the completion time of intermediate processes. Summary of the invention

[0008] In view of the defects in the prior art, the object of the present invention is to provide a method for predicting the progress of the assembly process of a single customized product.

[0009] According to one aspect of the present invention, a method for predicting the progress of a single customized product assembly process is provided, comprising:

[0010] Based on domain knowledge, calculate the lower bound of the time for all processes in assembling a single customized product;

[0011] Based on the lower bounds of the time for all processes, offline training of the data-driven completion time deviation prediction model is performed;

[0012] The trained prediction model is deployed in the actual production environment, real-time manufacturing data is input into the prediction model, and the completion time of the next process is predicted recursively.

[0013] Preferably, the calculation of the lower bound of the time for all processes of assembling a single customized product based on domain knowledge includes:

[0014] Define resource data, process data, work order data, planning data and execution data;

[0015] Based on the above defined data, calculate the completion time of each process;

[0016] Based on the completion time of each process, a lower bound of the completion time of each process is derived.

[0017] Preferably, the resource data includes one or more of the following:

[0018] W: Worker Set

[0019] w: Worker type

[0020] C w :w type worker collection

[0021] k: worker number, w: k ∈C w represents worker k of type w;

[0022] The process data includes one or more of the following:

[0023] I: The production process set of a single customized product workshop, whose length is n

[0024] i: process number

[0025] pred i : The set of preceding processes of process i

[0026] d i : Rated working hours for process i

[0027] part i :Material requirements for process i

[0028] qua i :Quality requirement of process i

[0029] r iw : The number of workers of type w required for process i

[0030] The work order data includes one or more of the following:

[0031] rea I : The time it takes for a work order to be sent to the production workshop for a single customized product

[0032] due i : Delivery time of intermediate products corresponding to process i

[0033] The planning data includes one or more of the following:

[0034] SP i: The planned start time of process i needs to satisfy the precedence constraint requirement SP i ≥FP j ,i∈I,j∈pred i

[0035] FP i : The planned end time of process i needs to meet the delivery date constraint FP i <due i ; At the same time, the planned end time needs to meet the rated working time constraint FP i =SP i +d i ;

[0036] W i : The set of workers assigned to process i

[0037] w pi : The set of workers of type w assigned to process i, |w pi |=r iw ; Worker allocation and process execution plan need to meet the worker quantity constraint, where x it is a 0-1 variable, if t∈[S pi ,F pi ], take 1, otherwise take 0.

[0038] O wk : The task list of worker k of type w, O wk =[o1,o2,...,o m ],o m ∈I; In the actual production process, workers complete the processes in their task list in turn; a worker can only complete one process at a time, and must meet The task list must also satisfy the pre-order constraint.

[0039] The execution data includes:

[0040] SE i : The actual start time of process i, where the actual start time is defined as the time when all workers and materials required for the process are ready;

[0041] FE i : The actual end time of process i, which is defined as the time when the process completes the quality inspection and releases the worker;

[0042] dE i : Actual working hours of process i, d Ei =F Ei -S Ei ;

[0043] w Eki : The time when worker k of type w arrives at process i, that is, if other conditions are met, the worker can start executing process i immediately;

[0044] part Ei : The time it takes for materials of process i to be fully prepared.

[0045] Preferably, the calculation of the completion time of each process based on the above-defined various data includes:

[0046] The actual start time SE of process i i The actual execution time SE i Calculate the actual completion time FE i :

[0047] FE i =SE i +dE i ;

[0048] Where: Actual start time SE of process i i By the planned start time SP i , personnel preparation time, material preparation time, and the completion time of the preceding process:

[0049]

[0050] It refers to the maximum actual completion time of all the preceding processes of process i, that is:

[0051]

[0052] Material preparation timepart Ei , at the planned start time part of process i Ei Decide:

[0053]

[0054] The actual start time SE of process i i Arranged as: Among them, the personnel preparation time for process i is W Ei The time available for execution of process i by all personnel involved in process i is jointly determined by:

[0055]

[0056] For a worker w k For example, let process i be its participating process list O wk The mth process o wkm , then the time when it can start executing process i is determined by the planned start time of process i and wk The time to complete the previous process in the process list is jointly determined. In addition, the influence of random disturbance factors needs to be considered, that is:

[0057]

[0058] The actual execution time of process i is given by the actual process operation time dJ i and rework time dB i Decision: dE i =dJ i +dB i ; Actual process operation time dJ i Rated operation time d of the process i There is a difference, namely: dJ i =d i +Δ i ,Δ i ≥0.

[0059] Preferably, the deducing the lower bound of the completion time of each process based on the completion time of each process includes:

[0060] When the disturbance factor is not considered, the lower bound of the actual start time of process i is derived as follows:

[0061] The logistics delivery completion time of process i is equal to its planned start time, that is:

[0062] bound lower (part Ei )=SP i

[0063] Without considering the influence of disturbance factors, the worker w k The earliest time to start executing process i is the maximum of the planned start time of process i and the completion time of the previous process in the worker's participating process list, that is:

[0064]

[0065] Derive the earliest start time for all workers to start executing process i, that is:

[0066]

[0067] Combining the derivation results, we get the lower bound of the actual start time of process i, namely:

[0068]

[0069] Without considering disturbance factors, the lower bound of the actual completion time of process i is derived as follows:

[0070] In the absence of extended process operation time and process rework, the actual operation time of process i is its rated operation time, that is:

[0071] bound lower (dE i ) = d i

[0072] The lower bound of the actual completion time of process i is obtained when the influence of disturbance factors is not considered:

[0073]

[0074] Get the lower bound of the actual completion time of each unfinished process in the actual production process of the workshop without considering the influence of future disturbance factors;

[0075] At any time t, all the processes of the workshop production are divided into three sets, namely the completed process set Opt finished 、Ongoing process set Opt execution 、Unstarted process set Opt notStart ;

[0076] For i∈Opt finished , its actual completion time is known, then:

[0077] bound lower (FE i )=max(SE i +d i ,t),i∈Opt finished

[0078] For the unstarted process set Opt notStart For the process in the process, the lower bound of its actual completion time needs to be calculated by step-by-step prediction;

[0079] Divide all processes into two sets, namely the set of processes with known lower bounds of completion time Opt bound , initialized to Opt finished ∪Opt execution , and the set of processes with unknown lower bounds of completion time Opt notbound , initialized to the unstarted process set Opt notStart ;

[0080] Existence theorem: There exists a process i∈Opt notbound ,satisfy and

[0081] Based on the theorem, Opt is calculated step by step notbound The lower bound of step i in , and update Opt bound With Optnotbound , and finally get the lower bound of the completion time of all processes at the current moment.

[0082] Preferably, the data-driven completion time deviation prediction model has an input of: a process diagram network modeled by the actual production status of a single customized product; and an output of: a deviation between the predicted actual completion time and the lower bound;

[0083] The deviation prediction model includes three SAGE layers, two LSTM layers and one fully connected layer group;

[0084] The SAGE layer is used for the topological structure of the input process graph network, automatically aggregating the features between different nodes, and realizing the associated extraction of node features;

[0085] The LSTM layer extracts implicit time series features based on the node features aggregated by the SAGE layer as hidden layer feature output;

[0086] The fully connected layer fits the nonlinear relationship between the hidden layer features output by the LSTM and the target variable, where the target variable refers to the deviation of the actual completion time from a lower bound.

[0087] Preferably, the actual production status of the single customized product is modeled as a process graph network, in which nodes represent processes and edges between nodes represent the association relationship between two processes.

[0088] Preferably, the nodes in the graph include actual production status characteristics of a single customized product, and the actual production status characteristics of a single customized product are specifically all of the following characteristics:

[0089] Network complexity NC, resource factor RF, resource intensity RS, rated working hours d i , resource requirements i 、Number of Predecessor Operations|pred i |、Number of subsequent processes|succ i |, earliest start time EST, latest start time LFT, maximum resource requirement GRD, minimum resource requirement SRD, earliest end time EFT, latest end time LST, maximum successor number MTS, minimum slack time MSLK, maximum position weight GRPW, maximum resource utilization GRU, weighted resource utilization and precedence WRUP, initial plan sequence value PPR, plan start time sequence value PFR, completed process proportion FPR, current time PT, current resource utilization RU, unfinished process working time RT, process slack time ST, process average slack time AST.

[0090] Preferably, the association relationship is used to characterize the precedence constraints and resource sharing relationships between different process nodes;

[0091] There are three types of association relationships:

[0092] The first type of association D1 is modeled based on a precedence constraint:

[0093]

[0094] The second type of relationship D2 is modeled based on resource sharing:

[0095]

[0096] The third type of relationship D3 models the relationship based on the initial plan:

[0097]

[0098] Among them: i, j represent the node number, pred j represents the set of predecessor processes of process j, O wk represents the task list of worker k of type w, SP sort A list of operations sorted in ascending order based on their planned start time.

[0099] Preferably, the trained prediction model is deployed in an actual production environment, real-time manufacturing data is input into the prediction model, and the completion time of the next process is predicted in a recursive manner, including:

[0100] The predicted completion time of the nth process is regarded as the true value and input into the training model to predict the deviation between the completion time of the n+1th process and the lower bound of the time.

[0101] The deviation is applied to the obtained lower bound of the time of the corresponding process to obtain the completion time of the (n+1)th process.

[0102] Compared with the prior art, the embodiments of the present invention have at least one of the following beneficial effects:

[0103] (1) The method for predicting the progress of the assembly process of a single customized product in the embodiment of the present invention is a method based on a hybrid framework, which combines domain knowledge with deep learning (the prediction model is based on a neural network, so it is based on deep learning), rather than a purely data-driven method. During the entire prediction process, the strict lower bound of the completion time of the next process can be calculated based on domain knowledge, and the deviation of the actual completion time from the strict lower bound can be obtained using the prediction model of deep learning.

[0104] (2) The method for predicting the progress of the assembly process of a single customized product in the embodiment of the present invention proposes a mathematical model of the OKP production process (specifically, parameter definition, calculation of the completion time of each process based on the defined parameters, and derivation of the lower bound of the calculation completion time based on the calculated actual completion time of the process). This mathematical model describes the dynamic characteristics of the OKP production process in real time and can derive a strict lower bound for the completion time of the next process.

[0105] (3) The method for predicting the progress of the assembly process of a single customized product in an embodiment of the present invention proposes a SAGE-LSTM method that combines a process diagram network and a recurrent neural network for production progress prediction. The SAGE network is used to extract hidden features from the process diagram network, and then the LSTM network is used to mine the nonlinear relationship between feature data and target variables.

[0106] (4) The single-piece customized product assembly process progress prediction method in the embodiment of the present invention proposes a recursive prediction method, which recursively obtains the completion time of the next process by step-by-step backward prediction, and finally obtains the predicted completion time of all processes, thereby realizing accurate prediction of the OKP production progress.

[0107] In summary, the method for predicting the progress of the assembly process of a single-piece customized product involved in the embodiment of the present invention can accurately predict the completion time of all unfinished processes in a single-piece customized production environment. The method effectively solves the problems existing in the prior art, including insufficient generalization performance in different production scenarios, difficulty in adapting to complex and changeable single-piece customized environments, difficulty in accurately predicting the production progress, and difficulty in predicting the completion time of intermediate processes. BRIEF DESCRIPTION OF THE DRAWINGS

[0108] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings:

[0109] Figure 1 It is an overall framework diagram of a method for predicting the progress of a single customized product assembly process in one embodiment of the present invention;

[0110] Figure 2 A diagram of the data driven model system architecture in a preferred embodiment of the present invention;

[0111] Figure 3 It is a schematic diagram of prediction results of applying a method for predicting the progress of assembly process of a single customized product to 200 cases in a specific embodiment of the present invention. DETAILED DESCRIPTION

[0112] The present invention is described in detail below in conjunction with specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those of ordinary skill in the art, several variations and improvements may be made without departing from the concept of the present invention. These all belong to the protection scope of the present invention.

[0113] In one embodiment of the present invention, a method for predicting the progress of a single customized product assembly process is provided. Figure 1 As shown, refer to the corresponding content of "Production Progress Forecast" in the middle of the figure, the main steps are as follows:

[0114] Step 1: Based on domain knowledge, calculate the lower bound of the time for all processes of assembling a single customized product;

[0115] Step 2: Based on the lower bounds of the time for all processes, perform offline training of the data-driven completion time deviation prediction model;

[0116] Step 3: Deploy the trained prediction model to the actual production environment, input the real-time manufacturing data into the prediction model, and recursively predict the completion time of the next process.

[0117] In the above-mentioned embodiment of the present invention, during the entire prediction process, the strict lower bound of the completion time of the next process can be calculated based on domain knowledge, and the deviation between the actual completion time and the strict lower bound can be obtained by using a data-driven completion time deviation prediction model, so as to perform applications such as dynamic plan adjustment, technical bottleneck analysis, and resource allocation.

[0118] In one embodiment of the present invention, step 1 is implemented: the lower bound calculation of the process completion time based on domain knowledge. In this embodiment, based on the production process characteristics of a single-piece customized product workshop, the uncertain disturbance factors affecting the process completion time in the production process of a single-piece customized product workshop are modeled in combination with domain knowledge, and the strict lower bound of the earliest completion time of all unfinished processes is calculated under a given scenario without considering future disturbance factors.

[0119] Specifically, step 1 can be divided into the following three steps:

[0120] Step 1.1, define the various data parameters needed.

[0121] Step 1.2, based on the data parameters defined in step 1.1, calculate the completion time of each process.

[0122] Step 1.3, based on the actual completion time of the process calculated in step 1.2, derive the lower bound of the completion time.

[0123] like Figure 1As shown, in a preferred embodiment, the implementation step 1.1 also includes data integration, which integrates the data into process data, work order data, resource data, planning data and execution data. These data can be obtained through data interfaces connected to major systems such as product data management PDM (Product Data Management), enterprise resource planning ERP (Enterprise Resource Planning), production information management system MES (Manufacturing Execution System), and Internet of Things IOT (Internet of Things).

[0124] (1) Resource data

[0125] W: Worker Set

[0126] w: Worker type

[0127] C w :w type worker collection

[0128] k: worker number, w: k ∈C w represents worker k of type w

[0129] (2) Process data

[0130] I: The production process set of a single customized product workshop, whose length is n

[0131] i: process number

[0132] pred i : The set of previous processes of process i

[0133] d i : Rated working hours for process i

[0134] part i :Material requirements for process i

[0135] qua i :Quality requirement of process i

[0136] r iw : The number of workers of type w required for process i

[0137] (3) Work order data

[0138] rea I : The time it takes for a work order to be sent to the production workshop for a single customized product

[0139] due i : The delivery time of the intermediate product corresponding to process i. For some intermediate products that do not have special delivery requirements, duei =due I

[0140] (4) Planning data

[0141] SP i : The planned start time of process i needs to satisfy the precedence constraint requirement SP i ≥FP j ,i∈I,j∈pred i

[0142] FP i : The planned end time of process i needs to meet the delivery date constraint FP i <due i At the same time, the planned end time needs to meet the rated working time constraint FP i =SP i +d i .

[0143] W i : The set of workers assigned to process i

[0144] w pi : The set of workers of type w assigned to process i, |w pi |=r iw Worker allocation and process execution planning need to meet the worker quantity constraints. where x it is a 0-1 variable, if t∈[S pi ,F pi ], take 1, otherwise take 0.

[0145] O wk : The task list of worker k of type w, O wk =[o1,o2,...,o m ],o m ∈I. In the actual production process, workers complete the processes in their task list in turn. Because a worker can only complete one process at a time, it must satisfy In addition, the task list must also satisfy the pre-order constraint.

[0146] (5) Execution data

[0147] SE i : The actual start time of process i, which is defined as the time when all workers and materials required for the process are ready.

[0148] FE i : The actual end time of process i, which is defined as the time when the process completes quality inspection and releases the worker.

[0149] dE i : Actual working hours of process i, d Ei =F Ei -S Ei .

[0150] w Eki : The time when worker k of type w arrives at process i, that is, if other conditions are met, the worker can start executing process i immediately.

[0151] part Ei : The time it takes for materials of process i to be fully prepared.

[0152] In another preferred embodiment, based on the above-mentioned implemented data, step 1.2 is implemented to calculate the completion time of each process based on the above-mentioned data parameters. Specifically:

[0153] The actual completion time of process i can be calculated by the actual start time and the actual execution time:

[0154] FE i =SE i +dE i

[0155] The actual start time of process i is determined by the planned start time, the time for personnel to be ready, the time for materials to be ready, and the completion time of the previous process. Because in the production execution process of a single-piece customized product workshop, special equipment such as welding machines and cranes are usually bound to workers, this topic does not consider the equipment availability separately, so:

[0156]

[0157] in It refers to the maximum actual completion time of all the preceding processes of process i, that is:

[0158]

[0159] The material distribution is planned by the warehousing and logistics management department based on the planning data. i Complete the delivery of complete sets of materials. Due to the existence of disturbance factors, logistics delivery may be delayed, namely:

[0160]

[0161] Based on the above formula, we can get:

[0162]

[0163] The time for all personnel to be ready for process i is determined by the time when all personnel involved in process i can start to execute process i, that is:

[0164]

[0165] For a worker w k For example, let process i be its participating process list O wk The mth process o wkm , then the time when it can start executing process i is determined by the planned start time of process i and w k The time to complete the previous process in the process list is jointly determined. In addition, the influence of random disturbance factors needs to be considered, that is:

[0166]

[0167] The actual execution time of process i is determined by the actual process operation time dJ i and rework time dB i It is decided that the process operation time in this topic includes the setup time and is not considered separately, so:

[0168] dE i =dJ i +dB i

[0169] Due to the existence of disturbance factors, the actual process operation time dJ i There is a difference with the rated operation time of the process i ,Right now:

[0170] dJ i =d i +Δ i ,Δ i ≥0

[0171] There is also a preferred embodiment, based on the actual completion time of the process calculated in step 1.2, derive the lower bound of the completion time.

[0172] When the influence of disturbance factors is not considered, the logistics distribution completion time of process i is equal to its planned start time, that is:

[0173] bound lower (part Ei )=SP i

[0174] Similarly, it can be calculated that, without considering the influence of disturbance factors, the worker w k The earliest time that process i can be started is the maximum of the planned start time of process i and the completion time of the previous process in the worker's participating process list, that is:

[0175]

[0176] The earliest start time at which all workers can start to execute process i can be derived together, that is:

[0177]

[0178] Combining the above derivation results, we can get the lower bound of the actual start time of process i, namely:

[0179]

[0180] That is, when the influence of disturbance factors is not considered, the actual completion time of process i is determined by its planned start time, the latest completion time of the previous process executed by all workers, and the latest completion time of all preceding processes. The production plan affects the actual execution process by determining the planned start time and worker resource allocation. Implicit disturbances are transmitted backward through the preceding constraint network and worker resource reuse, and are accumulated during the execution of each process.

[0181] Without considering the influence of disturbance factors, that is, without extending the process operation time and reworking the process, the actual operation time of process i is its rated operation time, that is:

[0182] bound lower (dE i ) = d i

[0183] The lower bound of the actual completion time of process i can be obtained without considering the influence of disturbance factors:

[0184]

[0185] Thus, we can get the lower bound of the actual completion time of each unfinished process in the actual production process of the workshop without considering the influence of future disturbance factors. At any time t, all the processes produced in the workshop can be divided into three sets, namely the completed process set Opt finished 、Ongoing process set Opt execution 、Unstarted process set Opt notStart .

[0186] For i∈Opt finished , its actual completion time is known; for i∈Opt finished , since its actual start time is known, we have:

[0187] bound lower (FE i )=max(SE i +d i ,t),i∈Opt finished

[0188] For the unstarted process set Opt notStart For the processes in the process, the lower bound of the actual completion time needs to be calculated by step-by-step prediction. For convenience, we divide all processes into two sets, namely the set of processes with known lower bounds of completion time Opt bound , initialized to Opt finished ∪Opt execution , and the set of processes with unknown lower bounds of completion time Opt notbound , initialized to the unstarted process set Opt notStart .

[0189] After proof, there is Theorem 1: There exists a process i∈Opt notbound ,satisfy and

[0190] Based on Theorem 1, we can gradually calculate Opt notbound The lower bound of step i in , and update Opt bound With Opt notbound , and finally get the lower bound of the completion time of all processes at the current moment (specifically: completed processes, ongoing processes, and unstarted processes).

[0191] In other embodiments, the proof process of Theorem 1 is provided: the pre-order constraint and the worker job sequence pre-order are modeled into the same network structure. Since the worker job sequence satisfies the pre-order constraint relationship, the final result is still a directed acyclic graph G1. notbound Draw a subgraph g of G1 as a node. Based on the property 1 of directed acyclic graph, we can get that the subgraph of directed acyclic graph is still a directed acyclic graph. Based on the properties of directed acyclic graph, we can get that there must be a node i with in-degree 0 in the directed acyclic graph, that is, and Because Opt notbound ∪Opt bound =Iand and and Therefore, we can get and

[0192] The above embodiment proposes a mathematical model of the OKP production process, describes the dynamic characteristics of the OKP production process in real time, and derives a strict lower bound of the completion time of the next process based on the mathematical model.

[0193] In a preferred embodiment of the present invention, step 2 is implemented to perform offline training of a data-driven completion time deviation prediction model. In this embodiment, step 2 consists of the following sub-steps:

[0194] Step 2.1: Input node feature design. The actual production status of a single customized product is modeled as a process graph network. Specifically, the nodes in the graph represent processes, and the edges between nodes represent the relationship between two processes.

[0195] This example proposes 26 state characteristics for the production schedule prediction problem of a single customized product, including network complexity (NC), resource factor (RF), resource intensity (RS), rated working hours (d i ), resource requirements (r i ), the number of preceding operations (|pred i |), the number of subsequent processes (|succ i |), earliest start time (EST), latest start time (LFT), maximum resource requirement (GRD), minimum resource requirement (SRD), earliest finish time (EFT), latest finish time (LST), maximum successor number (MTS), minimum slack time (MSLK), maximum position weight (GRPW), maximum resource utilization (GRU), weighted resource utilization and precedence (WRUP), initial plan sequence value (PPR), plan start time sequence value (PFR), completed process proportion (FPR), current time (PT), current resource utilization (RU), unfinished process time (RT), process slack time (ST), average process slack time (AST), covering three aspects: process information, planning information and execution information.

[0196] Step 2.2: Design of association relationships. In this embodiment, three association relationships between association nodes are designed to characterize the precedence constraints and resource sharing relationships between different process nodes:

[0197] The first type of association relationship D1 is modeled based on a precedence constraint, as shown in the following formula:

[0198]

[0199] The second type of association relationship D2 is modeled based on resource sharing, as shown in the following formula:

[0200]

[0201] The third type of association relationship D3 is modeled based on the initial plan, as shown in the following formula:

[0202]

[0203] Among them, SP sort A list of operations sorted in ascending order based on their planned start time.

[0204] Step 2.3: Data-driven model design. Figure 2The architecture of the SAGE-LSTM model used for production schedule prediction is described. It consists of three SAGE layers, two LSTM layers, and one fully connected layer. The SAGE layer is used to automatically aggregate features between different nodes based on the topological structure of the graph to achieve the correlation extraction of node features. Due to the sequential nature of the workshop production process, the LSTM network is used to extract implicit sequential features based on the node feature data aggregated by the SAGE layer. Finally, the nonlinear relationship between the hidden layer features output by the LSTM and the target variable is fitted through the fully connected layer.

[0205] like Figure 2 As shown, the input of the SAGE layer is a node matrix consisting of a process graph network (for example, if there are 10 processes, the dimension of the input node matrix is ​​10*26); in the SAGE layer, firstly, the neighbor nodes are sampled: for each node, samples are taken from its neighbors, which can be a fixed number of neighbor nodes or a sequence of neighbor nodes obtained by random walks and other methods. Then, the neighbor features are aggregated: for each node, the features of its sampled neighbor nodes are aggregated to generate a new representation of the node. Update the node representation: The aggregated neighbor features are combined with the original features of the target node (each node sampled from the neighbor nodes) to update the representation of the node.

[0206] In another preferred embodiment of the present invention, step 3: production progress prediction is implemented. The trained prediction model is deployed in the actual production environment, and the real-time manufacturing data (the 26 state characteristics mentioned above) is input into the prediction model. The completion time of the next process is recursively predicted by iterative prediction. The predicted completion time of the nth process is regarded as the true value and used to predict the completion time of the n+1th process, so as to realize the real-time and accurate prediction of the production progress.

[0207] In order to deeply and in detail understand the single-piece customized product assembly process progress prediction method in the above embodiment, a specific embodiment of the present invention is described in detail in combination with a specific data set.

[0208] (1) Construction of historical data set. The test cases come from the RCPSP data set in the PSPLIB database, which contains 2040 different use cases. The data set contains cases of four sizes (size refers to the number of processes), with activity numbers of 30, 60, 90, and 120, respectively, denoted as J30, J60, J90, and J120. Cases can be grouped according to the complexity of the project network, resource intensity, and resource factors. Each group has 10 different cases. In the actual single-piece customized manufacturing environment, the RF is relatively stable due to the fixed types of workers. The number of processes, network complexity, resource intensity, etc. will vary with different product designs. The training historical data set of this embodiment comes from four J30 case sets. The historical data set consists of multiple process graph networks, each of which represents a single-piece customized production project. The test historical data comes from four case sets of J30, J60, J90, and J120.

[0209] (2) Lower bound calculation. Based on the lower bound calculation method for process completion time proposed in the above example, the lower bound of the actual completion time of all unfinished processes at each slicing moment under each production project is calculated. In the process of calculating the lower bound of time, it is necessary to collect data from five aspects: process data, work order data, resource data, plan data, and execution data. In this embodiment, process data, work order data, and resource data are obtained through the RCPSP data set; plan data are obtained by solving the data set through the genetic algorithm; and execution data are obtained through simulation.

[0210] (3) Construction of training data set. In order to accurately predict the completion time of all processes, this embodiment adopts a recursive method to predict the production schedule. That is, the same prediction model is used to recursively predict the completion time of the next activity until the process is completed. In order to achieve the above goal, the process diagram network is divided into data to form a prediction data set. Specifically,

[0211] (3.1) Based on the temporal order, a workshop production graph network containing N processes is decomposed into n subgraphs, each subgraph contains m nodes, satisfying n = Nm.

[0212] (3.2) Take each subgraph as input data, the difference between the actual completion time of the next process and the theoretical lower bound calculated in step (2) as the label, construct attribute-label pairs, and generate a data set.

[0213] (4) Dataset division: J30 was randomly divided into training set, validation set, and test set in a ratio of 8:1:1, and the data in the J60, J90, and J120 case sets were divided into the test set.

[0214] (5) Initialize the model. Select appropriate hyperparameters for the SAGE-LSTM model, including the structure of the SAGE-LSTM, learning rate, batch size, number of hidden layer nodes, epoch, optimizer selection, etc. After the hyperparameters are defined, a random initialization model is established.

[0215] (6) Model training. The training set is divided into multiple batches for batch training. The loss function of each batch is calculated based on the mean square loss, and the weights of each parameter of the neural network are updated through the back propagation algorithm until the termination condition is reached, that is, the model training is completed.

[0216]

[0217] M is the number of data items; represents the predicted value; y i represents the truth value;

[0218] (7) Production progress prediction. The data in the test set is input into the trained model to calculate the deviation between the actual completion time and the lower bound. The deviation is then added to the calculated lower bound to obtain the final production progress prediction value.

[0219] In order to evaluate the performance of integrating data-driven and domain knowledge in the single-piece customized product assembly process progress prediction method proposed in the above embodiment of the present invention, three different production progress prediction methods are designed and compared, which are respectively recorded as a hybrid method, a plan-based method and a direct prediction method. Among them, the hybrid method is the single-piece customized product assembly process progress prediction method of the embodiment of the present invention. This method does not directly predict the completion time of the next process, but predicts the deviation between the completion time of the next process and the theoretical lower bound without considering the influence of future random factors. The plan-based method is similar to the hybrid method. It also does not directly predict the completion time of the next process, but takes the planned completion time as the lower bound to predict the deviation between the actual completion time of the next process and the planned completion time. Since the planned completion time is a static variable, this method is equivalent to normalizing the target variable. Unlike the above two methods, the direct prediction method directly uses the actual completion time of the next process as the prediction target. In order to ensure fair comparison, except for different prediction target values, the prediction models, training data sets and test data sets used by the above three methods are the same. The results are shown in Table 1.

[0220] Table 1 Experimental results of different methods

[0221]

[0222]

[0223] As can be seen from Table 1, the method of the embodiment of the present invention outperforms the other two comparison methods (plan-based method and direct prediction method) in 12 indicators of the four test sets, proving the superiority of the method of the embodiment of the present invention. The plan-based method and the direct prediction method performed slightly lower than the method of the embodiment of the present invention in the J30 data set, but their performance in the J60, J90, and J120 data sets was much lower than the method of the embodiment of the present invention. This is because pure data-driven methods are difficult to generalize to unseen data sets, which limits the application of pure data-driven methods. The method for predicting the progress of the assembly process of a single-piece customized product proposed in the embodiment of the present invention integrates domain knowledge, has good generalization to unseen cases, and can better adapt to the production progress prediction in the changeable single-piece customized workshop production environment.

[0224] In order to clearly reflect the performance of the production schedule prediction method proposed in the embodiment of the present invention, the predicted values ​​and actual values ​​of the final makespan of 200 random test cases are displayed in Figure 3 In. From Figure 3 It can be seen that the average prediction error of 200 random cases is 1.95, the average relative error is 1.47%, and the overlap between the predicted value and the actual value is high, indicating that the proposed single-piece customized product assembly process progress prediction method can achieve higher accuracy.

[0225] The above describes the specific embodiments of the present invention. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various modifications or variations within the scope of the claims, which does not affect the essence of the present invention. The above preferred features can be used in any combination without conflicting with each other.

Claims

1. A method for predicting the progress of a single-piece customized product assembly process, characterized in that: include: Based on domain knowledge, calculate the lower bound of the time for all processes in the assembly of a single customized product; Based on the lower bounds of the time for all processes, offline training of the data-driven completion time deviation prediction model is performed; The trained prediction model is deployed in an actual production environment, real-time manufacturing data is input into the prediction model, and the completion time of the next process is predicted recursively.

2. A method for predicting the progress of a single-piece customized product assembly process according to claim 1, characterized in that: The domain knowledge-based calculation of the lower bound of the time for all processes in assembling a single customized product includes: Define resource data, process data, work order data, planning data and execution data; Based on the above defined data, calculate the completion time of each process; Based on the completion time of each process, derive the lower bound of the completion time of each process.

3. The method for predicting the progress of the assembly process of a single customized product according to claim 2, characterized in that: The resource data includes one or more of the following: W: Worker Set w: Worker type C w :w type worker collection k: worker number, w k ∈C w represents worker k of type w; The process data includes one or more of the following: I: A set of production process steps for a single custom product workshop, whose length is n i: process number pred i : the set of preceding processes of process i d i : Rated working hours of process i part i :Material requirements for process i qua i :Quality requirement of process i r iw : The number of workers of type w required for process i The work order data includes one or more of the following: rea I : The time it takes for a work order to be sent to the production workshop for a single customized product due i : Delivery time of intermediate products corresponding to process i The planning data includes one or more of the following: SP i : The planned start time of process i needs to meet the pre-order constraint requirement SP i ≥FP j ,i∈I,j∈pred i FP i : The planned end time of process i needs to meet the delivery date constraint FP i <due i ; At the same time, the planned end time needs to meet the rated working hours constraint FP i =SP i +d i ; W i : The set of workers assigned to process i w pi : The set of workers of type w assigned to process i, |w pi |=r iw ; Worker allocation and process execution plan need to meet the worker quantity constraint, where x it is a 0-1 variable, if t∈[S pi ,F pi ], take 1, otherwise take 0. O wk : Task list of worker k of type w, O wk =[o1,o2,...,o m ],o m ∈I; In the actual production process, workers complete the processes in their task list in turn; a worker can only complete one process at a time, and must meet The task list must also satisfy the pre-order constraint. The execution data includes: SE i : The actual start time of process i, where the actual start time is defined as the time when all workers and materials required for the process are ready; FE i : The actual end time of process i, where the actual end time is defined as the time when the process completes the quality inspection and releases the worker; dE i : Actual working hours of process i, d Ei =F Ei -S Ei ; w Eki : The time when worker k of type w arrives at process i, that is, if other conditions are met, the worker can start executing process i immediately; part Ei : The time it takes for materials for process i to be fully prepared.

4. A method for predicting the progress of a single customized product assembly process according to claim 3, characterized in that: The calculation of the completion time of each process based on the above-mentioned various data is as follows: The actual start time SE of process i i The actual execution time SE i Calculate the actual completion time FE i : FAITH i =SE i +dE i ; Where: Actual start time SE of process i i By the planned start time SP i , personnel preparation time, material preparation time, and completion time of the preceding process: It refers to the maximum actual completion time of all preceding processes of process i, that is: Material preparation time part Ei , at the planned start time part of process i Ei Decide: The actual start time SE of process i i Arranged as: Among them, the personnel preparation time for process i is W Ei The time available for execution of process i by all personnel involved in process i is jointly determined by: For a worker w k For example, let process i be its participating process list O wk The mth process o wkm , then the time when it can start executing process i is determined by the planned start time of process i and w k The time to complete the previous process in the process list is jointly determined. In addition, the influence of random disturbance factors needs to be considered, namely: The actual execution time of process i is determined by the actual process operation time dJ i and rework time dB i Decision: dE i =dJ i +dB i ; Actual process operation time dJ i Rated operation time d of the process i There are differences, namely: dJ i =d i +Δ i ,Δ i ≥0.

5. The method for predicting the progress of the assembly process of a single customized product according to claim 4, characterized in that: Derivation of the lower bound of the completion time of each process based on the completion time of each process includes: Without considering disturbance factors, the lower bound of the actual start time of process i is derived as follows: The logistics delivery completion time of process i is equal to its planned start time, that is: bound lower (part Ei )=SP i Without considering the influence of disturbance factors, the worker w k The earliest time to start executing process i is the maximum of the planned start time of process i and the completion time of the previous process in the worker's participating process list, that is: Derive the earliest start time at which all workers begin to execute process i, that is: Combining the derivation results, we can get the lower bound of the actual start time of process i, namely: Without considering disturbance factors, the lower bound of the actual completion time of process i is derived as follows: In the absence of process operation time extension and process rework, the actual operation time of process i is its rated operation time, that is: bound lower (of i )=d i The lower bound of the actual completion time of process i without considering the influence of disturbance factors is obtained: Obtain the lower bound of the actual completion time of each unfinished process in the actual production process of the workshop, without considering the influence of future disturbance factors; At any time t, all the processes of the workshop production are divided into three sets, namely the completed process set Opt finished , ongoing process set Opt execution 、Unstarted process set Opt notStart ; For i∈Opt finished , its actual completion time is known, then: bound lower (FE i )=max(SE i +d i ,t),i∈Opt finished For the unstarted operation set Opt notStart For the process in the process, the lower bound of its actual completion time needs to be calculated through step-by-step prediction; Divide all processes into two sets, namely the set of processes with known lower bounds of completion time Opt bound , initialized to Opt finished ∪Opt execution , and the set of processes Opt whose lower bound of completion time is unknown notbound , initialized to the set of unstarted processes Opt notStart ; Existence Theorem: There exists an operation i∈Opt notbound ,satisfy and Based on the theorem, Opt is calculated step by step notbound The lower bound of step i in , and update Opt bound With Opt notbound , and finally get the lower bound of the completion time of all processes at the current moment.

6. The method for predicting the progress of the assembly process of a single customized product according to claim 1, characterized in that: The data-driven completion time deviation prediction model has as input: a process diagram network modeled by the actual production status of a single customized product; and outputs: the deviation between the predicted actual completion time and the lower bound; The deviation prediction model includes three SAGE layers, two LSTM layers, and one fully connected layer group; The SAGE layer is used to input the topological structure of the process graph network, automatically aggregate the features between different nodes, and realize the association extraction of node features; The LSTM layer extracts implicit time series features based on the node features aggregated by the SAGE layer and outputs them as hidden layer features; The fully connected layer fits the nonlinear relationship between the hidden layer features output by the LSTM and the target variable, where the target variable refers to the deviation of the actual completion time from a lower bound.

7. The method for predicting the progress of the assembly process of a single customized product according to claim 6, characterized in that: The actual production status of the single customized product is modeled into a process graph network, in which nodes represent processes and edges between nodes represent the association relationship between two processes.

8. The method for predicting the progress of the assembly process of a single customized product according to claim 7, characterized in that: The nodes in the graph include the actual production status characteristics of a single customized product, and the actual production status characteristics of a single customized product include all of the following characteristics: Network complexity NC, resource factor RF, resource intensity RS, rated working hours d i , resource requirements i 、Number of Predecessor Processes|pred i |、Number of subsequent processes|succ i |, earliest start time EST, latest start time LFT, maximum resource requirement GRD, minimum resource requirement SRD, earliest finish time EFT, latest finish time LST, maximum successor quantity MTS, minimum slack time MSLK, maximum position weight GRPW, maximum resource utilization GRU, weighted resource utilization and precedence WRUP, initial plan sequence value PPR, plan start time sequence value PFR, completed process proportion FPR, current time PT, current resource utilization RU, unfinished process working time RT, process slack time ST, average process slack time AST.

9. The method for predicting the progress of the assembly process of a single customized product according to claim 7, characterized in that: The association relationship is used to characterize the precedence constraints and resource sharing relationships between different process nodes; There are three types of association relationships: The first relationship D1 is modeled based on a precedence constraint: The second type of relationship D2 is modeled based on resource sharing: The third type of relationship D3 models the relationship based on the initial plan: Among them: i, j represent the node number, pred j represents the set of preceding processes of process j, O wk represents the task list of worker k of type w, SP sort A list of operations sorted in ascending order based on their planned start time.

10. The method for predicting the progress of the assembly process of a single customized product according to claim 1, characterized in that: The trained prediction model is deployed in an actual production environment, real-time manufacturing data is input into the prediction model, and the completion time of the next process is predicted recursively, including: The predicted completion time of the nth process is regarded as the true value and input into the training model to predict the deviation between the completion time of the n+1th process and the lower bound of the time. The deviation is applied to the obtained lower bound of the time of the corresponding process to obtain the completion time of the (n+1)th process.

Images