Airplane assembly line rhythm prediction and bottleneck identification method

Abstract

In view to the technical problems that the existing general neural network model has insufficient accuracy and lacks credibility when predicting the production rhythm of an aircraft assembly line, the present application provides a kind of aircraft assembly line rhythm prediction and bottleneck identification method, which takes the rated working hours of each operation, resource allocation scheme and material distribution scheme as input, establishes and trains a customized neural network model through mapping the aircraft assembly operation process, the customized neural network model includes SP-LSTM network model added with attention mechanism unit corresponding to each station, integrated gate unit and regression linear layer, which can accurately predict the production rhythm; at the same time, through the posterior analysis of the weight information in the customized neural network model, the identification of various production bottlenecks of the aircraft assembly line can be realized.

Description

Technical Field

[0001] This invention belongs to the field of aircraft assembly line productivity prediction, and specifically designs a method for predicting the cycle time and identifying bottlenecks in aircraft final assembly lines. Background Technology

[0002] Aircraft final assembly lines operate on a multi-variety, small-batch production model, and their production capacity often needs to be adjusted according to market order demand. Deviations between actual production takt time and the production plan can lead to idle resources on the assembly line, soaring material storage costs, and uneven task allocation, potentially delaying order delivery and increasing aircraft production costs. Furthermore, aircraft assembly is a complex systems engineering project involving multiple production factors such as operations, resources, and materials. The production parameters related to these factors interact, ultimately affecting the overall production takt time of the final assembly line. Therefore, the aircraft final assembly line production parameter configuration scheme, composed of key production parameters, also needs to be adjusted accordingly with changes in production targets. Moreover, some key parameters are subject to non-subjective random variations, such as the delivery frequency of externally supplied materials and work hours. Therefore, frequent evaluation and verification of parameter configuration schemes are necessary during aircraft final assembly operations. Specifically, it is necessary to use a specific parameter configuration scheme as input to predict the production takt time of the aircraft final assembly line after its continuous application.

[0003] With the deepening integration of artificial intelligence technology and the manufacturing industry, the use of neural network models to fit the nonlinear relationship between parameter configuration schemes and production capacity indicators of complex assembly systems has gained significant attention and application. However, due to the complexity of aircraft final assembly lines and the limited number of samples that can be collected due to the characteristics of small-batch production, general-purpose neural network models struggle to achieve satisfactory prediction accuracy under incomplete sample conditions. Furthermore, "black box" general-purpose neural network models cannot explain the basis for outputting specific production cycle times. For example, when outputting a poor production cycle time, it cannot explain which bottleneck factors were considered to arrive at such a result, leading to a lack of credibility in field applications. Therefore, against this backdrop, it is necessary to design a customized cycle time prediction and bottleneck identification method specifically for aircraft final assembly lines, incorporating operational information. Summary of the Invention

[0004] To address the technical problems of insufficient accuracy and lack of reliability in predicting production cycle time on aircraft final assembly lines using existing general neural network models, this invention proposes a method for predicting production cycle time and identifying bottlenecks on aircraft final assembly lines. This method takes the rated working hours, resource allocation schemes, and material delivery schemes of each operation as inputs. Through mapping and modeling the aircraft final assembly process, training, and posterior analysis of the weight information, it achieves the prediction of production cycle time on aircraft final assembly lines and the identification of various production bottlenecks.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] The method for predicting cycle time and identifying bottlenecks in aircraft final assembly lines is unique in that it includes the following steps:

[0007] Step 1: Compile basic information on aircraft final assembly site operations:

[0008] The basic operational information includes the total number of assembly stations, a set of key resource allocation schemes, a set of key material delivery schemes, and attribute information of all operations within each station. The set of key resource allocation schemes includes the quantity of each key resource allocated at each station. The set of key material delivery schemes includes the average and standard deviation of the delivery interval time for each key material. The attribute information of any operation within each station includes the average and standard deviation of the rated working hours for that operation, the set of key resource quantities and material quantities required for that operation, and the preceding operations for that operation.

[0009] Step 2: Construct the training set, validation set, and test set:

[0010] Historical data on all key resource configuration schemes, key material delivery schemes, and rated working hour parameters of all operations of the final assembly line are obtained. Historical data belonging to the same period are spliced ​​together to form the overall production parameter scheme of the final assembly line. After adding labels, a complete learning sample dataset is formed. After normalization, it is divided into training set, validation set, and test set. The labels are the final assembly line production cycle time of the overall production parameter scheme of the final assembly line for evaluation or actual recording.

[0011] Step 3: Construct a customized neural network model and train, validate, and test it using the training, validation, and test sets.

[0012] The customized neural network model includes a sub-neural network model, an integrated gate unit, and a regression linear layer;

[0013] There are multiple sub-neural network models, all of which are SP-LSTM network models with added attention mechanism units, and each corresponds one-to-one with a station. For each station: each task within the station corresponds to an LSTM unit, and the input of each LSTM unit is connected to the output of the LSTM unit corresponding to its predecessor task. The outputs of multiple parallel tasks are integrated by the IIG module and then output. The outputs of all LSTM units without subsequent tasks are integrated by the IG module to obtain the output of the sub-neural network model corresponding to that station.

[0014] There are multiple attention units, each corresponding to a task in each station, and they are placed before the LSTM unit corresponding to each task to process the input of the LSTM unit corresponding to each task.

[0015] The integrated gate unit is used to integrate the outputs of all sub-neural network models before outputting them;

[0016] The regression linear layer is used to transform the output vector of the integrated gate unit into a one-dimensional output, which corresponds to the production cycle of the aircraft final assembly line.

[0017] Step 4: Production cycle time prediction for the aircraft final assembly line:

[0018] Extract all work hour parameters, resource allocation schemes allocated to each station, and material delivery parameters consumed in each station from the basic information of aircraft final assembly site operations collected in step 1. Input these parameters into the customized neural network model constructed and trained in step 3 and the corresponding sub-neural network model of each station. The customized neural network model outputs the production cycle of the aircraft final assembly line.

[0019] Step 5: Bottleneck Identification:

[0020] By comparing and analyzing the weight information in the customized neural network model, bottleneck stations, critical operation paths, and scarce resources and materials can be identified.

[0021] Furthermore, the specific method by which the attention mechanism unit processes the input of the LSMT unit in step 3 is as follows:

[0022] When the attention unit has a preceding unit:

[0023] First, each attention mechanism unit uses the output of the preceding unit, the key resource allocation scheme of this unit, the key material delivery scheme of this unit, the quantity of each key resource required for the corresponding task of this unit, and the quantity of materials required for this unit as inputs to calculate the attention score of each key resource and material corresponding to the current task. Second, using all attention scores as inputs, the attention weights of each key resource and material are obtained through the Softmax function transformation. Finally, the attention weights of each key resource allocation scheme and material delivery scheme are multiplied one by one with the resource allocation scheme and material delivery scheme corresponding to the current task. Combined with the rated working hours average and standard deviation corresponding to the current task, the multiplication result, the rated working hours average and standard deviation of the current task are concatenated to form a complete input vector, which is then input to the LSTM unit corresponding to the subsequent task.

[0024] When the attention unit has no preceding unit:

[0025] An initial vector consisting entirely of zeros is used as the input to the attention mechanism unit instead of the input to the preceding unit. The length of the initial vector is the same as the length of the LSTM unit output.

[0026] Furthermore, the ensemble gate unit described in step 3 is used to integrate the hidden states h of the sub-neural network models corresponding to all stations. i and unit state C i and output vector h FIG The process is as follows:

[0027]

[0028] Among them, W FIG This represents the learnable weight matrix in a neural network model, used for dimensionality transformation; g i ω is an intermediate variable; i For the weight parameters to be analyzed in step 5, each ω i A sub-neural network model corresponding to one station.

[0029] Furthermore, the number of intermediate layers and the number of neurons in the intermediate layers of the regression linear layer mentioned in step 3 are hyperparameters of the customized neural network model, which are set according to the verification effect of the customized neural network model in practical applications.

[0030] Furthermore, the method for setting hyperparameters in step 3 is as follows: compare the root mean square error index of the entire customized neural network model under different hyperparameter settings for the same test data, and select the hyperparameter setting scheme with the smallest root mean square error index.

[0031] Furthermore, the method for identifying bottleneck stations in step 5 is as follows:

[0032] Comparative analysis of the internal weight data ω of integrated gate units i The station with the highest weight is the bottleneck station under the current input scheme;

[0033] The specific method for identifying critical task paths in step 5 is as follows:

[0034] Starting from the IG module within the sub-neural network model, the weight data within the IG module is compared, and then backpropagation continues along the path with the largest weight. If backpropagation reaches the IIG module within the sub-neural network model, the weight data within the IIG module is compared, and then backpropagation continues along the path with the largest weight until the attention mechanism unit without a preceding unit is reached. The jobs corresponding to the LSTM units traversed throughout the entire backpropagation process are combined to form the identified critical path.

[0035] The specific method for identifying scarce resources and materials in step 5 is as follows:

[0036] For each key resource and material, a judgment threshold is set. If the weight data of a key resource exceeds its threshold, it indicates that the resource is in short supply. If the weight data of a key material exceeds its threshold, it indicates that the material is in short supply.

[0037] This invention also provides a method for predicting the cycle time of an aircraft final assembly line and identifying bottlenecks, which is unique in that it includes a customized neural network model and a bottleneck identification unit.

[0038] The customized neural network model is used to predict the production cycle of the aircraft assembly line based on all the working time parameters of each station at the aircraft final assembly site, the number of resource allocation schemes allocated to each station, and the material delivery parameters consumed in each station.

[0039] The customized neural network model includes a sub-neural network model integrated gate unit and a regression linear layer;

[0040] There are multiple sub-neural network models, all of which are SP-LSTM network models with added attention mechanism units, and each corresponds one-to-one with a station. For each station: each task within the station corresponds to an LSTM unit, and the input of each LSTM unit is connected to the output of the LSTM unit corresponding to its predecessor task. The outputs of multiple parallel tasks are integrated by the IIG module and then output. The outputs of all LSTM units without subsequent tasks are integrated by the IG module to obtain the output of the sub-neural network model corresponding to that station.

[0041] There are multiple attention units, each corresponding to a task in each station, and they are placed before the LSTM unit corresponding to each task to process the input of the LSTM unit corresponding to each task.

[0042] The integrated gate unit is used to integrate the outputs of all sub-neural network models before outputting them;

[0043] The regression linear layer is used to transform the output vector of the integrated gate unit into a one-dimensional output, which corresponds to the production cycle of the aircraft final assembly line.

[0044] The bottleneck identification unit is used to compare and analyze the weight information in the customized neural network model to identify bottleneck positions, critical operation paths, and scarce resources and materials.

[0045] Furthermore, the bottleneck identification unit is a computer program instruction running on the processor, which is used to implement the following steps in parallel or serially:

[0046] 1) Identify bottleneck locations:

[0047] Comparative analysis of the internal weight data ω of integrated gate units iThe station with the highest weight is the bottleneck station under the current input scheme;

[0048] 2) Identify critical operation paths:

[0049] Starting from the IG module within the sub-neural network model, the weight data within the IG module is compared, and then backpropagation continues along the path with the largest weight. If backpropagation reaches the IIG module within the sub-neural network model, the weight data within the IIG module is compared, and then backpropagation continues along the path with the largest weight until the attention mechanism unit without a preceding unit is reached. The jobs corresponding to the LSTM units traversed throughout the entire backpropagation process are combined to form the identified critical path.

[0050] 3) Identify scarce resources and materials:

[0051] For each key resource and material, a judgment threshold is set. If the weight data of a key resource exceeds its threshold, it indicates that the resource is in short supply. If the weight data of a key material exceeds its threshold, it indicates that the material is in short supply.

[0052] Compared with the prior art, the present invention has the following beneficial effects:

[0053] 1. This invention, through mapping modeling and training of various production elements and links in the aircraft final assembly process, and through posterior analysis of the weight information therein, can accurately predict the production cycle time while identifying various production bottlenecks, thereby enriching the output information of the model, effectively improving the decision transparency and output credibility of the model, and promoting the practical application of neural network methods in the field of aircraft assembly systems.

[0054] 2. This invention custom-constructs a neural network model whose overall network structure and parameters can be mapped to the key elements of the aircraft final assembly process, thereby endowing the model with interpretability of structure and ex-post weight parameters. This allows the model to make fuller use of limited learning samples based on a preset data propagation process, resulting in more satisfactory prediction accuracy. Attached Figure Description

[0055] Figure 1 This is a work logic sequence diagram of an embodiment of the present invention.

[0056] Figure 2 This is a detailed implementation framework diagram of an embodiment of the present invention.

[0057] Figure 3 This is a diagram illustrating the neural network model architecture and forward propagation process of an embodiment of the present invention.

[0058] Figure 4 This is a flowchart illustrating the actual application of an embodiment of the present invention. Detailed Implementation

[0059] To make the objectives, solutions, and advantages of this invention more apparent and understandable, the invention will now be described in detail with reference to the accompanying drawings.

[0060] Reference Figure 2-3 The method for predicting the cycle time and identifying bottlenecks in an aircraft assembly line provided by this invention includes the following steps:

[0061] Step S1: Collect basic information on the aircraft final assembly site.

[0062] The basic information of the final assembly site includes the total number of assembly stations N, the set of key resource allocation schemes RA, the set of key material delivery schemes MD, and the attribute information of all operations within each station.

[0063] in:

[0064] The set of critical resource allocation schemes RA for the aircraft final assembly line is represented as RA = {ra1, ra2, ..., ra...} r ,…,ra R}, where r represents the critical resource index number, R represents the total number of critical resources, and ra r Let ra represent the configuration scheme for the r-th critical resource. r ={N S1_r N S2_r ,…,N Si_r ,…,N SN_r}, N Si_r The amount of the r-th critical resource allocated to the i-th station;

[0065] The critical material allocation scheme (MD) for the aircraft final assembly line is represented as MD = {MT} 1_avg ,MT 1_dev ,MT 2_avg ,MT 2_dev ,…,MT m_avg ,MT m_dev ,…},MT m_avg MT represents the average delivery interval time for the m-th critical material. m_dev This represents the standard deviation of the delivery interval time for the m-th critical material.

[0066] The j-th task within the i-th station is denoted as SiPj. The attribute information of this task includes: the average rated working hours T of this task. SiPj_avg The standard deviation T of the rated working hours for this operation SiPj_dev The set of key resource quantities required for this operation is rr. SiPj The required quantity of key materials for this operation is r. SiPj_material And the Pre-P work preceding this task. SiPjAmong them, the set of key resource quantities required for this operation is rr. SiPj ={Nr1,Nr2,…,Nr r ,…,Nr R}, Nr r This indicates the quantity of the r-th critical resource required for this task.

[0067] Step S2: Obtain historical data to form a learning sample dataset, and further divide it into training set, validation set and test set.

[0068] S201. Obtain historical data of all key resource allocation schemes and key material distribution schemes of the final assembly line, combine them with historical data of rated working hours parameters of all operations of the final assembly line, and combine the historical data of key resource allocation schemes, key material distribution schemes and rated working hours parameters of all operations belonging to the same period to form the overall production parameter scheme of the final assembly line.

[0069] S202. Take the assembly line production cycle time (represented as CT) from historical data that evaluates or actually records the overall production parameter scheme of the assembly line formed in step S201, and use it as a label to form a complete learning sample dataset together with the overall production parameter scheme of the assembly line formed in S201. Then, normalize all numerical features in the learning sample dataset. It is recommended to use the linear normalization method.

[0070] S203. Divide the normalized learning sample dataset into a training set, a validation set, and a test set according to a set ratio, with a suggested ratio of 8:1:1.

[0071] Step S3: Based on the basic information of aircraft assembly site operations collected in Step S1, construct a customized neural network model, design the network model hierarchy, structure and hyperparameters, and use the training set, validation set and test set to train, validate and test the network model.

[0072] S301. Construct a corresponding sub-neural network model for each station:

[0073] The sub-neural network model uses a serial-parallel long short-term memory neural network (SP-LSTM) to map the order of operations within the station. SP-LSTM is a publicly available network model in the prior art. The input to each sub-neural network model consists of parameters specific to the station corresponding to that sub-neural network model. These parameters include the operation time parameters of all operations at that station extracted from the overall operation time parameters, the number of resource allocation schemes allocated to that station extracted from the overall resource allocation scheme, and the distribution parameters of materials consumed at that station extracted from the overall material distribution scheme. The outputs of each sub-neural network model have no explicit meaning and are all sent to the integration gate unit. A single sub-neural network model includes multiple LSTM units, IG modules, and IIG modules. The IG and IIG modules contain weight data, and each weight data corresponds to a data propagation path. For any station: each operation at that station corresponds to an LSTM unit, and the input of each LSTM unit is connected to the output of the LSTM unit corresponding to its preceding operation. The outputs of multiple parallel operations are integrated by the IIG module and then output. The outputs of all LSTM units without a subsequent operation (i.e., the last operation) are integrated by the IG module to obtain the output of the sub-neural network model corresponding to that station.

[0074] S302, Requirements for constructing customized attention mechanism unit mapping tasks:

[0075] Customized attention mechanism units are used to map the resource occupation and material consumption of each task. There are multiple attention mechanism units, which correspond one-to-one with each task in the station. They are placed before each LSTM unit in the sub-neural network model to process the input of the LSTM unit corresponding to each task.

[0076] Each attention mechanism unit first takes the output of the preceding unit (LSTM unit or IIG module), the key resource configuration scheme of this unit, the key material delivery scheme of this unit, the quantity of each key resource required for the corresponding task of this unit, and the quantity of materials required for this unit as inputs to calculate the attention score of each key resource and material corresponding to the current task. Second, taking all attention scores as inputs, the unit transforms them through the Softmax function to obtain the attention weight of each key resource and material. Finally, each attention weight has a unique corresponding key resource configuration scheme item or material delivery scheme item. The attention weights of each key resource configuration scheme and material delivery scheme are multiplied one by one with the configuration scheme of the resource and the delivery scheme of the material corresponding to the current task. Combined with the rated working average and standard deviation of the current task, the multiplication result, the rated working average and standard deviation of the current task are concatenated to form a complete input vector, which is then input to the LSTM unit corresponding to the subsequent task.

[0077] It should be noted that in the above process, if a certain attention mechanism unit has no preceding unit, then an initial vector consisting entirely of zeros is used as the input of that attention mechanism unit to replace the input of the preceding unit. The length of this initial vector is the same as the length of the LSTM unit output.

[0078] To clearly illustrate the calculation process of an attention weight, let's take the attention mechanism unit before the LSTM unit corresponding to the j-th task SiPj inside the i-th station as an example. The calculation process of the attention weight of the r-th key resource is shown in equations (1)-(4). Assuming that the material required by the task SiPj is the m-th key material (the key material selected by the method of this invention is controlled to require only one of them in each task), the final output of the attention mechanism unit is shown in equation (5).

[0079] s resource_r =v T tanh(W resource_r ·Nr r +W resource_a ·N Si_r +W c ·C j-1 (1)

[0080] s material =v T tanh(W material_r ·r SiPj_material +W material_a ·(MT m_avg ,MT m_dev )+W c ·C j-1 (2)

[0081]

[0082]

[0083] x output =(T SiPj_avg ,T SiPj_dev ,ω resource_1 ·N Si_1 ,…,ω resource_r ·N Si_r ,…,ω resource_R ·N Si_R ,ω material ·MT m_avg ,ω material ·MT m_dev ) (5)

[0085] Where: s resource_r C represents the attention score of the r-th key resource;j-1 W represents the output of the preceding unit (LSTM unit or IIG module) of the preceding job of the j-th job, or an initial vector that is all zeros and has a length equal to the output length of the LSTM unit; resource_r W resource_a W rmaterial_r W material_a W c Both and v are learnable weight matrices in the sub-neural network model, used for dimensionality transformation; tanh is an activation function; ω resource_r ω represents the attention weight of the r-th key resource; material Indicates the attention weight of key materials; s resource_r s represents the attention score of the r-th key resource; material Indicates the attention score for key materials; x output This represents the complete input vector formed by concatenation.

[0086] S303, Sub-neural network model integrated output:

[0087] This invention integrates and outputs the outputs of all sub-neural network models by designing an integration gate unit (FIG). Specifically, this integration gate unit is used to integrate the hidden states h of the sub-neural network models corresponding to all stations. i and unit state C i and output vector h FIG The calculation process is shown in equation (4):

[0088]

[0089] Among them, W FIG This represents the learnable weight matrix in a neural network model, used for dimensionality transformation; g i ω is an intermediate variable; i For the weight parameters to be analyzed in step 5, each ω i A sub-neural network model corresponding to one station.

[0090] S304, Regression Linear Layer Design and Output:

[0091] The output vector h of the integrated gate unit in step S303 is... FIGThe input is a linear regression layer, which, after transformation, yields a final one-dimensional output. This one-dimensional output corresponds to the production cycle time on the aircraft assembly line. The specific number of intermediate layers and neurons in the linear regression layer are hyperparameters of the entire customized neural network model, which can be set based on the model's performance in practical applications. Specifically, the root mean square error (RMSE) of the entire customized neural network model under different hyperparameter settings is compared using the same test data, and the hyperparameter setting with the smallest RMS error is selected.

[0092] Step S4: Predicting the production cycle time of the aircraft final assembly line

[0093] Extract all work hour parameters, resource allocation schemes allocated to each station, and material delivery parameters consumed in each station from the basic information of aircraft assembly site operations collected in step S1. Input these parameters into the sub-neural network models corresponding to each station in the customized neural network model constructed and trained in step S3. The customized neural network model outputs the production cycle of the aircraft assembly line.

[0094] Step S5: After the customized neural network model outputs the production cycle time of the aircraft final assembly line, compare and analyze the weight information in the customized neural network model to support the identification of bottleneck stations, critical operation paths, scarce resources and materials.

[0095] S501, Bottleneck Location Identification

[0096] By analyzing the internal weight data ω of the integrated gate unit (FIG) i Comparative analysis supports the identification of bottleneck positions. Specifically, each weight within an integrated gate unit corresponds to an assembly position. The assembly position with the highest weight is the position that the entire customized neural network model needs to focus on most under the current input scheme, and therefore this position is identified as the bottleneck position.

[0097] S502, Critical Operation Path Identification

[0098] The critical path within a station refers to the path with the longest total task time in the task sequence network (i.e., a network that includes all tasks and their pre- and post-processing relationships). Within each sub-neural network model, the critical path is identified progressively according to the backpropagation sequence. Specifically, starting from the IG module within the sub-neural network model, the weight data within the IG module is compared, and then backpropagation continues along the path with the highest weight. If backpropagation reaches the IIG module within the sub-neural network model, the weight data within the IIG module is compared, and then backpropagation continues along the path with the highest weight. This process continues until the attention mechanism unit, which has no preceding units, is reached. Finally, the tasks corresponding to the LSTM units traversed throughout the entire backpropagation process constitute the identified critical path.

[0099] S503, Identification of Shortage Resources and Materials

[0100] Shortage resources and materials are identified by comparing the key resource weights (R in total) and key material weights (1) in the attention mechanism unit preceding each LSTM unit on the critical path. Each weight in the attention mechanism unit corresponds to a key resource or key material. Specifically, a judgment threshold needs to be set for each key resource and material. If the weight of a key resource exceeds its threshold, it indicates that the resource is in short supply; if the weight of a key material exceeds its threshold, it indicates that the material is in short supply. The specific values ​​of the key resource threshold and the key material threshold should be determined based on the model's performance in historical data. Specifically, the initial identification threshold for each weight is set to 1 / (R+1). All shortage resources and materials are identified in historical data, and these are reviewed by experienced on-site managers. Usually, the threshold is set low at the beginning, and a large number of resources or materials that are not actually in short supply on-site are identified. Based on this, the threshold value is gradually increased until the shortage resources and materials identified by the model in historical data are basically consistent with the actual situation.

[0101] Example:

[0102] This embodiment selects a certain type of aircraft final assembly line as an example. This aircraft final assembly line contains 60 operations, which are carried out according to the following... Figure 1 The logical sequence diagram shown is distributed across four discrete assembly stations. In specific applications of this embodiment, it is followed as follows: Figure 2 The method framework diagram shown is used.

[0103] Step S1: Select three key resources as key resource allocation schemes in the aircraft final assembly line: skilled operators, automatic drilling equipment, and laser measurement equipment; select 16 key materials as key material distribution schemes; and on this basis, calculate the rated working hours (including average and standard deviation), preceding operations, key material requirements, and key resource requirements for 60 operations.

[0104] Step S2: A total of 12,015 sets of learning samples were collected and prepared from the historical database of the aircraft assembly line. Then, 1,000 sets were randomly selected as the test set, 1,000 sets as the validation set, and the remaining 10,015 sets as the training set.

[0105] Step S3: Based on the operational sequence within each station, as well as the operational requirements and resource / material configuration parameters, construct a sub-neural network model mapping each assembly station; then, integrate the outputs of each sub-neural network model by establishing integrated gate units; finally, obtain the one-dimensional output of the overall neural network model through a linear regression layer. The architecture and forward propagation process of the constructed overall neural network model are as follows: Figure 3 As shown.

[0106] Step S4: For a customized neural network model that has already been trained, in practical applications, while the model outputs the production cycle time, the weight information in the neural network model can be interpreted through comparative analysis. This supports the identification of bottleneck stations, critical work paths, scarce resources, and scarce materials. Taking a production parameter input scheme as an example, the application process of the constructed model in this embodiment is demonstrated, such as... Figure 4 As shown ( Figure 4 The interpretable neural network model is the customized neural network model mentioned above. When a production parameter scheme is input into this neural network model, it first accurately outputs a predicted production cycle time of 43.176 days / unit (the actual labeled value is 42.673 days / unit). Then, it accurately identifies the first bottleneck station, and identifies S1P1, S1P2, S1P5, S1P7, S1P10, S1P20, and S1P22 as the critical paths. Finally, it identifies the scarce resources and materials for each bottleneck operation along the critical paths.

Claims

1. A method for predicting cycle time and identifying bottlenecks in aircraft final assembly lines, characterized in that, Includes the following steps: Step 1: Compile basic information on aircraft final assembly site operations: The basic operational information includes the total number of assembly stations, a set of key resource allocation schemes, a set of key material delivery schemes, and attribute information of all operations within each station. The set of key resource allocation schemes includes the quantity of each key resource allocated at each station. The set of key material delivery schemes includes the average and standard deviation of the delivery interval time for each key material. The attribute information of any operation within each station includes the average and standard deviation of the rated working hours for that operation, the set of key resource quantities and material quantities required for that operation, and the preceding operations for that operation. Step 2: Construct the training set, validation set, and test set: Historical data on all key resource configuration schemes, key material delivery schemes, and rated working hour parameters of all operations of the final assembly line are obtained. Historical data belonging to the same period are spliced ​​together to form the overall production parameter scheme of the final assembly line. After adding labels, a complete learning sample dataset is formed. After normalization, it is divided into training set, validation set, and test set. The labels are the final assembly line production cycle time of the overall production parameter scheme of the final assembly line for evaluation or actual recording. Step 3: Construct a customized neural network model and train, validate, and test it using the training, validation, and test sets. The customized neural network model includes a sub-neural network model, an integrated gate unit, and a regression linear layer; There are multiple sub-neural network models, all of which are SP-LSTM network models with added attention mechanism units, and each corresponds one-to-one with a station. For each station: each task within the station corresponds to an LSTM unit, and the input of each LSTM unit is connected to the output of the LSTM unit corresponding to its predecessor task. The outputs of multiple parallel tasks are integrated by the IIG module and then output. The outputs of all LSTM units without subsequent tasks are integrated by the IG module to obtain the output of the sub-neural network model corresponding to that station. There are multiple attention mechanism units, each corresponding to a task in each station, and they are set before the LSTM unit corresponding to each task to process the input of the LSTM unit corresponding to each task. The integrated gate unit is used to integrate the outputs of all sub-neural network models before outputting them; The regression linear layer is used to transform the output vector of the integrated gate unit into a one-dimensional output, which corresponds to the production cycle of the aircraft final assembly line. Step 4: Production cycle time prediction for the aircraft final assembly line: Extract all work hour parameters, resource allocation schemes allocated to each station, and material delivery parameters consumed in each station from the basic information of aircraft final assembly site operations collected in step 1. Input these parameters into the customized neural network model constructed and trained in step 3 and the corresponding sub-neural network model of each station. The customized neural network model outputs the production cycle of the aircraft final assembly line. Step 5: Bottleneck Identification: By comparing and analyzing the weight information in the customized neural network model, bottleneck stations, critical operation paths, and scarce resources and materials can be identified.

2. The method for predicting the cycle time and identifying bottlenecks in an aircraft final assembly line according to claim 1, characterized in that: The specific method by which the attention mechanism unit processes the input of the LSMT unit in step 3 is as follows: When the attention mechanism unit has a preceding unit: First, each attention mechanism unit uses the output of the preceding unit, the key resource allocation scheme of this unit, the key material delivery scheme of this unit, the quantity of each key resource required for the corresponding task of this unit, and the quantity of materials required for this unit as inputs to calculate the attention score of each key resource and material corresponding to the current task. Second, using all attention scores as inputs, the attention weights of each key resource and material are obtained through the Softmax function transformation. Finally, the attention weights of each key resource allocation scheme and material delivery scheme are multiplied one by one with the resource allocation scheme and material delivery scheme corresponding to the current task. Combined with the rated working hours average and standard deviation corresponding to the current task, the multiplication result, the rated working hours average and standard deviation of the current task are concatenated to form a complete input vector, which is then input to the LSTM unit corresponding to the subsequent task. When the attention mechanism unit has no preceding unit: An initial vector consisting entirely of zeros is used as the input to the attention mechanism unit instead of the input to the preceding unit. The length of the initial vector is the same as the length of the LSTM unit output.

3. The method for predicting the cycle time and identifying bottlenecks in an aircraft final assembly line according to claim 1 or 2, characterized in that: The ensemble gate unit described in step 3 is used to integrate the hidden states h of the sub-neural network models corresponding to all stations. i and unit state C i and output vector h FIG The process is as follows: Among them, W FIG This represents the learnable weight matrix in a neural network model, used for dimensionality transformation; g i ω is an intermediate variable; i For the weight parameters to be analyzed in step 5, each ω i A sub-neural network model corresponding to one station.

4. The method for predicting the cycle time and identifying bottlenecks in an aircraft final assembly line according to claim 3, characterized in that: The number of intermediate layers and the number of neurons in the intermediate layers of the regression linear layer mentioned in step 3 are hyperparameters of the customized neural network model, which are set according to the verification effect of the customized neural network model in practical applications.

5. The method for predicting the cycle time and identifying bottlenecks in an aircraft final assembly line according to claim 4, characterized in that: The method for setting hyperparameters in step 3 is as follows: compare the root mean square error (RMSE) of the entire customized neural network model under different hyperparameter settings for the same test data, and select the hyperparameter setting scheme with the smallest RMSE.

6. The method for predicting the cycle time and identifying bottlenecks in an aircraft final assembly line according to claim 4, characterized in that: The specific method for identifying bottleneck stations in step 5 is as follows: Comparative analysis of the internal weight data ω of integrated gate units i The station with the highest weight is the bottleneck station under the current input scheme; The specific method for identifying critical task paths in step 5 is as follows: Starting from the IG module within the sub-neural network model, the weight data within the IG module is compared, and then backpropagation continues along the path with the largest weight. If backpropagation reaches the IIG module within the sub-neural network model, the weight data within the IIG module is compared, and then backpropagation continues along the path with the largest weight until the attention mechanism unit without a preceding unit is reached. The jobs corresponding to the LSTM units traversed throughout the entire backpropagation process are combined to form the identified critical path. The specific method for identifying scarce resources and materials in step 5 is as follows: For each key resource and material, a judgment threshold is set. If the weight data of a key resource exceeds its threshold, it indicates that the resource is in short supply. If the weight data of a key material exceeds its threshold, it indicates that the material is in short supply.

7. An aircraft final assembly line cycle time prediction and bottleneck identification system, characterized in that: This includes customized neural network models and bottleneck identification units; The customized neural network model is used to predict the production cycle of the aircraft assembly line based on all the working time parameters of each station at the aircraft final assembly site, the number of resource allocation schemes allocated to each station, and the material delivery parameters consumed in each station. The customized neural network model includes a sub-neural network model integrated gate unit and a regression linear layer; There are multiple sub-neural network models, all of which are SP-LSTM network models with added attention mechanism units, and each corresponds one-to-one with a station. For each station: each task within the station corresponds to an LSTM unit, and the input of each LSTM unit is connected to the output of the LSTM unit corresponding to its predecessor task. The outputs of multiple parallel tasks are integrated by the IIG module and then output. The outputs of all LSTM units without subsequent tasks are integrated by the IG module to obtain the output of the sub-neural network model corresponding to that station. There are multiple attention mechanism units, each corresponding to a task in each station, and they are set before the LSTM unit corresponding to each task to process the input of the LSTM unit corresponding to each task. The integrated gate unit is used to integrate the outputs of all sub-neural network models before outputting them; The regression linear layer is used to transform the output vector of the integrated gate unit into a one-dimensional output, which corresponds to the production cycle of the aircraft final assembly line. The bottleneck identification unit is used to compare and analyze the weight information in the customized neural network model to identify bottleneck positions, critical operation paths, and scarce resources and materials.

8. The aircraft final assembly line cycle time prediction and bottleneck identification system according to claim 7, characterized in that: The bottleneck identification unit is a computer program instruction running on a processor, which is used to implement the following steps in parallel or serially: 1) Identify bottleneck locations: Comparative analysis of the internal weight data ω of integrated gate units i The station with the highest weight is the bottleneck station under the current input scheme; 2) Identify critical operation paths: Starting from the IG module within the sub-neural network model, the weight data within the IG module is compared, and then backpropagation continues along the path with the largest weight. If backpropagation reaches the IIG module within the sub-neural network model, the weight data within the IIG module is compared, and then backpropagation continues along the path with the largest weight until the attention mechanism unit without a preceding unit is reached. The jobs corresponding to the LSTM units traversed throughout the entire backpropagation process are combined to form the identified critical path. 3) Identify scarce resources and materials: For each key resource and material, a judgment threshold is set. If the weight data of a key resource exceeds its threshold, it indicates that the resource is in short supply. If the weight data of a key material exceeds its threshold, it indicates that the material is in short supply.

Images