Assembly completion state prediction and optimization method based on interpretable neural network

Abstract

The invention provides an assembly completion state prediction and optimization method based on an interpretable neural network, and belongs to the technical field of assembly state prediction. Comprising the following steps: constructing a system model of a station type assembly line, and obtaining historical operation information of the system model; wherein the system model comprises a plurality of assembly stations which are connected in sequence, and each assembly station is configured with a corresponding assembly task, assembly personnel and assembly materials; constructing an explainable neural network, and training, testing and verifying the explainable neural network by using the historical operation information and the logic relationship among the elements in the system model to obtain a final explainable neural network; and the final interpretable neural network is used to predict the assembly completion state of the station type assembly line, and the station type assembly line is optimized according to the prediction result. The working efficiency and flexibility of the station type assembly production line can be improved.

Description

Technical Field

[0001] This application relates to the field of assembly state prediction technology, and in particular to an assembly state prediction and optimization method based on an interpretable neural network. Background Technology

[0002] In highly developed modern manufacturing systems, the station-based assembly production model has become the core production organization form in discrete manufacturing fields (such as automobile manufacturing, electronics assembly, and aerospace) due to its flexibility and efficiency. This production organization form breaks down the overall assembly process into a series of interconnected or parallel assembly stations, each undertaking a specific assembly task. The smooth execution of all assembly tasks highly depends on the real-time, close, and efficient coordination among various manufacturing elements such as workers, materials, and tools at each assembly station.

[0003] With the deepening promotion and practice of lean manufacturing concepts in manufacturing systems, more stringent requirements have been placed on the on-time delivery of target products on assembly lines. On-site management personnel urgently need to monitor the real-time assembly status of target products at each assembly station in order to accurately predict their delivery time. This allows for dynamic adjustments to the assembly tasks across the entire assembly line, ensuring the on-time delivery of target products. Therefore, accurate prediction of target product delivery time has become a key indicator for measuring the management level of assembly lines and a crucial reference for optimizing assembly line scheduling strategies.

[0004] Driven by Industry 4.0, assembly lines have significantly improved their real-time perception and status monitoring capabilities by deeply integrating with advanced technologies such as digital twins and the Internet of Things. Against this backdrop, neural network-based prediction methods, with their rapid computational power and strong nonlinear modeling potential, have demonstrated significant advantages in manufacturing system prediction tasks.

[0005] However, there are still significant challenges in applying neural networks to predict the delivery time of target products, specifically the following issues:

[0006] First, highly nonlinear and complex interactions lead to inaccurate predictions: the operating environment of existing assembly lines is extremely complex. Dynamic and multi-layered complex interactions exist among assembly tasks, workers, and materials; simultaneously, frequent uncertainties such as material delivery delays, fluctuations in manual operation efficiency, and occasional equipment failures cause the expected delivery time of the target product to exhibit highly nonlinear characteristics. Existing neural network models struggle to fully capture and model these complex interactions and uncertainties, resulting in a significant reduction in the accuracy of predictions under complex operating conditions.

[0007] Secondly, the "black box" nature of neural network models leads to a lack of interpretability: existing neural network models (especially deep neural network models) are often considered "black boxes," and the internal logic and decision-making basis of their predictions are difficult to understand. However, in production scheduling optimization, managers not only need to know "when to complete," but also "why at this time" and "which key factors (such as specific station congestion, material shortages, and labor efficiency) affect the prediction results." Existing neural network models lack an explicit mapping to the actual organizational structure of the assembly production site and the logical relationships between its various elements, resulting in a lack of interpretability in their predictions and an inability to provide on-site managers with intuitive and reliable decision-making basis and optimization direction.

[0008] In summary, while existing neural network prediction methods possess potential, their accuracy and interpretability face significant challenges when applied to the complex scenario of a stationary assembly line. Therefore, it is necessary to propose a solution to address one or more of the problems existing in the aforementioned related technologies.

[0009] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of this application, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0010] This application provides a method for predicting and optimizing assembly completion state based on an interpretable neural network, the method comprising the following steps:

[0011] Construct a system model of a station-based assembly production line and obtain historical operation information from the system model; the system model includes multiple assembly stations connected in sequence, and each assembly station is configured with corresponding assembly tasks, assembly personnel and assembly materials;

[0012] An interpretable neural network is constructed, and historical task information and the logical relationships between the elements in the system model are used to train, test and verify the interpretable neural network to obtain the final interpretable neural network.

[0013] The final interpretable neural network is used to predict the assembly completion status of the station-type assembly line, and the station-type assembly line is optimized based on the prediction results.

[0014] Furthermore, the historical work information includes assembly task historical datasets, assembly personnel historical datasets, and assembly material historical datasets. All data in the assembly task historical dataset, assembly personnel historical dataset, and assembly material historical dataset contain corresponding time tags.

[0015] The assembly task history dataset contains the assembly tasks configured for each assembly station, along with their corresponding operation durations. It means that, among them, Indicates the first The first assembly station The assembly task has not yet started; at this time... Indicates that the first number was not recorded. The first assembly station The duration of each assembly task; Indicates the first The first assembly station An assembly task is being executed at this time. Indicates the first The first assembly station has completed its first execution. The duration of each assembly task; Indicates the first The first assembly station The assembly task has been completed. Indicates completion of the first The first assembly station The time required for each assembly task;

[0016] The assembly personnel historical dataset contains the number of personnel and shift durations for all job types configured at each assembly station. It means that, among them, Indicates the first The first assembly station is equipped with the first Number of personnel for each job type Indicates the first The first assembly station is equipped with the first Shift lengths for each job type;

[0017] The assembly material history dataset contains all types of materials configured at each assembly station, along with the corresponding delivery time intervals and quantities per delivery. It means that, among them, Indicates the first The first assembly station configuration similar materials Indicates the first The first assembly station configuration Delivery time interval for similar materials Indicates the first The first assembly station configuration The quantity of each type of material delivered in a single shipment.

[0018] Furthermore, the logical relationships between the elements in the system model include the assembly station relationship matrix between all assembly stations, the assembly task relationship matrix between all assembly tasks, the assembly personnel relationship matrix between all assembly personnel, the personnel-task relationship matrix between all assembly personnel and their corresponding assembly tasks, and the material-task relationship matrix between all assembly materials and their corresponding assembly tasks.

[0019] Furthermore, the steps involved in constructing an interpretable neural network and training, testing, and validating it using historical task information and the logical relationships between elements in the system model to obtain the final interpretable neural network include:

[0020] Construct an interpretable neural network, which consists of an assembly state diagram, a first layer of heterogeneous graph attention neural network, and a second layer of heterogeneous graph attention neural network connected in sequence.

[0021] An assembly state diagram is constructed using the assembly station relationship matrix, the assembly task relationship matrix, the assembly personnel relationship matrix, the personnel-task relationship matrix, and the material-task relationship matrix.

[0022] The historical datasets of assembly tasks, assembly personnel, and assembly materials were normalized separately, and all normalization results were integrated to obtain a preprocessed dataset.

[0023] The preprocessed dataset is divided into a training set, a test set, and a validation set;

[0024] The interpretable neural network is trained, tested, and validated using a preprocessed dataset to obtain the final interpretable neural network.

[0025] Furthermore, the assembly state diagram contains multiple assembly station sub-diagrams, and the assembly state diagram is represented as follows: ,in, This represents an assembly state diagram. Indicates the first A diagram of the assembly station locations. , Indicates the first Assembly task feature matrix in each assembly station sub-graph Indicates the first Assembly station location sub-map assembly material feature matrix Indicates the first The assembly worker feature matrix in the sub-graph of each assembly station location. Indicates the first Assembly task relationship matrix in the sub-graph of each assembly station location. Indicates the first Material-task relationship matrix in the assembly station sub-diagram Indicates the first Personnel-task relationship matrix in the sub-graph of each assembly station location This represents the time-stamped features corresponding to the preprocessed data;

[0026] Each assembly station sub-diagram corresponds to one of the assembly stations.

[0027] Furthermore, the first layer of the heterogeneous graph attention neural network includes a feature transformation module, a feature splitting module, and a feature aggregation module connected in sequence;

[0028] The feature transformation module performs feature dimension transformation on all received assembly station sub-graphs to obtain all assembly feature vectors corresponding to each assembly station sub-graph.

[0029] The feature decomposition module decomposes all assembly feature vectors corresponding to each assembly station sub-graph into task-task process path, material-task process path, personnel-task process path and task-time process path respectively.

[0030] The feature aggregation module aggregates the feature vectors output from each process path in each assembly feature vector to obtain the corresponding aggregated vector; it concatenates each assembly task with all the aggregated vectors obtained under the four process paths to obtain the corresponding aggregated vector matrix; it uses the first feature transformation matrix with learnable parameters to reduce the dimensionality of each aggregated vector matrix to obtain the one-dimensional vector corresponding to each assembly task; it concatenates the one-dimensional vectors corresponding to all assembly tasks in each assembly station subgraph to obtain the concatenation matrix corresponding to each assembly station subgraph, and the concatenation matrix includes multiple concatenation vectors.

[0031] The feature aggregation module includes a graph attention network and The activation function, with four process paths sequentially passing through the attention network and... The activation function is used to obtain the corresponding feature vector.

[0032] Furthermore, the expression for feature dimension transformation is:

[0033] (1)

[0034] in, Indicates the first The first assembly station location in the sub-diagram Assembly feature vectors corresponding to each assembly task. Indicates the first The first assembly station location in the sub-diagram Each assembly material has a corresponding assembly feature vector. Indicates the first The first assembly station location diagram Assembly feature vector corresponding to each assembler. Indicates the first In the assembly task feature matrix of the assembly station sub-graph, the first... Each assembly task feature vector Indicates the first In the assembly task feature matrix of the assembly station sub-graph, the first... Each assembly material feature vector Indicates the first In the assembly worker feature matrix of the assembly station location sub-map, the first... A feature vector of an assembly worker. Indicates the first The first feature transformation matrix with learnable parameters corresponding to the assembly task in each assembly station subgraph. Indicates the first The first feature transformation matrix with learnable parameters corresponding to the assembly materials in each assembly station sub-graph. Indicates the first The first feature transformation matrix with learnable parameters corresponding to the assembly personnel in each assembly station subgraph. express 3D real space, express 3D real space, This represents the output dimension of the feature transformation process.

[0035] Furthermore, the second layer of the heterogeneous graph attention neural network includes a linear transformation layer, a feature aggregation layer, and a feature fitting layer connected sequentially;

[0036] The linear transformation layer contains multiple second feature transformation matrices with learnable parameters. Each second feature transformation matrix with learnable parameters is used to compress the corresponding concatenated vector to obtain the corresponding one-dimensional feature vector.

[0037] The feature aggregation layer aggregates each one-dimensional feature vector into a corresponding multi-dimensional feature vector according to the assembly station relationship matrix, and concatenates all the multi-dimensional feature vectors to obtain a multi-dimensional concatenated feature vector.

[0038] The feature fitting layer uses all the second feature transformation matrices with learnable parameters to fit the mapping relationship between each multidimensional spliced ​​feature vector and the operation time to complete the assembly task, thereby obtaining the expected completion time of the target product.

[0039] Furthermore, the expression for a one-dimensional eigenvector is:

[0040] (2)

[0041] in, Indicates the first The one-dimensional feature vector is obtained by compressing the spliced ​​vectors corresponding to the assembly station sub-maps. Indicates the first The splicing vector corresponding to each assembly station location sub-graph Represents concatenated vectors The corresponding concatenated feature transformation matrix, express 3D real space, Indicates the first The number of all assembly tasks in the assembly station sub-diagram;

[0042] The expression for the predicted time to complete the corresponding assembly task in the assembly station sub-graph is:

[0043] (3)

[0044] in, Indicates the estimated completion time of the target product. This represents a multi-dimensional concatenated feature vector. This represents the multidimensional concatenated feature transformation matrix corresponding to the multidimensional concatenated feature vector. express 3D real space, Represents multidimensional concatenated feature vectors Dimensions.

[0045] Furthermore, the steps of using an ultimately interpretable neural network to predict the assembly completion status of a station-type assembly line and optimizing the station-type assembly line based on the prediction results include:

[0046] The supply frequency of assembly materials configured for each assembly station, as well as the number of personnel and shift duration for each job type, are optimized based on the expected completion time of the target product.

[0047] This application provides a method for predicting and optimizing assembly completion state based on an interpretable neural network, which has at least the following beneficial effects:

[0048] (1) This application constructs an interpretable neural network and uses the interpretable neural network to predict the completion status of the assembly line, so that on-site management personnel can make accurate predictions of the completion time based on the assembly status of the target product at any time.

[0049] (2) This application optimizes the supply frequency of the assembly materials configured for each assembly station according to the expected completion time of the target product, as well as the number of personnel and shift duration corresponding to the job, thereby enabling timely and effective adjustment and optimization of the assembly production line. Attached Figure Description

[0050] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application. It is obvious that the drawings described below are merely some embodiments of this application, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0051] Figure 1 The diagram illustrates the steps of an assembly completion state prediction and optimization method based on an interpretable neural network in an exemplary embodiment of this application.

[0052] Figure 2 This illustration shows a flowchart of an assembly completion state prediction and optimization method based on an interpretable neural network in an exemplary embodiment of this application.

[0053] Figure 3 A schematic diagram illustrating the relationship between the assembly stations in an exemplary embodiment of this application;

[0054] Figure 4 A schematic diagram illustrating the relationship between various assembly tasks in an exemplary embodiment of this application;

[0055] Figure 5 This is a schematic diagram of assembly station 1 in an aircraft assembly state diagram according to an exemplary embodiment of this application;

[0056] Figure 6 This diagram illustrates the first layer of a heterogeneous graph attention neural network in an exemplary embodiment of this application.

[0057] Figure 7 This diagram illustrates a second-layer heterogeneous graph attention neural network in an exemplary embodiment of this application.

[0058] Figure 8 A schematic diagram illustrating the bottleneck location analysis process in an exemplary embodiment of this application is shown;

[0059] Figure 9 This diagram illustrates the key element analysis process in an exemplary embodiment of this application. Detailed Implementation

[0060] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided to make this application more comprehensive and complete, and to fully convey the concept of the exemplary embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0061] Furthermore, the accompanying drawings are merely illustrative of this application and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0062] The following will provide a more detailed description of the assembly completion state prediction and adjustment method based on an interpretable neural network proposed in this example embodiment.

[0063] This example implementation provides a method for predicting and optimizing assembly completion status based on an interpretable neural network, such as... Figure 1 and Figure 2 As shown, the method may include the following steps:

[0064] Step S101 of this embodiment: as follows Figure 3 As shown, a system model of a station-type assembly production line is constructed, and historical operation information of the system model is obtained; wherein, the system model includes multiple assembly stations connected in sequence, and each assembly station is configured with corresponding assembly tasks, assembly personnel and assembly materials.

[0065] Furthermore, such as Figure 4 As shown, in this embodiment, taking an aircraft assembly line as an example, [the process involves]... Figure 4 As shown, the framework structure of the system model includes five assembly stations connected in sequence, as well as assembly tasks, assembly personnel, and assembly materials configured in each assembly station.

[0066] Furthermore, the relevant historical production datasets include assembly task history datasets, assembly personnel history datasets, and assembly material history datasets. For example... Figure 3 As shown in the figure, this embodiment takes an aircraft assembly line as an example for illustration. The assembly line includes 5 assembly stations, involving 95 assembly tasks, 24 types of assembly personnel and 22 types of assembly materials. The specific allocation of each assembly station is shown in Table 1 below:

[0067] Table 1. Distribution of different assembly stations on the aircraft assembly production line

[0068]

[0069] Table 1 clearly shows the assembly tasks, personnel, and materials at different assembly stations on the aircraft assembly production line.

[0070] Furthermore, Tables 2 to 6 show the configuration of different assembly stations on the aircraft assembly production line.

[0071] Table 2 Configuration of Assembly Station 1

[0072]

[0073] Table 3 Configuration of Assembly Station 2

[0074]

[0075] Table 4 Configuration of Assembly Station 3

[0076]

[0077] Table 5 Configuration of Assembly Station 4

[0078]

[0079] Table 6 Configuration of Assembly Station 5

[0080]

[0081] Tables 2 to 6 show that the historical dataset of assembly tasks includes the assembly tasks configured for each assembly station, as well as the corresponding operation duration. It means that, among them, Indicates the first The first assembly station The assembly task has not yet started; at this time... Indicates that the first number was not recorded. The first assembly station The duration of each assembly task; Indicates the first The first assembly station An assembly task is being executed at this time. Indicates the first The first assembly station has completed its first execution. The duration of each assembly task; Indicates the first The first assembly station The assembly task has been completed. Indicates completion of the first The first assembly station The time required for each assembly task.

[0082] The assembly personnel historical dataset contains the number of personnel and shift durations for all job types configured at each assembly station. It means that, among them, Indicates the first The first assembly station is equipped with the first Number of personnel for each job type Indicates the first The first assembly station is equipped with the first The shift schedule for each job type.

[0083] The assembly material history dataset contains all types of materials configured at each assembly station, along with the corresponding delivery time intervals and quantities per delivery. It means that, among them, Indicates the first The first assembly station configuration similar materials Indicates the first The first assembly station configuration Delivery time interval for similar materials Indicates the first The first assembly station configuration The quantity of each type of material delivered in a single shipment.

[0084] Furthermore, such as Figure 5 As shown, the logical relationships between the elements in the system model include the assembly station relationship matrix between all assembly stations, the assembly task relationship matrix between all assembly tasks, the assembly personnel relationship matrix between all assembly personnel, the personnel-task relationship matrix between all assembly personnel and their corresponding assembly tasks, and the material-task relationship matrix between all assembly materials and their corresponding assembly tasks.

[0085] Step S102 of this embodiment: Construct an interpretable neural network, and use historical task information and the logical relationships between the elements in the system model to train, test, and verify the interpretable neural network to obtain the final interpretable neural network. Step S102 of this embodiment may include the following sub-steps:

[0086] Sub-step S1021: Construct an interpretable neural network, which includes an assembly state diagram, a first-layer heterogeneous graph attention neural network, and a second-layer heterogeneous graph attention neural network connected in sequence.

[0087] Furthermore, such as Figure 6 As shown, the first layer of the heterogeneous graph attention neural network includes a feature transformation module, a feature splitting module, and a feature aggregation module connected in sequence.

[0088] The feature conversion module performs feature dimension conversion on all received assembly station sub-graphs to obtain all assembly element features corresponding to each assembly station sub-graph.

[0089] Here, the expression for feature dimension transformation is:

[0090] (1)

[0091] in, Indicates the first The first assembly station location diagram Assembly feature vectors corresponding to each assembly task. Indicates the first The first assembly station location diagram The assembly feature vector corresponding to each assembly material. Indicates the first The first assembly station location diagram Assembly feature vector corresponding to each assembler. Indicates the first In the assembly task feature matrix of the assembly station sub-graph, the first... Each assembly task feature vector Indicates the first In the assembly task feature matrix of the assembly station sub-graph, the first... Each assembly material feature vector Indicates the first In the assembly worker feature matrix of the assembly station location sub-map, the first... A feature vector of an assembly worker. Indicates the first The first feature transformation matrix with learnable parameters corresponding to the assembly task in each assembly station subgraph. Indicates the first The first feature transformation matrix with learnable parameters corresponding to the assembly materials in each assembly station sub-graph. Indicates the first The first feature transformation matrix with learnable parameters corresponding to the assembly personnel in each assembly station subgraph. express 3D real space, express 3D real space, This represents the output dimension of the feature transformation process.

[0092] The feature splitting module splits all assembly feature vectors corresponding to each assembly station sub-graph into task-task process path, material-task process path, personnel-task process path, and task-time process path.

[0093] The feature aggregation module aggregates the feature vectors output by each process path in each assembly feature vector to obtain the corresponding aggregated vector; it concatenates each assembly task with all aggregated vectors obtained under the four process paths to obtain the corresponding aggregated vector matrix; it uses the first feature transformation matrix with learnable parameters to reduce the dimensionality of each aggregated vector matrix to obtain the one-dimensional vector corresponding to each assembly task; it concatenates the one-dimensional vectors corresponding to all assembly tasks in each assembly station subgraph to obtain the concatenation matrix corresponding to each assembly station subgraph, and the concatenation matrix includes multiple concatenation vectors.

[0094] The feature aggregation module includes a graph attention network and The activation function, with four process paths sequentially passing through the attention network and... The activation function is used to obtain the corresponding feature vector.

[0095] Furthermore, such as Figure 7 As shown, the second layer of the heterogeneous graph attention neural network includes a linear transformation layer, a feature aggregation layer, and a feature fitting layer connected in sequence.

[0096] The linear transformation layer contains multiple second feature transformation matrices with learnable parameters. Each second feature transformation matrix with learnable parameters is used to compress the corresponding concatenated vector to obtain the corresponding one-dimensional feature vector.

[0097] Here, the expression for a one-dimensional feature vector is:

[0098] (2)

[0099] in, Indicates the first The one-dimensional feature vector is obtained by compressing the spliced ​​vectors corresponding to the assembly station sub-maps. Indicates the first The splicing vector corresponding to each assembly station location sub-graph Represents concatenated vectors The corresponding concatenated feature transformation matrix, express 3D real space, Indicates the first The number of all assembly tasks in the assembly station sub-graph.

[0100] The feature aggregation layer aggregates each one-dimensional feature vector into a corresponding feature vector based on the assembly station location relationship matrix. 3D feature vectors, and all By concatenating the 1D feature vectors, we obtain The feature vectors are concatenated in 8 dimensions. Each one-dimensional feature vector is 8-dimensional.

[0101] The feature fitting layer utilizes all learnable matrices to respectively fit each The mapping relationship between the 3D splicing feature vector and the operation time to complete the assembly task is fitted to obtain the expected completion time of the target product.

[0102] Here, the expression for the predicted time for completing the corresponding assembly task in the assembly station subgraph is:

[0103] (3)

[0104] in, Indicates the estimated completion time of the target product. This represents a multi-dimensional concatenated feature vector. This represents the multidimensional concatenated feature transformation matrix corresponding to the multidimensional concatenated feature vector. express 3D real space, .

[0105] Sub-step S1022: Construct the assembly state diagram using the assembly station relationship matrix, assembly task relationship matrix, assembly personnel relationship matrix, personnel-task relationship matrix, and material-task relationship matrix.

[0106] Furthermore, the assembly state diagram contains multiple assembly station sub-diagrams, and the assembly state diagram is represented as follows: ,in, This represents an assembly state diagram. Indicates the first A diagram of the assembly station locations. , Indicates the first Assembly task feature matrix in each assembly station sub-graph Indicates the first Assembly station location sub-map assembly material feature matrix Indicates the first The assembly worker feature matrix in the sub-graph of each assembly station location. Indicates the first Assembly task relationship matrix in the sub-graph of each assembly station location. Indicates the first Material-task relationship matrix in the assembly station sub-diagram Indicates the first Personnel-task relationship matrix in the sub-graph of each assembly station location This represents the time-stamped features corresponding to the preprocessed data.

[0107] Furthermore, each assembly station sub-diagram corresponds to one assembly station.

[0108] Sub-step S1023: Normalize the assembly task history dataset, assembly personnel history dataset, and assembly material history dataset respectively, and integrate all normalization results to obtain a preprocessed dataset.

[0109] Sub-step S1024: Divide the preprocessed dataset into a training set, a test set, and a validation set.

[0110] Furthermore, in this embodiment, the preferred ratio of the training set, validation set, and test set is 7:2:1.

[0111] Sub-step S1025: Use the preprocessed dataset to train, test, and validate the interpretable neural network to obtain the final interpretable neural network.

[0112] In step S103 of this embodiment, the assembly completion status of the station-type assembly production line is predicted using an ultimately interpretable neural network, and the station-type assembly production line is optimized based on the prediction results.

[0113] When the target product completion time does not meet requirements, adjustments can be made by referring to the interpretable parameters in the final interpretable neural network. For example... Figure 8 As shown, the attention parameters between assembly stations in the second-layer heterogeneous graph neural network can be visualized. The assembly station with the highest attention parameter is the bottleneck station and should be prioritized for optimization and adjustment. After identifying the bottleneck station, based on the weight parameters between different assembly stations in the first-layer heterogeneous graph neural network, the relationships between the three elements of task-task, personnel-task, and material-task in the corresponding assembly station can be visualized, as shown below. Figure 9 In the context of a, b, and c, a higher weight parameter means that the element is adjusted more preferentially.

[0114] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0115] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. In addition, those skilled in the art can combine and integrate the different embodiments or examples described in this specification.

[0116] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the scope of the technology disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application.

[0117] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein.

Claims

1. A method for predicting and optimizing assembly completion state based on interpretable neural networks, characterized in that, The method includes the following steps: A system model of a station-based assembly production line is constructed, and historical operation information of the system model is obtained; wherein, the system model includes multiple assembly stations connected in sequence, and each assembly station is configured with corresponding assembly tasks, assembly personnel and assembly materials; An interpretable neural network is constructed, and the historical operation information and the logical relationships between the elements in the system model are used to train, test and verify the interpretable neural network to obtain the final interpretable neural network. The final interpretable neural network is used to predict the assembly completion status of the station-type assembly line, and the station-type assembly line is optimized based on the prediction results.

2. The assembly completion state prediction and optimization method based on interpretable neural networks according to claim 1, characterized in that, The historical work information includes assembly task historical dataset, assembly personnel historical dataset, and assembly material historical dataset. All data in the assembly task historical dataset, the assembly personnel historical dataset, and the assembly material historical dataset contain corresponding time tags. The assembly task history dataset contains the assembly tasks configured for each assembly station, along with the corresponding operation durations. It means that, among them, Indicates the first The first assembly station The assembly task has not yet started; at this time... Indicates that the first number was not recorded. The first assembly station The duration of each assembly task; Indicates the first The first assembly station An assembly task is being executed at this time. Indicates the first The first assembly station has completed its first execution. The duration of each assembly task; Indicates the first The first assembly station The assembly task has been completed. Indicates completion of the first The first assembly station The time required for each assembly task; The assembly personnel historical dataset contains the number of personnel and shift durations for all job types configured at each assembly station. It means that, among them, Indicates the first The first assembly station is equipped with the first Number of personnel for each job type Indicates the first The first assembly station is equipped with the first Shift lengths for each job type; The assembly material history dataset contains all types of materials configured at each assembly station, along with the corresponding delivery time intervals and quantities per delivery. It means that, among them, Indicates the first The first assembly station configuration similar materials Indicates the first The first assembly station configuration Delivery time interval for similar materials Indicates the first The first assembly station configuration The quantity of each type of material delivered in a single shipment.

3. The assembly completion state prediction and optimization method based on interpretable neural networks according to claim 2, characterized in that, The logical relationships between the elements in the system model include the assembly station relationship matrix between all the assembly stations, the assembly task relationship matrix between all the assembly tasks, the assembly personnel relationship matrix between all the assembly personnel, the personnel-task relationship matrix between all the assembly personnel and their corresponding assembly tasks, and the material-task relationship matrix between all the assembly materials and their corresponding assembly tasks.

4. The assembly completion state prediction and optimization method based on interpretable neural networks according to claim 3, characterized in that, The steps of constructing an interpretable neural network and training, testing, and validating the interpretable neural network using the historical task information and the logical relationships between the elements in the system model to obtain the final interpretable neural network include: The interpretable neural network is constructed, comprising an assembly state diagram, a first layer of heterogeneous graph attention neural network, and a second layer of heterogeneous graph attention neural network connected in sequence. The assembly status diagram is constructed using the assembly station relationship matrix, the assembly task relationship matrix, the assembly personnel relationship matrix, the personnel-task relationship matrix, and the material-task relationship matrix. The assembly task history dataset, the assembly personnel history dataset, and the assembly material history dataset are normalized respectively, and all normalization results are integrated to obtain a preprocessed dataset. The preprocessed dataset is divided into a training set, a test set, and a validation set; The preprocessed dataset is used to train, test, and validate the interpretable neural network to obtain the final interpretable neural network.

5. The assembly completion state prediction and optimization method based on interpretable neural networks according to claim 4, characterized in that, The assembly state diagram includes multiple assembly station sub-diagrams, and the assembly state diagram is represented as follows: ,in, This represents an assembly state diagram. Indicates the first A diagram of the assembly station locations. , Indicates the first Assembly task feature matrix in each assembly station sub-graph Indicates the first Assembly station location sub-map assembly material feature matrix Indicates the first The assembly worker feature matrix in the sub-graph of each assembly station location. Indicates the first Assembly task relationship matrix in the sub-graph of each assembly station location. Indicates the first Material-task relationship matrix in the assembly station sub-diagram Indicates the first Personnel-task relationship matrix in the sub-graph of each assembly station location This represents the time-stamped features corresponding to the preprocessed data; Each of the assembly station sub-graphs corresponds to one of the assembly stations.

6. The assembly completion state prediction and optimization method based on interpretable neural networks according to claim 4, characterized in that, The first layer of the heterogeneous graph attention neural network includes a feature transformation module, a feature splitting module, and a feature aggregation module connected in sequence; The feature conversion module performs feature dimension conversion on all the received assembly station sub-graphs to obtain all assembly feature vectors corresponding to each assembly station sub-graph. The feature splitting module splits all the assembly element features corresponding to each assembly station sub-graph into task-task process path, material-task process path, personnel-task process path and task-time process path respectively. The feature aggregation module aggregates the feature vectors output by each process path in each assembly feature vector to obtain a corresponding aggregated vector; it concatenates each assembly task with all the aggregated vectors obtained under the four process paths to obtain a corresponding aggregated vector matrix; it reduces the dimensionality of each aggregated vector matrix using a first feature transformation matrix with learnable parameters to obtain a one-dimensional vector corresponding to each assembly task; and it concatenates the one-dimensional vectors corresponding to all assembly tasks in each assembly station subgraph to obtain a concatenation matrix corresponding to each assembly station subgraph, wherein the concatenation matrix includes multiple concatenation vectors. The feature aggregation module includes a graph attention network and The activation function, the four process paths sequentially pass through the attention network and the... The activation function is used to obtain the corresponding feature vector.

7. The assembly completion state prediction and optimization method based on interpretable neural networks according to claim 6, characterized in that, The expression for the feature dimension transformation is: (1) in, Indicates the first The first assembly station location diagram Assembly feature vectors for each assembly task. Indicates the first The first assembly station location diagram The assembly feature vector corresponding to each assembly material. Indicates the first The first assembly station location diagram Assembly feature vector corresponding to each assembler. Indicates the first In the assembly task feature matrix of the assembly station sub-graph, the first... Each assembly task feature vector Indicates the first In the assembly task feature matrix of the assembly station sub-graph, the first... Each assembly material feature vector Indicates the first In the assembly worker feature matrix of the assembly station location sub-map, the first... A feature vector of an assembly worker. Indicates the first The first feature transformation matrix with learnable parameters corresponding to the assembly task in each assembly station subgraph. Indicates the first The first feature transformation matrix with learnable parameters corresponding to the assembly materials in each assembly station sub-graph. Indicates the first The first feature transformation matrix with learnable parameters corresponding to the assembly personnel in each assembly station subgraph. express 3D real space, express 3D real space, This represents the output dimension of the feature transformation process.

8. The assembly completion state prediction and optimization method based on interpretable neural networks according to claim 6, characterized in that, The second layer of the heterogeneous graph attention neural network includes a linear transformation layer, a feature aggregation layer, and a feature fitting layer connected in sequence; The linear transformation layer contains multiple second feature transformation matrices with learnable parameters. Each of the second feature transformation matrices with learnable parameters is used to compress the corresponding concatenated vector to obtain the corresponding one-dimensional feature vector. The feature aggregation layer aggregates each one-dimensional feature vector into a corresponding multi-dimensional feature vector according to the assembly station relationship matrix, and concatenates all the multi-dimensional feature vectors to obtain a multi-dimensional concatenated feature vector. The feature fitting layer uses all the second feature transformation matrices with learnable parameters to fit the mapping relationship between each of the multidimensional spliced ​​feature vectors and the operation time for completing the assembly task, thereby obtaining the expected completion time of the target product.

9. The assembly completion state prediction and optimization method based on interpretable neural networks according to claim 8, characterized in that, The expression for the one-dimensional feature vector is: (2) in, Indicates the first The one-dimensional feature vector is obtained by compressing the spliced ​​vectors corresponding to the assembly station sub-maps. Indicates the first The splicing vector corresponding to each assembly station location sub-graph Represents concatenated vectors The corresponding concatenated feature transformation matrix, express 3D real space, Indicates the first The number of all assembly tasks in the assembly station sub-diagram; The expression for the estimated completion time of the target product is: (3) in, Indicates the estimated completion time of the target product. This represents a multi-dimensional concatenated feature vector. This represents the multidimensional concatenated feature transformation matrix corresponding to the multidimensional concatenated feature vector. express 3D real space, Represents multidimensional concatenated feature vectors Dimensions.

10. The assembly completion state prediction and optimization method based on interpretable neural networks according to claim 8, characterized in that, The steps of using the ultimately interpretable neural network to predict the assembly completion status of the station-type assembly line and optimizing the station-type assembly line based on the prediction results include: The supply frequency of assembly materials configured for each assembly station, as well as the number of personnel and shift duration corresponding to each job type, are optimized based on the expected completion time of the target product.

Images