Dynamic factory production line balancing method and system based on adaptive control

Abstract

The invention provides a dynamic factory production line balancing method and system based on adaptive control, and the method comprises the steps: obtaining a real-time operation state set of a production line, carrying out the work station correlation feature analysis processing of the real-time operation state set of the production line, generating a work station correlation feature matrix, carrying out the production line bottleneck recognition based on the work station correlation feature matrix, and carrying out the dynamic factory production line balancing. Generating a bottleneck station feature set, wherein the bottleneck station feature set comprises bottleneck station identifiers and corresponding operation parameter deviation features and material overstock features; and generating a production line dynamic adjustment strategy set according to the bottleneck station feature set, wherein the production line dynamic adjustment strategy set comprises a parameter adjustment strategy and a material distribution strategy for the bottleneck station. According to the invention, the timeliness and effectiveness of production line balance control can be improved.

Description

Technical Field

[0001] This invention relates to the field of production control, and more specifically, to a dynamic factory production line balancing method and system based on adaptive control. Background Technology

[0002] Production line balancing aims to maintain a high-efficiency and stable production rhythm by coordinating the operating status of each workstation and the material flow process, avoiding capacity waste caused by abnormal operation of local workstations or material blockages. Currently, it typically relies on manual experience or preset rules to identify production line bottlenecks and generate adjustment strategies by monitoring the operating parameters of individual workstations or statistically analyzing static material flow data, such as adjusting equipment parameters at specific workstations or increasing the frequency of material transfers. However, such methods often treat workstation operation and material flow as independent processes, failing to effectively link the mutual influence of operating parameters between different workstations and the temporal dependence of material flow. This results in insufficient analysis of the overall correlation of the production line, making it difficult to dynamically capture hidden bottlenecks caused by workstation coupling or material timing fluctuations. Consequently, the generated adjustment strategies lack specificity, and the stability of production line balancing results is poor. Summary of the Invention

[0003] This invention provides a dynamic factory production line balancing method and system based on adaptive control.

[0004] In a first aspect, embodiments of the present invention provide a dynamic factory production line balancing method based on adaptive control, comprising: acquiring a set of real-time operating states of the production line, the set of real-time operating states including the sequence of operating parameters of each workstation and the material flow timing information between workstations; performing workstation association feature parsing processing on the set of real-time operating states of the production line to generate a workstation association feature matrix, the workstation association feature matrix being used to characterize the coupling relationship of operating parameters and the dependency relationship of material flow between different workstations; identifying production line bottlenecks based on the workstation association feature matrix to generate a bottleneck workstation feature set, the bottleneck workstation feature set including a bottleneck workstation identifier and its corresponding operating parameter deviation features and material backlog features; and generating a set of dynamic adjustment strategies for the production line based on the bottleneck workstation feature set, the set of dynamic adjustment strategies for the production line including parameter adjustment strategies and material allocation strategies for bottleneck workstations.

[0005] Secondly, embodiments of the present invention provide a computer system, including: a memory storing a computer program; and a processor for loading the computer program to implement the dynamic factory production line balancing method based on adaptive control as described above.

[0006] This invention provides a dynamic factory production line balancing method based on adaptive control. It acquires a set of real-time production line operating states, including the sequence of operating parameters for each workstation and the material flow timing information between workstations. By simultaneously incorporating workstation operating data and material flow data, it comprehensively captures the dynamic process of the production line operation, avoiding the one-sided state description caused by relying on only a single type of data. After acquiring the set of real-time production line operating states, it performs workstation correlation feature analysis to generate a workstation correlation feature matrix. This matrix can characterize the coupling relationship of operating parameters and the dependence relationship of material flow between different workstations. By transforming scattered workstation data into systematic correlation information, it overcomes the limitations of traditional isolated analysis of workstation states, achieving a quantitative characterization of the overall correlation of the production line. Based on the workstation correlation feature matrix, it identifies production line bottlenecks and generates a bottleneck workstation feature set. This set includes bottleneck workstation identifiers and their corresponding operating parameter deviation features and material backlog features. By combining the correlation relationships between workstations for dynamic identification and simultaneously extracting two types of features—parameter anomalies and flow blockages—it can accurately locate the causes of bottlenecks and avoid misjudgments. Based on the bottleneck workstation feature set, a set of dynamic production line adjustment strategies is generated, including parameter adjustment strategies and material allocation strategies. By generating corresponding strategies for two types of features—parameter deviation and material backlog—coordinated solutions to bottleneck problems can be achieved, avoiding the one-sidedness of a single strategy. The set of dynamic production line adjustment strategies is input into the production line control system for production line balancing control operations. This ensures that the results of analysis, identification, and strategy generation directly affect the production line processes, guaranteeing real-time strategy implementation and improving the timeliness and effectiveness of production line balancing control. Attached Figure Description

[0007] Figure 1 This is a flowchart of a dynamic factory production line balancing method based on adaptive control provided in an embodiment of the present invention.

[0008] Figure 2 This is a schematic diagram of the composition of a computer system provided in an embodiment of the present invention. Detailed Implementation

[0009] Please see Figure 1 , Figure 1 A flowchart of a dynamic factory production line balancing method based on adaptive control is provided for embodiments of the present invention. This method can be executed by a computer system and may include the following steps: Step S100: Obtain the real-time operating status set of the production line. The real-time operating status set of the production line includes the sequence of operating parameters of each workstation and the material flow sequence information between workstations.

[0010] A real-time production line operation status set is a record of the current operating status of a factory production line. For example, in an automotive manufacturing production line scenario, the workstation operation parameter sequence is a sequence of various operating parameters for each workstation arranged in chronological order over a period of time. For instance, at an automotive engine assembly workstation, the workstation operation parameter sequence might include changes in parameters such as bolt tightening torque, part installation time, and equipment operating temperature over time. These parameters reflect the operating status and work efficiency of the workstation at different times. Meanwhile, the material flow sequence information between workstations records the specific time sequence of materials moving from one workstation to another, along with related time information. For example, the time it takes for a car body frame to move from the welding workstation to the painting workstation, and the dwell time at intermediate transition workstations, reflects the flow rhythm and time arrangement of materials throughout the entire production line.

[0011] To obtain a real-time set of production line operating statuses, various types of sensors can be installed at each workstation to obtain a sequence of operating parameters. For example, at an engine assembly workstation, torque sensors are installed to monitor the torque of bolt tightening in real time, time sensors are installed to record the installation time of each part, and temperature sensors are installed to obtain the operating temperature of the equipment. These sensors continuously collect data and transmit the collected data to a data acquisition system. The data acquisition system arranges these parameters into a sequence of workstation operating parameters according to the time sequence of data collection. For material flow sequence information, RFID (Radio Frequency Identification) technology can be used. RFID tags are attached to materials, and the tags contain the material's identification information and related flow information. RFID readers are set up at each workstation and along the material flow path. When materials with RFID tags pass by the readers, the readers automatically read the information in the tags and record the material's passage time and location. By organizing and analyzing this information, material flow sequence information can be obtained. In addition, barcode scanning technology can also be used. Barcodes are attached to materials and workstations, and the flow of materials is recorded by scanning the barcodes.

[0012] Step S200: Perform workstation association feature parsing on the real-time operating status set of the production line to generate a workstation association feature matrix. The workstation association feature matrix is ​​used to characterize the coupling relationship of operating parameters and the material flow dependency relationship between different workstations.

[0013] Workstation correlation feature analysis is a process of in-depth mining and analysis of the real-time operating status set of the production line, aiming to identify the correlation features between different workstations. For example, in an automobile manufacturing production line, the coupling relationship of operating parameters refers to the mutual influence and constraint between the operating parameters of different workstations. For instance, the assembly accuracy of the engine assembly workstation may affect the debugging time and results of the powertrain debugging workstation. If the installation accuracy of certain parts is insufficient during engine assembly, it may cause problems in the powertrain during debugging, requiring more time for adjustment and repair. Material flow dependency refers to the strict sequential and temporal dependence of material flow between different workstations. For example, only after the welding work of the car body is completed can the car body be transferred to the painting workstation for painting; and the completion time of the car body at the welding workstation directly affects the time it takes to reach the painting workstation.

[0014] A workstation association feature matrix is ​​a matrix used to quantify and represent the relationships between different workstations. Each element in the matrix reflects the strength of the association between two workstations. By generating a workstation association feature matrix, the interrelationships between various workstations on the production line can be understood more clearly and intuitively. For example, by analyzing the elements in the matrix, it can be determined which workstations are closely associated and which workstations have a greater impact on other workstations, thus enabling targeted adjustments and optimizations to the production line.

[0015] As one implementation method, step S200 involves performing workstation association feature parsing on the real-time operating status set of the production line to generate a workstation association feature matrix. Specifically, this can be implemented as follows: S210~S250: Step S210: Extract the workstation operation parameter sequence and material flow timing information of each workstation from the real-time operation status set of the production line to obtain the original feature set of the workstation.

[0016] The original feature set of a workstation is a set of features directly extracted from the real-time operating status set of the production line, without further processing. In an automotive manufacturing production line, the real-time operating status set collected by the data acquisition system contains a large amount of information, which needs to be filtered to extract the workstation operating parameter sequences and material flow timing information related to each workstation. Specifically, data filtering algorithms can be used to extract the operating parameter sequences and material flow timing information for each workstation from the real-time operating status set, based on the workstation's identification information and data type. For example, based on the identification of the engine assembly workstation, the workstation's operating parameter sequences, such as bolt tightening torque and part installation time, as well as the flow time information of engine-related materials at that workstation, can be extracted from the data set. Integrating this extracted information yields the original feature set of the workstation.

[0017] Step S220: Extract time series features from the workstation operation parameter sequence in the original feature set of the workstation to generate workstation operation sequence features, which include parameter fluctuation features and parameter trend features.

[0018] Time series feature extraction is a method for in-depth analysis of workstation operating parameter sequences, aiming to uncover hidden time-related features from the sequence. In automotive manufacturing production lines, parameter fluctuation characteristics reflect the changes in parameters over a short period, demonstrating their instability and randomness. For example, the bolt tightening torque at the engine assembly station may fluctuate within a short time due to factors such as slight equipment vibrations and operator habits. Parameter trend characteristics, on the other hand, reflect the direction and rate of change of parameters over a longer period, reflecting the overall trend of parameter change. For instance, with improvements in production processes and increased worker proficiency, the parts assembly time at the engine assembly station may show a gradual decreasing trend.

[0019] As one implementation method, step S220 involves extracting time series features from the workstation operation parameter sequence in the original feature set of the workstation to generate workstation operation sequence features. Specifically, this can be implemented as follows: steps S221~S226: Step S221: Perform outlier detection on the workstation operation parameter sequence, identify and remove outlier parameter values ​​in the sequence, and obtain the cleaned workstation operation parameter sequence.

[0020] Outlier detection is to ensure the accuracy and reliability of subsequent data analysis. Outliers may be caused by equipment failure, human error, or data transmission errors. In the automotive manufacturing production line, for the sequence of operating parameters of the engine assembly station, there may be situations where the tightening torque of a bolt is too high or the installation time of a part is too long. These data points may be outliers.

[0021] Outlier detection can be performed using statistical methods, such as the 3σ rule. First, the mean and standard deviation of the workstation operating parameter sequence are calculated. The mean represents the average level of the sequence, and the standard deviation represents the dispersion of the sequence. Then, according to the 3σ rule, data points deviating from the mean by more than three times the standard deviation are identified as outliers. For example, for a bolt tightening torque sequence, after calculating its mean and standard deviation, data points with torque values ​​less than the mean minus three times the standard deviation or greater than the mean plus three times the standard deviation are marked as outliers. Alternatively, machine learning-based methods, such as the Isolation Forest algorithm, can be used. The Isolation Forest algorithm constructs multiple decision trees, assigning data points to different leaf nodes. Data points with shorter paths in the decision trees are considered outliers. After identifying outliers, these outliers are removed from the workstation operating parameter sequence, resulting in a cleaned sequence. This ensures that the data used for subsequent analysis is more accurate and reliable.

[0022] Step S222: Perform trend decomposition on the cleaned workstation operating parameter sequence to separate the long-term trend component and short-term fluctuation component of the sequence.

[0023] Trend decomposition involves breaking down the sequence of cleaned workstation operating parameters into long-term trends and short-term fluctuations for separate analysis. In automotive manufacturing lines, the long-term trend component reflects the overall trend of parameter changes over a longer period, which may be influenced by factors such as improvements in production processes, equipment upgrades, and increased worker skills. For example, with continuous improvements in automotive production processes, the part assembly time at the engine assembly station may show a gradually decreasing long-term trend. The short-term fluctuation component reflects the random fluctuations of parameters in the short term, which may be caused by minor equipment malfunctions or changes in operator condition. For instance, the bolt tightening torque at the engine assembly station may experience small fluctuations throughout the day.

[0024] As one implementation method, step S222 involves performing trend decomposition on the cleaned workstation operating parameter sequence to separate the long-term trend component and short-term fluctuation component of the sequence. Specifically, this can be implemented as the following steps S2221~S2226: Step S2221: Perform sampling frequency analysis on the sequence of station operation parameters after cleaning, determine the sampling interval of the sequence, and set the initial scale parameter of adaptive wavelet decomposition based on the sampling interval.

[0025] Sampling frequency analysis is used to determine the sampling interval of the workstation's operating parameter sequence. In an automotive manufacturing line, different workstations may use different sampling frequencies for data acquisition. For example, an engine assembly workstation might collect bolt tightening torque data at regular time intervals. By analyzing the cleaned workstation operating parameter sequence, its sampling interval can be determined. Adaptive wavelet decomposition can automatically adjust the decomposition scale according to the data characteristics; the initial scale parameter is related to the sampling interval. Generally, the shorter the sampling interval, the smaller the initial scale parameter can be set. This is because shorter sampling intervals capture more detailed information, thus requiring a smaller scale for decomposition. For example, if the sampling interval is short, the initial scale parameter might be set to a relatively small value, allowing for more refined data analysis through wavelet decomposition. The specific relationship between the initial scale parameter and the sampling interval can be determined through experiments and experience.

[0026] Step S2222: Call the preset adaptive wavelet decomposition algorithm, starting from the initial scale parameter, to perform multi-scale wavelet decomposition on the cleaned workstation operation parameter sequence to obtain the wavelet coefficient set at different scales.

[0027] The preset adaptive wavelet decomposition algorithm can be the Mallat algorithm. In the automobile manufacturing production line, starting from the set initial scale parameters, multi-scale wavelet decomposition is performed on the station operation parameter sequence of the cleaned engine assembly station.

[0028] Multi-scale wavelet decomposition decomposes a sequence into wavelet coefficients at different scales, each scale representing a signal component of a different frequency. Smaller scales correspond to high-frequency signals, reflecting short-term fluctuations in the sequence; larger scales correspond to low-frequency signals, reflecting long-term trends. During the decomposition process, the algorithm continuously decomposes the sequence, obtaining a set of wavelet coefficients with each decomposition. As the decomposition scale increases, the resulting set of wavelet coefficients contains signal information at different frequencies. For example, when performing multi-scale wavelet decomposition on a bolt tightening torque sequence, the wavelet coefficients at smaller scales reflect the rapid fluctuations in torque over a short period, while the wavelet coefficients at larger scales reflect the overall trend of torque change over a longer period.

[0029] Step S2223: Calculate the energy value of the wavelet coefficients at each scale. The wavelet coefficients corresponding to the first K scales whose cumulative energy value reaches the preset energy threshold are determined as trend-related wavelet coefficients, and the wavelet coefficients corresponding to the remaining scales are determined as fluctuation-related wavelet coefficients, where K is a positive integer dynamically adjusted according to the sequence length.

[0030] Energy value is an indicator that measures the amount of information energy contained in wavelet coefficients. In automobile manufacturing production lines, it is necessary to calculate the energy values ​​of wavelet coefficients at different scales. The energy value can be calculated by summing the squares of the wavelet coefficients.

[0031] The preset energy threshold is a pre-defined proportion used to determine which wavelet coefficients are correlated with long-term trends and which are correlated with short-term fluctuations. By calculating and summing the energy values ​​of the wavelet coefficients at each scale, when the proportion of the accumulated energy value to the total energy value reaches the preset energy threshold, the wavelet coefficients at the first K scales are identified as trend-correlated wavelet coefficients. These wavelet coefficients primarily reflect the long-term trend information of the sequence. The wavelet coefficients at the remaining scales are identified as fluctuation-correlated wavelet coefficients, primarily reflecting the short-term fluctuation information of the sequence. The value of K is dynamically adjusted according to the length of the sequence. The longer the sequence, the richer the information it contains, and the more scales may be needed to reflect the long-term trend; therefore, the value of K may be larger. For example, for a long time series of parts installation at an engine assembly station, K may take a relatively large value; for a shorter sequence, K may take a smaller value.

[0032] Step S2224: Perform wavelet reconstruction on the trend-related wavelet coefficients to obtain the preliminary long-term trend component; perform wavelet reconstruction on the fluctuation-related wavelet coefficients to obtain the preliminary short-term fluctuation component.

[0033] Wavelet reconstruction is the process of recombinizing wavelet coefficients into a signal. In the automotive manufacturing production line, wavelet reconstruction is performed on the determined trend-related wavelet coefficients using the same wavelet basis and algorithm as the decomposition, such as the inverse transform of the Mallat algorithm. By performing the inverse transform on the trend-related wavelet coefficients, they are recombined into a signal, which is the preliminary long-term trend component, reflecting the overall changing trend of the workstation operating parameter sequence over a longer period of time.

[0034] Similarly, wavelet reconstruction of the wavelet coefficients related to fluctuation yields a preliminary short-term fluctuation component, which reflects the random fluctuation of the workstation's operating parameter sequence in the short term. For example, for the bolt tightening torque sequence at the engine assembly workstation, reconstruction of the trend-related wavelet coefficients yields a curve reflecting the long-term trend of torque variation; reconstruction of the fluctuation-related wavelet coefficients yields a curve reflecting the short-term random fluctuation of torque.

[0035] Step S2225: Calculate the residual sequence between the preliminary long-term trend component and the cleaned station operation parameter sequence. When the mean absolute value of the residual sequence is greater than the preset residual threshold, expand the scale range of the trend-related wavelet coefficients by one scale and reconstruct the wavelet until the mean absolute value of the residual sequence is less than or equal to the preset residual threshold, and obtain the final long-term trend component.

[0036] The residual sequence is the difference sequence between the initial long-term trend component and the cleaned sequence of station operating parameters. In an automotive manufacturing production line, the residual sequence of the initial long-term trend component and the cleaned sequence of station operating parameters for the engine assembly station is calculated. The residual sequence reflects the degree of difference between the initial long-term trend component and the original sequence.

[0037] The preset residual threshold is a pre-defined allowable error range. If the mean absolute value of the residual sequence is greater than the preset residual threshold, it indicates that the initial long-term trend component does not fit the original sequence well enough, and the scale range of the trend-related wavelet coefficients needs to be adjusted. Specifically, the scale range of the trend-related wavelet coefficients is expanded by one scale, i.e., an additional scale reflecting lower frequency information is added. Then, wavelet reconstruction is performed on the expanded trend-related wavelet coefficients to obtain a new initial long-term trend component. The residual sequence between the new initial long-term trend component and the cleaned workstation operating parameter sequence is recalculated, and the above process is repeated until the mean absolute value of the residual sequence is less than or equal to the preset residual threshold. The long-term trend component obtained at this point is the final long-term trend component, which better fits the original sequence and reflects the long-term trend of the sequence.

[0038] Step S2226: Subtract the cleaned station operation parameter sequence from the final long-term trend component to obtain the short-term fluctuation component.

[0039] In automotive manufacturing production lines, for the operating parameter sequence of the engine assembly station, subtracting the final long-term trend component from the cleaned sequence yields the short-term fluctuation component, which reflects the random fluctuation of the parameters after removing the long-term trend. For example, for a bolt tightening torque sequence, the cleaned sequence contains both long-term trend and short-term fluctuation information; subtracting the final long-term trend component leaves only the short-term random fluctuation portion. This method completes the trend decomposition of the cleaned station operating parameter sequence, obtaining the long-term trend component and the short-term fluctuation component, providing a foundation for subsequent feature analysis.

[0040] Step S223: Calculate the slope characteristics of the long-term trend component. The slope characteristics are used to characterize the long-term direction and rate of change of the parameter.

[0041] Slope characteristics are an important feature obtained through further analysis of long-term trend components, which can intuitively reflect the direction and rate of change of parameters over a long period. In the automotive manufacturing production line, for the long-term trend components of engine assembly stations, such as the long-term trend component of part installation time, linear regression can be used to calculate their slope.

[0042] Linear regression is a method that approximates the long-term trend component by fitting a straight line using the least squares method. Data points of the long-term trend component are used as sample points, and the least squares method is used to find a straight line that minimizes the sum of the distances from the sample points to this line. The slope of this line represents the slope characteristic of the long-term trend component. If the slope is positive, it indicates that the parameter shows an upward trend in the long run; for example, the part assembly time at an engine assembly station may gradually increase due to certain factors. If the slope is negative, it indicates that the parameter shows a downward trend in the long run; for example, the part assembly time may gradually decrease with improvements in production processes. The larger the absolute value of the slope, the faster the rate of change of the parameter.

[0043] Step S224: Calculate the amplitude and frequency characteristics of the short-term fluctuation component. The amplitude characteristic is the difference between the maximum and minimum values ​​of the fluctuation component, and the frequency characteristic is the number of fluctuation cycles per unit time.

[0044] Fluctuation amplitude and fluctuation frequency characteristics are two important features obtained from analyzing short-term fluctuation components. In an automotive manufacturing production line, for short-term fluctuation components at the engine assembly station, such as the short-term fluctuation component of bolt tightening torque, the maximum and minimum values ​​of the fluctuation component are first identified, and their difference is calculated to obtain the fluctuation amplitude characteristic. The fluctuation amplitude characteristic reflects the magnitude of the parameter's fluctuation range in the short term. For example, if the difference between the maximum and minimum values ​​of the short-term fluctuation component of bolt tightening torque is large, it indicates that the torque fluctuates significantly in the short term.

[0045] To determine the frequency characteristics of fluctuations, spectral analysis, such as Fourier transform, can be performed on the short-term fluctuation components. Fourier transform converts a signal in the time domain into a signal in the frequency domain. By analyzing the frequency domain signal, the main frequency components can be identified. The fluctuation frequency characteristics are obtained by calculating the number of fluctuation cycles per unit time. These characteristics reflect the frequency of parameter fluctuations in the short term. For example, a large number of fluctuation cycles per unit time indicates that the bolt tightening torque fluctuates frequently in the short term.

[0046] Step S225: Combine the slope feature, fluctuation amplitude feature, and fluctuation frequency feature into a feature vector to obtain the parameter trend feature.

[0047] In automotive manufacturing production lines, the slope, amplitude, and frequency characteristics of engine assembly stations are combined into a single feature vector. This feature vector integrates information about the long-term trend and short-term fluctuations of the parameters. For example, the slope characteristic reflects the long-term direction and rate of parameter change, the amplitude characteristic reflects the short-term fluctuation range, and the frequency characteristic reflects the frequency of short-term fluctuations. Combining these features into a multi-dimensional feature vector provides a more comprehensive description of the changing characteristics of the station's operating parameters.

[0048] Step S226: Use the original sequence of short-term fluctuation components as the parameter fluctuation feature, and combine the parameter trend feature and the parameter fluctuation feature to obtain the workstation running sequence feature.

[0049] In automotive manufacturing production lines, the original sequence of short-term fluctuation components of the engine assembly station is used as the parameter fluctuation feature, containing information on the random fluctuations of parameters in the short term. Combining the parameter trend feature and the parameter fluctuation feature yields the station's operational sequence feature. This feature integrates the long-term trend and short-term fluctuations of the parameters, providing a more comprehensive reflection of the changes in the station's operating parameters over time. For example, the parameter trend feature provides information on the long-term direction and rate of parameter change, while the parameter fluctuation feature provides information on the random fluctuations of parameters in the short term. Combining them allows for analysis of the station's operational status from multiple perspectives, providing a more accurate basis for production line optimization and adjustment.

[0050] Step S230: Extract material flow time sequence information from the original feature set of the workstation to generate material flow path features, which include path node sequence and flow duration features between nodes.

[0051] Material flow path feature extraction is a process of in-depth analysis of material flow time sequence information, aiming to uncover the path and time characteristics of materials flowing between workstations. In an automotive manufacturing production line, the path node sequence represents the various workstations a material passes through from its starting point to its destination. For example, a car tire starts from the warehouse, passes through the quality inspection station and the assembly station, and finally arrives at the vehicle assembly station; the path node sequence would then be [warehouse, quality inspection station, assembly station, vehicle assembly station]. The flow time feature between nodes represents the time spent by materials flowing between two adjacent workstations. By analyzing the material flow time sequence information, the flow time of a tire from the warehouse to the quality inspection station, and from the quality inspection station to the assembly station, can be recorded. Time series analysis methods can be used to extract the flow time between each node from the material flow time sequence information. For example, by calculating the recorded arrival and departure times of materials at each workstation, the flow time between nodes can be obtained. By combining the path node sequence and the flow duration feature between nodes, a material flow path feature is generated, which can clearly show the flow path and time arrangement of materials in the production line.

[0052] Step S240: Input the workstation running sequence features and material flow path features into the workstation association graph network model. Map each workstation to a graph node through the graph node initialization layer of the workstation association graph network model. Use the workstation running sequence features as graph node features and the material flow path features as graph edge features to construct the workstation association graph structure.

[0053] The workstation association graph network model is a graph neural network model used to analyze the relationships between workstations. In an automotive manufacturing production line, the graph node initialization layer is the first layer of the model, and its function is to map each workstation to a node in the graph. For example, the engine assembly workstation, body welding workstation, and painting workstation are each mapped to a node in the graph.

[0054] Using the operational sequence characteristics of each workstation as graph node features means that each node contains information about the changes in the workstation's operational parameters over time. In the engine assembly workstation node, this includes the long-term trends and short-term fluctuations of parameters such as bolt tightening torque and part installation time. Using material flow path characteristics as graph edge features means that the edges in the graph contain the path and time information of material flow between workstations. For example, the edge from the engine assembly workstation to the powertrain debugging workstation contains the sequence of path nodes for the engine to flow from the assembly workstation to the debugging workstation, as well as the flow time between nodes.

[0055] In this way, a graph structure reflecting the relationships between workstations is constructed. In this graph structure, the connections between nodes and the characteristics of the edges reflect the coupling relationships of operating parameters and the dependencies of material flow between workstations. For example, if there is material flow between two workstations and the flow time is short, it indicates that the relationship between these two workstations is relatively close, and the characteristics of the corresponding edges in the graph structure will also reflect this close relationship.

[0056] Step S250: Perform multi-layer feature aggregation on the workstation association graph structure through the graph convolutional layer of the workstation association graph network model to obtain graph aggregation features containing high-order association information between workstations. Convert the graph aggregation features into matrix form to obtain the workstation association feature matrix.

[0057] Graph convolutional layers are the core layers of the workstation association graph network model. Their function is to aggregate multi-layer features of the workstation association graph structure and mine high-order association information between workstations. In the automotive manufacturing production line, the relationships between workstations are not only simple direct relationships, but also indirect relationships generated through other workstations. These indirect associations constitute high-order association information.

[0058] As one implementation method, step S250 involves performing multi-layer feature aggregation on the workstation association graph structure through the graph convolutional layer of the workstation association graph network model to obtain graph aggregation features containing high-order association information between workstations. Specifically, this can be implemented as the following steps S251~S256: Step S251: Initialize the graph convolutional layer weight parameters of the workstation association graph network model, and set the number of graph convolutional layers to the preset number of layers.

[0059] In an automotive manufacturing production line, initializing the weight parameters of the graph convolutional layers in a workstation-related graph network model is crucial for determining the importance of each feature during feature aggregation. Weight parameters can be initialized using random initialization methods, such as the Xavier initialization method. The Xavier initialization method randomly generates the weight matrix based on the dimensions of the input and output, ensuring consistent variance between the input and output, which aids in model training and convergence.

[0060] The preset number of layers refers to the pre-defined number of graph convolutional layers. More graph convolutional layers allow the model to uncover richer high-order relationships, but also increase computational complexity and training time. By appropriately setting the number of graph convolutional layers, computational efficiency can be improved while maintaining model performance. For example, the number of graph convolutional layers can be set to an appropriate value based on the complexity of the production line and the depth of the relationships to be uncovered.

[0061] Step S252: Input the graph node features and graph edge features of the workstation association graph structure into the first graph convolutional layer, and perform matrix multiplication between the adjacency matrix of the first graph convolutional layer and the graph node features to obtain the first layer adjacency aggregation features.

[0062] In an automobile manufacturing production line, the graph node features and graph edge features of the constructed workstation association graph structure are input into the first graph convolutional layer. The adjacency matrix is ​​a matrix that represents the connection relationships between nodes in the graph. The elements in the matrix indicate whether there is a connection between two nodes and the strength of the connection.

[0063] By performing matrix multiplication between the adjacency matrix and the features of graph nodes, the features of adjacent nodes are aggregated. For example, for an engine assembly station node, its adjacency matrix elements represent its connections with other station nodes. Multiplying the adjacency matrix with the feature vector of the engine assembly station node transfers the feature information of adjacent nodes to that node, resulting in a first-level adjacency aggregation feature that contains information about adjacent nodes. This feature aggregation process can be understood as, in a graph structure, a node acquiring and integrating information from its adjacent nodes.

[0064] Step S253: Perform nonlinear activation processing on the first-layer adjacency aggregation features to obtain the first-layer activated features.

[0065] Nonlinear activation is used to introduce nonlinear factors and enhance the model's expressive power. In an automotive manufacturing line, the first-layer adjacency aggregation features can be processed using the ReLU (Modified Linear Unit) activation function. These first-layer adjacency aggregation features may contain negative values, which could indicate the inhibitory effect of certain features. The ReLU activation function removes these negative values, retaining only the positive values. This enhances the model's ability to distinguish features, enabling it to learn more complex patterns. After nonlinear activation, the resulting first-layer activation features contain more nonlinear information, helping the model better capture the relationships between workstations.

[0066] Step S254: Use the first layer activation features as input to the second layer graph convolutional layer, and combine the weight coefficients of the graph edge features to aggregate the first layer activation features again to obtain the second layer adjacency aggregation features. Perform non-linear activation processing on the second layer adjacency aggregation features to obtain the second layer activation features.

[0067] In an automotive manufacturing production line, the first layer of activation features is used as input to the second layer of graph convolutional layers. The weight coefficients of the graph edge features represent the importance of the edges in the feature aggregation process. Different material flow paths and time characteristics may have different impacts on the relationships between workstations, therefore, it is necessary to set weight coefficients for the graph edge features.

[0068] By combining the weight coefficients of graph edge features to further aggregate the first-layer activation features, the feature information of nodes is integrated in a broader scope. For example, considering the impact of material flow time and path on workstation association, the transmission strength of adjacent node features is adjusted by the weight coefficients of graph edge features. The second-layer adjacency aggregation features are obtained by calculating the weight coefficients of the first-layer activation features, the adjacency matrix, and the graph edge features.

[0069] The second-layer adjacency aggregation features are subjected to nonlinear activation processing, also using the ReLU activation function. This further enhances the model's nonlinear expressive power, resulting in a second-layer activation feature. This feature, based on the first-layer activation feature, further integrates more higher-order correlation information, enabling the model to have a deeper understanding of the relationships between workstations.

[0070] Step S255: Perform the above aggregation and activation processes sequentially according to the preset number of layers until the processing of the last graph convolutional layer is completed, and obtain the last layer activation features.

[0071] In an automobile manufacturing production line, the aggregation and activation processes in steps S252-S254 are repeated sequentially according to a pre-set number of graph convolutional layers. Each layer further aggregates and mines the correlation information between workstations based on the previous layer. As the number of layers increases, the model can capture increasingly higher-order correlation information. For example, after multiple layers of graph convolutional processing, the model can consider correlations indirectly generated through multiple intermediate workstations. After the final graph convolutional layer is completed, the final activation feature is obtained, which contains the richest high-order correlation information between workstations obtained after multi-layer feature aggregation.

[0072] Step S256: Concatenate the activation features of the last layer with the activation features of each intermediate layer to obtain a graph aggregation feature containing multi-scale correlation information.

[0073] In automobile manufacturing production lines, concatenating the activation features of the final layer with those of each intermediate layer is a way to integrate correlation information at different levels, resulting in a graph aggregation feature that contains multi-scale correlation information. The activation features at different layers reflect different levels of workstation correlation information; the final layer contains the highest-order correlation information, while the intermediate layer features contain relatively lower-order correlation information.

[0074] As one implementation method, step S256 involves concatenating the activation features of the last layer with the activation features of each intermediate layer to obtain a graph aggregation feature containing multi-scale correlation information. Specifically, this can be implemented as follows: steps S2561~S2566. Step S2561: Obtain the layer index of each graph convolutional layer in the workstation association graph network model, and sort the intermediate layer activation features in order from low to high according to the layer index to obtain the sorted intermediate layer activation feature sequence. The intermediate layer activation feature sequence contains the activation features of the first layer to the activation features of the (N-1)th layer, where N is the number of graph convolutional layers.

[0075] In an automotive manufacturing production line, the layer indices of each graph convolutional layer in the workstation-related graph network model are first obtained. The layer index indicates the positional order of each graph convolutional layer within the model. The activation features of the intermediate layers are then sorted according to their layer indices from lowest to highest. For example, if the graph convolutional layer count is 5, the activation features of the intermediate layers include those from the first, second, third, and fourth layers. These activation features are then arranged in ascending order of their layer indices to obtain a sorted sequence of intermediate layer activation features. This sorting process ensures that subsequent processing follows the correct hierarchical order, providing ordered input for feature concatenation.

[0076] Step S2562: Perform feature dimension expansion processing on each intermediate layer activation feature in the sorted intermediate layer activation feature sequence. Adjust the dimension of each intermediate layer activation feature to the same dimension as the last layer activation feature by zero padding, so as to obtain an intermediate layer activation feature sequence with uniform dimension.

[0077] In an automotive manufacturing line, activation features at different layers may have different dimensions. To concatenate the activation features of the last layer with those of intermediate layers, the intermediate layer activation features need to be expanded in dimension. Zero-padding is used to add zero elements to the vector of the intermediate layer activation features, making its dimension the same as that of the last layer. For example, if the dimension of the last layer activation features is 100, and the dimension of a certain intermediate layer activation feature is 80, then 20 zero elements are added to the end of the vector of that intermediate layer activation feature to expand its dimension to 100. This process yields a sequence of intermediate layer activation features with uniform dimensions, providing a consistent input format for subsequent feature concatenation operations.

[0078] Step S2563: Calculate the layer weight coefficients based on the layer depth of each graph convolutional layer. The larger the layer depth, the larger the corresponding layer weight coefficient. The layer weight coefficient of the first layer is the preset base weight, and the layer weight coefficient of the i-th layer is the product of the layer weight coefficient of the (i-1)-th layer and the preset growth factor. i is an integer greater than 1 and less than N.

[0079] In automotive manufacturing lines, layer depth represents the depth of a graph convolutional layer within the model. A greater layer depth indicates that the layer can capture higher-order correlation information. Calculating layer weight coefficients based on the layer depth of each graph convolutional layer aims to assign different levels of importance to the activated features of different layers during feature concatenation.

[0080] The layer weight coefficients of the first layer are preset base weights, which are initial values ​​set in advance. For the i-th layer (i greater than 1 and less than N), its layer weight coefficient is the product of the layer weight coefficient of the (i-1)-th layer and a preset growth factor. The preset growth factor is a value greater than 1, which causes the layer weight coefficients to gradually increase with the layer depth. For example, if the preset base weight is a small value, and the layer weight coefficients continuously increase with the number of layers, it means that activation features in higher layers have greater weight in feature concatenation because they contain more important higher-order relational information.

[0081] Step S2564: Multiply each intermediate layer activation feature in the dimensionally uniform intermediate layer activation feature sequence by the corresponding layer weight coefficient to obtain a weighted intermediate layer activation feature sequence.

[0082] In automobile manufacturing production lines, each intermediate layer activation feature in a dimensionally unified intermediate layer activation feature sequence is weighted and multiplied with its corresponding layer weight coefficient. This is done to adjust the importance of each layer's activation feature. For example, for an intermediate layer activation feature of a certain layer, it is multiplied with the corresponding layer weight coefficient, so that activation features of more important layers have a greater weight in subsequent concatenation. After weighted multiplication, a weighted intermediate layer activation feature sequence is obtained, where each element has undergone weight adjustment, more accurately reflecting the importance of each layer's activation feature in the overall associated information.

[0083] Step S2565: Concatenate the weighted intermediate layer activation feature sequence with the last layer activation feature according to the feature dimension to obtain the concatenated feature matrix.

[0084] In automobile manufacturing lines, the weighted intermediate layer activation feature sequence is concatenated with the final layer activation feature according to feature dimensions. This involves arranging the activation feature vectors from the weighted intermediate layer sequence sequentially and then concatenating them with the final layer activation feature vector. For example, if the weighted intermediate layer activation feature sequence contains multiple weighted activation feature vectors with uniform dimensions, they are arranged sequentially, and then the final layer activation feature vector is added to the end, resulting in a larger vector. This vector is then arranged into a matrix according to certain rules to obtain the concatenated feature matrix. This matrix integrates activation features from different layers, containing multi-scale correlation information from low to high order.

[0085] Step S2566: Perform feature compression processing on the spliced ​​feature matrix, and perform channel fusion on the spliced ​​feature matrix through a 1×1 convolution kernel to generate graph aggregation features containing multi-scale correlation information.

[0086] In automobile manufacturing production lines, feature compression of the concatenated feature matrix is ​​performed to reduce feature dimensionality while retaining important correlation information. A 1×1 convolution kernel is used to perform channel fusion on the concatenated feature matrix. A 1×1 convolution kernel is a special type of convolution kernel with a size of 1×1. By performing a convolution operation between the 1×1 convolution kernel and the concatenated feature matrix, multiple channels in the matrix are fused. Each channel may represent different types of correlation information, and channel fusion integrates this information. After processing with the 1×1 convolution kernel, a graph aggregation feature with lower dimensionality but containing multi-scale correlation information is obtained. This feature retains correlation information at different levels while reducing feature redundancy and improving information expression efficiency.

[0087] Step S300: Based on the workstation association feature matrix, identify production line bottlenecks and generate a bottleneck workstation feature set. The bottleneck workstation feature set includes the bottleneck workstation identifier and its corresponding operating parameter deviation features and material backlog features.

[0088] Production line bottleneck identification aims to pinpoint the critical workstations that limit overall production efficiency. In automotive manufacturing production lines, this is achieved through analysis of workstation correlation feature matrices. By mining the relationships between workstations reflected in the matrix elements, workstations that significantly impact other workstations and exhibit operational problems themselves are identified.

[0089] As one implementation method, step S300 involves identifying production line bottlenecks based on the workstation association feature matrix and generating a bottleneck workstation feature set. Specifically, this can be implemented as follows: S310~S360: Step S310: Perform matrix block processing on the workstation association feature matrix, dividing the workstation association feature matrix into multiple workstation feature sub-matrices according to workstation grouping, with each workstation feature sub-matrice corresponding to the association feature of one workstation.

[0090] In an automotive manufacturing production line, the workstation association feature matrix contains the association information between all workstations. To facilitate the analysis of the association features of each workstation, the workstation association feature matrix is ​​partitioned. The matrix is ​​divided into multiple sub-matrices according to the workstation's identifier, with each sub-matrix corresponding to the association features of one workstation. For example, the matrix can be divided according to workstations such as engine assembly and body welding, with each sub-matrix containing the association information between that workstation and other workstations. This partitioning method allows for independent analysis of the association features of each workstation, identifying its role and impact on the production line.

[0091] Step S320: Evaluate the feature importance of each workstation feature submatrix and calculate the correlation influence weight of each workstation feature submatrix. The correlation influence weight is used to characterize the degree of influence of the workstation on the operation of other workstations.

[0092] Feature importance assessment aims to determine the importance of each workstation in the production line and its influence on other workstations. In an automotive manufacturing production line, feature importance assessment is performed on the feature submatrix of each workstation, and its correlation influence weight is calculated.

[0093] As one implementation method, step S320 involves evaluating the feature importance of each workstation feature submatrix and calculating the correlation influence weight of each workstation feature submatrix. Specifically, this can be implemented as follows: steps S321~S324. Step S321: Perform singular value decomposition on each workstation feature submatrix to obtain a sequence of singular values ​​of the submatrix, and select the largest singular value in the sequence of singular values ​​of the submatrix as the principal eigenvalue of the workstation feature submatrix.

[0094] In an automobile manufacturing production line, singular value decomposition (SVD) is performed on the feature submatrix of each workstation. SVD is a method that decomposes a matrix into a product of three matrices, yielding a sequence of singular values ​​for each submatrix. These singular values ​​reflect the important characteristics and energy distribution of the matrix. The largest singular value in the sequence of singular values ​​is selected as the principal eigenvalue of the feature submatrix for that workstation. The principal eigenvalue represents the most important feature information in the feature submatrix of that workstation and can, to a certain extent, reflect the main related characteristics of that workstation and its potential impact on other workstations.

[0095] Step S322: Normalize the principal eigenvalues ​​of all workstation feature submatrices to obtain the proportion of principal eigenvalues ​​of each workstation. The proportion of principal eigenvalues ​​is the ratio of the principal eigenvalue of a single workstation to the sum of the principal eigenvalues ​​of all workstations.

[0096] In an automotive manufacturing production line, the principal eigenvalues ​​of all workstation eigenvalue submatrices are normalized. Normalization ensures that the principal eigenvalues ​​of different workstations are compared on the same scale. The proportion of each workstation's principal eigenvalue is calculated by dividing the principal eigenvalue of a single workstation by the sum of the principal eigenvalues ​​of all workstations. This proportion reflects the relative importance of that workstation among all workstations. For example, a higher proportion of a workstation's principal eigenvalue indicates greater importance in the production line's relationships and a potentially greater impact on other workstations.

[0097] Step S323: Extract the row vectors corresponding to each workstation from the workstation association feature matrix, calculate the L2 norm of the row vectors, and obtain the workstation association strength value.

[0098] In an automotive manufacturing production line, row vectors are extracted from the workstation association feature matrix for each workstation. Each row vector represents the association between that workstation and all other workstations. The L2 norm of the row vectors is calculated; the L2 norm is the square root of the sum of the squares of the vector elements. The workstation association strength reflects the overall degree of association between that workstation and other workstations. For example, the larger the L2 norm of a workstation's row vector, the stronger the association between that workstation and other workstations, indicating that it plays a more important connecting role in the production line.

[0099] Step S324: Perform feature fusion with the proportion of the main feature value and the workstation association strength value to obtain the fusion weight value. The fusion weight value is obtained by multiplying the proportion of the main feature value and the workstation association strength value and taking the average value. The fusion weight value is used as the association influence weight of the workstation feature submatrix.

[0100] In an automotive manufacturing production line, the proportion of principal eigenvalues ​​and the workstation correlation strength value are fused. The proportion of principal eigenvalues ​​reflects the relative importance of the workstation, while the workstation correlation strength value reflects the degree of correlation between the workstation and other workstations. The fused weight value is obtained by multiplying these two values ​​and taking the average. This fused weight value comprehensively considers both the importance and correlation strength of the workstation and is used as the correlation influence weight of the workstation's feature submatrix. This weight can more comprehensively assess the degree of influence of the workstation on the operation of other workstations.

[0101] Step S330: Sort each workstation by its associated influence weight from largest to smallest, and select a preset number of influential workstations as a candidate bottleneck workstation set.

[0102] In an automotive manufacturing production line, all workstations are ranked from highest to lowest based on their correlation influence weight. Workstations with higher correlation influence weights have a greater impact on the operation of other workstations and are more likely to become bottlenecks in the production line. A predetermined number of influential workstations are selected as a set of candidate bottleneck workstations. This predetermined number is a value set in advance based on the actual situation of the production line and the analysis requirements. For example, the workstations with the highest correlation influence weights are selected as candidate bottleneck workstations. These workstations occupy key positions in the correlation relationships of the production line, and further analysis is needed to determine whether they are truly bottleneck workstations.

[0103] Step S340: For each candidate bottleneck workstation in the candidate bottleneck workstation set, extract its workstation operation parameter sequence in the real-time operation status set of the production line, compare the deviation with the preset standard operation parameter sequence, and generate operation parameter deviation features.

[0104] In an automotive manufacturing production line, for each candidate bottleneck workstation in the candidate bottleneck workstation set, its workstation operating parameter sequence is extracted from the real-time operating status set of the production line. This sequence contains the changes of various operating parameters of the workstation over time during the actual production process. The preset standard operating parameter sequence is the parameter reference sequence for the workstation under normal operating conditions, representing the ideal operating state of the workstation.

[0105] As one implementation method, step S340 involves extracting the workstation operation parameter sequence in the real-time operation status set of the production line for each candidate bottleneck workstation in the candidate bottleneck workstation set, comparing the deviation with the preset standard operation parameter sequence, and generating operation parameter deviation features. Specifically, this can be implemented as the following steps S341~S346: Step S341: Extract the station operation parameter sequence of candidate bottleneck stations from the set of real-time operation status of the production line. The station operation parameter sequence includes parameter values ​​at multiple consecutive sampling times.

[0106] In automotive manufacturing production lines, data filtering and retrieval methods are used to extract the sequence of operating parameters for candidate bottleneck workstations from the real-time operating status set of the production line. This sequence contains the parameter values ​​of that workstation at multiple consecutive sampling times, such as the bolt tightening torque and part installation time at different times in the engine assembly workstation. These parameter values ​​are arranged in chronological order of sampling time, forming a time series that reflects the operating status of that workstation in the actual production process.

[0107] Step S342: Obtain the preset standard operating parameter sequence, which is the parameter reference sequence of the candidate bottleneck station under normal operating conditions.

[0108] The preset standard operating parameter sequence is formulated based on the design requirements, process standards, and historical normal operation data of the candidate bottleneck workstation. In the automotive manufacturing production line, the operating parameters of the workstation during multiple normal production processes can be statistically analyzed to obtain its average value, fluctuation range, and other characteristics, thereby determining the standard operating parameter sequence. This sequence represents the operating state of the workstation under ideal conditions.

[0109] Step S343: Perform time alignment processing on the workstation operation parameter sequence and the standard operation parameter sequence so that the two correspond one-to-one at the sampling time.

[0110] Because the sampling times of the workstation operating parameter sequence and the standard operating parameter sequence may not be completely consistent, time alignment is required. In automotive manufacturing production lines, interpolation or resampling methods can be used. For example, for two sequences with different sampling time intervals, linear interpolation can be used to estimate the corresponding parameter values ​​at the missing sampling times, ensuring a one-to-one correspondence between the workstation operating parameter sequence and the standard operating parameter sequence at the sampling times. This guarantees that subsequent deviation comparisons are performed at the same time points, improving the accuracy of the comparisons.

[0111] Step S344: Calculate the parameter difference between the aligned workstation operating parameter sequence and the standard operating parameter sequence at each sampling time to obtain the parameter difference sequence.

[0112] In an automotive manufacturing production line, the parameter difference at each sampling moment is calculated by subtracting the corresponding parameter values ​​from the aligned sequence of operating parameters for each workstation. For example, for the bolt tightening torque parameter, if at a certain sampling moment the torque value in the workstation's operating parameter sequence is A, and the torque value in the standard operating parameter sequence is B, then the parameter difference at that moment is AB. Arranging the parameter differences at all sampling moments in chronological order yields a parameter difference sequence, which reflects the difference between the actual operating parameters and the standard parameters for that workstation.

[0113] Step S345: Perform sliding window processing on the parameter difference sequence, calculate the mean and variance of the difference within each window, and obtain the window statistical feature sequence.

[0114] In automobile manufacturing production lines, sliding window processing is used to analyze the statistical characteristics of parameter differences over different time periods. A sliding window is a fixed-length time period that slides across the parameter difference sequence at certain step sizes.

[0115] As one implementation method, step S345 involves performing a sliding window process on the parameter difference sequence, calculating the mean and variance of the differences within each window, and obtaining the window statistical feature sequence. Specifically, this can be implemented as follows: steps S3451~S3456: Step S3451: Detect the fluctuation period of the parameter difference sequence, calculate the autocorrelation coefficient of the sequence through the autocorrelation function, and determine the lag step when the autocorrelation coefficient first reaches its maximum value as the fluctuation period length of the parameter difference sequence.

[0116] In automobile manufacturing production lines, detecting the fluctuation period of parameter difference sequences is crucial for understanding the periodic variation of these sequences. Autocorrelation functions, such as the Pearson autocorrelation function and the sample autocorrelation function, are used to measure the correlation between sequences. By calculating the autocorrelation coefficient of the parameter difference sequence, we can observe how it changes with the number of lag steps. The lag step length corresponding to the first time the autocorrelation coefficient reaches its maximum value represents the fluctuation period length of the parameter difference sequence. For example, if the autocorrelation coefficient reaches its maximum value after a lag of 5 sampling times, then the fluctuation period length of the parameter difference sequence is 5 sampling times, reflecting the fluctuation pattern of the parameter differences.

[0117] Step S3452: Determine the initial window size based on the fluctuation period length. The initial window size is set to twice the fluctuation period length.

[0118] In automobile manufacturing production lines, the initial window size is determined based on the length of the fluctuation cycle. Setting the initial window size to twice the length of the fluctuation cycle ensures that at least one complete fluctuation cycle is included within a single window, enabling a more comprehensive analysis of the statistical characteristics of the parameter differences. For example, if the fluctuation cycle length is 5 sampling times, then the initial window size is set to 10 sampling times. This window setting captures the fluctuations in the parameter differences, providing a suitable sample range for subsequent statistical calculations.

[0119] Step S3453: Using the initial window size as a reference, slide the window on the parameter difference sequence according to the preset step size to obtain multiple initial window segments.

[0120] In an automobile manufacturing production line, a sliding window is used as a baseline, sliding across the parameter difference sequence with a preset step size. The preset step size is a pre-defined value that determines the distance the window slides each time. For example, if the preset step size is one sampling time, the window moves one sampling time to the right each time. By sliding the window, the parameter difference sequence is divided into multiple initial window segments. Each window segment contains a certain number of parameter difference data points, which will be used for subsequent statistical calculations.

[0121] Step S3454: Calculate the standard deviation of the difference within each initial window segment. When the standard deviation is greater than the preset fluctuation threshold, increase the window size of the window segment by a preset ratio. When the standard deviation is less than the preset fluctuation threshold, decrease the window size of the window segment by a preset ratio to obtain a dynamically adjusted set of window segments.

[0122] In an automobile manufacturing production line, for each initial window segment, the standard deviation of its internal parameter differences is calculated. The standard deviation reflects the dispersion of the data within the window, that is, the magnitude of the fluctuation in the parameter differences. The preset fluctuation threshold is a pre-set standard value used to determine whether the fluctuation of the data within the window is too large or too small.

[0123] When the standard deviation is greater than the preset fluctuation threshold, it indicates that the data fluctuation within the window is large, and a larger window may be needed to capture more comprehensive fluctuation information. The window size of this segment is increased by a preset percentage, for example, 20%. When the standard deviation is less than the preset fluctuation threshold, it indicates that the data fluctuation within the window is small, and the window size can be reduced to improve computational efficiency. The window size of this segment is reduced by a preset percentage, for example, 10%. After such dynamic adjustments, a dynamically adjusted set of window segments is obtained. The sizes of these window segments are adaptively adjusted according to the data fluctuation, enabling more accurate analysis of the statistical characteristics of parameter differences.

[0124] Step S3455: For each window segment in the dynamically adjusted window segment set, calculate the arithmetic mean of all parameter differences within the segment as the difference mean, and calculate the average of the sum of squares of all parameter differences and the difference mean within the segment as the difference variance.

[0125] In the automobile manufacturing production line, statistical calculations are performed on each window segment in the dynamically adjusted set of window segments. The arithmetic mean of all parameter differences within a segment is calculated; that is, the sum of all parameter differences within the window is divided by the number of parameter differences to obtain the mean difference. The mean difference reflects the average level of parameter differences within that window.

[0126] The variance of the differences is calculated by averaging the sum of squares of all parameter differences and the mean of those differences within a segment. The variance reflects the dispersion of the parameter differences within the window. By calculating the mean and variance of the differences, the statistical characteristics of the parameter differences within the window can be described from different perspectives, providing data support for subsequent feature generation.

[0127] Step S3456: Arrange the mean and variance of the differences of each window segment in the window sliding order to obtain the window statistical feature sequence.

[0128] In an automotive manufacturing production line, the mean and variance of the differences for each window segment are arranged sequentially according to the window sliding order, forming a sequence, namely the window statistical characteristic sequence. This sequence contains statistical characteristics of the parameter differences over different time periods, such as the trend of the mean difference and the fluctuation of the variance. It reflects the temporal variation characteristics of the deviation between the operating parameters of the candidate bottleneck station and the standard parameters, providing an important data foundation for generating operating parameter deviation characteristics.

[0129] Step S346: Convert the window statistical feature sequence into vector form to obtain the running parameter deviation features.

[0130] In automotive manufacturing production lines, window statistical feature sequences are converted into vector form. Each element in the window statistical feature sequence is arranged in a specific order into a vector; this vector represents the operating parameter deviation feature. The operating parameter deviation feature, presented as a vector, synthesizes the statistical characteristics of parameter differences over different time periods, facilitating subsequent analysis and processing. Through this feature, one can intuitively understand the deviation between the operating parameters of a candidate bottleneck station and the standard parameters, determining whether the station exhibits operational anomalies.

[0131] Step S350: Simultaneously extract the material flow sequence information of the candidate bottleneck workstations, calculate the difference between the material input and output per unit time, and generate material backlog characteristics.

[0132] In automotive manufacturing production lines, for candidate bottleneck workstations, material flow timing information is extracted from the real-time operating status set of the production line. This information records the material input and output at different points in time for that workstation. The difference between the material input and output per unit time is calculated. If the difference is positive, it indicates that the material input exceeds the output within that unit time, potentially leading to material backlog; if the difference is negative, it indicates that the material output exceeds the input, and material flow is relatively smooth. By statistically analyzing the difference between the material input and output per unit time, a material backlog characteristic is generated. This characteristic can be represented by a numerical value or a vector, reflecting the degree of material backlog at the candidate bottleneck workstation. For example, the larger the value of the material backlog characteristic, the more severe the material backlog at that workstation, potentially affecting the overall production line efficiency.

[0133] Step S360: Combine the station identifier, corresponding operating parameter deviation characteristics, and material backlog characteristics of the candidate bottleneck station to obtain the bottleneck station feature set.

[0134] In automotive manufacturing production lines, the station identifier, corresponding operational parameter deviation characteristics, and material backlog characteristics of candidate bottleneck stations are combined to form a bottleneck station feature set. The station identifier uniquely identifies the bottleneck station; the operational parameter deviation characteristics reflect the deviation between the station's operating parameters and standard parameters; and the material backlog characteristics reflect the degree of material backlog at the station. Combining these three features provides a comprehensive description of the bottleneck station's status. For example, the station identifier can pinpoint the specific station, the operational parameter deviation characteristics can analyze any operational anomalies at that station, and the material backlog characteristics can reveal material flow issues at that station. This bottleneck station feature set provides detailed information for subsequent production line adjustment strategies.

[0135] Step S400: Generate a set of dynamic adjustment strategies for the production line based on the set of bottleneck workstation characteristics. The set of dynamic adjustment strategies includes parameter adjustment strategies and material allocation strategies for bottleneck workstations. Input the set of dynamic adjustment strategies into the production line control execution system, and the production line control execution system performs production line balancing control operations based on the set of dynamic adjustment strategies.

[0136] In automotive manufacturing production lines, generating a set of dynamic adjustment strategies based on the bottleneck workstation feature set is intended to address bottleneck issues and achieve production line balance and optimization. The bottleneck workstation feature set contains detailed information about the bottleneck workstation, such as operating parameter deviations and material backlog characteristics. Based on this information, reasonable adjustment strategies can be formulated.

[0137] As one implementation method, step S400, generating a set of dynamic adjustment strategies for the production line based on the bottleneck workstation feature set, can be specifically implemented as the following steps S410~S460: Step S410: For each bottleneck station in the bottleneck station feature set, input its operating parameter deviation features and material backlog features into the input layer of the adaptive adjustment strategy generation model.

[0138] In automotive manufacturing production lines, the adaptive adjustment strategy generation model is a model used to generate adjustment strategies based on the characteristic information of bottleneck workstations. For each bottleneck workstation in the bottleneck workstation feature set, its operating parameter deviation characteristics and material backlog characteristics are input into the input layer of the adaptive adjustment strategy generation model.

[0139] The operating parameter deviation characteristics reflect the deviation between the operating parameters of the bottleneck station and the standard parameters, while the material backlog characteristics reflect the degree of material backlog at the station. The input layer is the model's entry point; it receives this feature information and passes it to subsequent layers for processing. By inputting these features into the model, it can analyze the problems at the bottleneck station based on this information, providing a basis for generating adjustment strategies.

[0140] Step S420: The feature mapping layer of the model generated by the adaptive adjustment strategy maps the running parameter deviation features to the parameter adjustment demand vector and the material backlog features to the material allocation demand vector.

[0141] In automotive manufacturing production lines, the feature mapping layer of the adaptive adjustment strategy generation model transforms the input feature information into a vector form more suitable for model processing. For operational parameter deviation features, the feature mapping layer converts them into a parameter adjustment demand vector through a series of transformations and mapping operations. This vector represents the direction and extent of parameter adjustment required at the bottleneck workstation. For example, if the operational parameter deviation features indicate that the bolt tightening torque at a certain workstation is too high, then the parameter adjustment demand vector might indicate a need to reduce that torque.

[0142] For material backlog characteristics, the feature mapping layer maps them to a material allocation demand vector. This vector represents the material allocation needs of the bottleneck workstation, such as increasing or decreasing the material input or adjusting the direction of material allocation. Through the processing of the feature mapping layer, the original feature information is converted into a vector form that the model can understand and process, providing suitable input for subsequent strategy generation.

[0143] Step S430: Input the parameter adjustment demand vector and the material allocation demand vector into the strategy fusion layer, calculate the correlation weight between the two through the multi-head attention mechanism, and perform weighted fusion of the parameter adjustment demand vector and the material allocation demand vector based on the correlation weight to obtain the fused demand vector.

[0144] In the automotive manufacturing production line, parameter adjustment demand vectors and material allocation demand vectors are input into the strategy fusion layer of the adaptive adjustment strategy generation model. Multi-head attention is a mechanism used to calculate the correlation weights between vectors, and it can consider various correlation relationships between vectors.

[0145] In the strategy fusion layer, the multi-head attention mechanism first analyzes the parameter adjustment demand vector and the material allocation demand vector, calculating the correlation weight between them. The correlation weight reflects the degree of mutual influence between parameter adjustment and material allocation. For example, if the operating parameter deviation of a bottleneck workstation is closely related to the material backlog, then the correlation weight between their corresponding vectors will be relatively large.

[0146] Based on the calculated correlation weights, the parameter adjustment demand vector and the material allocation demand vector are weighted and fused. Each vector is multiplied by its corresponding correlation weight and then summed to obtain the fused demand vector. This fused demand vector integrates the demands from both parameter adjustment and material allocation, taking into account the relationships between them.

[0147] As one implementation method, step S430 involves inputting the parameter adjustment demand vector and the material allocation demand vector into the strategy fusion layer, calculating the correlation weight between the two through a multi-head attention mechanism, and then performing a weighted fusion of the parameter adjustment demand vector and the material allocation demand vector based on the correlation weight to obtain the fused demand vector. Specifically, this can be implemented as follows: steps S431-S436. Step S431: Perform dimension unification processing on the parameter adjustment demand vector and the material allocation demand vector to make them have the same characteristic dimensions, so as to obtain the dimension-unified parameter adjustment demand vector and the dimension-unified material allocation demand vector.

[0148] In automotive manufacturing production lines, parameter adjustment demand vectors and material allocation demand vectors may have different feature dimensions, which can complicate subsequent calculations and fusion. Therefore, it is necessary to unify their dimensions, which can be achieved using zero-padding or feature selection methods. If the dimension of the parameter adjustment demand vector is smaller than that of the material allocation demand vector, zero elements can be added to the end of the parameter adjustment demand vector to make its dimension the same as that of the material allocation demand vector. Conversely, if the dimension of the parameter adjustment demand vector is larger than that of the material allocation demand vector, a feature selection algorithm can be used to select the most important features in the parameter adjustment demand vector to make its dimension consistent with that of the material allocation demand vector. After dimension unification, the resulting parameter adjustment demand vector and material allocation demand vector are provided with input of the same dimension for multi-head attention mechanism calculations.

[0149] Step S432: Initialize the query matrix, key matrix, and value matrix of the multi-head attention mechanism, and use the parameter adjustment demand vector after dimension unification as the query vector, and the material allocation demand vector after dimension unification as the key vector and value vector.

[0150] In automotive manufacturing production lines, the query matrix, key matrix, and value matrix of a multi-head attention mechanism are key parameters used to calculate attention weights. These matrices need to be initialized before computation. Random initialization can be used to generate appropriate query, key, and value matrices based on the dimension of the input vector and the model's design requirements.

[0151] The parameter adjustment requirement vector, unified in dimensions, serves as the query vector, used to retrieve relevant information in subsequent calculations. The material allocation requirement vector, unified in dimensions, serves as both the key and value vectors. The key vector is used to match the query vector and calculate attention weights; the value vector is used for weighted summation based on these attention weights to obtain the final attention output. Through this setup, the multi-head attention mechanism can analyze the relationship between parameter adjustment requirements and material allocation requirements.

[0152] Step S433: Obtain query features by performing a linear transformation on the query vector using the query matrix, obtain key features by performing a linear transformation on the key vector using the key matrix, and obtain value features by performing a linear transformation on the value vector using the value matrix.

[0153] In the automotive manufacturing production line, the multi-head attention mechanism performs linear transformations on the query vector, key vector, and value vector. Query features are obtained by multiplying the query matrix and query vector. These query features represent the query vector in a new feature space, containing more information suitable for matching. Similarly, key features are obtained by multiplying the key matrix and key vector; value features are obtained by multiplying the value matrix and value vector. These linear transformations map the input vectors to a more suitable feature space, enabling more accurate calculation of attention weights in subsequent computations. Linear transformations increase the model's expressive power, allowing it to capture more complex relationships between input vectors.

[0154] Step S434: Calculate the dot product attention score of the query feature and the key feature, normalize the dot product attention score, and obtain the attention weight distribution.

[0155] In an automotive manufacturing production line, after obtaining query features and key features, a dot product attention score is calculated between them. The dot product attention score measures the similarity between the query features and key features; a higher score indicates greater similarity. An attention score matrix is ​​obtained by performing a dot product operation between the query features and key features. This attention score matrix is ​​then normalized, typically using the Softmax function. The Softmax function converts each element in the attention score matrix into a probability value, such that the sum of all elements is 1. After normalization, an attention weight distribution is obtained, which represents the degree of association between the query vector and the key vector; a larger weight indicates a stronger association.

[0156] Step S435: Weight the attention weight distribution and the value features to obtain the single-head attention output features.

[0157] In the automotive manufacturing production line, the value features are weighted and summed based on the calculated attention weight distribution. Each weight in the attention weight distribution is multiplied by its corresponding value feature element, and then all products are summed to obtain the single-head attention output feature. This feature integrates the correlation information between the query vector and the value vector, filtering and weighting the value features according to the attention weights, highlighting the value feature information closely related to the query vector. The single-head attention output feature reflects the relationship between parameter adjustment requirements and material allocation requirements under a single attention head.

[0158] Step S436: Linearly map the single-head attention output features through a preset fusion matrix to obtain the fusion requirement vector.

[0159] In automotive manufacturing production lines, a pre-defined fusion matrix is ​​used to linearly map the single-head attention output features. Multiplying the single-head attention output features by the fusion matrix yields a fused demand vector. The fusion matrix transforms the single-head attention output features into a vector form more suitable for subsequent processing. It can adjust the feature dimensions and representation, allowing the fused demand vector to better integrate parameter adjustment requirements and material allocation needs. Through this linear mapping, the information in the single-head attention output features is integrated and transformed to obtain a comprehensive demand vector, providing a unified input for subsequent production line adjustment strategies.

[0160] Step S440: Input the fusion requirement vector into the strategy generation layer of the strategy fusion layer, and generate a preliminary adjustment strategy by matching the adjustment strategy type corresponding to the fusion requirement vector through the preset strategy template library.

[0161] In the automotive manufacturing production line, the obtained fused demand vector is input into the strategy generation layer of the adaptive adjustment strategy generation model. The preset strategy template library is a library containing various adjustment strategy types, each corresponding to different production line problems and solutions.

[0162] The strategy generation layer matches the characteristics of the fused demand vector against a strategy template library. By comparing the feature similarity between the fused demand vector and each strategy type in the template library, it finds the most matching strategy type. For example, if the fused demand vector indicates that a bottleneck workstation requires simultaneous adjustments to operating parameters and material allocation, the strategy generation layer will find a corresponding strategy type in the template library that includes both parameter and material allocation adjustments. Based on the matched strategy type, a preliminary adjustment strategy is generated. This preliminary adjustment strategy is a general adjustment plan that provides direction for resolving the bottleneck workstation's problem.

[0163] Step S450: Verify the feasibility of the preliminary adjustment strategy. Check whether the preliminary adjustment strategy meets the physical constraints of the production line equipment and the safety operation specifications. If it does, it is determined as the final adjustment strategy. If it does not meet, return to the strategy generation layer to rematch the strategy template library until a final adjustment strategy that meets the constraints is generated.

[0164] In automobile manufacturing production lines, verifying the feasibility of initial adjustment strategies is a crucial step in ensuring their implementation in actual production. Physical constraints on production line equipment include maximum load capacity, operating speed range, and precision requirements. Safety operating procedures, on the other hand, are rules established to ensure the safety of personnel and equipment during the production process.

[0165] Check whether the preliminary adjustment strategy complies with these constraints and specifications. For example, if the preliminary adjustment strategy requires a device to operate at a speed exceeding its maximum load capacity, or violates certain provisions of the safe operating procedures, then this strategy is not feasible. If the preliminary adjustment strategy meets all constraints and specifications, then it is determined as the final adjustment strategy.

[0166] If the initial adjustment strategy fails to meet constraints and specifications, it is returned to the strategy generation layer to match other strategy types from the strategy template library. This process is repeated until a final adjustment strategy that meets the physical constraints and safe operating procedures of the production line equipment is generated. This ensures the feasibility and effectiveness of the adjustment strategy, preventing equipment damage or safety accidents in actual production.

[0167] Step S460: Classify and summarize the final adjustment strategies corresponding to all bottleneck workstations to obtain a set of dynamic adjustment strategies for the production line, including parameter adjustment strategies and material allocation strategies.

[0168] In the automotive manufacturing production line, a final adjustment strategy that satisfies the constraints has already been generated for each bottleneck workstation. The final adjustment strategies for all bottleneck workstations are then categorized and summarized.

[0169] As one implementation method, step S460 involves classifying and summarizing the final adjustment strategies corresponding to all bottleneck workstations to obtain a set of dynamic production line adjustment strategies that includes parameter adjustment strategies and material allocation strategies. Specifically, this can be implemented as follows: steps S461 to S466: Step S461: Identify the strategy type of the final adjustment strategy for each bottleneck workstation and determine whether the final adjustment strategy belongs to the parameter adjustment strategy or the material allocation strategy.

[0170] In automotive manufacturing production lines, the final adjustment strategy for each bottleneck workstation requires strategy type identification. By analyzing the specific content of the strategy, it's determined whether it adjusts workstation operating parameters or material allocation. For example, if the strategy includes adjusting equipment operating speed or torque, it's a parameter adjustment strategy; if it involves adjusting material input or allocation direction, it's a material allocation strategy. This identification helps distinguish different types of strategies, preparing for subsequent classification and summarization.

[0171] Step S462: If the final adjustment strategy belongs to the parameter adjustment strategy, extract the parameter adjustment object, adjustment direction and adjustment magnitude information in the strategy to generate a parameter adjustment strategy entry.

[0172] In automotive manufacturing production lines, when a final adjustment strategy is determined to be a parameter adjustment strategy, key information is extracted from the strategy. The parameter adjustment object is the specific parameter that needs adjustment, such as the speed or pressure of a piece of equipment. The adjustment direction indicates whether the parameter needs to be increased or decreased, such as increasing the speed or decreasing the pressure. The adjustment magnitude is the degree of parameter adjustment. This information is then organized to generate parameter adjustment strategy entries. For example, for a parameter adjustment strategy at an engine assembly station, the adjustment object is extracted as "bolt tightening torque," the adjustment direction as "reduction," and the adjustment magnitude as "a certain percentage." This information is combined into a single parameter adjustment strategy entry, which clearly describes the specific parameter adjustment content for that station.

[0173] Step S463: If the final adjustment strategy belongs to the material allocation strategy, extract the material source station, material destination station and material transfer rate information from the strategy to generate material allocation strategy entries.

[0174] In an automotive manufacturing production line, when the final adjustment strategy falls under the category of material allocation strategy, relevant information is extracted from the strategy. The material source station is the initial location of the material, the material destination station is the target location the material needs to reach, and the material transfer rate is the speed at which the material is transported between the two stations. This information is then organized to generate material allocation strategy entries. For example, for a material allocation strategy, the material source station might be extracted as "warehouse," the material destination station as "engine assembly station," and the material transfer rate as "rate x." This information is combined into a single material allocation strategy entry, which clearly defines the allocation and transfer of materials within the production line.

[0175] Step S464: Group all parameter adjustment strategy items according to workstation identifier to obtain parameter adjustment strategy groups.

[0176] In the automotive manufacturing production line, all parameter adjustment strategy items are grouped according to workstation identifiers. Each workstation identifier corresponds to a unique workstation, and grouping allows parameter adjustment strategy items targeting the same workstation to be grouped together. For example, all parameter adjustment strategy items targeting the engine assembly workstation are placed in one group. This grouping facilitates the unified management and execution of parameter adjustment strategies for each workstation and also makes it easier to compare and analyze the parameter adjustment status across different workstations.

[0177] Step S465: Group all material allocation strategy entries by material type to obtain material allocation strategy groups.

[0178] In automotive manufacturing production lines, all material allocation strategy items are grouped according to material type. Different types of materials may have different requirements and characteristics in their flow and allocation within the production line. For example, all material allocation strategy items related to engine parts are placed in one group, while those related to body parts are placed in another. By grouping by material type, material allocation can be managed and adjusted more effectively, improving the efficiency of material flow.

[0179] Step S466: Combine the parameter adjustment strategy group and the material allocation strategy group into a set to obtain the production line dynamic adjustment strategy set.

[0180] In automotive manufacturing production lines, parameter adjustment strategies and material allocation strategies are combined to form a set, known as the dynamic production line adjustment strategy set. This set includes parameter adjustment and material allocation strategies for all bottleneck workstations, providing a comprehensive production line adjustment solution. The dynamic production line adjustment strategy set is input into the production line control execution system, which then performs production line balancing control operations based on this set. The production line control execution system adjusts equipment parameters and optimizes material allocation and flow according to specific strategy entries in the strategy set, thereby achieving balanced and efficient production line operation. For example, based on entries in the parameter adjustment strategy set, the operating parameters of the equipment at the engine assembly workstation are adjusted; based on entries in the material allocation strategy set, the allocation and transfer of materials between different workstations are adjusted. Through this operation, bottlenecks in the production line are eliminated, improving the overall production efficiency and quality of the automotive manufacturing production line.

[0181] It is understood that the various algorithms and functions involved in the above descriptions of the embodiments of the present invention can all be obtained from relevant content in the prior art. To save space, they will not be elaborated on in the embodiments of the present invention. In addition, those skilled in the art can supplement the details based on common knowledge in the art when implementing the solutions of the present invention. For example, they can use normalization to eliminate dimensional conflicts before feature fusion, use interpolation to eliminate dimensional differences, reasonably set thresholds based on historical data, experience or business scenario requirements, train the model based on a general model training method, set the number of layers in the model structure based on actual needs, select activation functions, etc. The present invention will not provide redundant descriptions of overly detailed implementation processes here.

[0182] Please see Figure 2 , Figure 2This is a schematic diagram of a computer system provided in an embodiment of the present invention. The computer system includes at least a processor 101, a communication interface 102, and a memory 103. The processor 101, communication interface 102, and memory 103 can be connected via a bus or other means. The processor 101 (or Central Processing Unit, CPU) is the computing and control core of the computer system, capable of parsing various instructions and processing various data within the computer system. The communication interface 102 may optionally include a standard wired interface or a wireless interface (such as Wi-Fi, mobile communication interface, etc.), and can be used to send and receive data under the control of the processor 101; the communication interface 102 can also be used for data transmission and interaction within the computer system. The memory 103 is a storage device in the computer system used to store programs and data. It is understood that the memory 103 here can include the computer system's built-in memory, or it can include extended memory supported by the computer system. The memory 103 provides storage space, which stores the computer system's operating system; this invention does not limit this storage space.

[0183] In one embodiment, the processor 101 executes the dynamic factory production line balancing method based on adaptive control provided in the above embodiments of the present invention by running a computer program in the memory 103.

Claims

1. A dynamic factory production line balancing method based on adaptive control, characterized in that, include: Obtain a set of real-time operating statuses of the production line, which includes the sequence of operating parameters for each workstation and the material flow sequence information between workstations; The set of real-time operating statuses of the production line is subjected to workstation association feature parsing to generate a workstation association feature matrix. The workstation association feature matrix is ​​used to characterize the coupling relationship of operating parameters and the material flow dependency relationship between different workstations. Based on the workstation association feature matrix, production line bottlenecks are identified, and a bottleneck workstation feature set is generated. The bottleneck workstation feature set includes the bottleneck workstation identifier and its corresponding operating parameter deviation features and material backlog features. A set of dynamic adjustment strategies for the production line is generated based on the bottleneck workstation feature set. The set of dynamic adjustment strategies for the production line includes parameter adjustment strategies and material allocation strategies for the bottleneck workstation.

2. The method according to claim 1, characterized in that, The step of performing workstation association feature parsing on the set of real-time operating statuses of the production line to generate a workstation association feature matrix includes: Extract the workstation operation parameter sequence and material flow timing information of each workstation from the real-time operation status set of the production line to obtain the original feature set of the workstation; Time series features are extracted from the workstation operation parameter sequence in the original feature set of the workstation to generate workstation operation sequence features, which include parameter fluctuation features and parameter trend features. Material flow timing information in the original feature set of the workstation is extracted to generate material flow path features, which include path node sequence and flow duration features between nodes. The workstation runtime sequence features and the material flow path features are input into the workstation association graph network model. Each workstation is mapped to a graph node through the graph node initialization layer of the workstation association graph network model. The workstation runtime sequence features are used as graph node features and the material flow path features are used as graph edge features to construct the workstation association graph structure. The workstation association graph structure is subjected to multi-layer feature aggregation through the graph convolutional layer of the workstation association graph network model to obtain graph aggregation features containing high-order association information between workstations. The graph aggregation features are then converted into matrix form to obtain the workstation association feature matrix.

3. The method according to claim 1, characterized in that, The process of identifying production line bottlenecks based on the workstation association feature matrix and generating a bottleneck workstation feature set includes: The workstation association feature matrix is ​​processed by matrix block division, and the workstation association feature matrix is ​​divided into multiple workstation feature sub-matrices according to workstation grouping. Each workstation feature sub-matrice corresponds to the association feature of one workstation. For each workstation feature submatrix, feature importance is evaluated, and the correlation influence weight of each workstation feature submatrix is ​​calculated. The correlation influence weight is used to characterize the degree of influence of the workstation on the operation of other workstations. Based on the weight of the association influence of each workstation, sort them from largest to smallest, and select a preset number of influential workstations as the candidate bottleneck workstation set; For each candidate bottleneck workstation in the candidate bottleneck workstation set, extract its workstation operation parameter sequence in the real-time operation status set of the production line, compare the deviation with the preset standard operation parameter sequence, and generate operation parameter deviation features. Simultaneously, material flow timing information of the candidate bottleneck workstations is extracted, and the difference between the material input and output per unit time is calculated to generate material backlog characteristics. The bottleneck workstation feature set is obtained by combining the workstation identifier, corresponding operating parameter deviation characteristics, and material backlog characteristics of the candidate bottleneck workstation.

4. The method according to claim 3, characterized in that, The process of evaluating the feature importance of each workstation feature submatrix and calculating the correlation influence weight of each workstation feature submatrix includes: Singular value decomposition is performed on each workstation feature submatrix to obtain a sequence of singular values ​​of the submatrix. The largest singular value in the sequence of singular values ​​of the submatrix is ​​selected as the principal eigenvalue of the workstation feature submatrix. Normalize the principal eigenvalues ​​of all workstation feature submatrices to obtain the proportion of principal eigenvalues ​​of each workstation. The proportion of principal eigenvalues ​​is the ratio of the principal eigenvalue of a single workstation to the sum of the principal eigenvalues ​​of all workstations. Extract the row vector corresponding to each workstation from the workstation association feature matrix, calculate the L2 norm of the row vector, and obtain the workstation association strength value; The proportion of the principal feature value and the workstation association strength value are fused to obtain a fusion weight value, which is obtained by multiplying the proportion of the principal feature value and the workstation association strength value and taking the average value. The fusion weight value is used as the correlation influence weight of the workstation feature submatrix.

5. The method according to claim 1, characterized in that, The step of generating a set of dynamic production line adjustment strategies based on the bottleneck workstation feature set includes: For each bottleneck workstation in the bottleneck workstation feature set, its operating parameter deviation features and material backlog features are input into the input layer of the adaptive adjustment strategy generation model; The feature mapping layer of the model generated by the adaptive adjustment strategy maps the operating parameter deviation features into parameter adjustment demand vectors and the material backlog features into material allocation demand vectors. The parameter adjustment demand vector and the material allocation demand vector are input into the strategy fusion layer. The correlation weight between the two is calculated through a multi-head attention mechanism. Based on the correlation weight, the parameter adjustment demand vector and the material allocation demand vector are weighted and fused to obtain the fused demand vector. The fusion requirement vector is input into the strategy generation layer, and the adjustment strategy type corresponding to the fusion requirement vector is matched with the preset strategy template library to generate a preliminary adjustment strategy; The feasibility of the preliminary adjustment strategy is verified. It is checked whether the preliminary adjustment strategy meets the physical constraints of the production line equipment and the safety operation specifications. If it meets the requirements, it is determined as the final adjustment strategy. If it does not meet the requirements, it is returned to the strategy generation layer to rematch the strategy template library until a final adjustment strategy that meets the constraints is generated. The final adjustment strategies corresponding to all bottleneck workstations are categorized and summarized to obtain a set of dynamic production line adjustment strategies, including parameter adjustment strategies and material allocation strategies.

6. The method according to claim 5, characterized in that, The parameter adjustment demand vector and the material allocation demand vector are input into the strategy fusion layer. A multi-head attention mechanism is used to calculate their correlation weights. Based on these correlation weights, the parameter adjustment demand vector and the material allocation demand vector are weighted and fused to obtain a fused demand vector, including: The parameter adjustment demand vector and the material allocation demand vector are subjected to dimension unification processing so that they have the same characteristic dimension, resulting in a dimension-unified parameter adjustment demand vector and a dimension-unified material allocation demand vector. Initialize the query matrix, key matrix, and value matrix of the multi-head attention mechanism, and use the parameter adjustment demand vector after dimension unification as the query vector, and the material allocation demand vector after dimension unification as the key vector and value vector; The query features are obtained by linearly transforming the query vector using the query matrix, the key features are obtained by linearly transforming the key vector using the key matrix, and the value features are obtained by linearly transforming the value vector using the value matrix. Calculate the dot product attention score of the query feature and the key feature, and normalize the dot product attention score to obtain the attention weight distribution; The attention weight distribution and the value feature are weighted and summed to obtain the single-head attention output feature; The single-head attention output features are linearly mapped using a preset fusion matrix to obtain the fusion requirement vector.

7. The method according to claim 2, characterized in that, The step of performing multi-layer feature aggregation on the workstation association graph structure through the graph convolutional layer of the workstation association graph network model to obtain graph aggregation features containing high-order association information between workstations includes: Initialize the graph convolutional layer weight parameters of the workstation association graph network model, and set the number of graph convolutional layers to a preset number; The graph node features and graph edge features of the workstation association graph structure are input into the first graph convolutional layer. The first layer of adjacency aggregation features are obtained by performing matrix multiplication between the adjacency matrix of the first graph convolutional layer and the graph node features. The first layer of adjacency aggregation features are subjected to nonlinear activation processing to obtain the first layer of activation features; The first layer activation features are used as the input of the second layer graph convolutional layer. The first layer activation features are aggregated again by combining the weight coefficients of the graph edge features to obtain the second layer adjacency aggregation features. The second layer adjacency aggregation features are then subjected to nonlinear activation processing to obtain the second layer activation features. The above aggregation and activation processes are performed sequentially according to the preset number of layers until the processing of the last graph convolutional layer is completed, and the last layer of activation features is obtained. The activation features of the last layer are concatenated with the activation features of each intermediate layer to obtain a graph aggregation feature containing multi-scale correlation information.

8. The method according to claim 3, characterized in that, For each candidate bottleneck workstation in the candidate bottleneck workstation set, its workstation operating parameter sequence in the real-time operating status set of the production line is extracted, and the deviation is compared with the preset standard operating parameter sequence to generate operating parameter deviation features, including: Extract the station operation parameter sequence of candidate bottleneck stations from the set of real-time operating status of the production line, wherein the station operation parameter sequence includes parameter values ​​at multiple consecutive sampling times; Obtain a preset standard operating parameter sequence, which is a parameter reference sequence for the candidate bottleneck workstation under normal operating conditions; The workstation operation parameter sequence and the standard operation parameter sequence are time aligned so that they correspond one-to-one at the sampling time. The parameter difference sequence is obtained by calculating the parameter difference between the aligned workstation operating parameter sequence and the standard operating parameter sequence at each sampling time. The parameter difference sequence is processed by a sliding window, and the mean and variance of the difference within each window are calculated to obtain the window statistical feature sequence. The window statistical feature sequence is converted into vector form to obtain the running parameter deviation features.

9. The method according to claim 5, characterized in that, The final adjustment strategies corresponding to all bottleneck workstations are categorized and summarized to obtain a set of dynamic production line adjustment strategies, including parameter adjustment strategies and material allocation strategies, including: For each bottleneck workstation, the final adjustment strategy is identified by strategy type to determine whether the final adjustment strategy is a parameter adjustment strategy or a material allocation strategy. If the final adjustment strategy is a parameter adjustment strategy, then extract the parameter adjustment object, adjustment direction and adjustment magnitude information from the strategy to generate a parameter adjustment strategy entry; If the final adjustment strategy is a material allocation strategy, then extract the material source station, material destination station and material transfer rate information from the strategy to generate a material allocation strategy entry. All parameter adjustment strategy items are grouped by workstation identifier to obtain parameter adjustment strategy groups; All material allocation strategy entries are grouped by material type to obtain material allocation strategy groups; The parameter adjustment strategy group and the material allocation strategy group are combined into a set to obtain the production line dynamic adjustment strategy set.

10. A computer system, characterized in that, include: A memory, wherein a computer program is stored; A processor for loading the computer program to implement the dynamic factory production line balancing method based on adaptive control as described in any one of claims 1-9.

Images