Intelligent assembly line collaborative control system based on machine learning

Abstract

The application discloses an intelligent assembly line collaborative control system based on machine learning, comprising: a data acquisition module for acquiring multi-station operation data of the assembly line; a feature construction module for executing heterogeneous feature coding, generating state representation sequences and adjustment action sets; a trend prediction module for inputting the state representation sequences into an improved DTW-former model to perform collaborative trend prediction and outputing beat trend and cache evolution trend; a causal learning module for generating a causal diagram through an improved NOTEARS algorithm; a strategy generation module for generating screened collaborative adjustment strategies; an instruction mapping module for generating execution results; and a feedback updating module for completing closed-loop optimization according to feedback data. The application realizes unified modeling of multi-station operation states, collaborative trend prediction and causal constraint adjustment, and improves beat consistency, resource utilization efficiency and overall operation stability of the assembly line.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing and industrial automation control technology, and in particular to an intelligent assembly line collaborative control system based on machine learning. Background Technology

[0002] With the continuous development of intelligent manufacturing and industrial internet technologies, the demand for collaborative control of intelligent assembly lines, which are oriented towards multi-variety, small-batch, and dynamically adjustable cycle times, is constantly growing. Control technologies focusing on multi-workstation operation data modeling, cycle time trend prediction, and cross-workstation collaborative optimization have received widespread attention. Existing assembly line control systems mainly rely on fixed-cycle time control logic or local adjustment methods based on single-workstation feedback for production line operation management. However, the following problems are commonly found in actual industrial scenarios: The multi-station operation data in the assembly line comes from complex sources, including processing completion signals, equipment operating status signals, and buffer occupancy data, among other types of time-series data. The sampling periods and timestamp precision differ between different stations, making it difficult to form a stable and consistent multi-station collaborative state sequence using traditional time alignment and simple resampling methods. This results in inaccurate characterization of long-cycle and short-cycle cycle fluctuations. Regarding the dependencies and resource conflicts between stations, existing control strategies often rely on manual experience or static rule configuration, lacking the ability to systematically model the station topology and action-level resource occupancy relationships. Under cycle fluctuations or buffer backlog conditions, resource contention and mutual exclusion conflicts easily occur, reducing production line stability. In terms of trend prediction and causal analysis, traditional methods based on linear regression or simple recurrent neural networks struggle to characterize dynamic time misalignment and nonlinear propagation effects in non-stationary time sequences. The lack of a constraint learning mechanism for the causal adjacency relationships of stations leads to a lack of global consistency and interpretability in the generation of collaborative adjustment strategies, making it difficult to support the closed-loop optimization control requirements of the assembly line.

[0003] Therefore, how to provide a machine learning-based intelligent assembly line collaborative control system is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] One objective of this invention is to propose a machine learning-based intelligent assembly line collaborative control system. This invention utilizes multi-station operation data modeling technology, an improved DTW-former model collaborative trend prediction method, and an improved NOTEARS algorithm causal structure learning method to construct a collaborative control system comprising a data acquisition module, a feature construction module, a trend prediction module, a causal learning module, a strategy generation module, an instruction mapping module, and a feedback update module. It resamples and aligns the multi-station operation data of the assembly line to form long-period and short-period sequences, constructs a state representation sequence and a set of adjustment actions, and uses path feasible region constraints... The system predicts cycle time and buffer evolution trends by adjusting the gating coefficient. It learns the workstation causal adjacency matrix by adjusting the topological delay, the feasible region projection, and the residual consistency. It generates a causal graph and calculates the marginal causal gain of the adjustment action. It constructs a set of collaborative adjustment strategies, which are then mapped to discrete control commands after being filtered by action-level resource occupancy relationships and mutual exclusion constraints. The system updates the path feasible region, gating coefficient, projection intensity, and workstation causal adjacency matrix based on feedback data, forming a closed-loop optimized control structure. It has the advantages of high accuracy in collaborative trend prediction, clear characterization of causal relationships, controllable conflict of adjustment strategies, and significant improvement in assembly line operation stability and cycle time consistency.

[0005] According to an embodiment of the present invention, a machine learning-based intelligent assembly line collaborative control system includes: The data acquisition module is used to acquire multi-station operating data of the assembly line, and perform resampling and time alignment according to the control cycle to form long-cycle sequences and short-cycle sequences. The feature construction module is used to perform heterogeneous feature encoding on long-period sequences and short-period sequences, map discrete event quantities and continuous measurement quantities into state vectors of the same dimension, construct the workstation topology structure according to the assembly line process flow, and construct action-level resource occupancy relationships and mutual exclusion constraints based on the action execution time window and resource occupancy identifier in the multi-workstation operation data, and generate state representation sequences and adjustment action sets. The trend prediction module is used to input the state representation sequence into the improved DTW-former model to perform collaborative trend prediction. The improved DTW-former model includes a temporal attention layer and a DTW attention layer. In the DTW attention layer, a path feasible region is constructed, and gating coefficients are generated according to the path confidence to adjust the DTW attention weights. The output is the beat trend and the cache evolution trend. The causal learning module is used to construct a structural learning sample matrix based on the beat trend and cache evolution trend, learn the workstation causal adjacency matrix through the improved NOTEARS algorithm, perform topological delay joint feasible region projection, and adjust the projection intensity according to residual consistency to generate a causal graph. The strategy generation module is used to calculate the marginal causal gain of the adjustment actions in the set of adjustment actions based on the causal graph, and generate continuous coordinated adjustment coefficients by combining the beat trend and the cache evolution trend, construct a set of coordinated adjustment strategies, and perform feasibility screening based on the action-level resource occupancy relationship and mutual exclusion constraints to generate the screened coordinated adjustment strategies. The instruction mapping module is used to map the selected collaborative adjustment strategies into discrete control instructions, and send them to the corresponding workstation control units for execution, generating execution results; The feedback update module is used to generate feedback data based on the execution results, update the path feasible region and gating coefficient in the improved DTW-former model according to the feedback data, and update the projection intensity and workstation causal adjacency matrix in the improved NOTEARS algorithm to complete the closed-loop optimization of assembly line collaborative control.

[0006] Optionally, the data acquisition module specifically comprises: Collect processing completion signals, equipment operating status signals, buffer occupancy rate data, action execution time windows, resource occupancy indicators, and control cycle timestamps from multiple workstations on the assembly line; Based on the control cycle timestamp, the processing completion signal, equipment operating status signal, buffer occupancy rate data, action execution time window, and resource occupancy identifier execution time are aligned to obtain operating data under a unified time reference. According to the preset control cycle length, the running data under the unified time base is resampled to generate a running data sequence with equal time intervals; The operation data sequences with equal time intervals are spliced ​​together according to the workstation number order to construct a multi-workstation collaborative status sequence; The multi-station collaborative state sequence is segmented based on a preset time window length to generate long-period sequences and short-period sequences.

[0007] Optionally, the feature construction module specifically comprises: The processing completion signal features, equipment operation status features, buffer occupancy rate features, action execution time window features, and resource occupancy identifier features are extracted from long-period sequences and short-period sequences, respectively. Numerical normalization and vectorization encoding are performed to generate state vectors of the same dimension. The status vectors of the same dimension are concatenated according to the workstation number order to form a multi-workstation status vector matrix; Based on the assembly line process flow, a workstation topology matrix is ​​established, and the workstation topology matrix is ​​bound to the multi-workstation state vector matrix to generate a state representation sequence. An action-level resource occupancy relationship matrix is ​​constructed based on the action execution time window and resource occupancy identifier, and a mutual exclusion constraint matrix is ​​generated based on the resource occupancy conflict relationship. A set of adjustment actions is generated based on the action execution time window, and the set of adjustment actions is associated with and stored in the action-level resource occupancy relationship matrix and the mutual exclusion constraint matrix.

[0008] Optionally, the improved DTW-former model includes a temporal attention layer and a DTW attention layer, specifically: The state representation sequence is represented as a state matrix with a time length of T and a feature dimension of D; In the temporal attention layer, a first weight matrix, a second weight matrix, and a third weight matrix are set. The first weight matrix, the second weight matrix, and the third weight matrix are all trainable parameter matrices. The number of rows in the matrix is ​​equal to the feature dimension D, and the number of columns in the matrix is ​​equal to the attention embedding dimension H. The first weight matrix is ​​used to generate the query matrix, the second weight matrix is ​​used to generate the key matrix, and the third weight matrix is ​​used to generate the value matrix. The query matrix is ​​obtained by multiplying the state matrix and the first weight matrix; the key matrix is ​​obtained by multiplying the state matrix and the second weight matrix; and the value matrix is ​​obtained by multiplying the state matrix and the third weight matrix. The temporal attention weight matrix is ​​obtained by multiplying the query matrix and the key matrix by their transposes, dividing by the square root of H, and then normalizing. The temporal attention output matrix is ​​obtained by multiplying the temporal attention weight matrix and the value matrix. In the DTW attention layer, a dynamic time warping cumulative cost matrix is ​​constructed. The element in the i-th row and j-th column of the cumulative cost matrix is ​​equal to the local cost plus the minimum value among the elements in the (i-1)-th row and j-th column, the i-th row and (j-1)-th column, and the (i-1)-th row and (j-1)-th column. The local cost is equal to the Euclidean distance between the feature vector corresponding to the i-th time index and the feature vector corresponding to the j-th time index. Construct a feasible path region, which is a strip region. The strip width is equal to the integer value obtained by dividing the action execution time window length by the control cycle length. Elements in the cumulative cost matrix that are not in the strip region are assigned to infinity. The time index difference between any adjacent alignment points in the path is limited to not exceeding the strip width, and the time index in the path is located in the time index interval corresponding to the action execution time window. The path confidence is calculated based on the cumulative cost of the path endpoint. The path confidence is obtained after normalization. The gating coefficient is equal to 1 minus the path confidence. The DTW attention weight matrix is ​​obtained by multiplying the gating coefficient element-wise with the similarity matrix constructed based on the dynamic time regularization distance. The DTW attention output matrix is ​​obtained by multiplying the DTW attention weight matrix with the value matrix. The fusion representation matrix is ​​obtained by weighted summation of the temporal attention output matrix and the DTW attention output matrix. The fusion representation matrix is ​​then transformed linearly to generate the beat trend and cache evolution trend.

[0009] Optionally, the construction of the structure learning sample matrix based on the beat trend and cache evolution trend specifically involves: Align the beat trend and cache evolution trend according to the time index to obtain the beat trend vector and cache evolution trend vector corresponding to the same time index. Then, concatenate them according to the workstation number order to generate a trend feature vector. The trend feature vector is continuously truncated according to the preset sample length to form multiple sample segments; Multiple sample segments are superimposed to form a structure learning sample matrix. The rows of the structure learning sample matrix correspond to the sample segment indexes, and the columns correspond to the trend feature dimensions.

[0010] Optionally, the step of learning the workstation causal adjacency matrix using the improved NOTEARS algorithm, performing topological delay joint feasible region projection, and adjusting the projection intensity based on residual consistency to generate a causal graph specifically involves: The workstation causal adjacency matrix is ​​initialized based on the structure learning sample matrix. The workstation causal adjacency matrix is ​​a real number matrix with dimension N multiplied by N, where N represents the number of workstations. The sample fitting matrix is ​​calculated based on the workstation causal adjacency matrix and the structure learning sample matrix. The sample fitting matrix is ​​equal to the matrix multiplication of the structure learning sample matrix and the workstation causal adjacency matrix. The residual matrix is ​​calculated based on the sample fitting matrix and the structure learning sample matrix. The residual matrix is ​​equal to the structure learning sample matrix minus the sample fitting matrix. The workstation causal adjacency matrix is ​​updated based on the residual matrix. The updated matrix elements are equal to the original matrix elements minus the gradient value obtained by multiplying the learning rate by the residual matrix and the transpose of the structure learning sample matrix. Perform topological delay joint feasible region projection on the updated workstation causal adjacency matrix. The projection rules include setting matrix elements that do not satisfy the preceding and following dependencies in the process flow to 0, and setting matrix elements that do not satisfy the preset maximum propagation delay threshold to 0. Calculate the residual consistency index, which is equal to the sum of the squares of the differences between the corresponding elements of the current residual matrix and the previous residual matrix, divided by the total number of matrix elements. When the residual consistency index is greater than the preset consistency threshold, the projection intensity takes the first preset value; when the residual consistency index is less than or equal to the preset consistency threshold, the projection intensity takes the second preset value. The first preset value and the second preset value are different constants. The workstation causal adjacency matrix is ​​scaled based on the projection intensity. The scaled matrix elements are equal to the projection intensity multiplied by the projected matrix elements. A causal graph is generated based on the scaled workstation causal adjacency matrix. The nodes in the causal graph correspond to the workstation numbers, and the positions in the causal graph where the matrix elements are not equal to 0 correspond to directed edges.

[0011] Optionally, the step of calculating the marginal causal gain of the adjustment actions in the set of adjustment actions based on the causal graph, and generating continuous collaborative adjustment coefficients by combining the beat trend and the buffer evolution trend, and constructing a set of collaborative adjustment strategies, specifically involves: Based on the cause-effect graph, the set of working positions corresponding to the adjustment actions in the set of adjustment actions and the set of affected working positions are determined. The working positions are determined by the working position numbers in the set of adjustment actions, and the set of affected working positions is determined by the directed reachable nodes starting from the working positions in the cause-effect graph. An objective function is constructed for the cycle time trend and cache evolution trend of the workstations within the affected workstation set. The objective function consists of a cycle time trend term and a cache evolution trend term. Set an action intensity variable for the action in the set of action to be adjusted. The action intensity variable takes a value range of 0 to 1. The continuous coordinated adjustment coefficient is composed of the action intensity variable. The objective function is weighted based on the directed edge weights from the working station to the affected working station in the causal graph. The marginal causal gain is equal to the weighted value of the objective function when the action intensity variable is 1 minus the weighted value when the action intensity variable is 0. The continuous synergistic adjustment coefficient is assigned a value based on the marginal causal gain. The continuous synergistic adjustment coefficient is equal to the normalized value of the marginal causal gain. The normalized value is equal to the marginal causal gain minus the minimum value of the marginal causal gain and then divided by the maximum value of the marginal causal gain minus the minimum value of the marginal causal gain. A set of coordinated regulation strategies is constructed based on continuous coordinated regulation coefficients and a set of regulatory actions. Each coordinated regulation strategy in the set includes a regulatory action number and a continuous coordinated regulation coefficient.

[0012] Optionally, the step of performing feasibility screening based on action-level resource occupancy relationships and mutual exclusion constraints, and generating a screened collaborative adjustment strategy, specifically includes: The resource occupancy vector corresponding to any collaborative adjustment strategy in the collaborative adjustment strategy set is determined based on the action-level resource occupancy relationship matrix. The resource occupancy vector is determined by the row vector corresponding to the adjustment action number in the action-level resource occupancy relationship matrix. For any two coordinated adjustment strategies in the coordinated adjustment strategy set, a mutual exclusion constraint judgment value is determined. The mutual exclusion constraint judgment value is determined by the matrix elements in the mutual exclusion constraint matrix corresponding to the two adjustment action numbers. When the mutual exclusion constraint judgment value is equal to 1, it is determined that there is a conflict between the two coordinated adjustment strategies; when the mutual exclusion constraint judgment value is equal to 0, it is determined that there is no conflict between the two coordinated adjustment strategies. For conflicting collaborative regulation strategies, compare the continuous collaborative regulation coefficients and perform retention or elimination processing to generate the filtered collaborative regulation strategies.

[0013] Optionally, the instruction mapping module specifically comprises: Based on the adjustment action number, locate the corresponding adjustment action's action execution time window and resource usage identifier in the adjustment action set; Discrete control commands are generated based on the continuous coordinated adjustment coefficient. The discrete control commands include the workstation number, command type, and command amplitude. The workstation number is taken from the set of adjustment actions, the command type is mapped from the resource occupancy identifier, and the command amplitude is mapped from the continuous coordinated adjustment coefficient and the preset command quantization level. Discrete control commands are sorted according to the control cycle timestamp and sent to the corresponding workstation control unit to generate execution results.

[0014] Optionally, the feedback update module specifically comprises: Obtain the execution results and align them according to the control cycle timestamp to construct feedback data; The path feasible region and gating coefficients in the DTW-former model are updated and improved based on feedback data. The path feasible region update includes reestimation of the band width, and the gating coefficient update includes reestimation of the path confidence normalization interval. The projection intensity and workstation causal adjacency matrix in the improved NOTERAS algorithm based on feedback data are updated. The projection intensity update includes reassigning the first preset value and the second preset value, and the workstation causal adjacency matrix update includes reassigning the learning rate and the number of gradient update rounds. The updated path feasible region, gating coefficient, projection intensity, and workstation causal adjacency matrix are input into the collaborative trend prediction and causal learning process of the next control cycle.

[0015] The beneficial effects of this invention are: An improved DTW-former model is introduced into the trend prediction module. Through collaborative modeling of the temporal attention layer and the DTW attention layer, and by constructing a path feasible region and gating coefficient adjustment mechanism in the DTW attention layer, dynamic time alignment and nonlinear modeling of the beat trend and buffer evolution trend are realized, thereby improving the stability and accuracy of trend prediction. In the causal learning module, based on the structure learning sample matrix, the improved NOTEARS algorithm is used to perform topological delay joint feasible region projection and residual consistency adjustment, learn the workstation causal adjacency matrix and generate a causal graph, realize the constrained learning and propagation path characterization of multi-workstation causal relationships, and enhance the interpretability and global consistency of the collaborative adjustment strategy. By combining marginal causal gain calculation, action-level resource occupancy relationship and mutual exclusion constraint screening in the strategy generation module and feedback update module, as well as updating the feasible region of the path, gating coefficient and projection intensity, a closed-loop optimization control structure is constructed to achieve controllable conflict of collaborative adjustment strategy and adaptive parameter update, thereby improving the stability and cycle time consistency of the assembly line operation. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a schematic diagram of a machine learning-based intelligent assembly line collaborative control system proposed in this invention. Figure 2 This is a schematic diagram of the structure of the improved DTW-former model proposed in this invention; Figure 3 This is a schematic diagram of the structure of the improved NOTEARS algorithm proposed in this invention. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0018] refer to Figures 1-3 A machine learning-based intelligent assembly line collaborative control system includes: The data acquisition module is used to acquire multi-station operating data of the assembly line, and perform resampling and time alignment according to the control cycle to form long-cycle sequences and short-cycle sequences. The feature construction module is used to perform heterogeneous feature encoding on long-period sequences and short-period sequences, map discrete event quantities and continuous measurement quantities into state vectors of the same dimension, construct the workstation topology structure according to the assembly line process flow, and construct action-level resource occupancy relationships and mutual exclusion constraints based on the action execution time window and resource occupancy identifier in the multi-workstation operation data, and generate state representation sequences and adjustment action sets. The trend prediction module is used to input the state representation sequence into the improved DTW-former model to perform collaborative trend prediction. The improved DTW-former model includes a temporal attention layer and a DTW attention layer. The path feasible region is constructed in the DTW attention layer, and the gating coefficient is generated according to the path confidence to adjust the DTW attention weights. The output is the beat trend and the cache evolution trend. The causal learning module is used to construct structural learning samples based on the beat trend and cache evolution trend. It learns the workstation causal adjacency matrix through the improved NOTEARS algorithm, performs topological delay joint feasible region projection, and adjusts the projection intensity according to residual consistency to generate a causal graph. The strategy generation module is used to calculate the marginal causal gain of the adjustment action in the adjustment action set based on the causal graph, and generate continuous collaborative adjustment coefficients by combining the beat trend and cache evolution trend, construct a set of collaborative adjustment strategies, and perform feasibility screening based on the action-level resource occupancy relationship and mutual exclusion constraints to generate the screened collaborative adjustment strategies. The instruction mapping module is used to map the selected collaborative adjustment strategies into discrete control instructions, and send them to the corresponding workstation control units for execution, generating execution results; The feedback update module is used to generate feedback data based on the execution results, update the path feasible region and gating coefficients in the improved DTW-former model according to the feedback data, and update the projection intensity and workstation causal adjacency matrix in the improved NOTEARS algorithm to complete the closed-loop optimization of assembly line collaborative control.

[0019] In this embodiment, the data acquisition module specifically comprises: The system collects processing completion signals, equipment operating status signals, buffer occupancy rate data, action execution time window data, resource occupancy identifier data, and control cycle timestamp data from multiple workstations on the assembly line. The processing completion signal is represented by a binary status quantity, the equipment operating status signal is represented by a multi-bit status code, the buffer occupancy rate data is represented by the ratio of the occupied quantity to the buffer capacity, the occupancy rate is equal to the current quantity stored in the buffer divided by the buffer design capacity, the action execution time window is represented by a time interval consisting of the start time and the end time, and the resource occupancy identifier is represented by a two-dimensional identifier consisting of the resource number and the occupancy status. Based on the control cycle timestamp, the processing completion signal, equipment operation status signal, buffer occupancy rate data, action execution time window data, and resource occupancy identifier data are time aligned. Time alignment is achieved through a unified time reference sequence, which is arranged in ascending order according to the control cycle timestamp to form a set of time points. The operation data corresponding to any time point is selected by the minimum time difference matching method. The time difference is equal to the absolute value of the target time point minus the original sampling time point. If the time difference is less than the preset alignment threshold, it is directly matched. If the time difference is greater than the preset alignment threshold, an alignment value is generated by linear interpolation. The linear interpolation value is equal to the previous sampling value plus the time ratio multiplied by the difference between the subsequent sampling value and the previous sampling value. The time ratio is equal to the target time point minus the previous sampling time point and then divided by the subsequent sampling time point minus the previous sampling time point. The running data under a unified time base is resampled according to the preset control cycle length. The resampling constructs an equal time interval running data sequence with the control cycle length as the sampling interval. The time corresponding to the nth data point in the equal time interval running data sequence is the start time plus n multiplied by the control cycle length. The cache occupancy rate data is smoothed by the moving average method during the resampling process. The moving average is equal to the sum of the cache occupancy rates of the current time point and the previous k time points divided by k plus 1, where k is the preset window length. The data sequences running at equal time intervals are concatenated according to the workstation number in order. The concatenation method is vector-level horizontal concatenation. The running data of a single workstation at the same time point constitutes a one-dimensional feature vector. The one-dimensional feature vectors of different workstations at the same time point are connected in ascending order of workstation number to form a multi-workstation feature vector. The multi-workstation feature vectors are arranged over time to form a multi-workstation collaborative state sequence. The multi-workstation collaborative state sequence is a two-dimensional matrix structure, with the matrix rows corresponding to the time index and the matrix columns corresponding to the workstation feature dimensions. The multi-station collaborative state sequence is segmented based on a preset time window length. The time window length corresponds to the long cycle length and the short cycle length. The long cycle length is equal to the control cycle length multiplied by the number of long cycle steps, and the short cycle length is equal to the control cycle length multiplied by the number of short cycle steps. The segmentation method is to continuously extract segments according to the time index order. The starting index of the segmented interval is m multiplied by the step length, and the ending index is the starting index plus the window length minus 1. The step length is the control cycle length or the preset sliding step length. After segmentation, long cycle sequences and short cycle sequences are obtained. The long cycle sequence is used to characterize the rhythm trend change, and the short cycle sequence is used to characterize the instantaneous fluctuation characteristics.

[0020] In this embodiment, the feature construction module specifically comprises: For long-period sequences and short-period sequences, features of processing completion signal, equipment operating status, cache occupancy rate, action execution time window, and resource occupancy identifier are extracted respectively. The processing completion signal feature is represented by a binary vector, the equipment operating status feature is represented by a multi-bit status encoding vector, the cache occupancy rate feature is represented by a real number ratio, which is equal to the current cache quantity divided by the cache design capacity, the action execution time window feature is represented by the time difference between the start time and the end time, which is equal to the end time minus the start time, and the resource occupancy identifier feature is represented by a discrete encoding vector composed of the resource number and the occupancy status. Numerical normalization is performed on the features of the processing completion signal, equipment operating status, buffer occupancy rate, action execution time window, and resource occupancy identifier. The normalized feature value is equal to the original feature value minus the minimum value of the feature in the current sequence and then divided by the maximum value of the feature in the current sequence minus the minimum value, resulting in a normalized feature vector with a value range between 0 and 1. The normalized feature vector is vectorized and encoded, converting discrete encoded features into fixed-length vectors. Continuous features are kept as real-number vectors and dimension alignment is performed. Dimension alignment is done by padding with zeros. The length of the zero-padded vector is equal to the preset maximum feature dimension minus the current feature dimension, resulting in a state vector of the same dimension. The same-dimensional state vectors are concatenated according to the workstation number order. The concatenation method is that the time dimension remains unchanged and the feature dimension is connected horizontally. That is, the same-dimensional state vectors of different workstations under the same time index are connected in order according to the workstation number from small to large to form a multi-workstation state vector corresponding to a single time index. The multi-workstation state vectors are arranged in time order to form a multi-workstation state vector matrix. The matrix rows represent the time index and the matrix columns represent the concatenated feature dimensions. Based on the assembly line process flow, a workstation topology matrix is ​​established. The workstation topology matrix is ​​represented in the form of an adjacency matrix. A matrix element equal to 1 indicates the existence of material flow or control dependencies, while a matrix element equal to 0 indicates the absence of dependencies. The workstation topology matrix is ​​bound to a multi-workstation state vector matrix. The binding method is to introduce a topology mask matrix into the multi-workstation state vector matrix. The topology mask matrix is ​​obtained by extending the workstation topology matrix. The feature connections with dependencies are retained by matrix element-wise multiplication, and a state representation sequence is generated. An action-level resource occupancy relationship matrix is ​​constructed based on the action execution time window and resource occupancy identifier. The action-level resource occupancy relationship matrix is ​​represented in two-dimensional matrix form, where the matrix rows represent adjustment action numbers and the matrix columns represent resource numbers. A matrix element equal to 1 indicates that the adjustment action occupies the corresponding resource, and a matrix element equal to 0 indicates that no resource is occupied. A mutual exclusion constraint matrix is ​​generated based on the resource occupancy conflict relationship. The mutual exclusion constraint matrix is ​​represented in symmetric matrix form, where a matrix element equal to 1 indicates that there is a resource conflict relationship between two corresponding adjustment actions, and a matrix element equal to 0 indicates that there is no conflict relationship. An adjustment action set is generated based on the action execution time window. Each adjustment action in the adjustment action set consists of a workstation number, action start time, action end time, and resource occupancy number. The adjustment action set is generated by traversing the action execution time window data. The traversal condition is that the time window length is greater than 0. The generated adjustment action set is associated and stored with the action-level resource occupancy relationship matrix and the mutual exclusion constraint matrix. The association method is to use the adjustment action number as an index, and at the same time, the row index of the corresponding resource occupancy relationship matrix and the row and column index of the mutual exclusion constraint matrix.

[0021] In this embodiment, the improved DTW-former model includes a temporal attention layer and a DTW attention layer, specifically: The state representation sequence is represented as a state matrix, where the number of rows in the state matrix is ​​equal to the time length T, the number of columns is equal to the feature dimension D, and the element in the i-th row and k-th column of the state matrix represents the k-th eigenvalue of the state vector corresponding to the i-th time index. In the temporal attention layer, a first weight matrix, a second weight matrix, and a third weight matrix are set. The first weight matrix, the second weight matrix, and the third weight matrix are all real-number trainable parameter matrices. The number of rows in the matrix is ​​equal to the feature dimension D, and the number of columns in the matrix is ​​equal to the attention embedding dimension H. The first weight matrix is ​​used to map the state matrix to a query matrix, the second weight matrix is ​​used to map the state matrix to a key matrix, and the third weight matrix is ​​used to map the state matrix to a value matrix. The query matrix is ​​obtained by multiplying the state matrix and the first weight matrix. The key matrix is ​​obtained by multiplying the state matrix and the second weight matrix. The value matrix is ​​obtained by multiplying the state matrix and the third weight matrix. The temporal attention weight matrix is ​​obtained by matrix multiplication of the transpose of the query matrix and the key matrix. The elements of the intermediate matrix are equal to the dot product of the i-th row vector of the query matrix and the j-th row vector of the key matrix, and then divided by the square root of H to obtain the scaling matrix. The scaling matrix is ​​normalized to obtain the temporal attention weight matrix. The normalization method is to subtract the maximum value of each row element, take the exponent, and then divide by the sum of the exponents of the row. The temporal attention output matrix is ​​obtained by matrix multiplication of the temporal attention weight matrix and the value matrix. In the DTW attention layer, a dynamic time warping cumulative cost matrix is ​​constructed. The cumulative cost matrix has a dimension of T multiplied by T. The element in the i-th row and j-th column of the cumulative cost matrix is ​​obtained by recursion. The recursive formula is that the current element is equal to the local cost plus the minimum value among the three preceding terms. The local cost is equal to the square root of the sum of the squares of the differences between the corresponding elements of the i-th row vector and the j-th row vector of the state matrix. The three preceding terms include the element in the (i-1)-th row and j-th column, the element in the i-th row and (j-1)-th column, and the element in the (i-1)-th row and (j-1)-th column. Construct a feasible path region. The feasible path region adopts a strip region constraint method. The strip width is equal to the integer value obtained by dividing the action execution time window length by the control cycle length. Elements in the feasible path region satisfy that the absolute value i minus j is less than or equal to the strip width. Positions that do not satisfy this condition are assigned positive infinity in the cumulative cost matrix. At the same time, the time index of any alignment point in the path is limited to be within the time index interval corresponding to the action execution time window. The path confidence is calculated based on the cumulative cost of the path endpoint. The path confidence is equal to the cumulative cost of the path endpoint minus the minimum value in the cumulative cost matrix, then divided by the maximum value minus the minimum value. The gating coefficient is equal to 1 minus the path confidence. A similarity matrix is ​​constructed, and the elements of the similarity matrix are equal to the negative local cost and then normalized. The DTW attention weight matrix is ​​equal to the element-wise multiplication of the gating coefficient and the similarity matrix. The DTW attention output matrix is ​​equal to the matrix multiplication of the DTW attention weight matrix and the value matrix. The temporal attention output matrix and the DTW attention output matrix are weighted and summed according to the corresponding element positions to obtain the fusion representation matrix. The weighting coefficients are the preset constants α and 1 minus α. The fusion representation matrix is ​​linearly transformed to generate the beat trend and buffer evolution trend. The linear transformation is equal to matrix multiplication of the fusion representation matrix and the output weight matrix.

[0022] In this embodiment, the multi-station operation data of the assembly line has the characteristics of control cycle discreteness, action execution time window constraint, and resource occupation conflict coupling. The traditional DTW-former model only performs dynamic time warping calculation based on time series similarity, without considering the physical constraints of the control cycle and action execution time window on the path alignment range. This may cause the path alignment result to deviate from the actual process cycle propagation path, thus affecting the prediction accuracy of cycle time trend and buffer evolution trend. Based on this, a path feasible region constraint is introduced into the DTW attention layer, and the band width is set as the ratio of the action execution time window length to the control cycle length, and is included in the cumulative cost matrix. Positions that do not meet the time index difference constraint and action execution time window constraint are assigned an infinite value, so that the dynamic time warping process is recursively applied only within the region that meets the process time propagation condition. At the same time, path confidence is constructed and gating coefficients are generated. The gating coefficients are multiplied element-wise with the dynamic time warping similarity matrix to adjust the DTW attention weights, thereby suppressing abnormal alignment paths. By weighted and fused the temporal attention output and the DTW attention output under the path feasible region constraint, a fusion representation that simultaneously characterizes the evolution of long-cycle beats and short-cycle local disturbances is formed, thereby improving the ability of collaborative trend prediction to characterize the cumulative effect of workstation beat transmission delay and buffer fluctuation.

[0023] In this implementation, a structure learning sample matrix is ​​constructed based on the beat trend and cache evolution trend, specifically as follows: The beat trend and cache evolution trend are aligned one-to-one according to the time index. When the time index difference is 0, it is considered as data at the same time. The beat trend vector and cache evolution trend vector corresponding to the same time index are obtained. The beat trend vector and cache evolution trend vector corresponding to the same time index are horizontally concatenated in ascending order of workstation number. The concatenation method is vector-level feature dimension connection, that is, the new vector is equal to the beat trend vector followed by the cache evolution trend vector to generate a trend feature vector. The trend feature vector dimension is equal to the beat trend dimension plus the cache evolution trend dimension. The trend feature vector sequence is truncated by sliding according to the preset sample length L. The sliding step size is a preset integer. A single sample segment is formed by L consecutive trend feature vectors arranged in chronological order. A single sample segment is represented as a matrix with the number of rows equal to L and the number of columns equal to the trend feature dimension. Multiple sample segments are stacked vertically along the sample index direction. The stacking method is matrix row-level concatenation, that is, the matrix of the kth sample segment is appended as a whole below the matrix of the (k-1)th segment to form a structure learning sample matrix. The number of rows in the structure learning sample matrix is ​​equal to the number of sample segments multiplied by L, and the number of columns is equal to the trend feature dimension.

[0024] In this embodiment, the workstation causal adjacency matrix is ​​learned using the improved NOTEARS algorithm, topological delay joint feasible region projection is performed, and the projection intensity is adjusted based on residual consistency to generate a causal graph, specifically: The initial causal adjacency matrix of the workstations is an N-times N real number matrix, where N equals the number of workstations, and the initial values ​​of the matrix elements are 0 or small-range random real numbers. A linear structural equation model is constructed to represent the relationship between the structural learning sample matrix and the workstation causal adjacency matrix. The sample fitting matrix is ​​obtained by matrix multiplication of the structural learning sample matrix and the workstation causal adjacency matrix. The element in the i-th row and j-th column of the sample fitting matrix is ​​equal to the inner product of the vector in the i-th row of the structural learning sample matrix and the vector in the j-th column of the workstation causal adjacency matrix. The residual matrix is ​​equal to the structure learning sample matrix minus the sample fitting matrix, and the elements of the residual matrix are equal to the true value at the corresponding position minus the fitted value. Construct a loss function that equals the sum of squares of all elements in the residual matrix divided by the number of rows in the structure learning sample matrix, plus the sum of the absolute values ​​of the elements in the workstation causal adjacency matrix multiplied by the regularization coefficient. The gradient matrix is ​​obtained by taking the partial derivative of the loss function with respect to the workstation causal adjacency matrix. The elements of the gradient matrix are equal to the corresponding element values ​​of the residual matrix transpose multiplied by the structure learning sample matrix, plus the regularization coefficient sign term. The rule for updating the workstation causal adjacency matrix is ​​that the updated element is equal to the element before the update minus the learning rate multiplied by the corresponding element of the gradient matrix; The joint feasible region projection of topology and time delay includes topology constraints and propagation time delay constraints. The topology constraints are given by the workstation topology matrix. When the corresponding element in the workstation topology matrix is ​​equal to 0, the corresponding element in the workstation causal adjacency matrix is ​​assigned a value of 0. The propagation time delay constraints are achieved by calculating the shortest path time difference between workstations. The shortest path time difference is equal to the sum of the processing times of each workstation on the path. When the time difference is greater than the preset maximum propagation time delay threshold, the corresponding element in the workstation causal adjacency matrix is ​​assigned a value of 0. The residual consistency index is equal to the sum of the squared differences between the corresponding elements of the current residual matrix and the previous residual matrix, divided by the total number of matrix elements. The projection intensity is a preset constant γ. When the residual consistency index is greater than the consistency threshold, γ takes the first constant value. When the residual consistency index is less than or equal to the consistency threshold, γ takes the second constant value. The projected workstation causal adjacency matrix is ​​scaled by multiplying each element of the matrix by γ to obtain the final workstation causal adjacency matrix. A causal graph is constructed based on the final workstation causal adjacency matrix. Nodes in the causal graph represent workstation numbers, and non-zero elements in the matrix correspond to directed edges with edge weights equal to the corresponding matrix elements.

[0025] In this embodiment, the improved NOTEARS algorithm is used to learn the workstation causal adjacency matrix and construct a causal graph, establishing a clear mathematical mapping between the structure learning sample matrix and the causal relationships between workstations. The workstation causal adjacency matrix is ​​initialized as an N-by-N real matrix, and the sample fitting matrix and residual matrix are calculated using a linear structural equation model, making the causal strength estimation computable and iteratively optimized. A loss function containing residual squared terms and regularization terms is constructed to effectively suppress excessively large causal weights and avoid overfitting of the causal adjacency matrix. Furthermore, the gradient matrix is ​​obtained by taking the partial derivative of the loss function and updated according to the learning rate, achieving continuous convergence optimization of the causal weights. Further, a topological time delay connection is introduced. By projecting the feasible region, the workstation topology matrix and the shortest path time difference constraint are embedded into the causal adjacency matrix update process. Causal connections that do not meet the process flow dependency or exceed the preset maximum propagation delay threshold are forced to zero, thereby ensuring that the causal graph structure conforms to the actual assembly line physical and process constraints. At the same time, the projection intensity γ is adjusted by the residual consistency index, so that the projected workstation causal adjacency matrix is ​​adaptively scaled according to the model convergence degree, reducing the structural drift risk in the oscillation stage and improving the causal structure preservation capability in the stable stage. Finally, a causal graph with clear edge weights, clear direction and satisfying topology and time delay constraints is generated, providing a reliable structural foundation for the subsequent strategy generation module to calculate the marginal causal gain and build a set of coordinated adjustment strategies.

[0026] In this embodiment, the marginal causal gain of the adjustment action in the adjustment action set is calculated based on the causal graph, and a continuous collaborative adjustment coefficient is generated by combining the beat trend and the buffer evolution trend to construct a collaborative adjustment strategy set, specifically: Based on the cause-effect graph, the set of working positions corresponding to the adjustment actions in the set of adjustment actions and the set of affected working positions are determined. The working positions are determined by the working position numbers in the set of adjustment actions, and the set of affected working positions is determined by the directed reachable nodes starting from the working positions in the cause-effect graph. An objective function is constructed for the cycle time trend and cache evolution trend of the workstations within the affected workstation set. The objective function consists of a cycle time trend term and a cache evolution trend term. Set an action intensity variable for the action in the set of action to be adjusted. The action intensity variable takes a value range of 0 to 1. The continuous coordinated adjustment coefficient is composed of the action intensity variable. The objective function is weighted based on the directed edge weights from the working station to the affected working station in the causal graph. The marginal causal gain is equal to the weighted value of the objective function when the action intensity variable is 1 minus the weighted value when the action intensity variable is 0. The continuous synergistic adjustment coefficient is assigned a value based on the marginal causal gain. The continuous synergistic adjustment coefficient is equal to the normalized value of the marginal causal gain. The normalized value is equal to the marginal causal gain minus the minimum value of the marginal causal gain and then divided by the maximum value of the marginal causal gain minus the minimum value of the marginal causal gain. A set of coordinated regulation strategies is constructed based on continuous coordinated regulation coefficients and a set of regulatory actions. Each coordinated regulation strategy in the set includes a regulatory action number and a continuous coordinated regulation coefficient.

[0027] In this implementation, a feasibility screening is performed based on the action-level resource occupancy relationship and mutual exclusion constraints to generate a screened collaborative adjustment strategy, specifically: The action-level resource occupancy relationship matrix is ​​a binary matrix with dimension M multiplied by R, where M represents the number of adjustment actions in the set of adjustment actions, and R represents the number of resources. An element in the i-th row and j-th column of the matrix equal to 1 indicates that the i-th adjustment action occupies the j-th resource, and equal to 0 indicates that it does not occupy any resources. A single coordinated regulation strategy in the set of coordinated regulation strategies consists of a regulation action number and a continuous coordinated regulation coefficient, which is a real number between 0 and 1. Based on the action-level resource occupancy relationship matrix, the resource occupancy vector corresponding to any coordinated adjustment strategy is determined. The resource occupancy vector is equal to the row vector of the corresponding adjustment action number in the action-level resource occupancy relationship matrix. The position of the element with a value of 1 in the resource occupancy vector indicates that the corresponding resource is occupied. The mutual exclusion constraint matrix is ​​a symmetric binary matrix of dimension M multiplied by M. An element in the i-th row and j-th column equal to 1 indicates that there is a resource conflict between the i-th adjustment action and the j-th adjustment action, and equal to 0 indicates that there is no conflict. The total number of combinations of any two coordinated regulation strategies in the coordinated regulation strategy set is equal to the number of coordinated regulation strategies multiplied by the number of coordinated regulation strategies minus 1 and then divided by 2. For any two coordinated adjustment strategies, the adjustment action numbers i and j are read from the mutual exclusion constraint matrix as the mutual exclusion constraint judgment value. When the mutual exclusion constraint judgment value is equal to 1, it is determined that there is a conflict between the two coordinated adjustment strategies; when the mutual exclusion constraint judgment value is equal to 0, it is determined that there is no conflict. For conflicting collaborative regulation strategies, compare the continuous collaborative regulation coefficients, denoted as αi and αj. When αi is greater than αj, retain the collaborative regulation strategy with regulation action number i and remove the collaborative regulation strategy with regulation action number j. When αi is less than or equal to αj, retain the collaborative regulation strategy with regulation action number j and remove the collaborative regulation strategy with regulation action number i. After all conflict pairs have been screened, the remaining cooperative adjustment strategies constitute the set of cooperative adjustment strategies after screening. In the set of cooperative adjustment strategies after screening, the corresponding elements of any two cooperative adjustment strategies in the mutual exclusion constraint matrix are all equal to 0.

[0028] In this embodiment, the instruction mapping module specifically comprises: Retrieve the corresponding adjustment action record from the adjustment action set based on the adjustment action number. The adjustment action record includes the workstation number, the start time of the action execution time window, the end time of the action execution time window, and the resource occupation identifier. The workstation number is an integer number, and the resource occupation identifier is a discrete code corresponding to the resource number. The continuous coordinated adjustment coefficient is denoted as α, and the value of α ranges from 0 to 1. Let the preset instruction quantization level be Q, where Q is an integer greater than 1. The instruction amplitude level k is equal to the integer obtained by multiplying α by Q and rounding down. The instruction amplitude of the discrete control instruction is equal to k divided by Q and multiplied by the maximum allowable control amplitude. The maximum allowable control amplitude is the maximum adjustment amount allowed by the corresponding workstation control unit. The instruction type is determined by the resource occupancy identifier and control unit type mapping table. When the resource occupancy identifier is a certain resource number, the instruction type is mapped to the corresponding resource control type code. Discrete control instructions are represented as triples, where the first element of the triple is the workstation number, the second element is the instruction type code, and the third element is the instruction amplitude. The effective interval of the instruction is determined based on the start and end times of the action execution time window. The control cycle index n within the effective interval of the instruction satisfies that the control cycle timestamp is greater than or equal to the start time and less than or equal to the end time. Multiple discrete control commands are sorted in ascending order according to the control cycle timestamp. The sorting method is to compare the values ​​of the control cycle timestamps and put the command with the smaller timestamp first. The sorted discrete control commands are sent to the corresponding workstation control unit via fieldbus or industrial Ethernet. The workstation control unit adjusts the actuator according to the type and amplitude of the received command and generates the execution result, which includes the actual execution time, the actual execution amplitude, and the execution completion indicator.

[0029] In this embodiment, the feedback update module specifically includes: The execution results are obtained and matched with the corresponding adjustment action number according to the control cycle timestamp. The matching condition is that the workstation number in the execution result is consistent with the workstation number corresponding to the adjustment action number and the execution time falls within the corresponding action execution time window. The matched data forms feedback data. Feedback data is represented as a matrix, with rows corresponding to time indices and columns corresponding to workstation numbers. Each matrix element equals the actual execution amplitude minus the instruction amplitude in the discrete control command. The band width in the feasible path region is denoted as W. The actual alignment error is calculated in the feedback data. The actual alignment error is equal to the average absolute value of the difference between the predicted beat trend value and the actual beat value in the execution result. The updated band width value is equal to the original band width multiplied by one plus the error ratio coefficient multiplied by the actual alignment error. The error ratio coefficient is a preset constant. The path confidence normalization interval is determined by the minimum cumulative value and the maximum cumulative value. The new minimum and maximum cumulative values ​​are calculated from the feedback data. The minimum cumulative value is equal to the minimum value among the cumulative values ​​of all path endpoints in the current period, and the maximum cumulative value is equal to the maximum value among the cumulative values ​​of all path endpoints in the current period. The normalization interval is updated by replacement. The projection intensity is denoted as γ. The residual consistency index is calculated in the feedback data. The residual consistency index is equal to the sum of the squares of the differences between the corresponding elements in the current row and the previous row of the feedback data matrix and then divided by the number of elements. When the residual consistency index is greater than the consistency threshold, γ is equal to the first preset value. When the residual consistency index is less than or equal to the consistency threshold, γ is equal to the second preset value. The workstation causal adjacency matrix update includes recalculating the gradient update round number K, where K is equal to the preset maximum round number minus the residual consistency index multiplied by the proportional constant rounded down, and updating the learning rate η, where η is equal to the original learning rate multiplied by 1 minus the residual consistency index. The updated strip width, path confidence normalization interval, projection intensity γ, learning rate η, and gradient update round number K are input into the collaborative trend prediction and causal learning calculation process of the next control cycle.

[0030] Example 1: To verify the feasibility of the present invention in practice, it was applied to an intelligent assembly line for automotive parts. The assembly line has 6 workstations, which sequentially complete material loading and positioning, bolt tightening, pressing, visual inspection, sorting after completion, and buffer transfer. On-site issues include shared use of robots and tightening guns, AGVs occupying aisles, and fluctuations in buffer capacity. Production cycle time is prone to short-term congestion and resource conflicts when changing models or experiencing slight equipment fluctuations. This manifests as the buffer occupancy rate fluctuating repeatedly between 30% and 95%. The bolt tightening workstation and the pressing workstation compete for air supply and fixture resources at the same time, causing mutual exclusion conflicts and waiting. The control cycle is set to 1 second, the long-cycle sequence time window length is set to 300 seconds, and the short-cycle sequence time window length is set to 30 seconds. The data acquisition module collects processing completion signals, equipment operating status signals, buffer occupancy rate data, action execution time windows, resource occupancy identifiers, and control cycle timestamps, and performs resampling and time alignment. The feature construction module performs heterogeneous feature encoding on discrete event quantities and continuous measurement quantities, generating a state representation sequence and a set of adjustment actions. The set of adjustment actions includes four types of adjustment actions: "tightening cycle increment," "pressing waiting delay," "AGV release interval," and "buffer release threshold." The execution time window for each type of action is set to range from 10 seconds to 60 seconds, and is bound to a resource occupancy identifier. The trend prediction module inputs the state representation sequence into the improved DTW-former model and constructs a trend prediction module at the DTW attention layer. The feasible path region is determined by adjusting the DTW attention weights through gating coefficients, outputting the 30-second clock trend and cache evolution trend. The causal learning module constructs a structural learning sample matrix based on the clock trend and cache evolution trend, learns the workstation causal adjacency matrix through the improved NOTEARS algorithm, and generates a causal graph. The strategy generation module calculates the marginal causal gain of the adjustment action based on the causal graph, generates continuous collaborative adjustment coefficients by combining the clock trend and cache evolution trend, constructs a set of collaborative adjustment strategies, and completes feasibility screening based on action-level resource occupancy relationships and mutual exclusion constraints. The instruction mapping module maps the screened collaborative adjustment strategies into discrete control instructions and sends them to the workstation control unit. The feedback update module generates feedback data based on the execution results, updates the feasible path region, gating coefficients, projection intensity, and workstation causal adjacency matrix, forming a closed-loop optimization.

[0031] To quantify the comparison effect, data from two working days of the same shift on the same production line were selected. The traditional system uses a fixed cycle time and manual rule mutual exclusion processing. The system of this invention enables trend prediction, causal learning and strategy generation closed-loop control. The results are summarized in Tables 1 and 2.

[0032] Table 1. Comparison of Collaborative Trend Prediction and Cache Evolution Prediction Data

[0033] Analysis of Table 1 shows that the cycle time trend and cache evolution trend require more dynamic time alignment capabilities under short-term non-stationary fluctuations. Traditional systems exhibit significant error accumulation during model changeovers and AGV delays, with the cycle time trend MAE reaching 0.62s and the cache evolution trend RMSE reaching 13.4%. The system of this invention introduces a path feasible region in the DTW attention layer, restricting the aligned path from falling within the corresponding interval of the action execution time window. Furthermore, it suppresses the disturbance of low-confidence paths on the DTW attention weights through gating coefficients, resulting in a synchronous decrease in prediction error. The cycle time trend RMSE decreases from 0.88s to 0.45s, and the cache evolution trend MAE decreases from 9.6% to 4.2%. The median congestion lead time increases from 7s to 18s, indicating enhanced predictability of rapid increases in cache occupancy. The prediction stability variance is smaller, and prediction jitter converges within continuous control cycles, facilitating the acquisition of a more consistent structure learning sample matrix in subsequent causal learning and policy generation stages.

[0034] Table 2 Comparison of Coordinated Adjustment Strategies and Implementation Results

[0035] Analysis of Table 2 shows that the traditional system relies on fixed cycle times and manual rules for mutually exclusive processing. When encountering short-term congestion, it often adopts an "overall slowdown" approach, causing the average cycle time to rise to 8.9 seconds per item with greater fluctuations, the number of mutual exclusion conflicts to reach 27 times per shift, the cumulative waiting time to 41 minutes per shift, and the peak cache occupancy rate to approach 96%, accompanied by the risk of overflow. The system of this invention, in the causal learning module, generates a causal graph by combining topological delay with feasible region projection constraints on the causal adjacency matrix of workstations. Then, in the strategy generation module, it allocates continuous collaborative adjustment coefficients according to marginal causal gain and combines action-level resource occupancy relationships with mutual exclusion constraints to select collaborative adjustment strategies. The number of mutual exclusion conflicts is reduced to 9 times per shift, the waiting time is reduced to 14 minutes per shift, and the peak cache occupancy rate is reduced to 88%. Because the instruction mapping module maps the continuous coordinated adjustment coefficients into discrete control instructions and takes effect within the action execution time window, the adjustment actions are more in line with the available resource periods. The average cycle time drops to 8.2s / piece and the standard deviation converges to 0.9s. Unplanned downtime is reduced simultaneously, demonstrating the supporting effect of closed-loop optimization on the stability and cycle time consistency of the assembly line.

[0036] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A machine learning-based intelligent assembly line collaborative control system, characterized in that, include: The data acquisition module is used to acquire multi-station operating data of the assembly line, and perform resampling and time alignment according to the control cycle to form long-cycle sequences and short-cycle sequences. The feature construction module is used to perform heterogeneous feature encoding on long-period sequences and short-period sequences, map discrete event quantities and continuous measurement quantities into state vectors of the same dimension, construct the workstation topology structure according to the assembly line process flow, and construct action-level resource occupancy relationships and mutual exclusion constraints based on the action execution time window and resource occupancy identifier in the multi-workstation operation data, and generate state representation sequences and adjustment action sets. The trend prediction module is used to input the state representation sequence into the improved DTW-former model to perform collaborative trend prediction. The improved DTW-former model includes a temporal attention layer and a DTW attention layer. In the DTW attention layer, a path feasible region is constructed, and gating coefficients are generated according to the path confidence to adjust the DTW attention weights. The output is the beat trend and the cache evolution trend. The causal learning module is used to construct a structural learning sample matrix based on the beat trend and cache evolution trend, learn the workstation causal adjacency matrix through the improved NOTEARS algorithm, perform topological delay joint feasible region projection, and adjust the projection intensity according to residual consistency to generate a causal graph. The strategy generation module is used to calculate the marginal causal gain of the adjustment actions in the set of adjustment actions based on the causal graph, and generate continuous coordinated adjustment coefficients by combining the beat trend and the cache evolution trend, construct a set of coordinated adjustment strategies, and perform feasibility screening based on the action-level resource occupancy relationship and mutual exclusion constraints to generate the screened coordinated adjustment strategies. The instruction mapping module is used to map the selected collaborative adjustment strategies into discrete control instructions, and send them to the corresponding workstation control units for execution, generating execution results; The feedback update module is used to generate feedback data based on the execution results, update the path feasible region and gating coefficient in the improved DTW-former model according to the feedback data, and update the projection intensity and workstation causal adjacency matrix in the improved NOTEARS algorithm to complete the closed-loop optimization of assembly line collaborative control.

2. The intelligent assembly line collaborative control system based on machine learning according to claim 1, characterized in that, The data acquisition module is specifically: Collect processing completion signals, equipment operating status signals, buffer occupancy rate data, action execution time windows, resource occupancy indicators, and control cycle timestamps from multiple workstations on the assembly line; Based on the control cycle timestamp, the processing completion signal, equipment operating status signal, buffer occupancy rate data, action execution time window, and resource occupancy identifier execution time are aligned to obtain operating data under a unified time reference. According to the preset control cycle length, the running data under the unified time base is resampled to generate a running data sequence with equal time intervals; The operation data sequences with equal time intervals are spliced ​​together according to the workstation number order to construct a multi-workstation collaborative status sequence; The multi-station collaborative state sequence is segmented based on a preset time window length to generate long-period sequences and short-period sequences.

3. The intelligent assembly line collaborative control system based on machine learning according to claim 1, characterized in that, The feature construction module is specifically as follows: The processing completion signal features, equipment operation status features, buffer occupancy rate features, action execution time window features, and resource occupancy identifier features are extracted from long-period sequences and short-period sequences, respectively. Numerical normalization and vectorization encoding are performed to generate state vectors of the same dimension. The status vectors of the same dimension are concatenated according to the workstation number order to form a multi-workstation status vector matrix; Based on the assembly line process flow, a workstation topology matrix is ​​established, and the workstation topology matrix is ​​bound to the multi-workstation state vector matrix to generate a state representation sequence. An action-level resource occupancy relationship matrix is ​​constructed based on the action execution time window and resource occupancy identifier, and a mutual exclusion constraint matrix is ​​generated based on the resource occupancy conflict relationship. A set of adjustment actions is generated based on the action execution time window, and the set of adjustment actions is associated with and stored in the action-level resource occupancy relationship matrix and the mutual exclusion constraint matrix.

4. The intelligent assembly line collaborative control system based on machine learning according to claim 1, characterized in that, The improved DTW-former model includes a temporal attention layer and a DTW attention layer, specifically: The state representation sequence is represented as a state matrix with a time length of T and a feature dimension of D; In the temporal attention layer, a first weight matrix, a second weight matrix, and a third weight matrix are set. The first weight matrix, the second weight matrix, and the third weight matrix are all trainable parameter matrices. The number of rows in the matrix is ​​equal to the feature dimension D, and the number of columns in the matrix is ​​equal to the attention embedding dimension H. The first weight matrix is ​​used to generate the query matrix, the second weight matrix is ​​used to generate the key matrix, and the third weight matrix is ​​used to generate the value matrix. The query matrix is ​​obtained by multiplying the state matrix and the first weight matrix; the key matrix is ​​obtained by multiplying the state matrix and the second weight matrix; and the value matrix is ​​obtained by multiplying the state matrix and the third weight matrix. The temporal attention weight matrix is ​​obtained by multiplying the query matrix and the key matrix by their transposes, dividing by the square root of H, and then normalizing. The temporal attention output matrix is ​​obtained by multiplying the temporal attention weight matrix and the value matrix. In the DTW attention layer, a dynamic time warping cumulative cost matrix is ​​constructed. The element in the i-th row and j-th column of the cumulative cost matrix is ​​equal to the local cost plus the minimum value among the elements in the (i-1)-th row and j-th column, the i-th row and (j-1)-th column, and the (i-1)-th row and (j-1)-th column. The local cost is equal to the Euclidean distance between the feature vector corresponding to the i-th time index and the feature vector corresponding to the j-th time index. Construct a feasible path region, which is a strip region. The strip width is equal to the integer value obtained by dividing the action execution time window length by the control cycle length. Elements in the cumulative cost matrix that are not in the strip region are assigned to infinity. The time index difference between any adjacent alignment points in the path is limited to not exceeding the strip width, and the time index in the path is located in the time index interval corresponding to the action execution time window. The path confidence is calculated based on the cumulative cost of the path endpoint. The path confidence is obtained after normalization. The gating coefficient is equal to 1 minus the path confidence. The DTW attention weight matrix is ​​obtained by multiplying the gating coefficient element-wise with the similarity matrix constructed based on the dynamic time regularization distance. The DTW attention output matrix is ​​obtained by multiplying the DTW attention weight matrix with the value matrix. The fusion representation matrix is ​​obtained by weighted summation of the temporal attention output matrix and the DTW attention output matrix. The fusion representation matrix is ​​then transformed linearly to generate the beat trend and cache evolution trend.

5. The intelligent assembly line collaborative control system based on machine learning according to claim 1, characterized in that, The construction of the structure learning sample matrix based on the beat trend and cache evolution trend is as follows: Align the beat trend and cache evolution trend according to the time index to obtain the beat trend vector and cache evolution trend vector corresponding to the same time index. Then, concatenate them according to the workstation number order to generate a trend feature vector. The trend feature vector is continuously truncated according to the preset sample length to form multiple sample segments; Multiple sample segments are superimposed to form a structure learning sample matrix. The rows of the structure learning sample matrix correspond to the sample segment indexes, and the columns correspond to the trend feature dimensions.

6. The intelligent assembly line collaborative control system based on machine learning according to claim 1, characterized in that, The process involves learning the workstation causal adjacency matrix using the improved NOTEARS algorithm, performing topological delay joint feasible region projection, and adjusting the projection intensity based on residual consistency to generate a causal graph. Specifically: The workstation causal adjacency matrix is ​​initialized based on the structure learning sample matrix. The workstation causal adjacency matrix is ​​a real number matrix with dimension N multiplied by N, where N represents the number of workstations. The sample fitting matrix is ​​calculated based on the workstation causal adjacency matrix and the structure learning sample matrix. The sample fitting matrix is ​​equal to the matrix multiplication of the structure learning sample matrix and the workstation causal adjacency matrix. The residual matrix is ​​calculated based on the sample fitting matrix and the structure learning sample matrix. The residual matrix is ​​equal to the structure learning sample matrix minus the sample fitting matrix. The workstation causal adjacency matrix is ​​updated based on the residual matrix. The updated matrix elements are equal to the original matrix elements minus the gradient value obtained by multiplying the learning rate by the residual matrix and the transpose of the structure learning sample matrix. Perform topological delay joint feasible region projection on the updated workstation causal adjacency matrix. The projection rules include setting matrix elements that do not satisfy the preceding and following dependencies in the process flow to 0, and setting matrix elements that do not satisfy the preset maximum propagation delay threshold to 0. Calculate the residual consistency index, which is equal to the sum of the squares of the differences between the corresponding elements of the current residual matrix and the previous residual matrix, divided by the total number of matrix elements. When the residual consistency index is greater than the preset consistency threshold, the projection intensity takes the first preset value; when the residual consistency index is less than or equal to the preset consistency threshold, the projection intensity takes the second preset value. The first preset value and the second preset value are different constants. The workstation causal adjacency matrix is ​​scaled based on the projection intensity. The scaled matrix elements are equal to the projection intensity multiplied by the projected matrix elements. A causal graph is generated based on the scaled workstation causal adjacency matrix. The nodes in the causal graph correspond to the workstation numbers, and the positions in the causal graph where the matrix elements are not equal to 0 correspond to directed edges.

7. The intelligent assembly line collaborative control system based on machine learning according to claim 1, characterized in that, The method involves calculating the marginal causal gain of the adjustment actions in the set of adjustment actions based on the causal graph, and generating continuous collaborative adjustment coefficients by combining the beat trend and buffer evolution trend, thereby constructing a set of collaborative adjustment strategies. Specifically: Based on the cause-effect graph, the set of working positions corresponding to the adjustment actions in the set of adjustment actions and the set of affected working positions are determined. The working positions are determined by the working position numbers in the set of adjustment actions, and the set of affected working positions is determined by the directed reachable nodes starting from the working positions in the cause-effect graph. An objective function is constructed for the cycle time trend and cache evolution trend of the workstations within the affected workstation set. The objective function consists of a cycle time trend term and a cache evolution trend term. Set an action intensity variable for the action in the set of action to be adjusted. The action intensity variable takes a value range of 0 to 1. The continuous coordinated adjustment coefficient is composed of the action intensity variable. The objective function is weighted based on the directed edge weights from the working station to the affected working station in the causal graph. The marginal causal gain is equal to the weighted value of the objective function when the action intensity variable is 1 minus the weighted value when the action intensity variable is 0. The continuous synergistic adjustment coefficient is assigned a value based on the marginal causal gain. The continuous synergistic adjustment coefficient is equal to the normalized value of the marginal causal gain. The normalized value is equal to the marginal causal gain minus the minimum value of the marginal causal gain and then divided by the maximum value of the marginal causal gain minus the minimum value of the marginal causal gain. A set of coordinated regulation strategies is constructed based on continuous coordinated regulation coefficients and a set of regulatory actions. Each coordinated regulation strategy in the set includes a regulatory action number and a continuous coordinated regulation coefficient.

8. The intelligent assembly line collaborative control system based on machine learning according to claim 1, characterized in that, The process of performing feasibility screening based on action-level resource occupancy relationships and mutual exclusion constraints, and generating a screened collaborative adjustment strategy, specifically includes: The resource occupancy vector corresponding to any collaborative adjustment strategy in the collaborative adjustment strategy set is determined based on the action-level resource occupancy relationship matrix. The resource occupancy vector is determined by the row vector corresponding to the adjustment action number in the action-level resource occupancy relationship matrix. For any two coordinated adjustment strategies in the coordinated adjustment strategy set, a mutual exclusion constraint judgment value is determined. The mutual exclusion constraint judgment value is determined by the matrix elements in the mutual exclusion constraint matrix corresponding to the two adjustment action numbers. When the mutual exclusion constraint judgment value is equal to 1, it is determined that there is a conflict between the two coordinated adjustment strategies; when the mutual exclusion constraint judgment value is equal to 0, it is determined that there is no conflict between the two coordinated adjustment strategies. For conflicting collaborative regulation strategies, compare the continuous collaborative regulation coefficients and perform retention or elimination processing to generate the filtered collaborative regulation strategies.

9. The intelligent assembly line collaborative control system based on machine learning according to claim 1, characterized in that, The instruction mapping module is specifically as follows: Based on the adjustment action number, locate the corresponding adjustment action's action execution time window and resource usage identifier in the adjustment action set; Discrete control commands are generated based on the continuous coordinated adjustment coefficient. The discrete control commands include the workstation number, command type, and command amplitude. The workstation number is taken from the set of adjustment actions, the command type is mapped from the resource occupancy identifier, and the command amplitude is mapped from the continuous coordinated adjustment coefficient and the preset command quantization level. Discrete control commands are sorted according to the control cycle timestamp and sent to the corresponding workstation control unit to generate execution results.

10. The intelligent assembly line collaborative control system based on machine learning according to claim 1, characterized in that, The feedback update module is specifically as follows: Obtain the execution results and align them according to the control cycle timestamp to construct feedback data; The path feasible region and gating coefficients in the DTW-former model are updated and improved based on feedback data. The path feasible region update includes reestimation of the band width, and the gating coefficient update includes reestimation of the path confidence normalization interval. The projection intensity and workstation causal adjacency matrix in the improved NOTERAS algorithm based on feedback data are updated. The projection intensity update includes reassigning the first preset value and the second preset value, and the workstation causal adjacency matrix update includes reassigning the learning rate and the number of gradient update rounds. The updated path feasible region, gating coefficient, projection intensity, and workstation causal adjacency matrix are input into the collaborative trend prediction and causal learning process of the next control cycle.

Images