Task state recognition system and method in industrial agent for multiple industrial chains

Abstract

The invention discloses a task state recognition system and method in an industrial agent for multiple industrial chains, and the system at least comprises a task feature extraction module and an incomplete multi-view clustering module, the task feature extraction module constructs a shared potential subspace through a non-negative matrix factorization technology based on task attributes and topological features, and the incomplete multi-view clustering module is used for clustering the potential subspace; complementing the missing task data, and generating a basis matrix of each view and potential representations of all views; and the incomplete multi-view clustering module is used for solving the optimal potential representation and supplementing complete task attribute data for the multi-view task data with missing data and the topological connection relationship through an incomplete multi-view clustering algorithm based on the base moment and the potential representation generated by the task feature extraction module, so as to realize the recognition of the task state. According to the system and the method disclosed by the invention, the task state can still be accurately judged by virtue of the clustering result under the adverse condition that part of attribute data of the task is missing, so that stable and efficient operation of multiple industrial chains is powerfully guaranteed.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent computing, and mainly relates to a task status recognition system and method for industrial agents targeting multiple industrial chains. Background Art

[0002] In the context of multiple industrial chains in the modern industrial system, industrial agents bear the heavy responsibility of effectively managing and monitoring numerous tasks. Among them, tasks are woven into an extremely intricate multiple industrial chain network through various connection methods such as supply chain collaboration, technology collaboration, and production process connection. Task status recognition in multiple industrial chains refers to accurately identifying a set of tasks or specific status with similar execution patterns or similar business objectives from a large-scale task cluster based on the internal connection relationships between tasks in the multiple industrial chain network formed by collaborative operations within the task chain and undertaking different links across chains. The task status in multiple industrial chains is embodied as an organic structure composed of the task itself and the associated edges closely connected to it. Task status recognition follows the principle of close connection within tasks of the same class and relatively loose association between tasks of different classes, and implements structural differentiation for all tasks. Exploring the potential task status in the multiple industrial chain structure is of crucial significance for industrial agents to promptly detect abnormal tasks, reasonably plan task scheduling, and promote the efficient and stable operation of the entire industrial system.

[0003] With the deep penetration of data mining and analysis technologies in the process of industrial digitization, multi-view clustering algorithms have become one of the core means to analyze the structure of multiple industrial chains and the internal connections of tasks. Multi-view clustering aims to integrate data (i.e., views) from different data sources or different feature dimensions to achieve a more comprehensive and in-depth understanding and division of task groups. However, in actual multiple industrial chain scenarios, due to the complexity of tasks themselves and many challenges in data collection and management, the difficult problem of partial task data loss is often encountered. In multiple industrial chains, a task may be in multiple chains at the same time, and due to various factors, some attribute features are missing in some chains. For example, a production task covers the transportation of raw materials, the production arrangement of the factory, and the distribution of the final product, and thus involves three industrial chains: the raw material transportation chain, the production manufacturing chain, and the product distribution chain. The lack of GPS data of some transport vehicles will lead to the lack of relevant location and path information on the raw material transportation chain; the incomplete operation records of some production equipment make the equipment operation parameters and status information on the production manufacturing chain incomplete; the failure to upload some distribution feedback data in a timely manner results in the lack of customer feedback and delivery status information on the product distribution chain. Such data loss situations may be due to the incoherence of information recording during task execution, the failure of data collection equipment, or human negligence, or the concealment of some data for data security and business confidentiality considerations.

[0004] At the current stage, the existing multi-view clustering algorithms have shown significant limitations when dealing with the problem of missing data in some tasks of multiple industrial chains. Many traditional multi-view clustering algorithms assume that the data is complete and the information of each view is uniformly available, without fully considering the serious negative impact of missing data on the accuracy of task status recognition. Some algorithms attempt to simply fill in the missing data, but this approach often has the opposite effect because all the missing data filled in the same vector space is easily misclassified as the same task status, thus interfering with the normal task division. Some other algorithms choose to directly ignore the missing data, which obviously violates the original intention of comprehensive coverage of task status recognition, resulting in an inability to perform a complete and accurate status division for all tasks. There are also some algorithms that use a generative adversarial network (GAN) to generate the data of the missing view and optimize the division result through label guidance for the generator, but this method has extremely high requirements for the balanced training of the generator and the discriminator, and inevitably faces high computational complexity and long training time costs.

[0005] In addition, although some algorithms have proposed using latent subspaces to reconstruct view data for clustering division, they only focus on the attribute characteristics of the tasks themselves and seriously ignore the valuable topological characteristics in the multiple industrial chain environment. The topological structure in the industrial chain truly reflects the connection relationship and collaboration mode between tasks and is of irreplaceable importance to task status recognition. Therefore, how to design an innovative incomplete multi-view clustering algorithm that can skillfully introduce a topological supervision mechanism in the complex situation of missing data in some tasks of multiple industrial chains, then perform cyclic reconstruction on the missing view data, and finally achieve more accurate task clustering division and status recognition has become a key technical challenge that urgently needs to be overcome in the current industrial data analysis field. Summary of the Invention

[0006] The present invention precisely aims to fill the gaps in the existing technology and proposes a task status recognition system and method for industrial agents in multiple industrial chains, which at least includes a task feature extraction module and an incomplete multi-view clustering module. The task feature extraction module constructs a shared latent subspace based on task attributes and topological characteristics through non-negative matrix factorization technology, complements the missing task data, and generates the basis matrix of each view and the latent representation of all views. The incomplete multi-view clustering module, based on the basis matrix and latent representation generated by the task feature extraction module, uses an incomplete multi-view clustering algorithm for the multi-view task data with missing data and the topological connection relationship to solve the optimal latent representation and supplement the complete task attribute data, thereby realizing the recognition of task status. The incomplete multi-view clustering algorithm proposed by the present invention, which introduces topological supervision for cyclic reconstruction, provides strong technical support and guarantee for industrial agents to efficiently recognize task status, timely discover abnormal tasks and issue early warnings in multiple industrial chains, and promotes intelligent decision-making and task management in the industrial field.

[0007] To achieve the above object, the technical solution adopted by the present invention is: at least including a task feature extraction module and an incomplete multi-view clustering module,

[0008] The task feature extraction module: Through non-negative matrix factorization technology, based on task attributes and topological features, a shared latent subspace is constructed to complete the missing task data, and the basis matrix of each view and the latent representation of all views are generated;

[0009] The incomplete multi-view clustering module: Based on the basis matrix and latent representation generated by the task feature extraction module, for the multi-view task data and topological connection relationships with missing data, through the incomplete multi-view clustering algorithm, the optimal latent representation is solved and the task attribute data is supplemented to complete the recognition of the task state.

[0010] As an improvement of the present invention, in the task feature extraction module, the objective function of the non-negative matrix factorization technology is specifically:

[0011]

[0012] Among them, the basis matrix U of each view [k] and the latent representation P of all views need to maintain non-negativity to complete the matrix factorization process of the non-negative matrix factorization technology, is the selection matrix of the existing data on the kth layer of the industrial chain, The constraint term of ensures that the original task existing attribute data remains unchanged during the reconstruction process, and only the missing attribute data of the task is modified; at the same time, to avoid overfitting of the model during the training process, λ||U [k] || 2 is introduced as the L2 regularization term, and by constraining the Frobenius norm of the basis matrix U [k] , this regularization term can effectively adjust the complexity of the model, prevent the model from overfitting the training data, and enhance the generalization ability of the model.

[0013] As an improvement of the present invention, in the incomplete multi-view clustering module, the objective function of the incomplete multi-view clustering algorithm is specifically:

[0014]

[0015] Among them, S is the similarity matrix of the latent representation P, and the graph Laplacian matrix L corresponding to W [k] is expressed as: L [k] is expressed as: L [k] =D [k] -W [k] , where D [k] =diag(W[k] 1 N ) is the degree matrix, and its diagonal elements represent the weighted degree of enterprise i in the topological structure of this view. The corresponding regularized Laplacian matrix Ls of S is expressed as: In the formula, Ds represents the diagonal matrix of the similarity matrix S, which is defined as This construction ensures the semi - positive definiteness of the Laplacian matrix through symmetrization processing. represents the weights occupied by each layer of the industrial chain. The non - negative Shannon entropy is used to balance the importance of each view and coordinate the differences between them. Its penalty factor δ>0 is used to balance the influence of this collaborative effect. For the similarity matrix of the latent representation, for the Laplacian matrix, for the topological feature, for one.

[0016] As another improvement of the present invention, in the incomplete multi - view clustering module, it at least includes the cyclic reconstruction of the missing view, the supervision mechanism of the topological regularization term, the global coordination of the subspace latent representation, and the adaptive learning of the view weights.

[0017] The purpose of the cyclic reconstruction of the missing view is: to minimize the difference between the attribute feature X [k] and U [k] P, control all instance data of each view to converge in the latent subspace, and ensure that the latent representation P and the basis matrix U of each view [k] can reconstruct the original task data;

[0018] The supervision mechanism of the topological regularization term is specifically: add a topological regularization term tr(PL [k] P T ) to each view. This constraint term superimposes the topological information of each view into the optimization process of the shared representation P through the regularization term;

[0019] The global coordination of the subspace latent representation is specifically: add a global constraint term tr(PLsP T ). This graph Laplacian term retains the global attributes of the latent representation P to connect all task sample data in each view;

[0020] The adaptive learning of the view weights is specifically: introduce the non - negative Shannon entropy and set the penalty factor δ.

[0021] As yet another improvement of the present invention, the cyclic reconstruction of the missing view is specifically:

[0022]

[0023] Among them, in the process of cyclic reconstruction, by [k]Regularization is performed to prevent overfitting, and λ is the balance parameter of this regularization term.

[0024] To achieve the above object, the technical solution adopted by the present invention is also: for the task status recognition method in the industrial intelligent body of the multiple industrial chains, including the following steps:

[0025] S1: According to the task attribute feature X in the k-th layer of the industrial chain [k] Determine the relevant variables in each view, specifically including the attribute feature dimension d [k] , the latent subspace dimension r, the set M of task indices of the missing data in each view [k] , the topological connection relationship W of the task in each industrial chain [k] , the number C of clusters of the multi-view task data to be clustered, the number m of views, the total number N of tasks, the number of existing tasks in each view and the number of missing tasks

[0026] S2: Standardize the elements in the existing attribute feature view of each task to positive values by the min-max method, and the specific calculation method is:

[0027]

[0028] where represents the d-th attribute feature value of the i-th task in the k-th view, then represents the d-th attribute feature value of the j-th task in the k-th view;

[0029] S3: Determine the selection matrix of the existing task attributes [k] according to the set M of missing task indices in each view the filling matrix of the existing task attributes the selection matrix of the missing task attributes the filling matrix of the missing task attributes

[0030] S4: Randomly initialize the basis matrix U [k] of each view

[0031] and the latent shared representation P; [k] S5: Adopt the gradient descent strategy for alternating iterative optimization, and by fixing other variables, optimize the variables X [k] , U , P, S, separately to obtain the optimal solution of the objective function; in each iteration, update the value of the current variable by minimizing the sub-problem of the objective function with respect to the current variable until the preset termination condition is satisfied; define the parameter set in the t-th iteration asIn the update process of the (t + 1)-th optimization solution, the description of t in the subscripts of the following variables represents their iteration times; the superscript T represents the transpose matrix of the matrix variable; the superscript [k] represents that the variable is in the k-th layer of the industrial chain:

[0032] S6: Construct the objective function formula J of the initialized incomplete multi-view clustering CRTC-IMVC, and the specific formula is as follows:

[0033]

[0034] S7: Repeat steps S5 and S6 until the maximum number of iterations is reached or the objective function value converges, and obtain the optimal latent representation P * and the complete task attribute features x in each view [k] , and obtain the task x j The task status label(x j ) = kmeans(P, C), and the task status of each task in the entire multi-industrial chain network can be represented as γ i = {j: label(x j ) = i}, realizing the recognition of task status in the multi-industrial chain.

[0035] As an improvement of the present invention, in step S3, the existing attribute selection matrix is obtained by removing the columns corresponding to the missing samples or existing samples in the view from the identity matrix I ∈ R N ×N ; the filling matrix is formed by filling the columns corresponding to the existing data in the identity matrix into the columns corresponding to the zero matrix ; the filling matrix is formed by filling the columns corresponding to the missing data in the identity matrix into the columns corresponding to the zero matrix .

[0036] As another improvement of the present invention, the iterative optimization in step S5 specifically includes the following steps:

[0037] S51. Completion of missing data: Use the existing data and the latent shared representation to infer the values of the missing data, fix U [k] , P, S, and update X [k] , and the update formula is:

[0038]

[0039] Among them, represents the attribute feature matrix of the missing tasks in the k-th layer of the industrial chain, As the missing attribute selection matrix corresponding to the existing attribute selection matrix is used to accurately locate the missing attributes. Its construction method is to remove the columns corresponding to the existing samples in the view from the identity matrix I N ∈R N×N .

[0040] S52. Basis matrix update: Fix Update U [k] , and the update formula is:

[0041]

[0042] S53. Latent representation matrix update: Fix Update P, and the update formula is:

[0043]

[0044] S54. Similarity matrix update: Fix Update S, and the update formula is:

[0045]

[0046] where Q ij is an auxiliary variable, O ∈ R N×N is a zero matrix, l is the number of adjacent nodes most likely to be connected to P i , is the result of sorting Q i in ascending order;

[0047] S55. View weight update: Fix X [k] , U [k] , P, S, update and the update formula is:

[0048]

[0049] where, is an intermediate variable.

[0050] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0051] (1) The present invention solves the problem of task state division in the current multi-industrial chain environment with missing task attribute features through a new incomplete multi-view clustering algorithm, and adopts a cyclic reconstruction mechanism for missing data to effectively process and make up for the information loss caused by data missing, improving the integrity and availability of task data in the industrial chain.

[0052] (2) Innovatively, in the incomplete multi-view clustering algorithm of the present invention, a topological regularization term is added to each view, integrating the view topological information into the optimization process of the shared representation. By making the latent representation smoothly distributed among topologically adjacent samples, the model's ability to understand and capture the topological local structure is enhanced, enabling the shared representation to retain both multi-view topological information and the global attributes of all task attribute features, improving the robustness and consistency of the final representation, and thereby enhancing the accuracy of dividing the task state.

[0053] (3) Introducing Shannon entropy in incomplete multi-view clustering can not only effectively reduce the influence of noisy views but also coordinate the differences among views, fully considering the data quality issues in multiple industrial chains in the actual application scenario and the differences among industrial chains, thus enhancing the collaboration of different views and more accurately locating their potential task states. Brief Description of the Drawings

[0054] Figure 1 It is the overall framework diagram of the method of the present invention. Detailed Embodiments

[0055] The following further clarifies the present invention in conjunction with the drawings and detailed embodiments. It should be understood that the following detailed embodiments are only used to illustrate the present invention and not to limit the scope of the present invention.

[0056] Embodiment 1

[0057] For the task state recognition system in the industrial intelligent agent of multiple industrial chains, it includes at least two parts: a task feature extraction module and an incomplete multi-view clustering module. The core objective of the task feature extraction module is to extract the latent representation of the task from multi-view task data. By introducing the non-negative matrix factorization technique, this module combines the attribute features and topological connection relationships, constructs a shared latent subspace, completes the completion of missing task data, and generates the basis matrices of each view and the latent representation of all views. The incomplete multi-view clustering module aims to, based on the multi-view task data with missing data and the topological connection relationships, combine the basis matrices and latent representations generated by the task feature extraction module, and through the designed incomplete multi-view clustering algorithm, solve the optimal latent representation and supplement the complete task attribute data.

[0058] The task feature extraction module aims to accurately represent the attribute features, topological features, and the scenario of missing data in the multiple industrial chain network, making it highly effective and applicable in real application scenarios, and obtaining the latent representation of the task attribute features. Then, using the proposed incomplete multi-view clustering algorithm with topological supervision for cyclic reconstruction, the task state discovery objective with some missing task numbers in multiple industrial chains is completed.

[0059] The core objective of the task feature extraction module is to extract the latent representation of tasks from the attribute features and topological connection relationships based on multi-view task data using the Non-negative Matrix Factorization (NMF) technique. This module completes the complementation of missing task data by constructing a shared latent subspace and generates the basis matrices and latent representations of each view. Its main modeling modules are as follows:

[0060] There are m existing industrial chains, and there are N tasks on each industrial chain. The set of attribute features of the multi-view tasks is represented as X = {X [1] , X [2] , …, X [m] . The k-th view is represented by the matrix X [k] . d [k] is the number of attribute features in the k-th feature view. Due to the situation of partial task data missing, that is, the attribute information of some tasks in multiple industrial chains is missing on a certain layer of the industrial chain, the attribute features are also divided into existing attribute features and missing attribute features. Among them, the set of existing attribute features is The k-th view is represented by the inherent feature matrix of the existing tasks on the k-th industrial chain. d [k] is the number of features in the k-th layer of the industrial chain view, is the number of tasks existing in the k-th layer of the industrial chain view; and the set of missing tasks in this case is satisfies is the number of missing tasks in the k-th layer of the industrial chain view, satisfying The topological connection relationship of tasks on the k-th layer of the industrial chain is represented by the adjacency matrix W [k] . Among them, represents the connection relationship between task and task on the k-th layer of the industrial chain. If its value is 1, it means there is a connection relationship, and if it is 0, it means there is no connection relationship between these two tasks.

[0061] Based on the theory that multi-view data can all be mapped to the same subspace, any task data can be mapped into the latent subspace. Therefore, the missing task sample data and the complete task sample data share a common latent subspace. The basis matrix U [k] of each view and the latent representation P of all views are mainly extracted from the task data X based on the non-negative matrix factorization technique. To use the NMF technique, first, X [k] needs to be normalized so that each element in it is non-negative. The basic objective function in this case is expressed as:

[0062]

[0063] Among them, the basis matrix U of each view [k] and the latent representation P of all views need to be non - negative to complete the matrix factorization process of NMF. The complete reconstructed data can also be obtained from these two. is the selection matrix for the existing data on the k - th layer of the industrial chain. This matrix is obtained by removing the columns corresponding to the missing samples in that view from the identity matrix I ∈ R N×N . The constraint term of ensures that the original task - existing data remains unchanged during the reconstruction process, and only the task - missing data is modified. λ||U [k] || 2 is an L2 regularization term. Constraining the Frobenius norm of U [k] can effectively prevent the model from overfitting the data during training. The basis matrix U [k] extracted by this module provides a local interpretation of the task - status features in each view, capturing the multi - dimensional distribution of task attributes; the latent representation P provides a unified expression of the task - status features in the global latent subspace, revealing the cooperation mode and functional role of the task in the network.

[0064] In summary, the main functions of the task feature extraction module include multi - view task data processing, basis - matrix task data generation, and task latent - representation generation.

[0065] The core content of the incomplete multi - view clustering module is based on the multi - view task dataset X with missing data and the topological connection relationship W, and at the same time incorporates the basis matrix U [k] of each view and the latent representation P of all views obtained by the task feature extraction module. The designed incomplete multi - view clustering algorithm is used to obtain the optimal latent representation P * and the supplemented and complete task - attribute data. Since the final task - division result is determined by the latent representation P * , the attributes of P * need to be able to absorb the attribute features and topological features in the task data, and at the same time have the ability to connect global views. Therefore, the present invention re - defines the objective function of the incomplete multi - view clustering algorithm as follows:[[]]

[0066]

[0067] Among them, S is the similarity matrix of the latent representation P, and the Laplacian matrix L [k] = D [k] - W [k] is derived from the topological feature W [k] , and D [k]is a diagonal matrix, and its elements represent the number of connections for each task in the topological features of the corresponding view; the Laplacian matrix Ls = Ds - S is derived from the similarity matrix S, is a diagonal matrix.

[0068] In the incomplete multi-view clustering module, the design idea of the objective function at least includes the cyclic reconstruction of missing views, the supervision mechanism of the topological regularization term, the global coordination of the latent representation of the subspace, and the adaptive learning of view weights:

[0069] Cyclic reconstruction of missing views: The purpose of this term is to minimize the difference between the attribute feature X [k] and U [k] P, control all instance data of each view to converge smoothly in the latent subspace, and ensure that the latent representation P and the basis matrix U of each view [k] can effectively reconstruct the original task data. During the process, U [k] is regularized to prevent overfitting and improve the stability of the optimization process. λ is the balance parameter of this regularization term.

[0070] Supervision mechanism of the topological regularization term: Add a topological regularization term tr(PL [k] P T ) to each view. This constraint term superimposes the topological information of each view into the optimization process of the shared representation P through the regularization term. Specifically, topological regularization makes the latent representation of the data smoothly distributed among topologically adjacent samples. This smoothing effect means that points with a connection relationship in the topological feature data will have similar representations, thereby increasing the model's ability to understand the local structure in the topological information. In this way, the shared representation P can retain the topological information of multiple views at the same time, making the final representation more robust and consistent. α is the balance parameter of this regularization term.

[0071] Global coordination of the latent representation of the subspace: Add a global constraint term tr(PLsP T ). This graph Laplacian term retains the global attributes of the latent representation P to connect all task sample data in each view. β is the balance parameter of this regularization term. ||S|| 2 is the regularization of the similarity matrix S to avoid degenerate solutions, and during the process, ensure that only one element in each column of S is equal to 1, while all other elements are equal to 0. γ is the balance parameter of this regularization term.

[0072] Adaptive learning of view weights: Introduce non-negative Shannon entropy Used to balance the importance of each view and coordinate the differences between them. From an optimization perspective, if this item is removed, the importance of the clearly distinguishable view will approach 1 infinitely, while the importance of other views will be negligible. Additionally, directly minimizing this item will make the importance degrees of each view equal. By setting the penalty factor δ to balance these effects, the weights of different views can be adaptively adjusted.

[0073] Iteratively optimize the above clustering objective function until convergence to obtain the final latent representation P * . Finally, through common clustering algorithms such as Kmeans or spectral clustering, the task status attribution of each task can be obtained and can be represented as label(x j ). The task status of each task in the entire multi-industrial chain network can be represented as Υ i = {j: label(x j ) = i}. Thus, the incomplete multi-view clustering algorithm - CRTC-IMVC applicable to cyclic reconstruction with topological supervision in the multi-industrial chain is designed.

[0074] The main functions of the incomplete multi-view clustering module include the integration and optimization of two types of task feature data, the iterative optimization of task latent representation, and the complementation of task attribute features.

[0075] Through the design of the above two parts of the task feature extraction and incomplete multi-view clustering solutions, the discovery of task status in the case of missing task data in the multi-industrial chain can be completed. Subsequently, based on the recognition results, it can provide key guidance for resource allocation and monitoring in the task execution process. For example, when a specific task status is recognized, key resource tilting and monitoring deployment can be carried out for the tasks in this state. Because these tasks are judged to have similar behavior patterns or business objectives during the clustering process, it means that they have high similarities in the types and timing of resource requirements during the execution process, and also have commonalities in the risks and problems that may be encountered. Through centralized resource guarantee and refined monitoring, the success rate and efficiency of task execution can be effectively improved, ensuring the stable operation of the multi-industrial chain and the optimization of the overall efficiency.

[0076] In summary, the present invention provides a task status recognition system for industrial agents targeting multiple industrial chains, including a task feature extraction module and an incomplete multi-view clustering module. The task feature extraction module accurately represents task attributes and topological features, divides the task attribute set to fit the scenario of partial task attribute data missing, and then obtains the latent representation of task features through non-negative matrix factorization technology. The incomplete multi-view clustering module combines topological supervision and global constraints to optimize the task data reconstruction and clustering process. By iteratively optimizing the objective function, the best task latent representation is obtained to determine the task status and complete task attribute features. This system has significant advantages. Even in the adverse situation of partial task attribute data missing, it can still accurately determine the task status based on the clustering results. It helps industrial agents promptly detect abnormal tasks or those that require special attention, and quickly issue warning signals so that relevant personnel can handle them in a timely manner, effectively ensuring the stable and efficient operation of multiple industrial chains.

[0077] Embodiment 2

[0078] A method for recognizing task status in an industrial agent targeting multiple industrial chains, as Figure 1 shown, includes the following steps:

[0079] Step S1: According to X [k] determine the attribute feature dimension d in each view [k] , the latent subspace dimension r, the set M of task indices with missing data in each view [k] , the topological connection relationship W of tasks in each industrial chain [k] , the number C of clusters for the multi-view task data to be clustered, the number m of views, the total number N of tasks, the number of existing tasks in each view the number of missing tasks wherein, W [k] describes the correlation between tasks, reflects the collaboration mode between tasks in the industrial chain, and the number C of clusters defines the number of categories of task status, aiming to reveal the distribution characteristics of different task statuses (such as normal operation, abnormal shutdown, etc.).

[0080] Step S2: In order to eliminate the influence of the dimensional difference of feature values between different views on the model performance, each existing attribute feature view of each task is standardized to positive values by the min-max method. This standardization process ensures the comparability of different view features, avoids deviations caused by numerical range differences, and thus improves the accuracy of task status recognition. The specific calculation method is as follows.

[0081]

[0082] Step S3: Since there may be data missing in the actual industrial environment, the present invention designs a selection matrix and a filling matrix Separate the existing task data and the missing task data for processing. Specifically, according to the set of missing task indices M for each view [k] Determine the selection matrix of the existing task attributes The filling matrix of the existing task attributes The selection matrix of the missing task attributes The filling matrix of the missing task attributes where the selection matrix is obtained by removing the columns corresponding to the missing samples or the existing samples in the view from the identity matrix I ∈ R N×N ; the filling matrix is formed by filling the columns corresponding to the existing data in the identity matrix with the zero matrix in the corresponding columns; the filling matrix is formed by filling the columns corresponding to the missing data in the identity matrix with the zero matrix in the corresponding columns. The above mechanism can effectively handle the problem of missing data and provide complete data support for subsequent task status recognition.

[0083] Step S4: Randomly initialize the basis matrix U for each view [k] , the latent shared representation P. In addition, α, β, γ, δ, and λ are hyperparameters, and usually a grid search strategy can be used to determine them. This step is the initialization process of the incomplete multi-view clustering module.

[0084] Step S5: The present invention adopts a gradient descent strategy for alternating iterative optimization. By fixing other variables, the variables are optimized separately in turn to gradually approach the optimal solution of the objective function. In each iteration, by minimizing the sub-problem of the objective function with respect to the current variable, the value of the variable is updated until the preset termination condition is satisfied.

[0085] Define the parameter set at the t-th iteration as Then the update process of the (t + 1)-th optimization solution is as follows:

[0086] S51. Completion of missing data: Fix Update X [k] .

[0087] In a multi-industrial chain scenario, due to reasons such as sensor failures or communication interruptions, the attribute data of some tasks may be partially missing. Since the operating temperature or energy consumption data of some devices may not be collected in real time, in order to ensure the accuracy of task status recognition, we need to use the existing data and the latent shared representation to infer the values of the missing data. Its update formula is:

[0088]

[0089] S52. Basis matrix update: Fix Update U [k] .

[0090] The basis matrix U [k] is an important part of the task attribute features, which captures the main feature distribution of the task in the k-th view. By iteratively updating U [k] , we can better fit the task data X [k] , thus improving the accuracy of task status recognition. Its update formula is:

[0091]

[0092] S53. Latent representation matrix update: Fix Update P.

[0093] The latent representation matrix P is the core of task status recognition. It integrates information from multiple views and reveals the deep relationships between tasks. In a multi-industrial chain, the status of different devices may be affected by upstream and downstream tasks, and the latent representation matrix can capture this cross-view correlation. Its update formula is:

[0094]

[0095] S54. Similarity matrix update: Fix Update S.

[0096] The similarity matrix S is an important basis for task status division. It calculates the distance between tasks based on the latent representation matrix P and quantifies the similarity degree of task status. If two devices have similar operating modes, they may be classified into the same status category. Its update formula is:

[0097]

[0098] where O ∈ R N×N is a zero matrix, l is the number of adjacent nodes most likely to be connected to P i , and is the result after ascending order processing based on Q i .

[0099] S55. View weight update: Fix X [k] , U [k] , P, S, update

[0100] The view weight is introduced to enable this module to dynamically adjust the importance of each view and give priority to trusting those views that contribute more to task status recognition. Its update formula is:

[0101]

[0102] Step S6: Construct the objective function formula J of the initialized incomplete multi-view clustering CRTC-IMVC. This objective function comprehensively considers multiple factors such as data reconstruction error, topological relationship consistency, task similarity, and view weight assignment to ensure the comprehensiveness and accuracy of the task status recognition result. The specific formula is as follows:

[0103]

[0104] Step S7: Repeat the operations in Step S5 and S6 until the maximum number of iterations or the objective function value converges to obtain the optimal latent representation P * and the complete task attribute features X in each view [k] , and finally, the task x j to which it belongs, the task status label(x j ) = kmeans(P, C). The task status of each task in the entire multi-industrial chain network can be represented as Υ i = {j: label(x j ) = i}. This result intuitively shows the distribution of task status in the multi-industrial chain.

[0105] Industrial agents can clearly understand the implicit task status information in the entire multi-industrial chain based on the task status division result, which is beneficial for making the next behavior decision. For example, it can timely detect abnormal or tasks that need attention and issue warning reminders for handling; at the same time, it can also refer to the complete task attribute features after supplementation for further decision-making analysis and task attribute adjustment.

[0106] It should be noted that the above content only illustrates the technical idea of the present invention and cannot be used to limit the protection scope of the present invention. For those of ordinary skill in the art of this technology, without departing from the principle of the present invention, several improvements and retouches can still be made, and these improvements and retouches all fall within the protection scope of the claims of the present invention.

Claims

1. An industrial agent task status recognition system for multiple industrial chains, characterized in that: It at least includes a task feature extraction module and an incomplete multi-view clustering module. The task feature extraction module: By using non-negative matrix factorization technology, based on task attributes and topological features, constructs a shared latent subspace, completes the missing task data, and generates the basis matrices of each view and the latent representations of all views. The incomplete multi-view clustering module: Based on the basis matrices and latent representations generated by the task feature extraction module, for the multi-view task data with missing data and topological connection relationships, through the incomplete multi-view clustering algorithm, solves the optimal latent representation and supplements the complete task attribute data to realize the identification of task states.

2. The task status recognition system in the industrial agent for multiple industrial chains according to claim 1, characterized in that: In the task feature extraction module, the objective function of the non-negative matrix factorization technology is specifically: Among them, X [k] is the task attribute feature matrix in the k-th layer of the industrial chain, U [k] is the base matrix on each layer of the industrial chain, P is the common latent representation on all industrial chains, is the selection matrix of the existing data on the k-th layer of the industrial chain, represents the existing task attribute feature matrix on the k-th layer of the industrial chain, λ||U [k] || 2 is an L2 regularization term.

3. The task status recognition system in the industrial intelligent agent for multiple industrial chains according to claim 2, wherein: In the incomplete multi-view clustering module, the objective function of the incomplete multi-view clustering algorithm is specifically: Among them, S is the similarity matrix of the latent representation P, and L [k] is expressed as the Laplacian matrix, specifically L [k] = D [k] - W [k] , where W [k] is the topological feature matrix of the k-th view, and D [k] = diag(W [k] 1 N ) is the degree matrix, and its diagonal element represents the weighted degree of enterprise i in the topological structure of this view. The regularized Laplacian matrix Ls corresponding to S is expressed as: In the formula, Ds represents the diagonal matrix of the similarity matrix S, which is defined as represents the weights occupied by each layer of the industrial chain, and α, β, γ, λ, δ are the balance factors of each item.

4. The task status recognition system in the industrial agent for multiple industrial chains according to claim 3, wherein: In the incomplete multi-view clustering module, it at least includes the cyclic reconstruction of missing views, the supervision mechanism of the topological regularization term, the global coordination of the latent representation of the subspace, and the adaptive learning of view weights. The purpose of the cyclic reconstruction of the missing view is to minimize the attribute feature X [k] from U [k] the difference with P, control all instance data of each view to converge in the latent subspace, and ensure that the latent representation P and the basis matrix U of each view [k] can reconstruct the original task data; The supervision mechanism of the topological regularization term is specifically as follows: a topological regularization term tr(PL [k] P T ) is added to each view, and this constraint term superimposes the topological information of each view onto the optimization process of the shared representation P through the regularization term; The global coordination of the subspace potential representation is specifically as follows: adding a global constraint term tr(PLsP T ), this graph Laplacian term preserves the global properties of the potential representation P to connect all task sample data in each view; The adaptive learning of the view weights is specifically as follows: introducing non - negative Shannon entropy and setting a penalty factor δ.

5. The task status recognition system in the industrial intelligent agent for multiple industrial chains according to claim 4, characterized in that: The cyclic reconstruction of the missing views is specifically: Among them, during the loop reconstruction process, by regularizing U [k] overfitting is prevented, and λ is the balance parameter of this regularization term.

6. Using the system according to claim 1, the system executes a task status recognition method for industrial agents in multiple industrial chains, characterized in that , including the following steps: S1: Determine relevant variables in each view according to the task attribute characteristics X in the k-th layer of the industrial chain [k] wherein the relevant variables include the attribute feature dimension d [k] , the potential subspace dimension r, the set W of task indices of missing data in each view [k] , the topological connection relationship W of the task in each layer of the industrial chain [k] , the number of clusters C of the multi-view task data to be clustered, the number of views m, the total number of tasks N, the number of existing tasks in each view and the number of missing tasks S2: Standardize the elements within each task's existing attribute feature view to positive values by the min-max method. The specific calculation method is: Among them, represents the d-th attribute feature value of i tasks in the k-th view, while represents the d-th attribute feature value of j tasks in the k-th view. S3: Determine the selection matrix of existing task attributes according to the set M of missing task indices for each view [k] Determine the selection matrix of existing task attributes The filling matrix of existing task attributes The selection matrix of missing task attributes The filling matrix of missing task attributes S4: Randomly initialize each view basis matrix U [k] and the latent shared representation P; S5: Use the gradient descent strategy for alternating iterative optimization. By fixing other variables, sequentially optimize variables X [k] , U [k] , P, S, individually to obtain the optimal solution of the objective function; in each iteration, update the value of the current variable by minimizing the sub-problem of the objective function with respect to the current variable until the preset termination condition is met; S6: Construct the objective function formula J of the initialized incomplete multi-view clustering CRTC-IMVC. The specific formula is as follows: S7: Repeat steps S5 and S6 until the maximum number of iterations is reached or the objective function value converges, and obtain the best latent representation P * and the complete task attribute features X in each view [k] , and obtain the task x j The task status label(x j ) = kmeans(P, C), and the task status of each task in the entire multi-industrial chain network can be represented as Υ i = {j: label(x j ) = i}, realizing the recognition of task status in the multi-industrial chain 7. The method for task status recognition in the industrial agent for multiple industrial chains according to claim 6, wherein: In the said step S3, the existing attribute selection matrix is obtained by removing the columns corresponding to the missing samples or the existing samples in the view from the identity matrix I ∈ R N×N ; the filling matrix is formed by filling the columns corresponding to the existing data in the identity matrix with the zero matrix in the corresponding columns; the filling matrix is formed by filling the columns corresponding to the missing data in the identity matrix with the zero matrix in the corresponding columns.

8. The method for task status recognition in the industrial intelligent agent for multiple industrial chains according to claim 6, characterized in that: The iterative optimization of step S5 specifically includes the following steps: S51. Missing data completion: Estimate the values of missing data using existing data and latent shared representations, and fix U [k] , P, S, Update X [k] , and the update formula is: Among them, represents the attribute feature matrix of the missing tasks in the k-th layer of the industrial chain, represents the missing attribute selection matrix corresponding to the existing attribute selection matrix , which is used to accurately locate the missing attributes. Its construction method is to remove the columns corresponding to the existing samples in this view from the identity matrix I N ∈R N ×N ; S52. Base matrix update: Fix X [k] , P, S, Update U [k] , and the update formula is: S53. Potential representation matrix update: Fix X [k] , U [k] , S, Update P, and the update formula is: S54, Similarity matrix update: Fix X [k] , U [k] , P, Update S, and the update formula is: Among them, Q ij is an auxiliary variable, O ∈ R N×N is a zero matrix, and l is the number of adjacent nodes most likely to be connected to P i is the result after sorting Q i in ascending order.​ S55, View Weight Update: Fix X [k] , U [k] , P, S, Update The update formula is as follows: Among them, is an intermediate variable.

Images