An industrial big data processing system based on an industrial supply chain

Abstract

The application discloses an industrial big data processing system based on an industrial supply chain, comprising: a collection and preprocessing module, which is used for collecting multi-source heterogeneous data and performing preprocessing; a feature modeling module, which is used for establishing a self-supervised data twin, dynamically modeling a business link, and generating a local feature set through a self-supervised mechanism; a causal feature chain construction module, which is used for generating a causal feature sequence containing full-link disturbance response features; an optimized feature generation module, which is used for assigning weights of features of each business link and generating an optimized feature chain; a decision reasoning module, which is used for outputting demand prediction data, inventory allocation data, production scheduling data and risk early warning data; and a visual display module, which is used for visual real-time display. The application adopts full-link disturbance response aggregation and multi-layer intelligent modeling, realizes intelligent integration and dynamic optimization of supply chain full-process data, and has the advantages of fast abnormal response, controllable risk and high decision intelligence.

Description

Technical Field

[0001] This invention relates to the field of industrial big data processing, and in particular to an industrial big data processing system based on the industrial supply chain. Background Technology

[0002] Currently, supply chain data processing in the industrial sector mainly relies on traditional data collection and analysis methods. These methods tend to focus on single-point collection and processing of structured data, making it difficult to take into account the multi-source heterogeneous data generated from multiple stages such as procurement, production, warehousing, logistics, and sales. In particular, when dealing with unstructured data such as text and images, the standardization and feature extraction capabilities of existing systems are very limited, resulting in poor data fusion and end-to-end collaboration.

[0003] At the same time, most existing technologies use static rules or simple models for feature modeling and anomaly detection, which cannot dynamically reflect the complex causal relationships between supply chain links and the propagation of disturbances. In addition, the current system's data visualization methods are also relatively simple, making it difficult to intuitively display the temporal changes of causal features and weights, which is not conducive to management making real-time and accurate decisions on risks and key indicators throughout the entire supply chain process. Summary of the Invention

[0004] One objective of this invention is to propose an industrial big data processing system based on the industrial supply chain. This invention adopts full-link disturbance response aggregation and multi-layer intelligent modeling to realize intelligent integration and dynamic optimization of data throughout the entire supply chain process, and has the advantages of fast anomaly response, controllable risk and high decision-making intelligence.

[0005] An industrial big data processing system based on the industrial supply chain according to an embodiment of the present invention includes:

[0006] The data acquisition and preprocessing module is used to collect multi-source heterogeneous data from the procurement, production, warehousing, logistics and sales processes to obtain raw business data, and to preprocess the raw business data to form standardized business data.

[0007] The feature modeling module is used to build a self-supervised data twin for each business process, dynamically model the corresponding business process using standardized business data, and generate a local feature set through a self-supervised mechanism.

[0008] The causal feature chain construction module is used to causally concatenate the local feature sets of each business link according to business dependencies and time order, and use the end-to-end disturbance response aggregation mechanism to form a causal feature sequence containing end-to-end disturbance response features, while performing dimensionality reduction.

[0009] The optimized feature generation module is used to process the causal feature chain using an attention mechanism, assign weights to features in each business process, and generate an optimized feature chain.

[0010] The decision reasoning module is used to input the optimized feature chain into the multilayer perceptron and output demand forecast data, inventory allocation data, production scheduling data, and risk warning data.

[0011] The visualization module is used to visualize and display the principal component feature sequences and causal dependency weight matrices in real time during the optimization of the feature chain.

[0012] Optionally, modules can be integrated using the following methods:

[0013] Collect multi-source heterogeneous data from procurement, production, warehousing, logistics and sales processes to obtain raw business data, and preprocess the raw business data to form standardized business data;

[0014] A self-supervised data twin is established for each stage, and standardized business data is used to dynamically model the corresponding business stages. A local feature set is generated through a self-supervised mechanism.

[0015] According to the supply chain business process, the local feature sets of each business link are causally spliced ​​together according to business dependencies and time sequence to construct a causal feature chain;

[0016] The causal feature chain is processed by an attention mechanism, which assigns weights to features in each business process, performs feature redundancy and noise filtering, and generates an optimized feature chain.

[0017] The optimized feature chain is input into the multilayer perceptron, and the output includes demand forecast data, inventory allocation data, production scheduling data, and risk warning data.

[0018] The principal component feature sequences and causal dependency weight matrices during the feature chain optimization process will be visualized and displayed in real time.

[0019] Optionally, the raw business data includes structured data and unstructured data. The structured data includes tabular data, log data, and sensor sequence data. The unstructured data includes text data, image data, and audio data. The preprocessing includes data cleaning, format unification, missing value imputation, noise reduction, normalization, and time synchronization and alignment.

[0020] Optionally, the generation of the local feature set includes the following specific steps:

[0021] Establish self-monitoring data twins for each of the procurement, production, warehousing, logistics and sales stages, and input standardized business data into the corresponding self-monitoring data twins in chronological order;

[0022] Each self-supervised data twin uses a time series modeling method on the input standardized business data, combines historical standardized business data with current standardized business data, predicts the standardized business data at the next time point, and simultaneously inputs the current standardized business data into the autoencoder to generate latent space feature data, which corresponds one-to-one with the standardized business data.

[0023] Calculate the prediction error between the predicted standardized business data and the actual standardized business data, and calculate the autoencoder reconstruction error. Compare the prediction error with the prediction threshold and the reconstruction error with the reconstruction threshold. If any error exceeds the corresponding threshold, the standardized business data is marked as abnormal; otherwise, it is marked as normal.

[0024] For each business process, a weighted fusion method is used to assign fusion weights to the abnormal labels of the previous business process, the current business process, and the next business process. The three are multiplied by the fusion weights and then summed. If the fusion result is greater than the fusion threshold, the abnormal label of the current standardized business data is updated to abnormal; otherwise, it is updated to normal. The fusion result is used as the fused abnormal label data.

[0025] For standardized business data that is fused with abnormal label data, feature perturbation data is generated. The perturbation term is superimposed on the standardized business data and then input into the autoencoder to generate perturbed latent space feature data. The difference between the latent space feature data before and after perturbation is used to obtain the perturbation response feature data.

[0026] The current standardized business data, latent space feature data, fused anomaly label data, and disturbance response feature data are merged into a local feature set in chronological order.

[0027] Optionally, the construction of the causal feature chain includes the following specific steps:

[0028] Arrange the local feature sets of each business process in chronological order;

[0029] Based on the industry supply chain business processes, a causal weight matrix is ​​established;

[0030] Based on the local feature sets of all business processes, for each time point, the corresponding causal dependency weights in the causal weight matrix are dynamically adjusted according to the fused anomaly label data and disturbance response feature data in each local feature set.

[0031] The dynamic adjustment of the causal dependency weights in the causal weight matrix specifically includes:

[0032] Within a preset time window, calculate the average value of the fusion anomaly label data and the average norm of the disturbance response features for each business process.

[0033] The adjustment coefficient is calculated by weighted linear combination, and the adjustment coefficient is the sum of the average value of the fused anomaly labels multiplied by the first weight parameter and the average norm of the disturbance response features multiplied by the second weight parameter;

[0034] The initial weights of the corresponding elements in the causal weight matrix are multiplied by the adjustment coefficients to obtain the adjusted weight values.

[0035] Normalize each column of the adjusted causal weight matrix;

[0036] The local feature set of each link is multiplied by the adaptively optimized causal dependency weight at the time point, and the five weighted feature sets are summed to obtain the original causal feature vector at the time point. The original causal feature vectors of each time point are arranged in sequence to form the original causal feature sequence.

[0037] The original causal feature sequence is normalized to generate a normalized causal feature sequence;

[0038] Based on the normalized causal feature sequence, a full-link disturbance response aggregation mechanism is adopted to perform weighted summation of the disturbance response feature data of the local feature set of each link at each time point to obtain the full-link disturbance response feature at that time point. The full-link disturbance response feature is then superimposed on the corresponding time point causal feature vector of the normalized causal feature sequence to form a causal feature sequence containing the full-link disturbance response feature.

[0039] Principal component analysis is applied to causal feature sequences containing end-to-end perturbation response characteristics to generate dimensionality-reduced causal feature chains, including:

[0040] Calculate the covariance matrix of the normalized causal feature sequence and the end-link disturbance response feature;

[0041] Eigenvalue decomposition is performed on the covariance matrix to obtain eigenvalues ​​and corresponding eigenvectors;

[0042] Sort the eigenvalues ​​by size and select the top-ranked eigenvectors to form an eigenvector matrix.

[0043] Multiply the eigenvector matrix by the eigenvectors at each time point of the normalized causal feature sequence to obtain the principal component eigenvectors. The principal component eigenvectors at all time points form the principal component feature sequence.

[0044] The principal component feature sequence is used as the causal feature chain after dimensionality reduction.

[0045] Optionally, the generation of the optimized feature chain includes the following specific steps:

[0046] Input the causal feature chain and split it into a sequence of causal feature vectors at multiple time steps;

[0047] An attention mechanism is applied to the causal feature vector sequence to calculate the weight coefficient of the causal feature vector at each time step;

[0048] The weighting coefficients are obtained by first performing a linear transformation and adding a bias on the causal feature vector, then applying the hyperbolic tangent activation function, then performing an inner product with the attention vector, and finally normalizing it using an exponential function.

[0049] Based on the calculated weight coefficients, the causal feature vectors at each time step in the causal feature vector sequence are weighted and summed to obtain an overall weighted causal feature representation.

[0050] The weighted causal feature representation is subjected to sparse coding for feature selection and dimensionality reduction to obtain a sparse coefficient vector. The sparse coefficient vector is then used as the optimized causal feature representation to further generate an optimized feature chain.

[0051] Optionally, the real-time visualization includes the following specific steps:

[0052] The principal component feature sequence obtained by dimensionality reduction of the optimized feature chain through principal component analysis, and the causal dependency weight matrix are used as data inputs to the visualization platform. The principal component feature sequence reflects the main changing trend of the optimized feature chain, and the causal dependency weight matrix represents the degree of causal influence of each business link at different times.

[0053] Normalize the principal component feature sequences;

[0054] The causal dependency weight matrix is ​​visualized using a two-dimensional heatmap. The weight of each element in the matrix is ​​represented by the color intensity on the heatmap, and the heatmap is dynamically refreshed in chronological order.

[0055] The normalized principal component feature sequences and the causal dependency weight matrix in heatmap form are integrated and displayed on the same visualization platform.

[0056] The beneficial effects of this invention are:

[0057] This invention provides an industrial big data processing system based on the industrial supply chain. Through a modular architecture covering the entire process, it overcomes the limitations of existing technologies in multi-source heterogeneous data integration, intelligent feature modeling, causal analysis, intelligent reasoning, and visualization analysis. It achieves intelligent control and optimization of the entire industrial supply chain. The system can automatically collect multi-source data from procurement, production, warehousing, logistics, and sales. Through unified data cleaning, normalization, noise reduction, and time synchronization preprocessing, it standardizes and integrates structured and unstructured data, improving the efficiency and quality of data integration and providing a solid foundation for subsequent feature modeling. Relying on a self-supervised data twin, business data from each stage can be dynamically modeled in an unsupervised environment, automatically extracting key features, latent space representations, and anomaly labels. This enhances the ability to perceive complex supply chain business states. Furthermore, through anomaly label fusion and feature perturbation mechanisms, it enables timely detection and response to anomaly risks.

[0058] Furthermore, in terms of feature chain construction, this invention introduces the causal weight matrix and Granger causality test mechanism into the entire supply chain process. Through dynamic causal weight adjustment and full-link disturbance response aggregation, it effectively characterizes the multi-level dynamic dependencies between various business links, overcoming the shortcomings of traditional methods that can only perform unidirectional and static modeling. The combination of dimensionality reduction processing and attention mechanism not only achieves efficient compression and redundancy filtering of causal feature chains, but also improves the model's attention to and discrimination ability of key factors in the supply chain. Furthermore, the optimized feature chain is input into a multilayer perceptron for multi-objective reasoning. The system can simultaneously output demand forecasting, inventory allocation, production scheduling, and risk warning results, realizing collaborative intelligent decision-making for multiple links and multiple objectives in the supply chain. This provides strong technical support for enterprises to optimize resource allocation and risk control in uncertain environments.

[0059] Finally, this invention also integrates the visualization of principal component feature chains and causal dependency weight matrices, enabling business personnel to gain real-time insights into key indicators and dynamic changes in causal weights across the entire supply chain on a visualization platform, thereby improving the transparency and decision-making efficiency of supply chain management. Attached Figure Description

[0060] 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:

[0061] Figure 1 This is a schematic diagram of the overall structure of an industrial big data processing system based on the industrial supply chain proposed in this invention.

[0062] Figure 2 This is a schematic diagram illustrating the process of constructing and optimizing the causal feature chain and generating the feature chain for an industrial big data processing system based on the industrial supply chain proposed in this invention. Detailed Implementation

[0063] 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.

[0064] refer to Figure 1-2 An industrial big data processing system based on the industrial supply chain includes:

[0065] The data acquisition and preprocessing module is used to collect multi-source heterogeneous data from the procurement, production, warehousing, logistics and sales processes to obtain raw business data, and to preprocess the raw business data to form standardized business data.

[0066] The feature modeling module is used to build a self-supervised data twin for each business process, dynamically model the corresponding business process using standardized business data, and generate a local feature set through a self-supervised mechanism.

[0067] The causal feature chain construction module is used to causally concatenate the local feature sets of each business link according to business dependencies and time order, and use the end-to-end disturbance response aggregation mechanism to form a causal feature sequence containing end-to-end disturbance response features, while performing dimensionality reduction.

[0068] The optimized feature generation module is used to process the causal feature chain using an attention mechanism, assign weights to features in each business process, and generate an optimized feature chain.

[0069] The decision reasoning module is used to input the optimized feature chain into the multilayer perceptron and output demand forecast data, inventory allocation data, production scheduling data, and risk warning data.

[0070] The visualization module is used to visualize and display the principal component feature sequences and causal dependency weight matrices in real time during the optimization of the feature chain.

[0071] This invention uses a modular approach to uniformly collect and preprocess heterogeneous data from multiple sources in procurement, production, warehousing, logistics, and sales, improving data integration efficiency and standardization. The system achieves automatic fusion of structured and unstructured data, providing high-quality basic data for intelligent analysis and collaborative optimization of the entire supply chain. Through end-to-end data standardization processes, it provides strong support for data-driven supply chain optimization.

[0072] In this embodiment, the modules are interconnected using the following method:

[0073] Collect multi-source heterogeneous data from procurement, production, warehousing, logistics and sales processes to obtain raw business data, and preprocess the raw business data to form standardized business data;

[0074] A self-supervised data twin is established for each stage, and standardized business data is used to dynamically model the corresponding business stages. A local feature set is generated through a self-supervised mechanism.

[0075] According to the supply chain business process, the local feature sets of each business link are causally spliced ​​together according to business dependencies and time sequence to construct a causal feature chain;

[0076] The causal feature chain is processed by an attention mechanism, which assigns weights to features in each business process, performs feature redundancy and noise filtering, and generates an optimized feature chain.

[0077] The optimized feature chain is input into the multilayer perceptron, and the output includes demand forecast data, inventory allocation data, production scheduling data, and risk warning data.

[0078] The principal component feature sequences and causal dependency weight matrices during the feature chain optimization process will be visualized and displayed in real time.

[0079] This invention ensures efficient collaboration between various stages, from data acquisition and preprocessing to feature modeling, causal analysis, decision reasoning, and visualization, through clearly defined data flows and processing logic between modules. The systematic methodology enables the orderly flow and seamless connection of business data from different types and sources, effectively improving the overall controllability and scalability of the system. The closed-loop data interaction between modules achieves real-time synchronization and efficient transmission of information across the entire chain, providing solid technical support for intelligent management and dynamic optimization of multiple stages in complex supply chain scenarios.

[0080] In this embodiment, the original business data includes structured data and unstructured data. The structured data includes tabular data, log data, and sensor sequence data. The unstructured data includes text data, image data, and audio data. The preprocessing includes data cleaning, format unification, missing value imputation, noise reduction, normalization, and time synchronization and alignment.

[0081] This invention integrates structured and unstructured data into a unified data preprocessing workflow, achieving efficient cleaning, standardized formatting, noise reduction, and time alignment of various types of business data, effectively eliminating barriers between heterogeneous data. Combined with techniques such as missing value imputation, it improves the quality and completeness of raw business data. The standardized processing of multi-source data provides a solid foundation for subsequent feature modeling and causal chain construction, ensuring the accuracy and continuity of data flow throughout the supply chain and creating favorable conditions for intelligent analysis and decision-making.

[0082] In this embodiment, the generation of the local feature set includes the following specific steps:

[0083] Establish self-monitoring data twins for each of the procurement, production, warehousing, logistics and sales stages, and input standardized business data into the corresponding self-monitoring data twins in chronological order;

[0084] The self-supervised data twin refers to a model unit that takes standardized business data as input, performs time-series modeling and feature encoding on the standardized business data through a self-supervised learning mechanism, and outputs anomaly labels and latent space feature data.

[0085] Each self-supervised data twin uses a time series modeling method on the input standardized business data, combines historical standardized business data with current standardized business data, predicts the standardized business data at the next time point, and simultaneously inputs the current standardized business data into the autoencoder to generate latent space feature data, which corresponds one-to-one with the standardized business data.

[0086] Calculate the prediction error between the predicted standardized business data and the actual standardized business data, and calculate the autoencoder reconstruction error. Compare the prediction error with the prediction threshold and the reconstruction error with the reconstruction threshold. If any error exceeds the corresponding threshold, the standardized business data is marked as abnormal; otherwise, it is marked as normal.

[0087] The anomaly label refers to the mark made to determine whether the data is abnormal or normal based on the comparison result of the prediction error and reconstruction error of each standardized business data with the corresponding threshold. The anomaly label is a binary label, with 1 for abnormal and 0 for normal.

[0088] For each business process, a weighted fusion method is used to assign fusion weights to the abnormal labels of the previous business process, the current business process, and the next business process. The three are multiplied by the fusion weights and then summed. If the fusion result is greater than the fusion threshold, the abnormal label of the current standardized business data is updated to abnormal; otherwise, it is updated to normal. The fusion result is used as the fused abnormal label data.

[0089] For standardized business data that is fused with abnormal label data, feature perturbation data is generated. The perturbation term is superimposed on the standardized business data and then input into the autoencoder to generate perturbed latent space feature data. The difference between the latent space feature data before and after perturbation is used to obtain the perturbation response feature data.

[0090] The characteristic perturbation data is generated using random noise that follows a normal distribution;

[0091] This invention effectively simulates abnormal disturbances or sudden events that the supply chain may encounter in actual operation by superimposing disturbance terms onto standardized business data and inputting it into an autoencoder. By subtracting the latent space feature data before and after the disturbance, the resulting disturbance response feature data accurately reflects the sensitivity and response mechanism of the business system to external or internal disturbances. This approach not only enhances the model's ability to perceive potential abnormal states and risk signals, but also provides richer and more detailed feature information for subsequent causal modeling and dynamic risk analysis. Through this mechanism, the system can dynamically capture and quantify the transmission and impact of disturbances in various business links, improving the sensitivity of anomaly detection and the foresight of early warning, and providing strong data support for intelligent control and emergency management of the supply chain.

[0092] The current standardized business data, latent space feature data, fused anomaly label data, and disturbance response feature data are merged into a local feature set in chronological order.

[0093] This invention constructs self-supervised data twins at each stage of the supply chain, combining temporal modeling and autoencoder techniques to achieve dynamic feature extraction, anomaly identification, and latent space modeling of standardized business data. The integration of anomaly labeling and disturbance response mechanisms effectively improves the sensitivity to detecting abnormal business states. The weighted fusion of anomaly labels from multiple stages and the introduction of feature-based disturbance responses enhance the model's ability to identify complex supply chain state changes and risk signals, providing a scientific basis for dynamic risk monitoring and adaptive business adjustments throughout the entire process.

[0094] In this embodiment, the construction of the causal feature chain includes the following specific steps:

[0095] Arrange the local feature sets of each business process in chronological order;

[0096] Based on the industrial supply chain business process, a causal weight matrix is ​​set. The causal weight matrix is ​​a fifth-order square matrix, and each element represents the causal dependence weight of one business link on another business link.

[0097] The causal dependency weights are obtained by performing Granger causality tests on historical standardized business data of each business link to obtain a significance index of the causal relationship between the links, and the significance index is normalized as the causal dependency weights.

[0098] Based on the local feature set of all business links, for each time point, according to the fused anomaly label data and disturbance response feature data in each local feature set, the corresponding causal dependency weight in the causal weight matrix is ​​dynamically adjusted so that the business links with significant changes in fused anomaly label data or disturbance response feature data can obtain higher causal dependency weight in the causal weight matrix.

[0099] The dynamic adjustment of the causal dependency weights in the causal weight matrix specifically includes:

[0100] Within a preset time window, calculate the average value of the fusion anomaly label data and the average norm of the disturbance response features for each business process.

[0101] The adjustment coefficient is calculated by weighted linear combination, and the adjustment coefficient is the sum of the average value of the fused anomaly labels multiplied by the first weight parameter and the average norm of the disturbance response features multiplied by the second weight parameter;

[0102] The initial weights of the corresponding elements in the causal weight matrix are multiplied by the adjustment coefficients to obtain the adjusted weight values.

[0103] Normalize each column of the adjusted causal weight matrix;

[0104] The local feature set of each link is multiplied by the adaptively optimized causal dependency weight at the time point, and the five weighted feature sets are summed to obtain the original causal feature vector at the time point. The original causal feature vectors of each time point are arranged in sequence to form the original causal feature sequence.

[0105] The original causal feature sequence is normalized to generate a normalized causal feature sequence;

[0106] Based on the normalized causal feature sequence, a full-link disturbance response aggregation mechanism is adopted to perform weighted summation of the disturbance response feature data of the local feature set of each link at each time point to obtain the full-link disturbance response feature at that time point. The full-link disturbance response feature is then superimposed on the corresponding time point causal feature vector of the normalized causal feature sequence to form a causal feature sequence containing the full-link disturbance response feature.

[0107] The end-to-end disturbance response aggregation mechanism in this invention can dynamically integrate the disturbance response characteristics of each business link in the supply chain when data anomalies, risk events, or external disturbances occur. This mechanism effectively integrates the link-level impact of local anomaly labels and disturbances by weighted summation of the disturbance response characteristic data of each link at each time point, transforming scattered risk signals into a comprehensive response index across the entire chain. This not only accurately reflects the propagation path and impact range of disturbance events in the entire supply chain network, but also provides precise data support for anomaly early warning, risk transmission analysis, and dynamic adjustment of causal weights.

[0108] Principal component analysis is applied to causal feature sequences containing end-to-end perturbation response characteristics to generate dimensionality-reduced causal feature chains, including:

[0109] Calculate the covariance matrix of the normalized causal feature sequence and the end-link disturbance response feature;

[0110] Eigenvalue decomposition is performed on the covariance matrix to obtain eigenvalues ​​and corresponding eigenvectors;

[0111] Sort the eigenvalues ​​by size and select the top-ranked eigenvectors to form an eigenvector matrix.

[0112] Multiply the eigenvector matrix by the eigenvectors at each time point of the normalized causal feature sequence to obtain the principal component eigenvectors. The principal component eigenvectors at all time points form the principal component feature sequence.

[0113] The principal component feature sequence is used as the dimensionality-reduced causal feature chain, which describes the linear combination of normalized causal features and end-link perturbation response features in the principal component space.

[0114] This invention systematically models the causal relationships between various business processes through a causal weight matrix and Granger causality testing mechanism. Combined with dynamic adjustments to integrate anomaly label data and disturbance response feature data, it can capture and characterize complex dependencies and disturbance propagation between multiple processes in real time. Principal component analysis for dimensionality reduction and full-link feature aggregation effectively compress the high-dimensional feature space, highlighting key business control factors. This solution enhances the supply chain system's ability to model and respond to multi-source dynamic dependencies and optimizes data-driven risk identification and anomaly early warning effects.

[0115] In this embodiment, the generation of the optimized feature chain includes the following specific steps:

[0116] Input the causal feature chain and split it into a sequence of causal feature vectors at multiple time steps;

[0117] An attention mechanism is applied to the causal feature vector sequence to calculate the weight coefficient of the causal feature vector at each time step;

[0118] The weighting coefficients are obtained by first performing a linear transformation and adding a bias on the causal feature vector, then applying the hyperbolic tangent activation function, then performing an inner product with the attention vector, and finally normalizing it using an exponential function.

[0119] Based on the calculated weight coefficients, the causal feature vectors at each time step in the causal feature vector sequence are weighted and summed to obtain an overall weighted causal feature representation.

[0120] The weighted causal feature representation is subjected to sparse coding for feature selection and dimensionality reduction to obtain a sparse coefficient vector. The sparse coefficient vector is then used as the optimized causal feature representation to further generate an optimized feature chain.

[0121] This invention introduces an attention mechanism into the causal feature chain, enabling dynamic allocation of feature contributions from each business link in the supply chain. The sparse coding method completes feature selection and dimensionality reduction based on weighted feature representation, further removing irrelevant or redundant information, improving the model's ability to extract key features, optimizing feature chain generation, and enhancing the adaptability and generalization ability of the subsequent inference model to complex business state changes in the supply chain. This provides a high-quality input foundation for multi-objective collaborative optimization and intelligent decision-making.

[0122] In this embodiment, the real-time visualization display includes the following specific steps:

[0123] The principal component feature sequence obtained by dimensionality reduction of the optimized feature chain through principal component analysis, and the causal dependency weight matrix are used as data inputs to the visualization platform. The principal component feature sequence reflects the main changing trend of the optimized feature chain, and the causal dependency weight matrix represents the degree of causal influence of each business link at different times.

[0124] Normalize the principal component feature sequences;

[0125] The causal dependency weight matrix is ​​visualized using a two-dimensional heatmap. The weight of each element in the matrix is ​​represented by the color intensity on the heatmap, and the heatmap is dynamically refreshed in chronological order to show the trend of the causal weight of each business link changing over time.

[0126] The normalized principal component feature sequences and the causal dependency weight matrix in heatmap form are integrated and displayed on the same visualization platform. The platform supports multi-dimensional interactive queries, data retrieval by time and business process, and real-time synchronization and updating of all visualization content.

[0127] This invention improves the transparency of supply chain data and the scientific nature of management decisions by visualizing and displaying the principal component feature sequences and causal dependency weight matrices obtained after dimensionality reduction of the optimized feature chain through principal component analysis in real time. The normalized principal component feature sequences can intuitively reflect the dynamic change trend of the optimized feature chain, enabling managers to grasp the temporal evolution of key features across the entire chain in a timely manner. The causal dependency weight matrix is ​​displayed in a two-dimensional heatmap format, which not only clearly shows the dynamic changes in the degree of causal influence of each business link, but also supports real-time updates of the time series. Example

[0128] To verify the feasibility of this invention in a real industrial environment, it was applied to the supply chain intelligent management system of a home appliance manufacturing company. This company encompasses multiple stages including procurement, production, warehousing, logistics, and sales. Its daily operations generate a large amount of structured and unstructured data, including order records, production plans, equipment operation logs, sensor data, logistics documents, warehouse status, sales reports, and text, image, and voice information generated during business communications. Due to the numerous business stages and the diverse and varied data sources and formats, problems such as data silos, limited information sharing, slow predictive response, and difficulty in accurately identifying risks have long existed, directly impacting production scheduling, inventory optimization, and market response speed.

[0129] In applying this invention, the original business data of each link in procurement, production, warehousing, logistics and sales are automatically acquired through the acquisition and preprocessing module. All data are then converted, cleaned, denoised and normalized in a unified format to form high-quality standardized business data. Self-supervised data twins are deployed in each business link to perform dynamic feature modeling on the standardized data, extract key features, identify abnormal states, and generate latent space features and anomaly labels. By introducing the fusion of anomaly labels and disturbance response mechanisms, early perception of potential risks and sudden anomalies is achieved.

[0130] Subsequently, the system constructs a causal weight matrix based on the actual business processes of the enterprise's supply chain, uses historical data to verify causal relationships, and realizes causal splicing between local feature sets to form a causal feature chain. The causal weights can be dynamically adjusted according to the business status of each link, and have excellent responsiveness to sudden events and risk propagation. The causal feature chain is subjected to full-link perturbation response aggregation and principal component analysis dimensionality reduction to effectively extract the main control features of the supply chain, reduce data redundancy, and improve the effectiveness of feature representation. The feature chain re-input attention mechanism and sparse coding module are optimized to further focus on the key variables that have the greatest impact on actual decision-making, thereby improving intelligent reasoning and model generalization capabilities.

[0131] In actual production management, enterprises will optimize the feature chain input multilayer perceptron model. The system can make scientific predictions on future raw material demand, production capacity allocation, inventory scheduling and logistics distribution, and identify abnormal states and potential risks in the supply chain in advance. Through the visualization of principal component feature chains and causal dependency weight matrices, enterprise management can intuitively grasp the dynamic changes and causal dependencies of key indicators in each business link of the entire chain on the visualization platform, realizing full-process automation from data access, intelligent modeling, dynamic reasoning to decision visualization.

[0132] To verify the performance of the present invention in practice, it was compared with traditional methods, and the comparison results are shown in Table 1.

[0133] Table 1. Performance Comparison of Big Data Intelligent Processing Methods and Traditional Methods Based on Industrial Supply Chains

[0134] Comparison indicators Traditional methods Method of the present invention Prediction accuracy 78.6% 91.8% Improved inventory turnover efficiency 12.4% 28.7% Risk identification recall rate 65.3% 84.6% Model response time (seconds) 3.92 1.45 Anomaly detection false alarm rate 8.7% 3.2% Feature Dimension Compression Rate 35.1% 68.9%

[0135] In terms of prediction accuracy, the method of this invention is superior to the traditional method. The prediction accuracy of the traditional method is 78.6%, while the method of this invention reaches 91.8%, which is 13.2 percentage points higher. This improvement is due to the introduction of causal feature chain modeling and multi-source data fusion mechanism, which makes the model more accurate in modeling the correlation between upstream and downstream links, thereby enhancing the reliability and consistency of the overall prediction.

[0136] In terms of improving inventory turnover efficiency, the method of this invention achieved 28.7%, higher than the 12.4% of the traditional method. This optimization stems from the inference strategy that combines optimized feature chains with multilayer perceptrons, which can accurately assess the supply and demand status of each business node, thereby effectively guiding inventory flow and scheduling decisions, reducing resource redundancy and inventory backlog, and improving supply chain response speed.

[0137] The risk identification recall rate metric shows that the method of this invention is 84.6%, which is an improvement compared to the 65.3% of the traditional method. This improvement comes from the abnormal labeling mechanism of the self-supervised data twin and the enhanced expressive power of the disturbance response features, which enables the model to more comprehensively identify potential abnormal business states and enhance the system's security early warning capability in complex situations.

[0138] In terms of model response time, this invention significantly reduces it to 1.45 seconds, while the traditional method has a response time of 3.92 seconds. This invention achieves efficient screening through causal dependency matrix and feature sparsity processing, reducing the feature input dimension and model burden, thereby improving inference efficiency and real-time response capability.

[0139] The false alarm rate of the present invention is only 3.2%, which is lower than the 8.7% of the traditional method. This improvement is due to the multi-angle modeling of abnormal states by integrating anomaly labels and disturbance response mechanisms, which reduces the risk of misjudgment and improves the stability and reliability of the monitoring system.

[0140] Finally, in terms of feature dimension compression rate, this invention achieves a compression of 68.9%, which is more advantageous than the 35.1% of the traditional method. Thanks to the principal component analysis combined with the perturbation response aggregation strategy, the system can significantly reduce feature redundancy while maintaining the effectiveness of feature representation, thereby improving the model's generalization ability and saving computational resources.

[0141] 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. An industrial big data processing system based on the industrial supply chain, characterized in that, include: The data acquisition and preprocessing module is used to collect multi-source heterogeneous data from the procurement, production, warehousing, logistics and sales processes to obtain raw business data, and to preprocess the raw business data to form standardized business data. The feature modeling module is used to build a self-supervised data twin for each business process, dynamically model the corresponding business process using standardized business data, and generate a local feature set through a self-supervised mechanism. The causal feature chain construction module is used to causally concatenate the local feature sets of each business link according to business dependencies and time order, and use the end-to-end disturbance response aggregation mechanism to form a causal feature sequence containing end-to-end disturbance response features, while performing dimensionality reduction. The optimized feature generation module is used to process the causal feature chain using an attention mechanism, assign weights to features in each business process, and generate an optimized feature chain. The decision reasoning module is used to input the optimized feature chain into the multilayer perceptron and output demand forecast data, inventory allocation data, production scheduling data, and risk warning data. The visualization module is used to visualize and display the principal component feature sequences and causal dependency weight matrices in real time during the feature chain optimization process. The generation of the local feature set includes the following specific steps: Establish self-monitoring data twins for each of the procurement, production, warehousing, logistics and sales stages, and input standardized business data into the corresponding self-monitoring data twins in chronological order; Each self-supervised data twin uses a time series modeling method on the input standardized business data, combines historical standardized business data with current standardized business data, predicts the standardized business data at the next time point, and simultaneously inputs the current standardized business data into the autoencoder to generate latent space feature data, which corresponds one-to-one with the standardized business data. Calculate the prediction error between the predicted standardized business data and the actual standardized business data, and calculate the autoencoder reconstruction error. Compare the prediction error with the prediction threshold and the reconstruction error with the reconstruction threshold. If any error exceeds the corresponding threshold, the standardized business data is marked as abnormal; otherwise, it is marked as normal. For each business process, a weighted fusion method is used to assign fusion weights to the abnormal labels of the previous business process, the current business process, and the next business process. The three are multiplied by the fusion weights and then summed. If the fusion result is greater than the fusion threshold, the abnormal label of the current standardized business data is updated to abnormal; otherwise, it is updated to normal. The fusion result is used as the fused abnormal label data. For standardized business data that is fused with abnormal label data, feature perturbation data is generated. The perturbation term is superimposed on the standardized business data and then input into the autoencoder to generate perturbed latent space feature data. The difference between the latent space feature data before and after perturbation is used to obtain the perturbation response feature data. The current standardized business data, latent space feature data, fused anomaly label data, and disturbance response feature data are merged into a local feature set in chronological order.

2. The industrial big data processing system based on the industrial supply chain according to claim 1, characterized in that, The modules are connected in the following way: Collect multi-source heterogeneous data from procurement, production, warehousing, logistics and sales processes to obtain raw business data, and preprocess the raw business data to form standardized business data; A self-supervised data twin is established for each stage, and standardized business data is used to dynamically model the corresponding business stages. A local feature set is generated through a self-supervised mechanism. According to the supply chain business process, the local feature sets of each business link are causally spliced ​​together according to business dependencies and time sequence to construct a causal feature chain; The causal feature chain is processed by an attention mechanism, which assigns weights to features in each business process, performs feature redundancy and noise filtering, and generates an optimized feature chain. The optimized feature chain is input into the multilayer perceptron, and the output includes demand forecast data, inventory allocation data, production scheduling data, and risk warning data. The principal component feature sequences and causal dependency weight matrices during the feature chain optimization process will be visualized and displayed in real time.

3. The industrial big data processing system based on the industrial supply chain according to claim 2, characterized in that, The raw business data includes structured and unstructured data. The structured data includes tabular data, log data, and sensor sequence data. The unstructured data includes text data, image data, and audio data. The preprocessing includes data cleaning, format unification, missing value imputation, noise reduction, normalization, and time synchronization and alignment.

4. The industrial big data processing system based on the industrial supply chain according to claim 2, characterized in that, The construction of the causal feature chain includes the following specific steps: Arrange the local feature sets of each business process in chronological order; Based on the industry supply chain business processes, a causal weight matrix is ​​established; Based on the local feature sets of all business processes, for each time point, the corresponding causal dependency weights in the causal weight matrix are dynamically adjusted according to the fused anomaly label data and disturbance response feature data in each local feature set. The dynamic adjustment of the causal dependency weights in the causal weight matrix specifically includes: Within a preset time window, calculate the average value of the fusion anomaly label data and the average norm of the disturbance response features for each business process. The adjustment coefficient is calculated by weighted linear combination, and the adjustment coefficient is the sum of the average value of the fused anomaly labels multiplied by the first weight parameter and the average norm of the disturbance response features multiplied by the second weight parameter; The initial weights of the corresponding elements in the causal weight matrix are multiplied by the adjustment coefficients to obtain the adjusted weight values. Normalize each column of the adjusted causal weight matrix; The local feature set of each link is multiplied by the adaptively optimized causal dependency weight at the time point, and the five weighted feature sets are summed to obtain the original causal feature vector at the time point. The original causal feature vectors of each time point are arranged in sequence to form the original causal feature sequence. The original causal feature sequence is normalized to generate a normalized causal feature sequence; Based on the normalized causal feature sequence, a full-link disturbance response aggregation mechanism is adopted to perform weighted summation of the disturbance response feature data of the local feature set of each link at each time point to obtain the full-link disturbance response feature at that time point. The full-link disturbance response feature is then superimposed on the corresponding time point causal feature vector of the normalized causal feature sequence to form a causal feature sequence containing the full-link disturbance response feature. Principal component analysis is applied to causal feature sequences containing end-to-end perturbation response characteristics to generate dimensionality-reduced causal feature chains, including: Calculate the covariance matrix of the normalized causal feature sequence and the end-link disturbance response feature; Eigenvalue decomposition is performed on the covariance matrix to obtain eigenvalues ​​and corresponding eigenvectors; Sort the eigenvalues ​​by size and select the top-ranked eigenvectors to form an eigenvector matrix. Multiply the eigenvector matrix by the eigenvectors at each time point of the normalized causal feature sequence to obtain the principal component eigenvectors. The principal component eigenvectors at all time points form the principal component feature sequence. The principal component feature sequence is used as the causal feature chain after dimensionality reduction.

5. The industrial big data processing system based on the industrial supply chain according to claim 2, characterized in that, The generation of the optimized feature chain includes the following specific steps: Input the causal feature chain and split it into a sequence of causal feature vectors at multiple time steps; An attention mechanism is applied to the causal feature vector sequence to calculate the weight coefficient of the causal feature vector at each time step; The weighting coefficients are obtained by first performing a linear transformation and adding a bias on the causal feature vector, then applying the hyperbolic tangent activation function, then performing an inner product with the attention vector, and finally normalizing it using an exponential function. Based on the calculated weight coefficients, the causal feature vectors at each time step in the causal feature vector sequence are weighted and summed to obtain an overall weighted causal feature representation. The weighted causal feature representation is subjected to sparse coding for feature selection and dimensionality reduction to obtain a sparse coefficient vector. The sparse coefficient vector is then used as the optimized causal feature representation to further generate an optimized feature chain.

6. The industrial big data processing system based on the industrial supply chain according to claim 2, characterized in that, The real-time visualization process includes the following specific steps: The principal component feature sequence obtained by dimensionality reduction of the optimized feature chain through principal component analysis, and the causal dependency weight matrix are used as data inputs to the visualization platform. The principal component feature sequence reflects the main changing trend of the optimized feature chain, and the causal dependency weight matrix represents the degree of causal influence of each business link at different times. Normalize the principal component feature sequences; The causal dependency weight matrix is ​​visualized using a two-dimensional heatmap. The weight of each element in the matrix is ​​represented by the color intensity on the heatmap, and the heatmap is dynamically refreshed in chronological order. The normalized principal component feature sequences and the causal dependency weight matrix in heatmap form are integrated and displayed on the same visualization platform.

Images