Supply chain data analysis system based on deep learning technology

Abstract

The invention relates to the technical field of data analysis, in particular to a supply chain data analysis system based on a deep learning technology. The system comprises a multi-modal causal characterization unit, a dynamic digital twinning unit and an AI control compiling unit. According to the method, multi-modal data is fused into a unified framework, the problems of manual definition and low timeliness are solved, the problem of data fault is solved, visual causal path visualization is constructed, potential risks in a supply chain are actively prevented and controlled, the decision accuracy is improved, and the method is suitable for popularization and application. Then, the dynamic digital twinning unit reduces cross-link error accumulation through a dynamic prediction and updating mechanism, so that an enterprise can quickly respond when a change occurs, the supply chain loss is reduced, finally, the AI control compiling unit is adopted to efficiently make a decision and execute, the problems of slow progress and inaccuracy caused by manual operation are reduced, and the efficiency is improved. And a bidirectional cognitive channel between the supply chain management system and the IA equipment is realized.

Description

Technical Field

[0001] The present invention relates to the field of data analysis technology, and in particular to a supply chain data analysis system based on deep learning technology. Background Art

[0002] As the complexity of global supply chains increases dramatically, traditional data analysis methods face severe challenges in dealing with multi-dimensional and highly dynamic supply chain scenarios. The existing technology architecture has three faults:

[0003] First, data discontinuity: Multi-source heterogeneous data (structured databases, unstructured documents, and real-time sensor streams) is difficult to represent uniformly. Traditional ETL tools (such as Informatica) require manual definition of database schemas when integrating unstructured text (contracts, reports) with sensor stream data, resulting in poor timeliness.

[0004] Second, model gaps: Single-task models break the deep connections between supply chain links (such as the implicit impact of procurement decisions on logistics carbon emissions) and fail to model the causal chain of procurement-production-logistics, leading to the accumulation of errors across links.

[0005] Third, execution fault: The lack of a two-way cognitive channel between the analysis system and IA equipment leads to decision lags and accumulated execution deviations. In view of this, through the three-layer architecture innovation of multimodal causal representation learning, dynamic digital twins and cognitive IA interfaces, we can achieve a paradigm shift in supply chain analysis from "post-statistics" to "pre-deduction" and from "human assistance" to "autonomous decision-making". We propose a supply chain data analysis system based on deep learning technology. Summary of the Invention

[0006] The purpose of the present invention is to provide a supply chain data analysis system based on deep learning technology to solve the problems raised in the above background technology.

[0007] To achieve the above objectives, the present invention provides a supply chain data analysis system based on deep learning technology, including a multimodal causal representation unit, a dynamic digital twin unit, and an AI control compilation unit;

[0008] The multimodal causal representation unit is used to fuse structured data, unstructured text and real-time streaming data, and construct a cross-modal causal graph to reveal the causal relationship between different data modalities;

[0009] The dynamic digital twin unit uses a dual-clock driving mechanism to build a supply chain status model based on the causal relationship between the multimodal data output by the multimodal causal representation unit, inputs any multimodal data into the supply chain status model, outputs the future status of the supply chain, and presets minute-level increments to dynamically deduce the status of future time periods;

[0010] The AI control compilation unit is used to establish a linear relationship between the future state and the decision-making instructions, call out the decision-making instructions corresponding to the future state of the dynamic digital twin unit in the required time period, compile the decision-making instructions into equipment control parameters through the industrial knowledge graph and verify the execution feasibility, and feed back the execution results to the multimodal causal representation unit.

[0011] As a further improvement of the present technical solution, the multimodal causal representation unit includes a cross-modal fusion module and a causal discovery module;

[0012] The cross-modal fusion module is used to collect data normalization, extract semantic features, use a deep learning model to map data from different modalities into a unified feature space, and capture the association between data from different modalities through an attention mechanism;

[0013] The causal discovery module is used to apply a gradient-intervention causal discovery algorithm to calculate the causal relationship between data of different modalities to generate a cross-modal causal graph, and calculate the average causal effect under gradient intervention through a Bayesian network to form a causal relationship matrix between variables.

[0014] As a further improvement of the present technical solution, the causal discovery module also includes a multi-reference path assessment module, which is used to traverse the causal relationship matrix to collect all paths within a length threshold, calculate the path score based on a comprehensive scoring index, and the scoring index includes causal strength, modal consistency and timeliness weights, construct a Pareto front, and select a non-inferior solution path set.

[0015] As a further improvement of this technical solution, building a supply chain status model in the dynamic digital twin unit includes the following steps:

[0016] The historical data of multiple variables are collected in chronological order, as well as the historical data of other variables in the non-inferior solution path set in their causal relationship to form a training set. Each sample includes the historical data of the input variable and the future state of the target variable. The supply chain status model is trained to reflect the multimodal data and the future state of the supply chain on one of the non-inferior solution paths. Any variable is input into the supply chain status model to predict the current state of other variables in the non-inferior solution path set. Multiple data of variables within a time period are input into the supply chain status model to predict the target data at the bottom of the non-inferior solution path in the future time period to reflect the future state of the supply chain.

[0017] As a further improvement of this technical solution, the dynamic digital twin unit also includes a prediction period input module, which is used to input a time period step that is smaller than the future time period into the supply chain state model and call out the future state of the supply chain that matches the time period step.

[0018] As a further improvement of the present technical solution, the AI control compilation unit includes a decision instruction establishment module and an AI control execution module;

[0019] The decision instruction establishment module is used to collect decision instructions for multi-source data from historical data as a sample set, use a regression model to train the sample set, map the multi-source data state to the decision instruction, and output the decision instruction corresponding to the current state of other variables on the non-inferior solution path set output by the dynamic digital twin unit, as well as the decision instruction corresponding to the future state of the supply chain;

[0020] The control execution module is used to integrate knowledge in the industrial field to construct a knowledge graph, including equipment parameters, operating manuals, process flows, etc., use semantic relationships to represent decision instructions, and establish mapping rules from decision instructions to equipment control parameters in the knowledge graph, so that AI can automatically drive the corresponding equipment according to the control parameters.

[0021] As a further improvement of the present technical solution, the AI control compilation unit also includes an AI decision-making module. When the AI decision-making module receives the decision instruction from the decision instruction establishment module, it converts the control parameters into intelligent voice feedback to the user through AI. If the user feedback is in agreement, the control execution module is driven to execute the corresponding control parameters. If the user feedback is paused, the drive of the decision instruction establishment module is intercepted.

[0022] As a further improvement of the present technical solution, the dynamic digital twin unit also includes a variable data analysis module, which is used to collect the non-inferior solution path positions of the variables input into the supply chain status model, and quantify the degree of influence of the current variables on other variables in the non-inferior solution path set, so that the dynamic digital twin unit predicts the current status of other variables on the non-inferior solution path set and sorts them according to the degree of influence.

[0023] Compared with the prior art, the present invention has the following beneficial effects:

[0024] In this supply chain data analysis system based on deep learning technology, the multimodal causal representation unit can process structured data, unstructured text and real-time streaming data, and integrate them into a unified framework, solving the problems of manual definition and low timeliness of traditional ETL tools, and overcoming the data gap problem, providing strong support for supply chain risk warning, and building intuitive causal path visualization, which helps to proactively prevent and control potential risks in the supply chain and improve the accuracy of decision-making. Then, the dynamic digital twin unit reduces the accumulation of cross-link errors through dynamic prediction and update mechanisms, enabling enterprises to respond quickly when changes occur and reduce supply chain losses. Finally, AI is used to control the compilation unit for efficient decision-making and execution, reducing the slow progress and inaccuracy caused by human operations, and realizing a two-way cognitive channel between the supply chain management system and IA equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a block diagram of the overall structural principle of the present invention;

[0026] Figure 2 The figure is a schematic diagram showing the overall structure of the present invention in detail.

[0027] The meaning of each number in the figure is:

[0028] 100. Multimodal causal representation unit; 110. Cross-modal fusion module; 120. Causal discovery module;

[0029] 200, dynamic digital twin unit;

[0030] 300, AI control compilation unit; 310, decision instruction establishment module; 320, AI control execution module. DETAILED DESCRIPTION

[0031] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.

[0032] Example 1

[0033] See also Figure 1-Figure 2 As shown, this embodiment provides a supply chain data analysis system based on deep learning technology, including a multimodal causal representation unit 100, a dynamic digital twin unit 200 and an AI control compilation unit 300;

[0034] The multimodal causal representation unit 100 is used to integrate structured data (such as order information in a database), unstructured text (such as contracts and reports), and real-time streaming data (such as sensor data), and construct a cross-modal causal graph to reveal the causal relationship between different data modalities (such as the mutual influence of procurement, production, logistics, etc.). This provides support for supply chain risk warning, avoids the manual definition and low timeliness of traditional ETL tools, and solves the data gap problem.

[0035] The multimodal causal representation unit 100 includes a cross-modal fusion module 110 and a causal discovery module 120;

[0036] The cross-modal fusion module 110 is used to collect and normalize data from structured databases, unstructured documents, and real-time sensors, extract semantic features, and specifically standardize structured data to ensure consistent data formats. It uses pre-trained models such as BERT to encode unstructured text, extract semantic features, perform time series analysis on real-time streaming data, identify patterns and anomalies, and prepare for subsequent analysis. It uses deep learning models (such as BERT, CNN, and LSTM) to map data from different modalities to a unified feature space and capture the association between data from each modality through an attention mechanism. The attention mechanism calculates the attention weight between data from different modalities. The formula is:

[0037] Among them, A(Q, K, V) is the weight value. Each modality is linearly transformed into Q, K, and V through an independent linear layer to ensure that the embedding spaces of different modalities are aligned, so that Q and K correspond to different modal data of linear transformation respectively. V is the weight pair value of modal data Q and modal data K. The cross-modal information is aggregated by weighting the weight pair value (V) and outputting the fusion features related to modal data Q and modal data K. k is the dimension of the modal data K, and T is the matrix transpose operation. The purpose is to align the dimensions to calculate cross-modal similarity and ultimately generate dynamic attention weights. The attention mechanism is a key step in achieving multimodal interaction and information fusion. It focuses on the modal data K and V through the modal data Q. Assuming that the modal data includes production plan adjustment and logistics delay risk, the attention mechanism focuses on the production plan adjustment through the logistics delay risk, calculates the similarity between the logistics delay risk and the production plan adjustment, and reflects the correlation strength between cross-modal elements, providing a data basis for the subsequent generation of cross-modal causal diagrams.

[0038] The causal discovery module 120 is used to apply the causal discovery algorithm of gradient intervention, calculate the causal relationship between data of different modalities to generate a cross-modal causal graph, and calculate the average causal effect under gradient intervention through the Bayesian network to form a causal relationship matrix between variables and identify key causal paths.

[0039] Specifically, the gradient intervention method is a causal discovery method that aims to estimate the causal influence between variables through gradient calculation. It combines gradient calculation in machine learning and intervention analysis in causal inference, and can effectively identify causal relationships between multimodal data. Suppose there is a deep learning model f(X, Z), where X is the intervention variable and Z is another variable that may affect Y. The intervention model is expressed as: Y = f(X, Z) + ε, where ε is the error term. The average causal effect (ACE) is calculated to calculate the expected change in Y when X is changed. The causal effect is estimated by calculating the gradient of Y with respect to X. The gradient represents the expected rate of change of Y when X changes. The causal influence of X on Y is evaluated based on the size of the gradient to form a causal adjacency matrix. The learned causal adjacency matrix is then converted into a Bayesian network structure. The top m paths with the largest information flow values are selected as critical paths, such as raw material price fluctuations → purchase order volume → production plan adjustments → logistics delay risks. This forms a critical causal path, provides intuitive causal path visualization, and facilitates proactive prevention and control of supply chain risks.

[0040] When multiple variables (multimodal data) act together on a supply chain, it is necessary to quantify the joint contribution of each variable to the result and identify the dominant causal path. Therefore, the causal discovery module 120 also includes a multi-reference path assessment module. The multi-reference path assessment module is used to traverse the causal relationship matrix to collect all paths within a length threshold, calculate the path score based on comprehensive scoring indicators, including causal strength, modal consistency, and timeliness weights, construct a Pareto frontier, select a set of non-inferior solution paths, that is, retain paths with high scores. By filtering and storing the causal graph, memory usage is saved, and an intuitive causal path display is provided to help identify key influencing factors in the supply chain. By predicting changes in the critical path, measures can be taken in advance to prevent and control potential risks, such as:

[0041] Critical Path 1: Raw material price fluctuations (structured) → Changes in purchase contract terms (text) → Supplier delivery delays (flow data) → Production plan adjustments (structured) → Logistics delay risk (target);

[0042] Critical Path 2: Abnormal equipment temperature (streaming data) → Decreased production line yield (structured) → Urgent replenishment demand (structured) → Logistics delay risk (target).

[0043] Next, the dynamic digital twin unit 200 uses a dual-clock driving mechanism to build a supply chain status model based on the causal relationship between the multimodal data output by the multimodal causal representation unit 100. It inputs any multimodal data into the supply chain status model, outputs the future state of the supply chain, and presets minute-level increments to dynamically update the state of the future time period, reducing the accumulation of errors across links and ensuring the accuracy and timeliness of dynamic updates.

[0044] like Figure 2 As shown, building a supply chain status model in the dynamic digital twin unit 200 includes the following steps:

[0045] Collect historical data of multiple variables in chronological order, as well as historical data of other variables in the non-inferior solution path set on its causal relationship to form a training set, ensuring that each data point has an accurate timestamp for subsequent time series analysis. Each sample includes the historical data of the input variable and the future state of the target variable. Train the supply chain status model to reflect the multimodal data and the future state of the supply chain on one of the non-inferior solution paths. Input any variable into the supply chain status model to predict the current state of other variables on the non-inferior solution path set. Input multiple data of variables within the time period into the supply chain status model to predict the target data at the bottom of the non-inferior solution path in the future time period to reflect the future state of the supply chain. For example, input the flow data of supplier delivery delays into the supply chain status model. On the one hand, it is predicted that the production plan adjustment structure Data and logistics delay risk data are convenient for timely real-time adjustment of the "result" behind the current variable on the non-inferior solution path to avoid the inability to change other links in advance due to problems in one link, resulting in supply chain losses. On the other hand, based on multiple data of variables in a certain time period, the changes of this variable in this time period are reflected and input into the supply chain status model to predict the future state of the supply chain. For example, multiple data of the supplier delivery delay data flow within 1 week are input into the supply chain status model. The target data of the logistics delay risk data in the historical data within a week is reflected to reflect the target data of the logistics delay risk data in the next week, that is, the target data of the logistics delay risk data in the future time period is output to reflect the future state of the supply chain, realize the prediction of the supply chain, and further prevent it in advance.

[0046] In order to further improve the mobility, the dynamic digital twin unit 200 also includes a prediction period input module. The prediction period input module is used to input a period step that is smaller than the future time period into the supply chain status model, and call out the future state of the supply chain that matches the period step. By setting the future period step, when it is necessary to understand the future period of a certain period, it is only necessary to directly call out the corresponding period step in the predicted future time period supply chain status (target data). There is no need to call out all the predicted data in large quantities, thereby reducing the running memory pressure. For example, the future time period of the future state of the future supply chain is 1 year, but currently only the data within one week needs to be analyzed for layout. Therefore, the input supply chain status model period step is one week, and the sample training set of the one-week time period is directly called out for training, thereby reducing the training operation intensity.

[0047] Furthermore, the AI control compilation unit 300 is used to establish a linear relationship between the future state and the decision-making instructions, call out the decision-making instructions corresponding to the future state of the dynamic digital twin unit 200 in the required time period, compile the decision-making instructions into equipment control parameters through the industrial knowledge graph and verify the execution feasibility, and feed back the execution results to the multimodal causal representation unit 100 to establish a two-way cognitive channel between the analysis system and the IA equipment, thereby reducing decision lag and execution deviation.

[0048] The AI control compilation unit 300 includes a decision instruction establishment module 310 and an AI control execution module 320;

[0049] The decision instruction establishment module 310 is used to collect decision instructions for multi-source data from historical data as a sample set, use the regression model to train the sample set, map the multi-source data state to the decision instruction, and output the decision instruction corresponding to the current state of other variables on the non-inferior solution path set output by the dynamic digital twin unit 200, as well as the decision instruction corresponding to the future state of the supply chain;

[0050] The control execution module 320 is used to integrate knowledge in the industrial field to construct a knowledge graph, including equipment parameters, operating manuals, process flows, etc., use semantic relationships to represent decision instructions, and establish mapping rules from decision instructions to equipment control parameters in the knowledge graph, so that AI automatically drives the corresponding equipment according to the control parameters. For example, "production plan adjustment" is mapped to "equipment temperature parameters". If the production plan is adjusted to increase the cycle, the equipment can reduce production intensity and reduce equipment operation, which means there is enough time to dissipate heat, reduce the heat dissipation temperature of the equipment, and save resources. If the production plan is adjusted to reduce the cycle, it means that the equipment needs to run for a long time to catch up, the temperature will rise sharply, and the heat dissipation intensity of the equipment needs to be increased to improve the heat dissipation effect and avoid affecting the service life.

[0051] Taking into account that when the multi-source data state mapping output is not adjusted completely according to the decision instructions, the AI control compilation unit 300 also includes an AI decision module. The AI decision module is used to convert the control parameters into intelligent voice feedback to the user through AI when receiving the decision instructions of the decision instruction establishment module 310. If the user feedback is in agreement, the control execution module 320 is driven to execute the corresponding control parameters. If the user feedback is paused, the drive of the decision instruction establishment module 310 is intercepted. This not only enables AI to control parameters autonomously and avoid slow progress and inaccurate operations caused by human operations, but also establishes a two-way communication mechanism to ensure real-time data exchange between AI and users, avoid conflicts between AI drive and user operations, and improve security.

[0052] Example 2

[0053] Since variable data will have a downward impact along the non-inferior solution path and will hardly have an upward impact, when any variable is input into the supply chain status model and the current status of other variables on the non-inferior solution path set is predicted, the initial variable causing the problem cannot be accurately judged, resulting in decision lag and execution deviation. Therefore, the dynamic digital twin unit 200 also includes a variable data analysis module. The variable data analysis module is used to collect the non-inferior solution path position of the variable input into the supply chain status model, and quantify the degree of influence of the current variable on other variables in the non-inferior solution path set, so that the dynamic digital twin unit 200 predicts the current status of other variables on the non-inferior solution path set in order of influence. Specifically, the position of the current variable in the causal graph is queried, all non-inferior solution paths are traversed, all paths containing the variable are found, and the specific position of the variable in the path (such as the starting point, intermediate node or end point) is determined. The degree of influence between variables can be determined by causal strength, timeliness weights, etc. The causal effect of the current input variable on other variables in the path is comprehensively evaluated by re-using the time series analysis of historical data, and the gradient intervention method is used again to calculate the causal effect of the current input variable on other variables in the path. The causal strength is evaluated according to the gradient size. The larger the value, the stronger the impact. The time series analysis is performed using historical data to verify the accuracy of the causal strength and timeliness weights. The causal strength and timeliness weights are weighted and summed to obtain the comprehensive impact score of the variable on other variables. According to the impact score calculated above, the other variables in the path are sorted. The higher the impact score, the higher the priority, which is beneficial when predicting the current state of other variables on the non-inferior solution path set. It is necessary to sort them according to the degree of influence of the input variables in order to give priority to key variables. Assuming that the input variable is "supplier delivery delay", the goal is to predict its current state of variables such as production plan adjustment and logistics delay risk, and sort them by the degree of influence. Find the non-inferior solution path containing "supplier delivery delay" in the causal graph, for example:

[0054] Path 1: Supplier delivery delay → production plan adjustment → logistics delay risk

[0055] Path 2: Supplier delivery delay → urgent replenishment demand → logistics delay risk;

[0056] Calculate the causal strength and timeliness weight of "supplier delivery delay" on "production plan adjustment" and "urgent replenishment demand". Assuming the causal strength is 0.8 and 0.6 respectively, and the timeliness weight is 0.9 and 0.7 respectively, the comprehensive score is:

[0057] S(Supplier delivery delay → Production plan adjustment) = 0.8*0.5+0.9*0.5=0.85

[0058] S(supplier delivery delay → urgent replenishment demand) = 0.6*0.5+0.7*0.5=0.65

[0059] According to the comprehensive score ranking, "production plan adjustment" takes precedence over "urgent replenishment demand", which means that the production plan adjustment is executed first through the decision instruction corresponding to the output of the decision instruction establishment module 310, which not only improves the prediction accuracy, but also optimizes the decision efficiency, ensuring the response speed and accuracy of the supply chain management system.

[0060] In summary, the supply chain data analysis system based on deep learning technology of the present invention provides a comprehensive and efficient solution for optimizing supply chain management by integrating a multimodal causal representation unit 100, a dynamic digital twin unit 200 and an AI control compilation unit 300. The multimodal causal representation unit 100 can process structured data, unstructured text and real-time streaming data, and integrate them into a unified framework, solving the problems of manual definition and low timeliness of traditional ETL tools, and overcoming the data fault problem, providing strong support for supply chain risk warning, and constructing an intuitive causal path visualization, which helps to actively prevent and control potential risks in the supply chain and improve the accuracy of decision-making. Then, the dynamic digital twin unit 200 reduces the accumulation of cross-link errors through a dynamic prediction and update mechanism, enabling enterprises to respond quickly when changes occur and reduce supply chain losses. Finally, the AI control compilation unit 300 is used for efficient decision-making and execution, reducing the slow progress and inaccuracy caused by human operations, and realizing a two-way cognitive channel between the supply chain management system and the IA equipment.

[0061] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are merely preferred examples of the present invention and are not intended to limit the present invention. Various changes and improvements may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and improvements fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. Supply chain data analysis system based on deep learning technology, characterized by: It includes a multimodal causal representation unit (100), a dynamic digital twin unit (200), and an AI control compilation unit (300); The multimodal causal representation unit (100) is used to fuse structured data, unstructured text and real-time streaming data, and construct a cross-modal causal graph to reveal the causal relationship between different data modalities; The dynamic digital twin unit (200) constructs a supply chain state model using a dual-clock driving mechanism based on the causal relationship between the multimodal data output by the multimodal causal representation unit (100), inputs any multimodal data into the supply chain state model, outputs the future state of the supply chain, and presets minute-level increments to dynamically deduce the state of the future period; The AI control compilation unit (300) is used to establish a linear relationship between the future state and the decision instruction, call out the decision instruction corresponding to the future state of the dynamic digital twin unit (200) in the required time period, compile the decision instruction into equipment control parameters through the industrial knowledge graph and verify the feasibility of execution, and feed back the execution result to the multimodal causal representation unit (100).

2. The supply chain data analysis system based on deep learning technology according to claim 1, characterized in that: The multimodal causal representation unit (100) includes a cross-modal fusion module (110) and a causal discovery module (120); The cross-modal fusion module (110) is used to collect data for normalization processing, extract semantic features, map data of different modalities to a unified feature space using a deep learning model, and capture the association between data of each modality through an attention mechanism; The causal discovery module (120) is used to apply a gradient intervention causal discovery algorithm to calculate the causal relationship between data of different modalities to generate a cross-modal causal graph, and calculate the average causal effect under gradient intervention through a Bayesian network to form a causal relationship matrix between variables.

3. The supply chain data analysis system based on deep learning technology according to claim 2, characterized in that: The causal discovery module (120) further includes a multi-reference path evaluation module, which is used to traverse the causal relationship matrix to collect all paths within a length threshold, calculate the path score based on a comprehensive scoring index, the scoring index includes causal strength, modal consistency and timeliness weight, construct a Pareto front, and select a non-inferior solution path set.

4. The supply chain data analysis system based on deep learning technology according to claim 3 is characterized by: Constructing a supply chain status model in the dynamic digital twin unit (200) includes the following steps: Historical data of multiple variables are collected in chronological order, as well as historical data of other variables in the non-inferior solution path set in their causal relationship to form a training set. Each training set includes the historical data of the input variable and the future state of the target variable. The supply chain status model is trained to reflect the multimodal data and the future state of the supply chain on one of the non-inferior solution paths. Any variable is input into the supply chain status model to predict the current state of other variables in the non-inferior solution path set. Multiple data of variables within a time period are input into the supply chain status model to predict the target data at the bottom of the non-inferior solution path in the future time period to reflect the future state of the supply chain.

5. The supply chain data analysis system based on deep learning technology according to claim 4 is characterized by: The dynamic digital twin unit (200) further comprises a forecast period input module, which is used to input a period step that is smaller than the future time period into the supply chain state model, and retrieve a future state of the supply chain that matches the period step.

6. The supply chain data analysis system based on deep learning technology according to claim 5, characterized in that: The AI control compiling unit (300) includes a decision instruction establishing module (310) and an AI control executing module (320); The decision instruction establishment module (310) is used to collect decision instructions for multi-source data from historical data as a sample set, use a regression model to train the sample set, map the multi-source data state to the decision instruction, and output the decision instruction corresponding to the current state of other variables on the non-inferior solution path set output by the dynamic digital twin unit (200), as well as the decision instruction corresponding to the future state of the supply chain; The control execution module (320) is used to integrate knowledge in the industrial field to construct a knowledge graph, including equipment parameters, operation manuals, process flows, etc., use semantic relationships to represent decision instructions, and establish mapping rules from decision instructions to equipment control parameters in the knowledge graph, so that AI automatically drives the corresponding equipment according to the control parameters.

7. The supply chain data analysis system based on deep learning technology according to claim 6, characterized in that: The AI control compilation unit (300) further includes an AI decision module, which is used to convert control parameters into intelligent voice feedback to the user through AI when receiving the decision instruction of the decision instruction establishment module (310). If the user feedback is in agreement, the control execution module (320) is driven to execute the corresponding control parameters. If the user feedback is paused, the driving of the decision instruction establishment module (310) is intercepted.

8. The supply chain data analysis system based on deep learning technology according to claim 7, characterized in that: The dynamic digital twin unit (200) further includes a variable data analysis module, which is used to collect the non-inferior solution path positions of the variables input into the supply chain status model, and quantify the degree of influence of the current variable on other variables in the non-inferior solution path set, and sort the current states of other variables on the non-inferior solution path set predicted by the dynamic digital twin unit (200) according to the degree of influence.

Images